ACS Publications. Most Trusted. Most Cited. Most Read
Natural Language Processing Methods for the Study of Protein–Ligand Interactions
More

OR SEARCH CITATIONS

My Activity
Publications

Figure 1Loading Img
Download Hi-Res ImageDownload to MS-PowerPointCite This:J. Chem. Inf. Model. 2025, 65, 5, 2191-2213
  • Open Access
Review

Natural Language Processing Methods for the Study of Protein–Ligand Interactions
Click to copy article linkArticle link copied!

  • James Michels
    James Michels
    Department of Computer and Information Science, University of Mississippi, University, Mississippi 38677, United States
    More by James Michels
  • Ramya Bandarupalli
    Ramya Bandarupalli
    Department of BioMolecular Sciences, School of Pharmacy, University of Mississippi, University, Mississippi 38677, United States
    More by Ramya Bandarupalli
  • Amin Ahangar Akbari
    Amin Ahangar Akbari
    Department of BioMolecular Sciences, School of Pharmacy, University of Mississippi, University, Mississippi 38677, United States
    More by Amin Ahangar Akbari
  • Thai Le
    Thai Le
    Department of Computer Science, Indiana University, Bloomington, Indiana 47408, United States
    More by Thai Le
  • Hong Xiao*
    Hong Xiao
    Department of Computer and Information Science and Institute for Data Science, University of Mississippi, University, Mississippi 38677, United States
    *E-mail: [email protected]
    More by Hong Xiao
  • Jing Li*
    Jing Li
    Department of BioMolecular Sciences, School of Pharmacy, University of Mississippi, University, Mississippi 38677, United States
    *E-mail: [email protected]
    More by Jing Li
    Orcidhttps://orcid.org/0000-0003-3277-6818
  • Erik F. Y. Hom*
    Erik F. Y. Hom
    Department of Biology and Center for Biodiversity and Conservation Research, University of Mississippi, University, Mississippi 38677, United States
    *E-mail: [email protected]
    More by Erik F. Y. Hom
    Orcidhttps://orcid.org/0000-0003-0964-0031
Open PDF

Journal of Chemical Information and Modeling

Cite this: J. Chem. Inf. Model. 2025, 65, 5, 2191–2213
Click to copy citationCitation copied!
https://doi.org/10.1021/acs.jcim.4c01907
Published February 24, 2025

Copyright © 2025 The Authors. Published by American Chemical Society. This publication is licensed under

CC-BY-NC-ND 4.0 .

Abstract

Click to copy section linkSection link copied!
High Resolution Image
Download MS PowerPoint Slide

Natural Language Processing (NLP) has revolutionized the way computers are used to study and interact with human languages and is increasingly influential in the study of protein and ligand binding, which is critical for drug discovery and development. This review examines how NLP techniques have been adapted to decode the “language” of proteins and small molecule ligands to predict protein–ligand interactions (PLIs). We discuss how methods such as long short-term memory (LSTM) networks, transformers, and attention mechanisms can leverage different protein and ligand data types to identify potential interaction patterns. Significant challenges are highlighted including the scarcity of high-quality negative data, difficulties in interpreting model decisions, and sampling biases in existing data sets. We argue that focusing on improving data quality, enhancing model robustness, and fostering both collaboration and competition could catalyze future advances in machine-learning-based predictions of PLIs.

This publication is licensed under

CC-BY-NC-ND 4.0 .
  • cc licence
  • by licence
  • nc licence
  • nd licence
Copyright © 2025 The Authors. Published by American Chemical Society

Subjects

what are subjects

1. Introduction

Click to copy section linkSection link copied!
The study of protein–ligand interactions (PLIs) lies at the heart of cellular function and regulation, orchestrating a complex interplay of molecular processes essential for life. These interactions govern fundamental biological activities, including enzyme catalysis, (1,2) cellular signaling, (3) membrane transport, (4) immune response (5) and transcription factor regulation. (6) At the molecular level, PLIs control cellular homeostasis through metabolic feedback loops, (7) facilitate signal transduction cascades across membranes, (3) mediate immune system recognition of foreign molecules, (5) and regulate gene expression through the control of ligand-dependent activity of transcription factors. (6) The remarkable specificity of these interactions is achieved through a combination of structural complementarity and physicochemical properties and enables precise control of cellular functions.
Understanding PLIs has become instrumental in modern drug discovery development, (8,9) providing a rational framework for designing drugs with maximal efficacy and minimal side effects. Structure-based drug design efforts optimize lead compound development through the strategic modification of chemical groups to enhance binding affinity and specificity. (10) Beyond pharmaceutical applications, protein–ligand engineering efforts are revolutionizing both agricultural biotechnology and industrial bioprocessing. For instance, the engineering of crop proteins has led to improved nutrient utilization efficiency (11) and studies of protein–ligand interactions have been crucial in the development of enzymes with enhanced catalytic activity. (12)
Experimentally, methods like X-ray crystallography and cryo-electron microscopy (13) provide atomic-resolution structural information while biophysical approaches like isothermal titration calorimetry and surface plasmon resonance can provide binding thermodynamic and kinetic data for protein–ligand interactions. (14) Although these experimental methods provide high-quality data benchmarks, (15) they are typically resource- and labor-intensive and thus low-throughput. Computational approaches that simulate the underlying physics and chemistry of PLIs, such as molecular docking (16) or dynamics simulations, (17) can be less resource intensive but nevertheless demand significant computational and time investment. (18)
Recent advancements in machine learning (ML) and deep learning have opened new avenues for effective PLI prediction by leveraging large-scale data sets. ML-based approaches can rapidly assess compound–protein pairs by \learning' from diverse biochemical, topological, and physicochemical properties (19−23) at a pace far quicker than that using traditional methods. ML models have already delivered promising predictive performance for drug–target interaction and binding affinity, supporting early stage target identification and lead optimization. (24,25) As the excitement for ML use in the biological sciences grows, the prediction of protein–ligand interactions appears increasingly possible given recent advances in both ML and Natural Language Processing (NLP), (26,27) the computational study of language. (28)
NLP centers on the computational analysis and manipulation of language constructs (28) to bridge the gap between human communication and computer automation. NLP has experienced significant recent breakthroughs as demonstrated by the proliferation of widely used chatbots such as OpenAI’s ChatGPT, (29,30) Anthropic’s Claude, (31) and Microsoft’s Bing Copilot. (32) NLP has been further used to summarize texts, deduce author sentiment, solve symbolic math problems, and even generate programming code. (33−36) The effectiveness of NLP is predicated on languages having a structured symbolic syntax and set of rules to assemble basic units known as “tokens” (e.g., characters, words, or punctuation) to form higher-order constructs such as sentences or paragraphs. (37) The structured outputs of such a system reflect the grammar, conventions, and styles of the associated language. In NLP, tokens are transformed to encode “meanings” through mathematical vectors such that tokens of similar meaning are positioned closer together in the representational vector space. (38) By analyzing a large collection of data, NLP methods aim to infer emergent relationships between tokens that define the “rules” of a language. Once learned, this inferred set of rules can be used to perform predictive tasks such as separating tokens into categories, translating text from one language to another, and even predicting whether a literary work will be a commercial success. (39−41)
NLP can provide a complementary perspective to a simply biochemical view of biomolecules by treating protein and compound sequences as “languages” composed of amino acid and chemical tokens. Through the use NLP-inspired models, researchers can capture subtle sequence patterns, secondary structural motifs, and functional domains that correlate with ligand-binding specificity and affinity. (26,27) Integrating NLP approaches with ML for PLI prediction has shown early promise, as models pretrained on large protein or compound databases learn contextual embeddings that can enhance pattern recognition for predictions of ligand binding affinity and specificity. By analyzing common surrounding tokens of given amino acids or atoms, biological roles may be inferred, such as whether an amino acid plays an important role in secondary structure. (42,43) Early comparisons with traditional methods have shown encouraging performance improvements, highlighting the potential of NLP-based models to refine both the accuracy and interoperability of predictions and ultimately, help expedite drug discovery. For practical uses, NLP methods have been used for a variety of predictive tasks, including to inferr disease-gene associations, (44) predict tumor gene expression patterns, (45) and assign functional annotations to various protein-coding genes. (21) Despite impressive advances, the creation of these NLP models is associated with a sizable computational burden (46−51) and it remains a challenge to understand what and which specific features of the input sequence data are responsible for predictive success.
In this review, we explain how NLP offers new ways to understand and predict PLIs. We first describe the relationship between common protein and ligand text representations vis-á-vis the characteristics of human language. Next, we present a paradigm of data collection for PLI studies and employ a table of data sources organized loosely by tasks for which they may be best suited. We then introduce and discuss three major NLP-associated methods often employed in machine-learning-based PLI studies: the Recurrent Neural Network (including variants like Long Short-Term Memory (LSTM)), the Transformer, and Attention Mechanisms. We provide several tables to convey how published studies have employed these major architectures to predict PLIs. What is in common between these major methods is their efficacy in capturing long-distance relationships between atoms and/or amino acids that are crucial for binding; we contextualize their use by presenting a conceptual framework for predicting PLIs that is followed by many NLP-PLI studies.
We conclude with a discussion of the limitations of using NLP in studying PLIs and with the data currently available for training machine learning models. Current approaches have shown promising results, but there are still significant challenges related to data variety, model interpretability, and bias. NLP offers valuable strategies for exploratory analysis and has taken a place in the foundation of such efforts but is not a standalone solution; integrating insights from other disciplines, such as computer vision, and domain-specific knowledge may be crucial for advancing PLI research in the future. We emphasize the need for high-quality, well-balanced data sets and suggest that new strategies, such as high-throughput simulations, could provide a pathway to overcoming current data limitations. Moreover, we emphasize the importance of integrating domain expertise, such as structure-based insights.

2. The Languages of Life

Click to copy section linkSection link copied!
Human languages are ever-evolving, (52) often ambiguous, (53) and idiosyncratic, which make them not ideal for computational study given the importance of context. (54) Human languages are generally hierarchical, composed of layers on the order of words, phrases, sentences, etc. by which information is communicated. (55) Human languages also demonstrate complex local behaviors that diverge from a hierarchal perspective, including long-distance dependencies (56) (e.g., subject and pronoun), as well as common substructure constructs like idioms (“raining cats and dogs”) or groups of objects that function as a unit (“knife and fork”). (57) Recursion is another linguistic aspect of human language that goes beyond simple hierarchy, for example, “I believe that you suppose that...”. (58) In general, these meta-linguistic occurrences are consistent with a view of language in which the linear order of words gives rise to a construct that embodies information. (57) While biochemical texts are distinct from human languages, there is remarkable similarity between the two regarding the hierarchal-and-sequential nature of construction as well as how local and global information is encoded (Figure 1). Nevertheless, the hierarchies of construction are not directly analogous as both protein sequences and molecular texts have significant structural and ontological distinctions that should be accounted for during computational processing. A comparison between human languages and the most common forms of text-based representations of proteins and molecules is presented below.

Figure 1

Figure 1. Language of protein sequences and the ligand SMILES representation: NLP methods can be applied to text representations to infer local and global properties of human language, proteins, and molecules alike. Local properties are inferred from subsequences in text: (left) for human language, this includes a part of speech or role a word serves; (middle) for protein sequences, this includes motifs, functional sites, and domains; and (right) for SMILES strings, this can include functional groups and special characters used in SMILES syntax to indicate chemical attributes. Similarly, global properties can theoretically be inferred from a text in its entirety.

High Resolution Image
Download MS PowerPoint Slide

2.1. The “Language” of Proteins

Protein sequences are akin to human language in that they possess a hierarchical order of construction and embody embedded information. Human language text is inherently ordered with characters of an alphabet assembled linearly and grouped into words, phrases, and sentences that convey an emergent message. Protein sequences similarly obey a hierarchy of assembly, with amino acids (AAs) serving as the alphabet. When AAs are strung together, secondary structural motifs, domains, and quaternary (multidomain-interacting) structures may emerge with properties that contribute to function. (59,60) While external factors such as post-translational modifications and cellular state can play a substantial role in dictating protein three-dimensional structure and function, the AA sequence represents the essential blueprint that ontologically defines the properties of a protein. (61−63) This fact has served as the foundation for bioinformatic analysis of proteins. (64) Individual AAs and common subsequences contribute to the “information” of the overall protein just as words contribute to the meaning of a text.
However, protein sequences are not entirely analogous in their hierarchy as compared to human languages, and “words” are not easily identified or demarcated. In linguistics, a “word” is a complete unit of meaning that a reader can recognize. It would be dubious to assume that AAs are equivalent to “words” because the roles of individual AAs are highly dependent on their context and environment. The meaning of a word may be independent of its surroundings; however, an amino acid carries “meaning” highly dependent on its three-dimensional context. Protein motifs or domains are also not comparable to “words”, since not all regions of a protein are independent of one another (65) and motifs and domains are not completely independent units. This lack of word-equivalence for protein sequences has driven “sub-word” identification methods that identify strings that act similarly to words. (66) Protein sequences also differ from human languages in the length scale of interactions and the number of long-distance interactions that contribute to a 3D structure. While human language often features distant dependencies, such as between subject and pronoun or text that foreshadows later content, these relations can be easily deduced by a reader and remain relatively sparse on a per-sentence basis. In contrast, AAs may have numerous distant relationships that are difficult to predict (67−69) without the assistance of computational or experimental tools. These characteristics allow a protein sequence to encode multiple layers of complex information, including 3D structure, structural dynamics, and/or binding interactions. (59,60) In essence, a sequence is not just a static representation, but rather a sophisticated programmatic embodiment that determines both structure and behavior of a protein.

2.2. The “Language” of Ligands

The chemical structures of molecules can be similarly translated into text-based notations and analyzed computationally. (70) However, unlike the elements of human text and protein sequences, the chemical connectivity patterns of molecules are not one-dimensional. Nevertheless, text-based schema has been developed to represent chemical information in a manner convenient for computational analysis, (71) with the Simplified Molecular-Input Line-Entry System (SMILES) format being one of the most widely used. (72)
SMILES strings are text representations constructed over a depth-first traversal of a two-dimensional molecular graph (Figure 1), with atoms, atomic properties, bonds, and structural properties represented by characters following an established set of conversion rules. Given the memory-efficient and somewhat human-readable format of SMILES, it has become a standard in chemical databases and computational tools, (73−75) and the most commonly used text representation in PLI studies. Although SMILES lacks an intuitive way to determine a chemical equivalent of a “word”, there is a well-defined grammar to denote properties and substructures of a molecule. Moreover, the same molecule can be represented by multiple different SMILES strings, (72) which is similar to how there could be multiple sentence constructions to convey the same idea in human languages. In NLP applications, incorporating tokens with the same meaning into the training process can yield a robust predictive model. (76) The use of multiple SMILES per molecule has been leveraged to guide ML models to discern which parts of a ligand contribute to drug potency. (77)
The SMILES format is dissimilar from human languages in a similar way as for protein sequences. First, the lengths of SMILES strings could vary far more than in human languages, ranging from listing each atom of a small molecule to those constituting entire proteins. The SMILES format is less practical to use for larger molecules, however, since structural graphs can provide a more compact and accurate representation of atoms in a large three-dimensional structure. A disadvantage of using SMILES in general is that it is difficult to intuitively discern “word” equivalents within the string. Individual branches separated by parentheses could be viewed as words, (78) but this is only practical for small branching groups. Moreover, the handling of nesting parentheses in SMILES for large molecules can be problematic and has become a major limiting factor in ML models designed to generate novel molecules. (79) The sum of these SMILES shortcomings has led to the development of alternative chemical representations for computational studies such as DeepSMILES and SELFIES. (80,81) Although promising, these alternative forms have rarely been used in ML-based PLI studies to date. The question remains whether a three-dimensional molecule can be truly mapped to a text representation in a way that preserves all relevant structural information for use in predicting PLIs.

3. Protein–Ligand Interaction Data and Data Sets

Click to copy section linkSection link copied!
Protein–ligand binding is a complex process dictated by many factors including protein states, hydrophobicity/hydrophilicity, and conformational flexibility. (82) The question of how to represent a protein and ligand in a computational space is critical and multifaceted. A wealth of information has been collected experimentally and generated through simulation studies on the properties of proteins and ligands, but these data are highly variable with regard to type, quality, and quantity. This section catalogs several primary data representations used in PLI studies. We also discuss the availability, selection, and curation of available data for machine-learning-based training and evaluation.
Protein and ligand representations are typically sequence- or structure-based. Unlike sequence-based text formats, structure-based information can appear in multiple forms, e.g., atomic coordinates of protein–ligand complexes or contact maps. Some structural information can be artificially reconstructed from sequence-based formats through algorithms such as AlphaFold for proteins (83) and RDKit for ligands. (84) PLI studies using machine-learning methods will typically select either sequence-based or structure-based inputs, although there is a growing use of mixed input data types. (85,86) For example, a mixed-data study may represent proteins by AA sequences but ligands by atomic coordinates, a choice based in part on the fact that highly accurate 3D chemical structures are easier to obtain than those of proteins and that full-atom representations of ligands are not memory intensive.
Other data can also be incorporated to augment ground-truth information about PLIs. For example, molecular weights, polarity, and bioactive properties can be incorporated into models to further improve the prediction of PLIs. (87,88) Studies have included molecular weights, ligand polar surface area, and protein aromaticity, (87) or bioactive properties of chemical and clinical relevance (88) have resulted in improved predictions of binding affinity. Leveraging multiple-sequence alignment or phylogenetic information to identify coevolutionary trends among AAs and sites of covalent modification has been shown to dramatically improve the accuracy of structural predictions of protein–ligand complexes. (89) The use of non-sequence/non-structural data can enable models to yield better predictive performance for characterizing protein and ligand and their interactions. (87)
Data for the study of PLIs can be manually curated by domain experts or sourced from existing data sets. Widely used public databases such as ChEMBL, (90) PubChem, (73) and DrugBank (91) play a critical role in the development and evaluation of drug–protein interaction models. These databases contain a wealth of information on ligands, proteins, and their interactions, supporting various predictive tasks. For instance, PubChem contains over 119 million ligands and is a cornerstone resource for general-purpose regression and classification models. Similarly, DrugBank focuses on the human proteome and offers curated data tailored to drug discovery, while ChEMBL provides comprehensive data on protein–ligand interactions, including SMILES-based ligand information.
Many databases are also inherently interconnected. For example, data sets involving structural information often reference available structures in the Protein Data Bank. (92) Similarly, sequence-based data sets frequently link back to UniProt (93) for protein sequence data. This interconnectedness emphasizes the importance of selecting a data set with the intended predictive task in mind. For tasks requiring high-quality, targeted data─such as predicting kinase activity─specialized data sets like Davis (94) or KIBA (95) are preferable. These data sets offer focused, curated information that aligns with specific biological questions. Conversely, general data sources like ChEMBL or PubChem are more suitable for deriving models aimed at uncovering generalizable underlying rules.
Given a protein–ligand representation, several predictive tasks are possible. Classification studies seek to categorize PLIs into distinct groups, for example, whether a protein–ligand pair binds or not. These models are relatively simple and allow for input from various sources. Regression studies use a continuous functional metric to characterize PLIs such as a binding affinity/dissociation constant (Kd) or inhibition constant (IC50). Continuous target variables allow for the involvement of numerical values derived directly from ’ground-truth’ experimental data in both training and evaluation. Databases like PDBBind (96) contain functional metrics such as Kd and IC50 but not all protein and ligand pairings cataloged have such metrics available, for example, complexes identified from X-ray crystallography, Cryo-EM, or NMR screening studies. (13,97) Since regression studies require quantitative PLI data and not merely whether a protein and ligand interact, relevant data set sizes may be smaller than those for classification. However, gathering such data is a laborious process in terms of both time and laboratory resources. Additionally, while functional metrics associated with regression studies can be used to predict exact values, the same data can support classification tasks, such as predicting binding versus non-binding rather than a specific binding affinity value.
Table 1 provides a comprehensive overview of existing PLI data sets and databases, summarizing their characteristics and suitability for various predictive tasks. Preassembled data sets are appealing for their convenience, though aligning the data set’s scope and quality with one’s modeling goals and the nature of the scientific inquiry is essential. Such a task-driven approach ensures robust model performance and meaningful predictions.
Table 1. Data Sets and Databases for PLI Predictiona
Data set NameYearProteinsLigandsInteractionsProtein CategoryLigand CategoryTask
Functional Data Available
Protein Data Bank (PDB) (92)2000220,777––General (Structure)General (Structure)C
BRENDA (103)20028,42338,623–EnzymesGeneralR, C
PDBBindb, (96)2004––23,496General (Structure)GeneralR, C
DrugBankb, (91)20064,94416,56819,441Human ProteomeGeneralC
BindingDB (104)20072,294505,0091,059,214GeneralGeneralR, C
PubChem (73,92)2009248,623119,108,078250,633GeneralGeneralR, C
Davis (94)20114426830,056Kinases (Sequence)Kinase Inhibitors (SMILES)R
PSCDB (105)2011––839Human ProteomeGeneralR, C
ChEMBL (90)201215,3982,399,74320,334,684General (Protein ID)General (SMILES)R, C
DUD-E (106)201210222,8862,334,372GeneralGeneralR, C
Iridium Databaseb, (107)2012––233GeneralGeneralR, C
KIBA (95)201446752,498246,088Kinases (Protein ID)Kinase Inhibitors (SMILES)R
Natural Ligand Database (NLDB)b, (108)20163,248–189,642Enzymes (Structure)GeneralR, C
PDID (109)20163,746511,088,789Human ProteomeGeneralR, C
dbHDPLSb, (110)2019––8,833General (Structure)GeneralC
CovPDBb, (111)20227331,5012,294General (Structure)GeneralC
PSnpBindb, (112)202273132,261640,074GeneralGeneralR, C
Protein Binding Atlasb, (112) Portal20231,71630,360129,333Drug TargetsDrug MoleculesR, C
Protein–Ligand Binding Database (PLDB)b, (113)2023125561,831Carbonic Anhydrases, Heat Shock ProteinsGeneralR
BioLiP2 (114)2023426,209–823,510General (Structure)GeneralR, C
PLAS-20kb, (115)2024––20,000EnzymesGeneralR, C
Functional Data Unavailable
Database of Interacting Proteins (116)200428,850–81,923Various Species–C
Protein Small-Molecule Classification Databaseb, (117)20094,9168,690–General (Structure)General (Structure)C
CavitySpaceb, (118)202223,391–23,391General (Structure)GeneralC
a

Note: Data sets categorized as “General” provide broad information without focusing on specific categories of proteins or ligands. Data types (e.g., sequence, structure), are denoted in parentheses. Categories labeled with “Protein ID” include protein IDs from established databases. Data sets may receive periodic updates. Suggested tasks are denoted as “R” for regression and “C” for classification. "−" indicated that exact information is either not included in the source or is not readily obtainable.

b

Protein–ligand complexes are available with the data set.

A secondary but still crucial consideration is the splitting of the data into training, validation, and test sets for use in a model. The training set constitutes the majority of the data from which a model’s parameters are learned;the validation set is used to tune the model’s configuration (controled by "hyperparameters"); (98) and the test set is a separate set of data points used to determine model performance. (98) There are several ways to create data splits aside from the simple option of randomly dividing the data. For example, a model may be designed with data splits that ensure different proteins or ligands are included in the training/validation/test sets such that proteins and ligands are not shared between them. (99) Evaluating a PLI prediction model on thes sets would then provide data on a model’s performance on unknown proteins and ligands that are outside of its training set. Competitions often provide a specific, well-designed test set data split as a benchmark, an approach used for other predictive challenges such as the Critical Assessment of Structural Prediction (CASP) (100) and the Critical Assessment of Prediction of Interactions (CAPRI). (101,102)

4. Machine Learning and NLP for PLIs

Click to copy section linkSection link copied!
Machine learning is a field of study where algorithms are used to uncover hidden patterns from data sets without explicit rule-based programming. Desired outcomes of specific processes are referred to as tasks (e.g., classification, regression, etc.), and depending upon the tasks, a suitable machine learning model is chosen, which includes decision trees, support vector machines, neural networks (NNs), and deep learning architectures. (98) NLP tasks often rely on deep learning and neural network architectures, which can both process the immense amounts of language-related data available and model the complex and often conflicting rules of human languages.. (119) Due to the parallels between the representation of language constructs and those of proteins and ligands, NLP-oriented machine learning approaches will be the focal point of this review article.
The general workflow for any ML-based study can be broadly characterized into three stages: data preparation, model creation, and model evaluation (Figure 2). For PLI studies, data preparation typically entails selecting the types and formats of protein and ligand data (e.g., sequence and/or structural). ML model creation may involve the following three tasks, although the boundary between these tasks could be fuzzy at times: (i) Extract: the “extraction” of vector “embeddings” from the protein and ligand input data, which can be used in computational operations (described in Section 4.2), (ii) Fuse: the fusion of protein and ligand vector embeddings, and (iii) Predict: the prediction of a PLI target property as a model’s output. The predictive capability of the model would be ideally validated against results from other studies and/or real-world measurements in a model evaluation stage. While data preparation and extraction steps have typically been the focus of most research efforts, every component of the workflow is crucial to successful PLI prediction.

Figure 2

Figure 2. Summary of the data preparation, model creation, and model evaluation workflow. Model Creation for PLI studies follows an Extract-Fuse-Predict Framework: input protein and ligand data are extracted and embedded, combined, and passed into a machine learning model to generate predictions.

High Resolution Image
Download MS PowerPoint Slide

4.1. The Extract-Fuse-Predict Framework

A variety of models for PLI prediction have been constructed in recent years, and these models tend to fall into four general categories: (1) sequence-based, where protein sequences and SMILES are used to represent protein and ligand, respectively; (2) structure-based, where structural information is included in the representation of both protein and ligand; (3) mixed representations, where both structural and sequence information are involved; and (4) sequence-structure-plus, which substantially incorporates other ground-truth information beyond sequence and structural data (such as molecular weights or polar surface area (87)). Tables 2, 3, 4, and 5 summarize several representative NLP-based PLI prediction studies across these categories over the past five years. Although PLI studies could be categorized in other ways─for example by the ML model used (neural network, decision tree, etc.) or by the predictive task type (classification vs. regression)─we have chosen to emphasize a categorization based on input data type since the computational methods used for sequence text and structural data comprise a major difference.
Table 2. Sequence-Based PLI Prediction Modelsa
 Extraction  
Model NameProtein ExtractorLigand ExtractorFusionPrediction
LSTM
Affinity2Vec (140)ProtVecSeq2SeqHeterogeneous NetworkGradient-Boosting Trees (R)
DeepLPI (141)ResNetResNetConcatenation with LSTMFCN (C, R)
FusionDTA (142)BiLSTMBiLSTMConcatenation with Linear AttentionFCN (R)
Transformer
Shin et al. (181)CNNTransformerConcatenationFCN (R)
MolTrans (182)TransformerTransformerInteraction Matrixb with CNNFCN (C)
ELECTRA-DTA (180)CNN with Squeeze-and-Excite MechanismCNN with Squeeze-and-Excite MechanismConcatenationFCN (R)
MGPLI (184)Transformer, CNNTransformer, CNNConcatenationFCN (C)
SVSBI (183)Transformer, LSTM, and AutoEncoderTransformer, LSTM, and AutoEncoderk-embedding fusioncFCN, Gradient-Boosting Treesd (R)
Non-Transformer Attention
DeepCDA (121)CNN with LSTMCNN with LSTMTwo-Sided AttentiondFCN (R)
HyperAttention- DTI (151)CNNCNNCross-Attention, ConcatenationFCN (C)
ICAN (150)VariousVariousCross-Attention, Concatenation1D CNN (C)
Other NLP Methods
GANsDTA (202)GAN DiscriminatorGAN DiscriminatorConcatenation1D CNN (R)
Multi-PLI (203)CNNCNNConcatenationFCN (C, R)
ChemBoost (124)VariousSMILESVecConcatenationGradient-Boosting Trees (R)
a

Note: A model’s task of Classification (C) and/or Regression (R) is denoted beside the “Prediction” column entries in parentheses. Definitions for specific terms may be found in the Glossary (Table 6). Terms Defined by the Cited Authors:

b

Interaction Matrix: Output from dot product operations to measure interactions between protein subsequence and ligand substructure pairs.

c

k-embedding fusion: The use of machine learning to find an optimal combination of lower-order embeddings via different integrating operations.

d

Two-sided Attention: Attention mechanism that computes scores using the products of both pairs of protein/ligand fragments and protein/ligand feature vectors.

Table 3. Structure-Based PLI Prediction Modelsa
 Extraction  
Model NameProtein ExtractorLigand ExtractorFusionPrediction
Transformer
UniMol (122)Transformer-Based EncoderTransformer-Based EncoderConcatenationTransformer-Based Decoder (R)
Other Attention
Lim et al. (160)GNNGNNAttentionFCN (C)
Jiang et al. (152)GCNGCNConcatenationFCN (R)
GEFA (153)GCNGCNConcatenationFCN (R)
Knutson et al. (155)GATGATConcatenationFCN (C, R)
AttentionSite-DTI (158)GCN with AttentionGCN with AttentionConcatenation, Self-AttentionFCN (C, R)
HAC-Net (156)GCN with Attention AggregationGCN with AttentionCombined Graph RepresentationFCN (R)
BindingSite-AugmentedDTI (157)GCN with AttentionGCN with AttentionConcatenation, Self-AttentionVarious (R)
PBCNet (154)GCNMessage-Passing NNAttentionFCN (R)
a

Note: A model’s task of Classification (C) and/or Regression (R) is denoted beside the “Prediction” column entries in parentheses. Definitions for specific terms may be found in the Glossary (Table 6).

Table 4. Mixed Representation PLI Prediction Modelsa
  Extraction  
Model NameInput TypeProteinLigandFusionPrediction
LSTM
Zheng et al. (204)P: Struct. L: SeqDynamic CNNb with AttentionBiLSTM with AttentionConcatenationFCN (C)
DeepGLSTM (85)P: Seq L: Struct.BiLSTM with FCNGCNConcatenationFCN (R)
Transformer
Transformer- CPI (86)P: Seq L: Struct.Transformer EncoderGCNTransformer DecoderFCN (C)
DeepPurpose (201)P: Seq L: Either4 Various Encoders5 Various EncodersConcatenationFCN (C, R)
CAT-CPI (185)P: Seq L: ImageTransformer EncoderTransformer EncoderConcatenationCNN and FCN (C)
Non-Transformer Attention
Tsubaki et al. (205)P: Seq L: Struct.CNNGNNAttention and ConcatenationFCN (C)
DeepAffinity (206)P: Seq L: Struct.RNN-CNN with AttentionRNN-CNN with AttentionConcatenationFCN (R)
MONN (207)P: Seq L: Struct.CNNGCNPairwise Interaction Matrix,c AttentionLinear Regression (C, R)
GraphDTA (197)P: Seq L: Struct.CNN4 GNN VariantsConcatenationFCN (R)
CPGL (208)P: Seq L: Struct.LSTMGAT with AttentionTwo-Sided Attention,d ConcatenationLogistic Regression (C)
CAPLA (161)P: Both L: Struct.Dilated Convolutional BlockDilated Convolutional Block with Cross-Attention to Binding PocketCross-Attention, ConcatenationFCN (R)
a

Note: A model’s task of Classification (C) and/or Regression (R) is denoted beside the “Prediction” column entries in parentheses. Definitions for specific terms may be found in the Glossary (Table 6). The input representations for sequence and structure are abbreviated for brevity. Terms Defined by the Cited Authors:

b

Dynamic CNN: ResNet-based CNN modified to handle inputs of variable lengths by padding the sides of the input with zeroes.

c

Pairwise Interaction Matrix: A [number of atoms]-by-[number of residues] matrix in which each element is a binary value indicating if the corresponding atom-residue pair has an interaction. (207)

d

Two-sided Attention: Attention mechanism that uses dot product operations between protein AA and ligand atom pairs, while taking matrices of learned weights into account.

Table 5. Sequence-Structure-Plus PLI Prediction Modelsa
 Extraction  
Model NameProtein ExtractorLigand ExtractorAdditional Features UsedFusionPrediction
LSTM
HGDTI (209)BiLSTMBiLSTMDisease and Side Effect InformationConcatenationFCN (C)
ResBiGAAT (87)Bidirectional GRU with AttentionBidirectional GRU with AttentionGlobal Protein FeaturesConcatenationFCN (R)
Transformer
Gaspar et al. (125)Transformer or LSTMECFC4 FingerprintsMultiple Sequence Alignment InformationConcatenationRandom Forest (C)
HoTS (210)CNNFCNBinding RegionTransformer BlockFCN (C, R)
PLA-MoRe (88)TransformerGIN and AutoEncoderBioactive PropertiesConcatenationFCN (R)
AlphaFold 3 (89)Attention-Based EncoderbAttention-Based EncoderbPost-Translational Modifications, Multiple Sequence Alignment InformationAttentionDiffusion Transformerc
Other NLP Methods
MultiDTI (123)CNN with FCNCNN with FCNDisease and Side Effect InformationHeterogeneous NetworkFCN (C)
a

Note: A model’s task of Classification (C) and/or Regression (R) is denoted beside the “Prediction” column entries in parentheses. Definitions for specific terms may be found in the Glossary (Table 6). Terms Defined by the Cited Authors:

b

Atom Attention Encoder: An attention-based encoder that uses cross-attention to capture local atom features.

c

Diffusion Transformer: A transformer-based model that aims to remove noise from predicted atomic coordinates until a suitable final structure is output.

Table 6. Glossary of Terms That Appear in the Tables
TermDefinition
AutoEncoderA neural network tasked with compressing and reconstructing input data, often used for feature learning. (262)
BiLSTMBidirectional Long Short-Term Memory, a variant of LSTM where two passes are made over the input sequence, one reading in forward order, and one in reverse order.
CNNConvolutional Neural Network, a type of neural network that processes grid-like data, such as images, through a gradually-optimized filter that slides across input data to discern important features.
Dilated Convolutional BlockConvolutional Neural Network operations with defined gaps between kernels, which can capture larger receptive fields with fewer parameters.
ECFC4 FingerprintA molecular fingerprint that encodes information about the presence of specific substructures within a diameter of 4 bonds from each atom. (263)
FCNFully-Connected Network, a feedforward Neural Network where each neuron in one layer connects to every layer in the next. FCNs can also be referred to as Multi-Layer Perceptrons.
GAN DiscriminatorAn NN part of Generative Adversarial Networks (GAN) that learns important features to distinguish between real and artificial data.
GATGraph Attention Network, a type of Graph Neural Network that uses attention mechanisms to deciding the value of neighboring nodes to a given node when updating a node’s information. (264)
GCNGraph Convolutional Network, a type of Graph Neural Network that aggregates neighboring node features through a first-order approximation on a local filter of the graph. (265)
GINGraph Isomorphism Network, a type of Graph Neural Network that uses a series of functions to ensure embeddings are the same no matter what order nodes are presented in. (266)
Gradient-Boosting TreesA machine learning technique where many decision trees are trained in order, such that the next tree learns from the misclassified samples of the previous tree. All trees are then used to “vote” on results of each input.
GRUGated Recurrent Unit, a simplified version of Long Short-Term Memory that similarly uses a gating mechanism to retain and forget information, but is less complex than Long Short-Term Memory. (137)
Heterogeneous NetworkA graph where nodes and edges represent different types of information, often used to convey complex relationships in biological systems (e.g., drug, target, side-effect, etc.).
Message-Passing NNType of Graph Neural Network that computes individual messages to be passed between nodes so that representations for each node contain information from its neighbors. (267)
ProtVecA method for representing protein sequences as dense vectors using skip-gram neural networks. (268)
Random ForestA machine learning method where many decision trees are constructed, and the result of the ensemble is the mode of the individual tree predictions.
ResNetShort for Residual Network. A neural network architecture that speeds up training by learning functions to substitute for layer operations, allowing for the “skipping” of layers and faster training. (269)
Seq2SeqA machine learning method used for language translation in NLP, featuring an encoder-decoder structure. (266)
SMILESVecPrevious work from authors. 8-character ligand SMILES fragments are assigned a vector through a single-layer neural network, and an input SMILES string’s vector is equal to the mean of fragment vectors present in that input SMILES. (270)
Squeeze-And-Excite MechanismMechanism for Convolutional Neural Networks that uses global information to adapt the model to emphasize more important features. (271)

4.2. Extraction of Embeddings

NLP approaches deconstruct text into individual tokens or “units of meaning” for use in computational operations and inferences via a process referred to as “tokenization”. (37) Schema for tokenization, aside from character-based and word-based, can also be subword-based. Subword-based tokenization breaks down text into units smaller than words to create a wider vocabulary; it is commonly selected when the definition of a “word” is unclear, as subwords can be used as a means to discover “words”. (66,120) Common ways to assemble subwords include methods such as “n-grams”, where each subword has a select fixed-length value n (e.g., “Sma”, “mar”, “art”, etc. for n = 3 and the word "Smart"). While subword tokenization has been attempted in PLI studies for both protein (e.g., amino acid k-mers such as “KHR”, “LKL”, “KGY”) and ligand (e.g., “CCCC”,“[C@@H]”), (121−125) the current trend is to use amino acids and/or individual atoms directly as tokens.
To be processed computationally, tokens must be translated into a numerical form through a process known as “embedding”. There are many types of token embedding, but they are generally designed to capture either a particular token meaning, frequency, or both (126,127) and represented by a multidimensional vector. The direction of a token’s vector embedding effectively represents its “meaning” and its magnitude represents the strength by which that meaning is conveyed. In isolation, each token could possess multiple meanings (e.g., the word “run” has multiple meanings (128)), and so context may be necessary to impart an intended meaning. NLP methods have been demonstrated to be highly effective at extracting patterns that convey context-dependent meanings from a large corpus of text. Embeddings that capture semantic meaning and relationships can then be used for many other tasks aside from predicting whether a protein interacts with a ligand, such as predicting protein and ligand solubilities. (129,130)
Token embedding is typically accomplished using a neural network (NN) architecture that approximates nonlinear relationships between the “inputs” of the network (the data) and its “outputs” (the predictions). (131) Neurons in an artificial NN receive, integrate, and transmit signals to other neurons through a nonlinear response function and are arranged in layers. Information is passed from an input layer through one or more intermediate “hidden” layers to an output layer. (98) Interconnection weights that govern the strength of influence of one neuron on another are crucial parameters of an NN. A wide variety of NNs have been applied to studying PLIs although not all are commonly used in NLP. Nevertheless, two types of NNs commonly associated with NLP are Recurrent Neural Networks (RNNs) (132,133) and attention-based NN models. (134) Below, we highlight the details necessary to understand how RNNs, attention, and other non-NLP-driven NNs have been used to glean global patterns essential for PLI predictive tasks. For reference, Figure 3 presents simplified framework diagrams of RNN, transformer, and attention operations.

Figure 3

Figure 3. Framework diagrams for RNN (and its variant LSTM), transformer, and attention with arrows representing a flow of information. (A) The "unrolled" structure of an RNN and the recurrent units, where hidden states propagate across time steps. The recurrent unit takes the current token Xt as input, combines it with the value of the current hidden state ht, and computes their weighted sum before generating the response Ot and an updated hidden state ht+1. Weighted sums depend upon the associated network weights Wxh, Whh, or Woh, which connect input to hidden state, hidden state to hidden state, and hidden state to output, respectively. LSTM differs in that a memory state is updated during each iteration, facilitating long-term dependency learning. (B) A simplified framework of a transformer's encoder-decoder architecture, and associated attention mechanism. A scaled product of the Query and Key vectors yields attention weights that can provide interpretability, with the new embedding vector (or the output vector) updated based on this specific key.

High Resolution Image
Download MS PowerPoint Slide

4.2.1. Recurrent Neural Networks

RNNs (132) are specialized in processing sequential data in which the order of the data is significant. Consider an input data sequence x1, x2, ..., xt–1, xt, xt+1, ... in which individual tokens xt are ordered by a time-step t, and the input sequence embodies a particular yet unknown pattern over the length of the sequence. In traditional NNs, information flows from the input layer to the output in a single pass, making it difficult to decipher any interdependencies between earlier and subsequent tokens. To remedy this, the RNN architecture introduces recurrent units through which the processing of the input sequence at the current time-step will also update “hidden states” that serve as memory, nonlinearly capturing the information of all input tokens up to the current time-step. The recurrent unit derives its name from the fact that the hidden state participates in the computation both as an input and as an output for each input in the sequence.
In other words, given the network weights, the ordered sequence of input tokens will determine a network output sequence O1, O2, ..., Ot–1, Ot, Ot+1, ..., and the hidden states h1, h2, ..., ht–1, ht, ht+1... Thus, the hidden states are functionally equivalent to the hidden layers of traditional NNs but differ by updating recurrently, where information is carried over from previous time-steps to the current time-step. Consequently, the dependencies between tokens of the sequential inputs can be captured implicitly by the hidden state.
RNNs can be represented in an unfolded, or “unrolled” state (see Figure 3 A). In this representation, an input sequence can be considered as a mapping between preceding input values and values of subsequent elements in the same sequence, due to the inherent patterns existing within all elements. (135) For example, given a protein sequence for which each AA is a token, an RNN would process the sequence of AAs one at a time to create and maintain a mapping for the next AA in the sequence accounting for all input tokens seen so far. The mapping, encoded in the network weights of RNN, may be found via the backpropagation process, through which the shared weights are adjusted so that the “errors” between the computed outputs of the RNN and the expected outputs as presented in the input sequence are calculated and minimized. (136) The process of using backpropagation to adjust the network weights so that the desired outputs of an NN are achieved is the so-called training process in machine learning, with the resulting collection of weights being called a model.
A good example of an RNN applied to the study of PLIs is provided by Abdelkader et al.’s ResBiGAAT model, (87) which was designed to use a variant of bidirectional RNN layers to embed input strings (protein sequences or SMILES). ResBiGAAT's bidirectional RNN, which processed the input sequence of tokens both forwards and backwards in different passes, enables it to identify relations between a given token and both its previous and subsequent tokens. While effective in many NLP tasks, early RNNs commonly suffered diminishing returns with increasing text length. This was due to a simplistic network architecture in which there was systematic and non-discriminatory retention of information from all tokens, including outlier tokens that contribute little informationally to the underlying pattern. A variant of an RNN was chosen in ResBiGAAT that features a gating mechanism to specifically update and forget information from previous time steps; (137) the RNN used was also modified to include residual connections that enable information to be transmitted directly between layers without the need for calculating intermediate layers. This enabled several RNN layers to be stacked together with a relatively insignificant increase in convergence time. This use of RNN, alongside several other changes, allowed ResBiGAAT to outperform a selection of baselines at the time of publication in 2023.
To address the diminishing returns of early RNNs, gating mechanisms were developed to control the flow of information into the hidden state. The primary example of this is Long Short-Term Memory (LSTM) networks, (138) a popular variant of RNNs in which three gates are introduced into each recurrent unit: input gate, forget gate, and output gate (Figure 3A). The signature component of LSTMs is the forget gate, which selectively inhibits information not concordant with previously learned patterns found from processing prior tokens. (138) In addition, the input gate controls the level of input information added to the cell state, and the output gate governs the amount of information output at each step. Combined, the gating mechanisms selectively handle memory functionality, enabling effective encoding of long-term dependencies. For example, in the task of predicting protein secondary structures, LSTM has been shown to attenuate the contribution of AAs that do not correlate with any defined secondary structural element, yielding a small but definitively improved performance over the then state-of-the-art. (42,43) Unlike human languages where sentence structures possess distinct temporal orders, sequence-based representation of proteins and ligands may exhibit temporal or spatial symmetry, leading to researchers utilizing bidirectional LSTMs (BiLSTMs) to capture both preceding and subsequent tokens in a sequence string by applying an LSTM to text in both original and reverse order, and concatenating each of the resulting embeddings end-to-end. (139)
LSTMs and BiLSTMs are promising embedding approaches for predicting binding affinities of proteins and ligands. (140−142) However, their effectiveness is constrained by the computational inefficiency of the LSTM/BiLSTM architectures when processing large-scale data sets. Most successful applications of LSTM to date have been applied to only relatively small training data sets, on the order of a few thousand proteins and ligand pairs. This limitation mainly arises from the inherently non-parallel design where the tokens are being processed step-by-step, which makes training on large data sets slow and computationally expensive. Thus, NN architectures that leverage parallelization will be important to ensure reasonable training and prediction runtimes.

4.2.2. Attention-Based Architectures

Protein lengths can vary dramatically, from Insulin with 51-AAs to “giant proteins” that can exceed 85,000 AAs. (143) To use large amounts of sequence data to effectively process and predict PLIs for which long-distance interactions may be impactful, several alternatives to RNN have been proposed. The “neural attention”─or simply “attention”─mechanism is an important recent breakthrough by which “attention weights” are dynamically calculated to quantify the relative contribution of different input tokens or elements to a predictive end goal. (134)
In the context of attention, (134) the input sequence of data is tokenized and represented (or embedded) as key-value pairs. A specific, previous section of the input (or a key) is said to be “attended to” when the model gives it a heavier weight in the process of updating the representation (i.e., the query) with each new input token. The attention weight is stored in a matrix, and is determined via a normalizing function and a similarity comparison between the key and the query, the latter of which may change dynamically as the representation of the input stream is updated. The attention mechanism is highly general, and can be applied to inputs such assequences and images, with the keys being potentially any embedding that is relevant to the current task. (144−146) In many NN architectures, attention can also incorporate hidden states into the calculation, allowing a more sophisticated mechanism for capturing longer-range correlations in deeper layers. (134,146)
Attention mechanisms have proven highly compatible with traditional protein sequence analysis approaches in identifying long-distance interactions between AAs of a protein. (147) In PLI studies, attention mechanisms can dynamically adjust the contribution of specific AAs or ligand atoms to a predictive outcome by amplifying interaction sites with higher attention scores and downplaying less relevant ones (Figure 4). This process mirrors the biological intuition that certain residues and atoms are more critical for binding in a protein–ligand complex than others. The use of attention mechanisms has enabled the identification of AAs in proteins and atoms in a ligand that are highly cross-correlated and appear to physically interact (Figure 4), (148,149) although the degree of success in identifying interacting sites remains to be assessed. Attention has also provided an effective way to “fuse” protein and ligand representations in binding prediction models. (86,121,142,150,151)

Figure 4

Figure 4. Sample attention weights for relating protein and ligand. The heatmaps on the left help visualize the weighted importance of select protein residues and ligand atoms in a PLI. Structural views of the protein–ligand binding pocket are shown in the middle, with insets of the 2D ligand structures on the right. The colored residues and red color highlights indicate AAs in the protein binding pocket and ligand atoms with high attention scores. Reproduced with permission from Figure 7 of Wu et al. (148) Used with permission under license CC BY 4.0. Copyright 2023 The Author(s). Published by Elsevier Ltd.

High Resolution Image
Download MS PowerPoint Slide
Attention is a versatile mechanism that can also be applied to structural information such as the spatial coordinates of individual atoms or contact maps of protein–ligand complexes. (152−154) The structural information of proteins and ligands can be well-represented by a graph with nodes representing AAs or atoms, and edges representing chemical bonds or amino acid contacts. Edges may also represent other predefined relationships or constraints between nodes. Integrating attention mechanisms into Graph Neural Networks (GNNs), a class of NNs specialized for processing graphs, has been increasingly used for the study of PLIs. (155−158) GNNs use “message-passing” whereby each node’s embedding is updated iteratively based on information from connected nodes. (159) Each connection can be assigned a weight that quantifies the likelihood of interdependence between connected nodes. For example, a cysteine residue may have a higher weight for a nearby cysteine than a nearby glycine due to the potential to form a disulfide bond between cysteines. GNNs are often augmented further, for example, by the addition of an attention mechanism to prioritize connected nodes during message-passing. (152,153,156−158,160) An example of attention’s application to PLI studies is Jin et al.’s CAPLA model, (161) which used a “cross-attention” mechanism to directly correlate tokens within the protein and ligand to one another. The resulting attention weights can display the degree by which each unit relates to one another in order to provide a degree of interpretability, as determined by posthoc evaluation of the attention mechanism.

4.2.3. Transformers

While attention mechanisms have been quite beneficial for the predictive success of NLP methods, the “transformer” architecture pioneered in 2017 has also been instrumental in advancing these capabilities. (134) Transformers are a type of NN architecture that divides attention mechanisms into multiple parallel operations, each applying a different set of weights to the input data sequence. Several relationships between tokens are captured and processed simultaneously, dramatically improving the efficiency with which human text can be processed. The transformer architecture is the foundation of popular large language models such as ChatGPT (30) and was a key component of DeepMind’s AlphaFold system. (83,162) Transformers have become widely used in bioinformatics, for DNA, RNA, and protein sequence analysis, as well as gene-based disease predictions and PLI predictions. (163)
Transformers are designed to solve the problem of “sequence transduction” or the conversion of an input sequence of ordinal data into a predicted output sequence, such as a translated text or a vector representation. (164) In NLP, this is called machine translation, whereby the input sequence, for example, could be a sentence in English and the output sequence is its French counterpart. The transformer is an extension of the so-called “encoder-decoder” architecture (Figure 3B), a state-of-the-art sequence-transduction method commonly used today. (134,137,165,166) The premise of encoder-decoders is that sequentially ordered input data (e.g., English text, protein sequences, SMILES) can be “compressed” or encoded by a lower-dimensional fixed-length vector with minimal information loss. “Encoding” is the process of compressing informative features into a reduced vector representation, effectively capturing implicit rules or structures contained within the data. Typically, in this reduced representation (called the “latent” space), inputs with similarly informative characteristics appear close to one another. These compressed vectors can subsequently be “decoded” or expanded to an output representation of choice to complete the transduction task. These transduction tasks naturally align with the goal of text translation from one language to another. (137,167) Importantly, transformers differ from traditional encoder-decoder models by incorporating the attention mechanism. (134) Attention allows latent representations to vary in length, thus eliminating a fundamental constraint of encoder-decoder models: that every input sequence, regardless of length, be represented by a fixed-length vector in the latent space. Transformers are widely used today (27,163,168) (especially for long input sequences) given their inherent parallel architecture, which makes processing data sets with billions of items feasible. As compared to LSTMs, transformers are architecturally more complex and tend to achieve better performance. (169−172) Even so, transformers may not be the most effective approach, particularly when dealing with small data sets on the order of thousands of items. (173−175) In the biological domain, transformers have been applied to the prediction of protein–protein binding affinities, (176) post-translational modifications, (177) and quantum chemical properties of small molecules. (178)
Early applications of transformers for the study of PLIs involved simply retraining existing models designed for human language inputs; (168,179) surprisingly, these transformers surpassed existing state-of-the-art models for predicting binding affinities. (180,181) As new transformers were developed specifically to handle protein sequence data, predictive performance for PLIs improved. (182−184) These developments included preemptively dividing the texts into subsequences to determine which amino acids contribute to binding and merging embeddings from different transformers to provide multiple representational perspectives. Transformers have been further modified for use with additional data types, such as protein structures and images, as well as for predicting PLI properties beyond binding affinity, e.g., binding poses. (122,185) One such example leverages algebraic topology (186) by converting protein–ligand complex structures into unique one-dimensional sequences. (187) This novel approach was notably able to synthesize embeddings directly for the complex itself and demonstrates space for innovation in further developing the transformer architecture for PLI problems.
So far, transformer-based models have demonstrated mastery at manipulating language constructs for tasks involving reasoning, coding, vision, and mathematics at a level that mirrors human performance. (188) This success has also been extended to molecular biology with the advent of Protein Language Models (PLMs). (20,177,189) Through discerning the probabilities of amino acid appearances given a location and surrounding context, PLMs infer a notion of syntax and semantics for proteins from data sets of protein sequences on the order of millions. (190,191) Once a PLM is trained, the embeddings outputed from the last hidden layers can be transferred to any protein-related prediction task. While the embedding vectors are not fully explainable as to what information is contained within, the inferred semantic information is sufficiently preserved in the vector for PLMs to be highly effective in protein-related tasks. PLMs have demonstrated greater efficacy than sequence-based RNNs or LSTMs in predicting specific protein properties, such as structure, function, and cellular localization. (192,193) PLMs also present an opportunity to draw conclusions about small protein families that may not have enough evolutionary information available to perform traditional MSA-based approaches. (194) Although PLMs have not been spotlighted as much as breakthrough structure prediction projects such as AlphaFold, (83) they do see practical use for highly specialized tasks. Such cases include predicting if amino acid variations may preclude genetic disease (195) or identifying cellular sublocalization of peroxisomal proteins. (196)
An example of a transformer encoder is Qian et al.’s CAT-CPI model, (185) which applies a transformer to extract features from protein sequences and images of molecules. For protein sequences, Qian et al. experimented with several different tokenization strategies to be used in conjunction with the transformer to assemble protein subsequences based on frequency among the total corpus of protein sequences. A second transformer encoder was applied to discern long-distance relationships between pixels in the input images of molecules, gathering a different type of information entirely. The use of transformers for two different formats of input demonstrated the variety of use cases for an architecture as versatile as the transformer.

4.3. Fusion of Protein–Ligand Representations: Concatenation or Cross-Attention

Once candidate interacting protein and ligand embeddings are extracted, they need to be fused for an interaction pattern to emerge. Methods for extracting embeddings from protein and ligand sequence data have been the primary focus of the field to date such that approaches for fusion have been somewhat neglected untilrecently. A naive method for fusion is to simply concatenate protein and ligand embedding vectors end-to-end. More refined approaches, though, could involve advanced data structures like graphs, whereby information such as coordinates of protein and ligand is used not only to build a graph representation but is also incorporated into an attention mechanism to account for local factors such as polarity or size. (154,156,197) A mechanism of “cross-attention” could be incorporated into the fusion approach whereby the importance between the different token representations of the protein and ligand are directly calculated (150,151,161) in an attempt to mirror the underlying interaction of a protein with a ligand. (155) Cross-attention has been shown to be at least as competitive in predictive PLI tasks as other fusion methods, (197) and an improvement over the use of separate, independent attention mechanisms for both protein and ligand. (198)
While fusion appears to be a natural and important component for NLP studies of PLIs, some models circumvent the idea of fusion altogether and use protein-only or ligand-only representations explicitly. For example, Wang et al.’s CSConv2D algorithm only embeds ligand information. (199) An individual model is trained separately for each protein to predict that protein’s compatible ligands, resulting in the creation of hundreds of models. Although the task was to predict PLIs, protein information was only incorporated indirectly by labeling ligands during model training as either binding to a given protein or not. Nonetheless, protein-only or ligand-only models are rare, with most contemporary NLP-PLI models considering both protein and ligand together through a fusion step.
Mixed-data approaches aimed at combining different data types for protein and/or ligand (e.g., sequence + structure; sequence + image; (185) or both sequence and structure for protein + structure for ligand (161)) have further spurred study into which input formats are best for protein and ligand. Mixed-data models may use a variety of architectures such as an LSTM or transformer for a protein sequence and a GNN for ligand structures. (85,86) Combining multiple state-of-the-art embeddings for both sequence and structure has outperformed sequence-only baselines. (86) Despite the increased complexity involved in handling sequence and structural data simultaneously, mixed-data models are advantageous for both the ease-of-use of protein sequences and the completeness of ligand structural representations.
Although underexplored, combining multiple embeddings for each protein and ligand input in the fusion process may be beneficial. It has been suggested that different protein encoders for extracting features may gather different but relevant information to improve predictive outcomes. (200) In the DeepPurpose algorithm, Huang et al. pursued a library approach that offered 15 different protein and ligand embeddings (including transformer and RNN) to be combined and fed into a small NN to generate binary binding and/or continuous binding affinity predictions. (201) This menu-option system enables users to compare feature extractors and find the best protein and ligand embeddings for their research. Another approach is to combine multiple embeddings through operations such as component-wise multiplication or component-wise difference, as each embedding could represent a different set of features. (183,200) Shen et al.’s SVSBI algorithm (183) demonstrated how a higher-order embedding, by concatenating three different transformer embeddings, could outperform several state-of-the-art baselines (including those based on individual transformers alone) in the prediction of binding affinity.

4.4. Prediction of Target Variables

Ultimately, specific research questions must motivate the relevant PLI target variables that will be predicted by constructed ML models. These models often consist of one or more fully connected layers with relatively fewer parameters than the NNs used for feature extraction or fusion. The purpose of these layers is to utilize the latent protein and ligand features to predict an output target variable such as binding affinity or a binary indication of whether a pairing interacts. Thus, the fused protein and ligand embeddings are passed through these final layers to compute the prediction. Embeddings that effectively capture important underlying features can also be applied to predict other useful properties beyond binding affinity such as protein and ligand solubility. (129,130)

4.5. Evaluation

Evaluation is typically performed by comparing statistical metrics between models on the same test data sets. Evaluation metrics vary by task: classification predictions can be assessed via metrics such as precision, recall, and F1 score metrics whereas regression predictions are often evaluated relative to the ground-truth test data via concordance index and mean square error metrics. (98,211) Premade data sets such as PDBBind (96) are frequently bundled with both training and test data sets to enable fair comparisons with other established models. Models aiming to be generalizable across several types of PLIs should ideally be evaluated on several different sets of proteins and ligands.
While ML models can be assessed through the aforementioned statistical metrics, the practical utility of PLI predictive models and their predictive accuracy in real-world cases is best determined by domain experts. (212) For example, if a model is designed to predict binding affinities, a set of predictions generated in silico would be best confirmed through in vitro experimentation. PLI prediction models could also gain credibility if predictions are validated through physics-based simulation techniques such as molecular docking and molecular dynamics simulations. (213,214) For instance, Chatterjee et al.’s AI-Bind predicted interactions between SARS-CoV-2 viral proteins and human targets, used molecular docking and in vitro/clinical results to confirm these predictions in agreement with existing literature. (214) Similarly, Kalakoti et al.’s TransDTI employed a transformer-based architecture and corroborated predictions for MAP2k and TGF-β inhibitors with molecular dynamics simulations. (213) These methods confirm the accuracy of predicted interactions and align with existing biological knowledge, demonstrating both predictive reliability and practical relevance. Such experimental and simulation-based validation can justify a model’s use in the setting where it can be most effective and create opportunities for future interdisciplinary collaboration between ML practitioners and domain experts in computational and experimental biology.

5. Challenges and Future Directions

Click to copy section linkSection link copied!
Advances in generative AI and NLP have revolutionized how we tackle tasks related to human language. Early successes of NLP methods in discerning the “rules” of protein structure (as exemplified by AlphaFold (83)) suggest significant potential for NLP to transform our approach to studying PLIs. While many innovations in the NLP computational toolkit for PLIs have emerged in recent years, several practical hurdles remain, limiting the impact and potential insights derivable from the ML approaches. This section presents an overview of the many challenges confronting the PLI field and suggests various avenues to address them.

5.1. Lack of “True Negatives”

A common challenge in today’s data-driven ML paradigm is the limited availability of abundant, high-quality, and labeled data. (215) In PLI studies, there is a particular lack of bona fide “negative examples”, i.e., data for ligand-like molecules that do not bind a protein of interest that are critical for model training. If a model is trained on only positive data without any means to adjust for it, there would consequently be a sizable bias toward labeling all test data as positive. For instance, this could be an enzyme paired with a molecule that is obviously not a compatible substrate. In “supervised” ML, (216) models are trained on data with labels of whether a protein–ligand pair is binding or non-binding, and protein–ligand data spanning the full spectrum of interaction/no-interaction are necessary for models to ‘learn’. When a similar situation is encountered in other ML tasks, a common approach is to select random data points not explicitly labeled as “positive” and assume them as “negative”. However, given the complexity and specificity of PLIs, these are often trivial negative examples, since molecules that do not interact with a protein of interest and are dissimilar to the “true” ligands embody little information from which ML models can learn. Manually curating protein–ligand pairs that display weak interaction or lower binding affinity is an option for addressing this problem, although this is time-consuming and labor-intensive.
Unfortunately, the availability of informative negative PLI data requires deliberate efforts of domain experts who recognize the importance of generating, curating, and reporting such data, which are rarely publicized or emphasized in the literature regardless of data type. (217−219) This scarcity of negative examples has been observed in several fields. (220) Learning from positive data only or from a mix of positive and unlabeled data is an active field of study, with attempts to apply “unsupervised” and “semi-supervised” methods (221) (see (202,222) for examples related to PLI prediction). Compared with supervised models, un/semi-supervised models typically require larger data sets of tens to hundreds of thousands of PLIs and are more computationally intensive. (202) In cases where negative data does exist albeit at a significantly reduced quantity, classification studies of PLIs can adjust the distribution of ligands to ensure equal proportions of positive and negative examples; this has been shown to mitigate over representational bias of positive data. (223) Future studies should resolve the lack of readily available non-interacting protein–ligand pairs, perhaps through mining the scientific literature for meaningful non-binding pairs.

5.2. Diversity Bias in PLI Data Sets

Many PLI data sets display an underlying bias concerning either the diversity or types of proteins and ligands, hindering the effectiveness of ML algorithms. Training with insufficiently different data points can lead to poor predictive performance when a model is deployed for real-world examples. For example, binding affinity predictors trained on the popular PDBBind data set (96) with both protein and ligand information represented performed no better than those trained on only protein or only ligand information as inputs, (99) suggesting that some implicit non-informative patterns within the proteins and ligands of the PDBBind data set were learned rather than information concerning the mechanics of binding. The commonly used DUD-E (106) data set of bioactive compounds and respective protein targets demonstrates a similar problem: classification models that appeared highly accurate were found to differentiate binders/non-binders based primarily on their different shape classes and not the embedding of any relevant information about the protein–ligand interface. (99,224) Existing literature suggests that this is a problem of quality over quantity, as memorization-related biases in PLI models are not alleviated by merely increasing the data set size or removing overrepresented items. (225) The presence of bias is understandable, given how idiosyncratic research interests in biological or pharmaceutical fields shape the particular proteins and subsets of ligands studied and the type of PLI data generated and made available.
Given that models trained on biased data often fail in practical, real-world prediction tasks, the creation of high-quality, well-balanced, and unbiased PLI data sets is essential to the future of ML-based PLI studies. One way around the experimental challenges of generating sufficient protein–ligand data may be through high-throughput molecular dynamics simulations and/or docking studies using AlphaFold-predicted (83) protein structures. In particular, methods that can accurately estimate binding affinities, such as free-energy perturbation (226) or umbrella sampling, (227,228) appear promising. Although current simulation methods remain time-intensive, advancements in high-performance computing and the growing availability of GPU-based resources are making this approach increasingly feasible (229) and the benefits may be worth investing in this pursuit. These approaches, unrestricted by experimental technical limitations, could be systematically deployed at scale to generate protein–ligand complex structures and binding information, particularly for historically understudied protein classes and ligand categories. These procedures could also be automated, requiring far less human intervention than laboratory experiments, to yield valuable binding pocket information for improved structure-based ML predictions.

5.3. Interpretable and Generalizable Design in PLI Predictions

The open-data movement and the broad accessibility of machine-learning tools have catalyzed the development of numerous predictive models to discern patterns within data. However, these models often rely on complicated weighted operations that are challenging to interpret. Many ML studies fail to consider designing human-friendly interpretations of how their models’ predictions are calculated. While interpretability is not a requirement for a high-performing model, a lack of interpretability can be a hurdle to the acceptance of such models as users may doubt the trustworthiness of a “black-box” model. (48) One potential approach for bridging the “explainability” gap is the use of attention weights to corroborate existing protein–ligand contacts (cf. Figure 4). (86,121,142,150,151) Attention weights highlight regions in PLI models that converge with higher weight values but may result in “false positives” whereby higher binding weights are inadvertently assigned to non-binding regions. Unfortunately, a systematic assessment of “false positives” in attention weights has yet to be performed, leaving it unclear whether they are a reliable metric. (98) Such false positives are one facet of a larger debate on whether attention weights provide sufficient explanatory power for PLI models. (230−232)
While NLP presents attention mechanisms as one possible avenue, other methods of explainability are starting to be explored for interpretable PLI predictions. One example is a game-theory approach to compute “Shapley values”, which quantify the importance of individual features by evaluating each feature’s contribution to the final prediction across all possible combinations of those features. (233,234) Visualizations are another intuitive approach to aid our understanding of predictive models. For example, graph visualization can depict the predicted bonds between an interacting protein and ligand, and “saliency maps” (235) can highlight specific subregions of protein and ligand that are the most influential in a prediction, by discerning how subtle perturbations in individual input features affect the output. Several avenues for interpretability remain to be tested, (236) but none have been established as standard. Determining a reliable interpretability method for PLI prediction models will be critical for the field.
Another important aspect of modeling protein–ligand interactions is generalizability─or how well a model performs on unseen data. During the evaluation of machine learning models, test sets are typically selected with a presumed a priori understanding of the expected sample distribution to ensure accurate evaluation. However, the true sample distribution may differ, and it is important that a model can accommodate potentially unseen variations of input data. Although many PLIs have been identified to date, the full scope and distribution of all possible protein interactions remains unknown. However, there exist several means through which generalizability can be improved, including the production of additional data novel examples, reducing diversity bias, different strategies for splitting data into training and test sets, or alternative training schema. (214,237)
A similar task for which highly generalizable models have emerged is protein language modeling, where patterns are observed from analyzing massive data sets of protein sequences for purposes such as predicting protein stability or studying the evolutionary relationships between proteins. (189,238) While protein language models (PLMs) have achieved great predictive success, they require immense amounts of diverse data. Although the total number of unique tokens is much smaller than for human languages, protein sequences may contain far more tokens in total for data sets than for human languages. For example, UniProt’s UniRef50, (93,239) totals over 9.5 billion amino acids in length, and assuming that each AA is a token, that is a substantially larger corpus than most NLP data sets. (194) Currently, there may not be enough data available forPLI studies to train on the same scale as in PLM studies. However, with high-throughput analysis and the natural progression of PLI prediction studies, this may eventually be feasible.

5.4. The Insufficiency of an NLP-Only Approach for PLI Studies?

While NLP offers beneficial strategies for the study of PLIs, it is not a panacea, and there may be opportunities from other disciplines within computer science to contribute to the study of PLIs. For example, computer vision techniques may be favorable to use in handling structural information over NLP techniques designed to handle text. (240) Complementary approaches to NLP such as multimodal methods that integrate information from images and textual descriptions, can be applied to capture richer representations. (185,204) More advanced architectures, such as those exploring generative modeling, offer further avenues for integrating diverse data sources. (202) While the success of such hybrid strategies has yet to exceed the performance of other neural networks in PLI predictive tasks, (241) they demonstrate the potential for innovation by taking inspiration from other subdomains of ML and computer science beyond NLP.
Approaches informed by a deep domain-specific understanding have led to the practical success of ML methods. This has been demonstrated by the AlphaFold initiative in which nuanced awareness guided which biological features merited focus. (83,162) For example, the researchers behind AlphaFold-Multimer’s protein–protein interaction prediction algorithm (242) created an interface-aware protocol that crops protein structures to reduce computational burden and decrease the representation of non-interfacial amino acids while maintaining an important balance of interacting and non-interacting regions. Although AlphaFold-Multimer performs very well on predicting protein complexes in several cases, (243,244) preliminary results suggest that the more recently released AlphaFold 3 may offer further improvements. (89)
Whereas AlphaFold2 used a highly specified geometry-based module to generate protein structures, AlphaFold3 (89) incorporates a diffusion model similar to those that are popular in image generation tasks. (245,246) The diffusion model (247) of AlphaFold3 begins with a “noise” cloud of atoms placed at random and then iteratively converges to an accurate representation of the input sequences’ 3D structure. This initial inclusion of “noise” induces the model to refine the local structure rather than quickly converging to a local minimum. Whereas the previous AlphaFold2 geometry-based module was specific to proteins alone, a simplified diffusion model allows for the prediction of protein interactions with biological objects such as nucleic acids and small molecules. AlphaFold3 has been a significant advance, outperforming both molecular docking tools and diffusion-based-only models on structure prediction tasks. (248) The recent release of AlphaFold3’s open-source code makes the model highly accessible, allowing researchers to examine new predictions of protein interactions.
Due to the recency of AlphaFold3's release in May 2024, independent validation of AlphaFold3’s predictions has thus far been limited. While studies have looked into AlphaFold3’s limitations on protein–protein interactions (249) and protein-nucleic acid interactions, (250,251) the limitations associated with predicting PLIs are unclear. Studies to date suggest that AlphaFold3 has difficulties predicting accurate ligand-binding poses, pocket shape, and the assembly of domains for flexible proteins. (252) In a case study of flexible domains of receptor proteins, AlphaFold3 was shown to generate plausible but not the most stable conformations of proteins. AlphaFold3 predictions represent just onestable, averaged conformation based on inferred patterns within the training data, while ground-truth experimental methods like cryo-EM capture stable states that may be influenced by particular environmental contexts. Other possible limitations include how AlphaFold3 predicts some categories of interactions more accurately than others, the possibility of model hallucinations, and restrictive hardware requirements to run the model. (251,253) AlphaFold3 is a powerful tool that is effective for general use, but only time will tell how it, along with other competing tools like RoseTTAFold (254) and OpenFold, (255) will perform in future PLI studies.
The study of PLIs may eventually outgrow NLP methods, but for the foreseeable future, advances in NLP have established a strong foundation for processing texts representing biological objects. NLP still plays a key role in text-driven tasks such as the de novo generation of SMILES strings for automated molecular design. (256−258) Regardless, machine learning-based PLI studies will need to rely on close collaborations between experts in both biological and computational domains to catalyze further innovations in what is an interdisciplinary goal.

6. Conclusion

Click to copy section linkSection link copied!
Natural language processing (NLP), a subdiscipline of machine learning (ML), offers myriad tools for both experimental and computational researchers to accelerate exploratory studies in structural biology. The prediction of protein–ligand interactions (PLIs) can be reimagined through NLP by treating protein and ligand representations like text. Protein sequences resemble readable text with inherent meaning to be inferred, while chemical text formats such as theSMILES allows for limited NLP application to small molecules. Current efforts seek to leverage multiple or augmented SMILES representations to address these limitations.
Approaches to tackling PLI prediction tasks using sequence-only data, structural data, or a combination of both, have all yielded successful predictions, although the advantage of one input data type over others remains unclear. Sequence-only data approaches are simple and amenable to NLP but requires a significant abstraction of chemical information; structural data is informationally rich but computationally expensive to handle, while combining both sequence and structural data types offers balance at the expense of complexity.
The transformer architecture, in general, and attention mechanisms, in particular, have yielded the most promising NLP-based PLI prediction results to date. Incorporating complementary data (e.g., multiple sequence alignments, ligand polarities, etc.) can improve predictive success but at a significant increase in computational cost. After data selection and preparation, all methods have followed a general ML Extract-Fuse-Predict model creation framework of: (i) extracting feature embeddings for protein and ligand, (ii) fusing protein and ligand embeddings, and (iii) making predictions based on the created ML model.
The first step of data set selection is crucial for any ML-based study of PLIs, and no single data set can satisfy all needs, with many suffering from missing data or the lack of negative data. Data sets must align with specific research goals, requiring thoughtful consideration as to what inputs, formats, and target variable(s) are selected for the ML model. Appropriate tokenization and embedding methods, which convert proteins and ligands into numerical representations, are vital for a successful model. Atoms or amino acids typically serve as tokens, and neural networks (NNs) have helped identify hidden patterns more quickly. NLP-inspired NNs, such as Long Short-Term Memory NNs, along with attention mechanisms and transformer architectures, have shown particular promise for understanding PLIs. A modular approach combining multiple embeddings can capture diverse perspectives, improving prediction accuracy, especially for the prediction of binding affinities. After appropriate embeddings are obtained, graph-based methods and cross-attention mechanisms have been shown to be effective in combining data from diverse sources.
NLP has been central to ML studies of PLIs and has yielded promising results, although many challenges remain. Explaining ML model predictions is essential for their trustworthiness and acceptance. Current explanatory metrics, such as attention weights and Shapley values, offer some degree of interpretability but remain to be fully validated. A major challenge is the lack of well-annotated non-binding protein–ligand pairs, or “negative data”. Unsupervised methods or manually curated selections of non-binding pairs are potential solutions. Popular PLI data sets may contain biases that cause models to “memorize” idiosyncratic patterns rather than “learn” the true mechanics of PLIs. Ensuring balanced training data sets (positive vs. negative data, number of proteins vs. ligands, etc.) would be essential to avoid such bias.
As protein and ligand sequence representations differ from human language, it may be difficult to capture their complexity with NLP methods alone, especially as much of the variation in protein function can often be explained by simple amino acid interactions rather than complex higher-order interactions. (259) While NLP has contributed significantly to the advance of PLI studies, future improvements may come from both modifying machine learning architectures and incorporating nuanced biological domain knowledge. For instance, the researchers behind AlphaFold-Multimer’s protein–protein interaction prediction algorithm (242) created an interface-aware protocol that crops protein structures to reduce computational burden and the representation of non-interfacial amino acids while maintaining an important balance of interacting and non-interacting regions. Some researchers have also integrated mass spectrometry data to improve model predictions of protein complexes. (260) More recently in AlphaFold3, (89) a diffusion layer has been added to Alphafold’s previous workflow to enable the study of PLIs. Time will tell to what degree AlphaFold3 will advance predictions of PLIs but progress in PLI research will undoubtedly require interdisciplinary collaborations between computer scientists, chemists, and biologists.
Although it is best practice to evaluate model performance against ground-truth experimental results or results from physics-based computer simulations, few studies to date have benchmarked their model predictions in this way. Formal competition may prove to be a promising avenue for future advances in PLI prediction. Other grand challenges, such as protein folding and protein assembly, have had significant progress facilitated through competitions like Critical Assessment of Structural Prediction (CASP) (100) and Critical Assessment of Prediction of Interactions (CAPRI). (101,102) These well-adjudicated competitions use unpublished test sets for objective model comparisons. Milestone algorithms like AlphaFold (261) and RosettaFold (254) were formed, improved, and refined through the crucible of such contests. Creating a dedicated competition devoted to protein–ligand interactions could similarly inspire innovation and catalyze seminal algorithmic advances for PLI prediction.

Data Availability

Click to copy section linkSection link copied!

No data or software was generated for this review.

Author Information

Click to copy section linkSection link copied!
  • Corresponding Authors
  • Authors
    • James Michels - Department of Computer and Information Science, University of Mississippi, University, Mississippi 38677, United States
    • Ramya Bandarupalli - Department of BioMolecular Sciences, School of Pharmacy, University of Mississippi, University, Mississippi 38677, United States
    • Amin Ahangar Akbari - Department of BioMolecular Sciences, School of Pharmacy, University of Mississippi, University, Mississippi 38677, United States
    • Thai Le - Department of Computer Science, Indiana University, Bloomington, Indiana 47408, United States
  • Author Contributions

    JM: Conceptualization (lead); investigation (lead); project administration (lead); visualization (lead); writing─original draft preparation; writing─review and editing (co-lead). RB: Investigation (supporting); visualization (supporting); review and editing (equal). AAA: Investigation (supporting); review and editing (equal). TL: Conceptualization (supporting); funding acquisition (equal). HX: Conceptualization (supporting); supervision (supporting); review and editing (equal). JL: Conceptualization (supporting); funding acquisition (equal); review and editing (equal). EH: Conceptualization (supporting); funding acquisition (equal); project administration (supporting); supervision (lead); writing─review and editing (co-lead).

  • Notes
    The authors declare no competing financial interest.

Acknowledgments

Click to copy section linkSection link copied!

This work was supported in part by NIGMS/NIH Institutional Development Award (IDeA) #P20GM130460 to J.L, NSF award #1846376 to E.F.Y.H, and University of Mississippi Data Science/AI Research Seed Grant award #SB3002 IDS RSG-03 to J.M., J.L., T.L, and E.F.Y.H.

References

Click to copy section linkSection link copied!

This article references 271 other publications.

  1. 1
    Songyang, Z.; Cantley, L. C. Recognition and specificity in protein tyrosine kinase-mediated signalling. Trends Biochem. Sci. 1995, 20, 470– 475,  DOI: 10.1016/S0968-0004(00)89103-3
    Google Scholar
    1
    Recognition and specificity in protein tyrosine kinase-mediated signaling
    Songyang, Zhou; Cantley, Lewis C.
    Trends in Biochemical Sciences (1995), 20 (11), 470-5CODEN: TBSCDB; ISSN:0968-0004. (Elsevier Trends Journals)
    A review, with 46 refs. There are several factors that contribute to the specificities of protein tyrosine kinases (PTKs) in signal transduction pathways. While protein-protein interaction domains, such as the Src homol. (SH2 and SH3) domains, regulate the cellular localization of PTKs and their substrates, the specificities of PTKs are ultimately detd. by their catalytic domains. The use of peptide libraries has revealed the substrate specificities of SH2 domains and PTK catalytic domains, and has suggested cross-talk between these domains.
  2. 2
    Johnson, L. N.; Lowe, E. D.; Noble, M. E.; Owen, D. J. The Eleventh Datta Lecture. The structural basis for substrate recognition and control by protein kinases. FEBS Lett. 1998, 430, 1– 11,  DOI: 10.1016/S0014-5793(98)00606-1
    Google Scholar
    2
    The structural basis for substrate recognition and control by protein kinases
    Johnson, Louise N.; Lowe, Edward D.; Noble, Martin E. M.; Owen, David J.
    FEBS Letters (1998), 430 (1,2), 1-11CODEN: FEBLAL; ISSN:0014-5793. (Elsevier Science B.V.)
    A review with 49 refs. Protein kinases catalyze phospho transfer reactions from ATP to serine, threonine or tyrosine residues in target substrates and provide key mechanisms for control of cellular signaling processes. The crystal structures of 12 protein kinases are now known. These include structures of kinases in the active state in ternary complexes with ATP (or analogs) and inhibitor or peptide substrates (e.g. cAMP dependent protein kinase, phosphorylase kinase and insulin receptor tyrosine kinase); kinases in both active and inactive states (e.g., CDK2/cyclin A, insulin receptor tyrosine kinase and MAPK); kinases in the active state (e.g. casein kinase 1, Lck); and kinases in inactive states (e.g. twitchin kinase, calcium calmodulin kinase 1, FGF receptor kinase, c-Src and Hck). This paper summarizes the detailed information obtained with active phosphorylase kinase ternary complex and reviews the results with ref. to other kinase structures for insights into mechanisms for substrate recognition and control.
  3. 3
    Kristiansen, K. Molecular mechanisms of ligand binding, signaling, and regulation within the superfamily of G-protein-coupled receptors: molecular modeling and mutagenesis approaches to receptor structure and function. Pharmacol. Ther. 2004, 103, 21– 80,  DOI: 10.1016/j.pharmthera.2004.05.002
    Google Scholar
    3
    Molecular mechanisms of ligand binding, signaling, and regulation within the superfamily of G-protein-coupled receptors: molecular modeling and mutagenesis approaches to receptor structure and function
    Kristiansen, Kurt
    Pharmacology & Therapeutics (2004), 103 (1), 21-80CODEN: PHTHDT; ISSN:0163-7258. (Elsevier Science B.V.)
    A review. The superfamily of G-protein-coupled receptors (GPCRs) could be subclassified into 7 families (A, B, large N-terminal family B-7 transmembrane helix, C, Frizzled/Smoothened, taste 2, and vomeronasal 1 receptors) among mammalian species. Cloning and functional studies of GPCRs have revealed that the superfamily of GPCRs comprises receptors for chem. diverse native ligands including endogenous compds. like amines, peptides, and Wnt proteins (i.e., secreted proteins activating Frizzled receptors); endogenous cell surface adhesion mols.; and photons and exogenous compds. like odorants. The combined use of site-directed mutagenesis and mol. modeling approaches have provided detailed insight into mol. mechanisms of ligand binding, receptor folding, receptor activation, G-protein coupling, and regulation of GPCRs. The vast majority of family A, B, C, vomeronasal 1, and taste 2 receptors are able to transduce signals into cells through G-protein coupling. However, G-protein-independent signaling mechanisms have also been reported for many GPCRs. Specific interaction motifs in the intracellular parts of these receptors allow them to interact with scaffold proteins. Protein engineering techniques have provided information on mol. mechanisms of GPCR-accessory protein, GPCR-GPCR, and GPCR-scaffold protein interactions. Site-directed mutagenesis and mol. dynamics simulations have revealed that the inactive state conformations are stabilized by specific interhelical and intrahelical salt bridge interactions and hydrophobic-type interactions. Constitutively activating mutations or agonist binding disrupts such constraining interactions leading to receptor conformations that assocs. with and activate G-proteins.
  4. 4
    West, I. C. What determines the substrate specificity of the multi-drug-resistance pump?. Trends Biochem. Sci. 1990, 15, 42– 46,  DOI: 10.1016/0968-0004(90)90171-7
    Google Scholar
    There is no corresponding record for this reference.
  5. 5
    Vivier, E.; Malissen, B. Innate and adaptive immunity: specificities and signaling hierarchies revisited. Nat. Immunol. 2005, 6, 17– 21,  DOI: 10.1038/ni1153
    Google Scholar
    5
    Innate and adaptive immunity: specificities and signaling hierarchies revisited
    Vivier, Eric; Malissen, Bernard
    Nature Immunology (2005), 6 (1), 17-21CODEN: NIAMCZ; ISSN:1529-2908. (Nature Publishing Group)
    A review. The conventional classification of known immune responses by specificity may need re-evaluation. The immune system can be classified into two subsystems: the innate and adaptive immune systems. In general, innate immunity is considered a nonspecific response, whereas the adaptive immune system is thought of as being very specific. In addn., the antigen receptors of the adaptive immune response are commonly viewed as 'master sensors' whose engagement dictates lymphocyte function. Here the authors propose that these ideas do not genuinely reflect the organization of immune responses and that they bias the authors' view of immunity as well as the authors' teaching of immunol. Indeed, the level of specificity and mode of signaling integration used by the main cellular participants in the adaptive and innate immune systems are more similar than previously appreciated.
  6. 6
    Desvergne, B.; Michalik, L.; Wahli, W. Transcriptional regulation of metabolism. Physiol. Rev. 2006, 86, 465– 514,  DOI: 10.1152/physrev.00025.2005
    Google Scholar
    6
    Transcriptional regulation of metabolism
    Desvergne, Beatrice; Michalik, Liliane; Wahli, Walter
    Physiological Reviews (2006), 86 (2), 465-514CODEN: PHREA7; ISSN:0031-9333. (American Physiological Society)
    A review. Our understanding of metab. is undergoing a dramatic shift. Indeed, the efforts made towards elucidating the mechanisms controlling the major regulatory pathways are now being rewarded. At the mol. level, the crucial role of transcription factors is particularly well-illustrated by the link between alterations of their functions and the occurrence of major metabolic diseases. In addn., the possibility of manipulating the ligand-dependent activity of some of these transcription factors makes them attractive as therapeutic targets. The aim of this review is to summarize recent knowledge on the transcriptional control of metabolic homeostasis. We first review data on the transcriptional regulation of the intermediary metab., i.e., glucose, amino acid, lipid, and cholesterol metab. Then, we analyze how transcription factors integrate signals from various pathways to ensure homeostasis. One example of this coordination is the daily adaptation to the circadian fasting and feeding rhythm. This section also discusses the dysregulations causing the metabolic syndrome, which reveals the intricate nature of glucose and lipid metab. and the role of the transcription factor PPARγ in orchestrating this assocn. Finally, we discuss the mol. mechanisms underlying metabolic regulations, which provide new opportunities for treating complex metabolic disorders.
  7. 7
    Atkinson, D. E. Biological feedback control at the molecular level: Interaction between metabolite-modulated enzymes seems to be a major factor in metabolic regulation. Science 1965, 150, 851– 857,  DOI: 10.1126/science.150.3698.851
    Google Scholar
    There is no corresponding record for this reference.
  8. 8
    Huang, S.-Y.; Zou, X. Advances and challenges in protein-ligand docking. Int. J. Mol. Sci. 2010, 11, 3016– 3034,  DOI: 10.3390/ijms11083016
    Google Scholar
    8
    Advances and challenges in protein-ligand docking
    Huang, Sheng-You; Zou, Xiaoqin
    International Journal of Molecular Sciences (2010), 11 (), 3016-3034CODEN: IJMCFK; ISSN:1422-0067. (Molecular Diversity Preservation International)
    A review. Mol. docking is a widely-used computational tool for the study of mol. recognition, which aims to predict the binding mode and binding affinity of a complex formed by two or more constituent mols. with known structures. An important type of mol. docking is protein-ligand docking because of its therapeutic applications in modern structure-based drug design. Here, we review the recent advances of protein flexibility, ligand sampling, and scoring functions - the three important aspects in protein-ligand docking. Challenges and possible future directions are discussed in the conclusion.
  9. 9
    Chaires, J. B. Calorimetry and thermodynamics in drug design. Annu. Rev. Biophys. 2008, 37, 135– 151,  DOI: 10.1146/annurev.biophys.36.040306.132812
    Google Scholar
    9
    Calorimetry and thermodynamics in drug design
    Chaires, Jonathan B.
    Annual Review of Biophysics (2008), 37 (), 135-151CODEN: ARBNCV ISSN:. (Annual Reviews Inc.)
    A review. Modern instrumentation for calorimetry permits direct detn. of enthalpy values for binding reactions and conformational transitions in biomols. Complete thermodn. profiles consisting of free energy, enthalpy, and entropy may be obtained for reactions of interest in a relatively straightforward manner. Such profiles are of enormous value in drug design because they provide information about the balance of driving forces that cannot be obtained from structural or computational methods alone. This perspective shows several examples of the insight provided by thermodn. data in drug design.
  10. 10
    Serhan, C. N. Signalling the fat controller. Nature 1996, 384, 23– 24,  DOI: 10.1038/384023a0
    Google Scholar
    There is no corresponding record for this reference.
  11. 11
    McAllister, C. H.; Beatty, P. H.; Good, A. G. Engineering nitrogen use efficient crop plants: the current status: Engineering nitrogen use efficient crop plants. Plant Biotechnol. J. 2012, 10, 1011– 1025,  DOI: 10.1111/j.1467-7652.2012.00700.x
    Google Scholar
    There is no corresponding record for this reference.
  12. 12
    Goldsmith, M.; Tawfik, D. S. Enzyme engineering: reaching the maximal catalytic efficiency peak. Curr. Opin. Struct. Biol. 2017, 47, 140– 150,  DOI: 10.1016/j.sbi.2017.09.002
    Google Scholar
    12
    Enzyme engineering: reaching the maximal catalytic efficiency peak
    Goldsmith, Moshe; Tawfik, Dan S.
    Current Opinion in Structural Biology (2017), 47 (), 140-150CODEN: COSBEF; ISSN:0959-440X. (Elsevier Ltd.)
    A review. The practical need for highly efficient enzymes presents new challenges in enzyme engineering, in particular, the need to improve catalytic turnover (kcat) or efficiency (kcat/KM) by several orders of magnitude. However, optimizing catalysis demands navigation through complex and rugged fitness landscapes, with optimization trajectories often leading to strong diminishing returns and dead-ends. When no further improvements are obsd. in library screens or selections, it remains unclear whether the maximal catalytic efficiency of the enzyme (the catalytic 'fitness peak') has been reached; or perhaps, an alternative combination of mutations exists that could yield addnl. improvements. Here, we discuss fundamental aspects of the process of catalytic optimization, and offer practical solns. with respect to overcoming optimization plateaus.
  13. 13
    Vajda, S.; Guarnieri, F. Characterization of protein-ligand interaction sites using experimental and computational methods. Curr. Opin. Drug Discovery Devel. 2006, 9, 354– 362
    Google Scholar
    13
    Characterization of protein-ligand interaction sites using experimental and computational methods
    Vajda, Sandor; Guarnieri, Frank
    Current Opinion in Drug Discovery & Development (2006), 9 (3), 354-362CODEN: CODDFF; ISSN:1367-6733. (Thomson Scientific)
    A review. The ability to identify the sites of a protein that can bind with high affinity to small, drug-like compds. has been an important goal in drug design. Accurate prediction of druggable sites and the identification of small compds. binding in those sites have provided the input for fragment-based combinatorial approaches that allow for a more thorough exploration of the chem. space, and that have the potential to yield mols. that are more lead-like than those found using traditional high-throughput screening. Current progress in exptl. and computational methods for identifying and characterizing druggable ligand binding sites on protein targets is reviewed herein, including a discussion of successful NMR, x-ray crystallog. and tethering technologies. Classical geometric and energy-based computational methods are also discussed, with particular focus on two powerful technologies, i.e., computational solvent mapping and grand canonical Monte Carlo simulations (as used by Locus Pharmaceuticals Inc). Both methods can be used to reliably identify druggable sites on proteins and to facilitate the design of novel, low-nanomolar-affinity ligands.
  14. 14
    Du, X.; Li, Y.; Xia, Y.-L.; Ai, S.-M.; Liang, J.; Sang, P.; Ji, X.-L.; Liu, S.-Q. Insights into protein-ligand interactions: Mechanisms, models, and methods. Int. J. Mol. Sci. 2016, 17, 144,  DOI: 10.3390/ijms17020144
    Google Scholar
    14
    Insights into protein-ligand interactions: mechanisms, models, and methods
    Du, Xing; Li, Yi; Xia, Yuan-Ling; Ai, Shi-Meng; Liang, Jing; Sang, Peng; Ji, Xing-Lai; Liu, Shu-Qun
    International Journal of Molecular Sciences (2016), 17 (2), 144/1-144/34CODEN: IJMCFK; ISSN:1422-0067. (MDPI AG)
    Mol. recognition, which is the process of biol. macromols. interacting with each other or various small mols. with a high specificity and affinity to form a specific complex, constitutes the basis of all processes in living organisms. Proteins, an important class of biol. macromols., realize their functions through binding to themselves or other mols. A detailed understanding of the protein-ligand interactions is therefore central to understanding biol. at the mol. level. Moreover, knowledge of the mechanisms responsible for the protein-ligand recognition and binding will also facilitate the discovery, design, and development of drugs. In the present review, first, the physicochem. mechanisms underlying protein-ligand binding, including the binding kinetics, thermodn. concepts and relationships, and binding driving forces, are introduced and rationalized. Next, three currently existing protein-ligand binding models-the "lock-and-key", "induced fit", and "conformational selection"-are described and their underlying thermodn. mechanisms are discussed. Finally, the methods available for investigating protein-ligand binding affinity, including exptl. and theor./computational approaches, are introduced, and their advantages, disadvantages, and challenges are discussed.
  15. 15
    Fan, F. J.; Shi, Y. Effects of data quality and quantity on deep learning for protein-ligand binding affinity prediction. Bioorg. Med. Chem. 2022, 72, 117003,  DOI: 10.1016/j.bmc.2022.117003
    Google Scholar
    15
    Effects of data quality and quantity on deep learning for protein-ligand binding affinity prediction
    Fan, Frankie J.; Shi, Yun
    Bioorganic & Medicinal Chemistry (2022), 72 (), 117003CODEN: BMECEP; ISSN:0968-0896. (Elsevier B.V.)
    Prediction of protein-ligand binding affinities is crucial for computational drug discovery. A no. of deep learning approaches have been developed in recent years to improve the accuracy of such affinity prediction. While the predicting power of these systems have advanced to some degrees depending on the dataset used for model training and testing, the effects of the quality and quantity of the underlying data have not been thoroughly examd. In this study, we employed erroneous datasets and data subsets of different sizes, created from one of the largest databases of exptl. binding affinities, to train and evaluate a deep learning system based on convolutional neural networks. Our results show that data quality and quantity do have significant impacts on the prediction performance of trained models. Depending on the variations in data quality and quantity, the performance discrepancies could be comparable to or even larger than those obsd. among different deep learning approaches. In particular, the presence of proteins in the training data leads to a dramatic increase in prediction accuracy. This implies that continued accumulation of high-quality affinity data, esp. for new protein targets, is indispensable for improving deep learning models to better predict protein-ligand binding affinities.
  16. 16
    Sousa, S. F.; Ribeiro, A. J. M.; Coimbra, J. T. S.; Neves, R. P. P.; Martins, S. A.; Moorthy, N. S. H. N.; Fernandes, P. A.; Ramos, M. J. Protein-Ligand Docking in the New Millennium A Retrospective of 10 Years in the Field. Curr. Med. Chem. 2013, 20, 2296– 2314,  DOI: 10.2174/0929867311320180002
    Google Scholar
    There is no corresponding record for this reference.
  17. 17
    Morris, C. J.; Corte, D. D. Using molecular docking and molecular dynamics to investigate protein-ligand interactions. Mod. Phys. Lett. B 2021, 35, 2130002,  DOI: 10.1142/S0217984921300027
    Google Scholar
    17
    Using molecular docking and molecular dynamics to investigate protein-ligand interactions
    Morris, Connor J.; Corte, Dennis Della
    Modern Physics Letters B (2021), 35 (8), 2130002CODEN: MPLBET; ISSN:0217-9849. (World Scientific Publishing Co. Pte. Ltd.)
    A review. Mol. docking and mol. dynamics (MD) are powerful tools used to investigate protein-ligand interactions. Mol. docking programs predict the binding pose and affinity of a protein-ligand complex, while MD can be used to incorporate flexibility into docking calcns. and gain further information on the kinetics and stability of the protein-ligand bond. This review covers state-of-the-art methods of using mol. docking and MD to explore protein-ligand interactions, with emphasis on application to drug discovery. We also call for further research on combining common mol. docking and MD methods.
  18. 18
    Lecina, D.; Gilabert, J. F.; Guallar, V. Adaptive simulations, towards interactive protein-ligand modeling. Sci. Rep. 2017, 7, 8466,  DOI: 10.1038/s41598-017-08445-5
    Google Scholar
    18
    Adaptive simulations, towards interactive protein-ligand modeling
    Lecina Daniel; Gilabert Joan F; Guallar Victor; Guallar Victor
    Scientific reports (2017), 7 (1), 8466 ISSN:.
    Modeling the dynamic nature of protein-ligand binding with atomistic simulations is one of the main challenges in computational biophysics, with important implications in the drug design process. Although in the past few years hardware and software advances have significantly revamped the use of molecular simulations, we still lack a fast and accurate ab initio description of the binding mechanism in complex systems, available only for up-to-date techniques and requiring several hours or days of heavy computation. Such delay is one of the main limiting factors for a larger penetration of protein dynamics modeling in the pharmaceutical industry. Here we present a game-changing technology, opening up the way for fast reliable simulations of protein dynamics by combining an adaptive reinforcement learning procedure with Monte Carlo sampling in the frame of modern multi-core computational resources. We show remarkable performance in mapping the protein-ligand energy landscape, being able to reproduce the full binding mechanism in less than half an hour, or the active site induced fit in less than 5 minutes. We exemplify our method by studying diverse complex targets, including nuclear hormone receptors and GPCRs, demonstrating the potential of using the new adaptive technique in screening and lead optimization studies.
  19. 19
    Ikebata, H.; Hongo, K.; Isomura, T.; Maezono, R.; Yoshida, R. Bayesian molecular design with a chemical language model. J. Comput. Aided Mol. Des. 2017, 31, 379– 391,  DOI: 10.1007/s10822-016-0008-z
    Google Scholar
    There is no corresponding record for this reference.
  20. 20
    Rives, A.; Meier, J.; Sercu, T.; Goyal, S.; Lin, Z.; Liu, J.; Guo, D.; Ott, M.; Zitnick, C. L.; Ma, J.; Fergus, R. Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences. Proc. Natl. Acad. Sci. U. S. A. 2021,  DOI: 10.1073/pnas.2016239118
    Google Scholar
    There is no corresponding record for this reference.
  21. 21
    Cao, Y.; Shen, Y. TALE: Transformer-based protein function Annotation with joint sequence–Label Embedding. Bioinformatics 2021, 37, 2825– 2833,  DOI: 10.1093/bioinformatics/btab198
    Google Scholar
    There is no corresponding record for this reference.
  22. 22
    Wang, S.; Guo, Y.; Wang, Y.; Sun, H.; Huang, J. SMILES-BERT: Large Scale Unsupervised Pre-Training for Molecular Property Prediction. Proceedings of the 10th ACM International Conference on Bioinformatics, Computational Biology and Health Informatics. New York, NY, USA 2019, 429– 436
    Google Scholar
    There is no corresponding record for this reference.
  23. 23
    Chithrananda, S.; Grand, G.; Ramsundar, B. ChemBERTa: Large-Scale Self-Supervised Pretraining for Molecular Property Prediction. arXiv , 2020.
    Google Scholar
    There is no corresponding record for this reference.
  24. 24
    Kumar, N.; Acharya, V. Machine intelligence-driven framework for optimized hit selection in virtual screening. J. Cheminform. 2022, 14, 48,  DOI: 10.1186/s13321-022-00630-7
    Google Scholar
    There is no corresponding record for this reference.
  25. 25
    Erikawa, D.; Yasuo, N.; Sekijima, M. MERMAID: an open source automated hit-to-lead method based on deep reinforcement learning. J. Cheminform. 2021, 13, 94,  DOI: 10.1186/s13321-021-00572-6
    Google Scholar
    There is no corresponding record for this reference.
  26. 26
    Zhou, M.; Duan, N.; Liu, S.; Shum, H.-Y. Progress in neural NLP: Modeling, learning, and reasoning. Engineering (Beijing) 2020, 6, 275– 290,  DOI: 10.1016/j.eng.2019.12.014
    Google Scholar
    There is no corresponding record for this reference.
  27. 27
    Patwardhan, N.; Marrone, S.; Sansone, C. Transformers in the real world: A survey on NLP applications. Inf. 2023, 14, 242,  DOI: 10.3390/info14040242
    Google Scholar
    There is no corresponding record for this reference.
  28. 28
    Bijral, R. K.; Singh, I.; Manhas, J.; Sharma, V. Exploring Artificial Intelligence in Drug Discovery: A Comprehensive Review. Arch. Comput. Methods Eng. 2022, 29, 2513– 2529,  DOI: 10.1007/s11831-021-09661-z
    Google Scholar
    There is no corresponding record for this reference.
  29. 29
    Ray, P. P. ChatGPT: A comprehensive review on background, applications, key challenges, bias, ethics, limitations and future scope. Internet of Things and Cyber-Physical Systems 2023, 3, 121– 154,  DOI: 10.1016/j.iotcps.2023.04.003
    Google Scholar
    There is no corresponding record for this reference.
  30. 30
    Radford, A.; Narasimhan, K.; Salimans, T.; Sutskever, I. Improving language understanding by generative pre-training. https://www.mikecaptain.com/resources/pdf/GPT-1.pdf, Accessed: 2023–10–27.
    Google Scholar
    There is no corresponding record for this reference.
  31. 31
    Goodside, R, Papay, Meet Claude: Anthropic’s Rival to ChatGPT. https://scale.com/blog/chatgpt-vs-claude, 2023.
    Google Scholar
    There is no corresponding record for this reference.
  32. 32
    Bing Copilot. Bing Copilot; https://copilot.microsoft.com/.
    Google Scholar
    There is no corresponding record for this reference.
  33. 33
    Rahul; Adhikari, S.; Monika NLP based Machine Learning Approaches for Text Summarization. 2020 Fourth International Conference on Computing Methodologies and Communication (ICCMC) 2020, 535– 538
    Google Scholar
    There is no corresponding record for this reference.
  34. 34
    Nasukawa, T.; Yi, J. Sentiment analysis: capturing favorability using natural language processing. Proceedings of the 2nd international conference on Knowledge capture. New York, NY, USA 2003, 70– 77
    Google Scholar
    There is no corresponding record for this reference.
  35. 35
    Lample, G.; Charton, F. Deep Learning for Symbolic Mathematics. arXiv , 2019.
    Google Scholar
    There is no corresponding record for this reference.
  36. 36
    Feng, Z.; Guo, D.; Tang, D.; Duan, N.; Feng, X.; Gong, M.; Shou, L.; Qin, B.; Liu, T.; Jiang, D.; Zhou, M. CodeBERT: APre-Trained Model for Programming and Natural Languages. arXiv , 2020.
    Google Scholar
    There is no corresponding record for this reference.
  37. 37
    Mielke, S. J.; Alyafeai, Z.; Salesky, E.; Raffel, C.; Dey, M.; Gallé, M.; Raja, A.; Si, C.; Lee, W. Y.; Sagot, B.; Tan, S. Between words and characters: A Brief History of Open-Vocabulary Modeling and Tokenization in NLP. arXiv , 2021.
    Google Scholar
    There is no corresponding record for this reference.
  38. 38
    Camacho-Collados, J.; Pilehvar, M. T. From word to sense embeddings: A survey on vector representations of meaning. J. Artif. Intell. Res. 2018, 63, 743– 788,  DOI: 10.1613/jair.1.11259
    Google Scholar
    There is no corresponding record for this reference.
  39. 39
    Ashok, V. G.; Feng, S.; Choi, Y. Success with style: Using writing style to predict the success of novelsd.
    Google Scholar
    There is no corresponding record for this reference.
  40. 40
    Barberá, P.; Boydstun, A. E.; Linn, S.; McMahon, R.; Nagler, J. Automated text classification of news articles: A practical guide. Polit. Anal. 2021, 29, 19– 42,  DOI: 10.1017/pan.2020.8
    Google Scholar
    There is no corresponding record for this reference.
  41. 41
    Wang, H.; Wu, H.; He, Z.; Huang, L.; Church, K. W. Progress in machine translation. Engineering (Beijing) 2022, 18, 143– 153,  DOI: 10.1016/j.eng.2021.03.023
    Google Scholar
    There is no corresponding record for this reference.
  42. 42
    Sønderby, S. K.; Winther, O. Protein Secondary Structure Prediction with Long Short Term Memory Networks. arXiv , 2014.
    Google Scholar
    There is no corresponding record for this reference.
  43. 43
    Guo, Y.; Li, W.; Wang, B.; Liu, H.; Zhou, D. DeepACLSTM: deep asymmetric convolutional long short-term memory neural models for protein secondary structure prediction. BMC Bioinformatics 2019, 20, 341,  DOI: 10.1186/s12859-019-2940-0
    Google Scholar
    There is no corresponding record for this reference.
  44. 44
    Bhasuran, B.; Natarajan, J. Automatic extraction of gene-disease associations from literature using joint ensemble learning. PLoS One 2018, 13, e0200699,  DOI: 10.1371/journal.pone.0200699
    Google Scholar
    There is no corresponding record for this reference.
  45. 45
    Pang, M.; Su, K.; Li, M. Leveraging information in spatial transcriptomics to predict super-resolution gene expression from histology images in tumors. bioRxiv , 2021, 2021.11.28.470212.
    Google Scholar
    There is no corresponding record for this reference.
  46. 46
    Bouatta, N.; Sorger, P.; AlQuraishi, M. Protein structure prediction by AlphaFold2: are attention and symmetries all you need?. Acta Crystallogr. D Struct Biol. 2021, 77, 982– 991,  DOI: 10.1107/S2059798321007531
    Google Scholar
    There is no corresponding record for this reference.
  47. 47
    Skolnick, J.; Gao, M.; Zhou, H.; Singh, S. AlphaFold 2: Why It Works and Its Implications for Understanding the Relationships of Protein Sequence, Structure, and Function. J. Chem. Inf. Model. 2021, 61, 4827– 4831,  DOI: 10.1021/acs.jcim.1c01114
    Google Scholar
    47
    AlphaFold 2: Why It Works and Its Implications for Understanding the Relationships of Protein Sequence, Structure, and Function
    Skolnick, Jeffrey; Gao, Mu; Zhou, Hongyi; Singh, Suresh
    Journal of Chemical Information and Modeling (2021), 61 (10), 4827-4831CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    AlphaFold 2 (AF2) was the star of CASP14, the last biannual structure prediction expt. Using novel deep learning, AF2 predicted the structures of many difficult protein targets at or near exptl. resoln. Here, the authors present the authors' perspective of why AF2 works and show that it is a very sophisticated fold recognition algorithm that exploits the completeness of the library of single domain PDB structures. It also learned local side chain packing rearrangements that enable it to refine proteins to high resoln. The benefits and limitations of its ability to predict the structures of many more proteins at or close to at. detail are discussed.
  48. 48
    Adadi, A.; Berrada, M. Peeking inside the black-box: a survey on explainable artificial intelligence (XAI). IEEE Access 2018, 6, 52138,  DOI: 10.1109/ACCESS.2018.2870052
    Google Scholar
    There is no corresponding record for this reference.
  49. 49
    Box, G. E. P. Science and Statistics. J. Am. Stat. Assoc. 1976, 71, 791– 799,  DOI: 10.1080/01621459.1976.10480949
    Google Scholar
    There is no corresponding record for this reference.
  50. 50
    Geirhos, R.; Jacobsen, J.-H.; Michaelis, C.; Zemel, R.; Brendel, W.; Bethge, M.; Wichmann, F. A. Shortcut learning in deep neural networks. Nature Machine Intelligence 2020, 2, 665– 673,  DOI: 10.1038/s42256-020-00257-z
    Google Scholar
    There is no corresponding record for this reference.
  51. 51
    Outeiral, C.; Nissley, D. A.; Deane, C. M. Current structure predictors are not learning the physics of protein folding. Bioinformatics 2022, 38, 1881– 1887,  DOI: 10.1093/bioinformatics/btab881
    Google Scholar
    51
    Current structure predictors are not learning the physics of protein folding
    Outeiral, Carlos; Nissley, Daniel A.; Deane, Charlotte M.
    Bioinformatics (2022), 38 (7), 1881-1887CODEN: BOINFP; ISSN:1367-4811. (Oxford University Press)
    Predicting the native state of a protein has long been considered a gateway problem for understanding protein folding. Recent advances in structural modeling driven by deep learning have achieved unprecedented success at predicting a protein's crystal structure, but it is not clear if these models are learning the physics of how proteins dynamically fold into their equil. structure or are just accurate knowledge-based predictors of the final state. In this work, we compare the pathways generated by state-of-the-art protein structure prediction methods to exptl. data about protein folding pathways. The methods considered were AlphaFold 2, RoseTTAFold, trRosetta, RaptorX, DMPfold, EVfold, SAINT2 and Rosetta. We find evidence that their simulated dynamics capture some information about the folding pathway, but their predictive ability is worse than a trivial classifier using sequence-agnostic features like chain length. The folding trajectories produced are also uncorrelated with exptl. observables such as intermediate structures and the folding rate const. These results suggest that recent advances in structure prediction do not yet provide an enhanced understanding of protein folding.
  52. 52
    Steels, L. Modeling the cultural evolution of language. Phys. Life Rev. 2011, 8, 339– 356,  DOI: 10.1016/j.plrev.2011.10.014
    Google Scholar
    There is no corresponding record for this reference.
  53. 53
    Maurya, H. C.; Gupta, P.; Choudhary, N. Natural language ambiguity and its effect on machine learning. Int. J. Modern Eng. Res. 2015, 5, 25– 30
    Google Scholar
    There is no corresponding record for this reference.
  54. 54
    Tenney, I.; Xia, P.; Chen, B.; Wang, A.; Poliak, A.; McCoy, R. T.; Kim, N.; Van Durme, B.; Bowman, S. R.; Das, D.; Pavlick, E. What do you learn from context? Probing for sentence structure in contextualized word representations. arXiv , 2019.
    Google Scholar
    There is no corresponding record for this reference.
  55. 55
    Miyagawa, S.; Berwick, R. C.; Okanoya, K. The emergence of hierarchical structure in human language. Front. Psychol. 2013, 4, 71,  DOI: 10.3389/fpsyg.2013.00071
    Google Scholar
    There is no corresponding record for this reference.
  56. 56
    Liu, H.; Xu, C.; Liang, J. Dependency distance: A new perspective on syntactic patterns in natural languages. Phys. Life Rev. 2017, 21, 171– 193,  DOI: 10.1016/j.plrev.2017.03.002
    Google Scholar
    There is no corresponding record for this reference.
  57. 57
    Frank, S. L.; Bod, R.; Christiansen, M. H. How hierarchical is language use?. Proc. Biol. Sci. 2012, 279, 4522– 4531,  DOI: 10.1098/rspb.2012.1741
    Google Scholar
    There is no corresponding record for this reference.
  58. 58
    Oesch, N.; Dunbar, R. I. M. The emergence of recursion in human language: Mentalising predicts recursive syntax task performance. J. Neurolinguistics 2017, 43, 95– 106,  DOI: 10.1016/j.jneuroling.2016.09.008
    Google Scholar
    There is no corresponding record for this reference.
  59. 59
    Ferruz, N.; Höcker, B. Controllable protein design with language models. Nature Machine Intelligence 2022, 4, 521– 532,  DOI: 10.1038/s42256-022-00499-z
    Google Scholar
    There is no corresponding record for this reference.
  60. 60
    Ofer, D.; Brandes, N.; Linial, M. The language of proteins: NLP, machine learning & protein sequences. Comput. Struct. Biotechnol. J. 2021, 19, 1750– 1758,  DOI: 10.1016/j.csbj.2021.03.022
    Google Scholar
    60
    The language of proteins: NLP, machine learning & protein sequences
    Ofer, Dan; Brandes, Nadav; Linial, Michal
    Computational and Structural Biotechnology Journal (2021), 19 (), 1750-1758CODEN: CSBJAC; ISSN:2001-0370. (Elsevier B.V.)
    A review. Natural language processing (NLP) is a field of computer science concerned with automated text and language anal. In recent years, following a series of breakthroughs in deep and machine learning, NLP methods have shown overwhelming progress. Here, we review the success, promise and pitfalls of applying NLP algorithms to the study of proteins. Proteins, which can be represented as strings of amino-acid letters, are a natural fit to many NLP methods. We explore the conceptual similarities and differences between proteins and language, and review a range of protein-related tasks amenable to machine learning. We present methods for encoding the information of proteins as text and analyzing it with NLP methods, reviewing classic concepts such as bag-of-words, k-mers/n-grams and text search, as well as modern techniques such as word embedding, contextualized embedding, deep learning and neural language models. In particular, we focus on recent innovations such as masked language modeling, self-supervised learning and attention-based models. Finally, we discuss trends and challenges in the intersection of NLP and protein research.
  61. 61
    Ptitsyn, O. B. How does protein synthesis give rise to the 3D-structure?. FEBS Lett. 1991, 285, 176– 181,  DOI: 10.1016/0014-5793(91)80799-9
    Google Scholar
    There is no corresponding record for this reference.
  62. 62
    Yu, L.; Tanwar, D. K.; Penha, E. D. S.; Wolf, Y. I.; Koonin, E. V.; Basu, M. K. Grammar of protein domain architectures 2019, 116, 3636– 3645,  DOI: 10.1073/pnas.1814684116
    Google Scholar
    There is no corresponding record for this reference.
  63. 63
    Petsko, G. A.; Ringe, D. Protein Structure and Function; Primers in Biology; Blackwell Publishing: London, England, 2003.
    Google Scholar
    There is no corresponding record for this reference.
  64. 64
    Shenoy, S. R.; Jayaram, B. Proteins: sequence to structure and function-current status. Curr. Protein Pept. Sci. 2010, 11, 498– 514,  DOI: 10.2174/138920310794109094
    Google Scholar
    64
    Proteins: sequence to structure and function - current status
    Shenoy, Sandhya R.; Jayaram, B.
    Current Protein and Peptide Science (2010), 11 (7), 498-514CODEN: CPPSCM; ISSN:1389-2037. (Bentham Science Publishers Ltd.)
    A review. In an era that has been dominated by structural biol. for the last 30-40 yr, a dramatic change of focus toward sequence anal. has spurred the advent of the genome projects and the resultant diverging sequence/structure deficit. The central challenge of computational structural biol. is therefore to rationalize the mass of sequence information into biochem. and biophys. knowledge and to decipher the structural, functional, and evolutionary clues encoded in the language of biol. sequences. In investigating the meaning of sequences, 2 distinct anal. themes have emerged: (1) in the 1st approach, pattern recognition techniques are used to detect similarity between sequences and hence to infer related structures and functions; (2) in the 2nd, ab initio prediction methods are used to deduce 3-dimensional structure, and ultimately to infer function, directly from the linear sequence. Here, the authors attempt to provide a crit. assessment of what one may and may not expect from the biol. sequences and to identify major issues yet to be resolved. The presentation is organized under several subtitles such as protein sequences, pattern recognition techniques, protein tertiary structure prediction, membrane protein bioinformatics, human proteome, protein-protein interactions, metabolic networks, potential drug targets based on simple sequence properties, disordered proteins, the sequence-structure relation, and chem. logic of protein sequences.
  65. 65
    Takahashi, M.; Maraboeuf, F.; Nordén, B. Locations of functional domains in the RecA protein. Overlap of domains and regulation of activities. Eur. J. Biochem. 1996, 242, 20– 28,  DOI: 10.1111/j.1432-1033.1996.0020r.x
    Google Scholar
    There is no corresponding record for this reference.
  66. 66
    Liang, W.; KaiYong, Z. Detecting “protein words” through unsupervised word segmentation. arXiv , 2014.
    Google Scholar
    There is no corresponding record for this reference.
  67. 67
    Kuntz, I. D.; Crippen, G. M.; Kollman, P. A.; Kimelman, D. Calculation of protein tertiary structure. J. Mol. Biol. 1976, 106, 983– 994,  DOI: 10.1016/0022-2836(76)90347-8
    Google Scholar
    There is no corresponding record for this reference.
  68. 68
    Rodrigue, N.; Lartillot, N.; Bryant, D.; Philippe, H. Site interdependence attributed to tertiary structure in amino acid sequence evolution. Gene 2005, 347, 207– 217,  DOI: 10.1016/j.gene.2004.12.011
    Google Scholar
    There is no corresponding record for this reference.
  69. 69
    Eisenhaber, F.; Persson, B.; Argos, P. Protein structure prediction: recognition of primary, secondary, and tertiary structural features from amino acid sequence. Crit. Rev. Biochem. Mol. Biol. 1995, 30, 1– 94,  DOI: 10.3109/10409239509085139
    Google Scholar
    69
    Protein structure prediction: recognition of primary, secondary, and tertiary structural features from amino acid sequence
    Eisenhaber F; Persson B; Argos P
    Critical reviews in biochemistry and molecular biology (1995), 30 (1), 1-94 ISSN:1040-9238.
    This review attempts a critical stock-taking of the current state of the science aimed at predicting structural features of proteins from their amino acid sequences. At the primary structure level, methods are considered for detection of remotely related sequences and for recognizing amino acid patterns to predict posttranslational modifications and binding sites. The techniques involving secondary structural features include prediction of secondary structure, membrane-spanning regions, and secondary structural class. At the tertiary structural level, methods for threading a sequence into a mainchain fold, homology modeling and assigning sequences to protein families with similar folds are discussed. A literature analysis suggests that, to date, threading techniques are not able to show their superiority over sequence pattern recognition methods. Recent progress in the state of ab initio structure calculation is reviewed in detail. The analysis shows that many structural features can be predicted from the amino acid sequence much better than just a few years ago and with attendant utility in experimental research. Best prediction can be achieved for new protein sequences that can be assigned to well-studied protein families. For single sequences without homologues, the folding problem has not yet been solved.
  70. 70
    Garfield, E. Chemico-linguistics: computer translation of chemical nomenclature. Nature 1961, 192, 192,  DOI: 10.1038/192192a0
    Google Scholar
    There is no corresponding record for this reference.
  71. 71
    Wigh, D. S.; Goodman, J. M.; Lapkin, A. A. A review of molecular representation in the age of machine learning. Wiley Interdiscip. Rev. Comput. Mol. Sci. 2022,  DOI: 10.1002/wcms.1603
    Google Scholar
    There is no corresponding record for this reference.
  72. 72
    Weininger, D. SMILES, a chemical language and information system. 1 Introduction to methodology and encoding rules. J. Chem. Inf. Comput. Sci. 1988, 28, 31– 36,  DOI: 10.1021/ci00057a005
    Google Scholar
    72
    SMILES, a chemical language and information system. 1. Introduction to methodology and encoding rules
    Weininger, David
    Journal of Chemical Information and Computer Sciences (1988), 28 (1), 31-6CODEN: JCISD8; ISSN:0095-2338.
    The SMILES (simplified mol. input line entry system) chem. notation system is described for information processing. The system is based on principles of mol. graph theory and it allows structure specification by use of a very small and natural grammar well suited for high-speed machine processing. The system is easy to use, has high machine compatibility, and allows many computer applications, including notation generation, const. speed database retrieval, substructure searching, and property prediction models.
  73. 73
    Wang, Y.; Xiao, J.; Suzek, T. O.; Zhang, J.; Wang, J.; Bryant, S. H. PubChem: a public information system for analyzing bioactivities of small molecules. Nucleic Acids Res. 2009, 37, W623– 33,  DOI: 10.1093/nar/gkp456
    Google Scholar
    73
    PubChem: a public information system for analyzing bioactivities of small molecules
    Wang, Yanli; Xiao, Jewen; Suzek, Tugba O.; Zhang, Jian; Wang, Jiyao; Bryant, Stephen H.
    Nucleic Acids Research (2009), 37 (Web Server), W623-W633CODEN: NARHAD; ISSN:0305-1048. (Oxford University Press)
    PubChem (http://pubchem.ncbi.nlm.nih.gov) is a public repository for biol. properties of small mols. hosted by the US National Institutes of Health (NIH). PubChem BioAssay database currently contains biol. test results for more than 700 000 compds. The goal of PubChem is to make this information easily accessible to biomedical researchers. In this work, we present a set of web servers to facilitate and optimize the utility of biol. activity information within PubChem. These web-based services provide tools for rapid data retrieval, integration and comparison of biol. screening results, exploratory structure-activity anal., and target selectivity examn. This article reviews these bioactivity anal. tools and discusses their uses. Most of the tools described in this work can be directly accessed at http://pubchem.ncbi.nlm.nih.gov/assay/. URLs for accessing other tools described in this work are specified individually.
  74. 74
    Degtyarenko, K.; de Matos, P.; Ennis, M.; Hastings, J.; Zbinden, M.; McNaught, A.; Alcántara, R.; Darsow, M.; Guedj, M.; Ashburner, M. ChEBI: a database and ontology for chemical entities of biological interest. Nucleic Acids Res. 2007, 36, D344– 50,  DOI: 10.1093/nar/gkm791
    Google Scholar
    There is no corresponding record for this reference.
  75. 75
    Wishart, D. S.; Knox, C.; Guo, A. C.; Cheng, D.; Shrivastava, S.; Tzur, D.; Gautam, B.; Hassanali, M. DrugBank: a knowledgebase for drugs, drug actions and drug targets. Nucleic Acids Res. 2008, 36, D901– 6,  DOI: 10.1093/nar/gkm958
    Google Scholar
    75
    DrugBank: a knowledgebase for drugs, drug actions and drug targets
    Wishart, David S.; Knox, Craig; Guo, An Chi; Cheng, Dean; Shrivastava, Savita; Tzur, Dan; Gautam, Bijaya; Hassanali, Murtaza
    Nucleic Acids Research (2008), 36 (Database Iss), D901-D906CODEN: NARHAD; ISSN:0305-1048. (Oxford University Press)
    DrugBank is a richly annotated resource that combines detailed drug data with comprehensive drug target and drug action information. Since its first release in 2006, DrugBank has been widely used to facilitate in silico drug target discovery, drug design, drug docking or screening, drug metab. prediction, drug interaction prediction and general pharmaceutical education. The latest version of DrugBank (release 2.0) has been expanded significantly over the previous release. With ∼4900 drug entries, it now contains 60% more FDA-approved small mol. and biotech drugs including 10% more exptl.' drugs. Significantly, more protein target data has also been added to the database, with the latest version of DrugBank contg. three times as many non-redundant protein or drug target sequences as before (1565 vs. 524). Each DrugCard entry now contains more than 100 data fields with half of the information being devoted to drug/chem. data and the other half devoted to pharmacol., pharmacogenomic and mol. biol. data. A no. of new data fields, including food-drug interactions, drug-drug interactions and exptl. ADME data have been added in response to numerous user requests. DrugBank has also significantly improved the power and simplicity of its structure query and text query searches. DrugBank is available at http://www.drugbank.ca.
  76. 76
    Wang, X.; Hao, J.; Yang, Y.; He, K. Natural language adversarial defense through synonym encoding. Proceedings of the Thirty-Seventh Conference on Uncertainty in Artificial Intelligence 2021, 823– 833
    Google Scholar
    There is no corresponding record for this reference.
  77. 77
    Bjerrum, E. J. SMILES Enumeration as Data Augmentation for Neural Network Modeling of Molecules. arXiv , 2017.
    Google Scholar
    There is no corresponding record for this reference.
  78. 78
    Lee, I.; Nam, H. Infusing Linguistic Knowledge of SMILES into Chemical Language Models. arXiv , 2022.
    Google Scholar
    There is no corresponding record for this reference.
  79. 79
    Skinnider, M. A. Invalid SMILES are beneficial rather than detrimental to chemical language models. Nature Mach. Intell. 2024, 6, 437,  DOI: 10.1038/s42256-024-00821-x
    Google Scholar
    There is no corresponding record for this reference.
  80. 80
    O’Boyle, N.; Dalke, A. DeepSMILES: An Adaptation of SMILES for Use in Machine-Learning of Chemical Structures. ChemRxiv , 2018.
    Google Scholar
    There is no corresponding record for this reference.
  81. 81
    Krenn, M.; Häse, F.; Nigam, A.; Friederich, P.; Aspuru-Guzik, A. Self-referencing embedded strings (SELFIES): A 100% robust molecular string representation. Mach. Learn.: Sci. Technol. 2020, 1, 045024,  DOI: 10.1088/2632-2153/aba947
    Google Scholar
    There is no corresponding record for this reference.
  82. 82
    Gohlke, H.; Mannhold, R.; Kubinyi, H.; Folkers, G. In Protein-Ligand Interactions; Gohlke, H., Ed.; Methods and Principles in Medicinal Chemistry; Wiley-VCH Verlag: Weinheim, Germany, 2012.
    Google Scholar
    There is no corresponding record for this reference.
  83. 83
    Jumper, J. Highly accurate protein structure prediction with AlphaFold. Nature 2021, 596, 583– 589,  DOI: 10.1038/s41586-021-03819-2
    Google Scholar
    83
    Highly accurate protein structure prediction with AlphaFold
    Jumper, John; Evans, Richard; Pritzel, Alexander; Green, Tim; Figurnov, Michael; Ronneberger, Olaf; Tunyasuvunakool, Kathryn; Bates, Russ; Zidek, Augustin; Potapenko, Anna; Bridgland, Alex; Meyer, Clemens; Kohl, Simon A. A.; Ballard, Andrew J.; Cowie, Andrew; Romera-Paredes, Bernardino; Nikolov, Stanislav; Jain, Rishub; Adler, Jonas; Back, Trevor; Petersen, Stig; Reiman, David; Clancy, Ellen; Zielinski, Michal; Steinegger, Martin; Pacholska, Michalina; Berghammer, Tamas; Bodenstein, Sebastian; Silver, David; Vinyals, Oriol; Senior, Andrew W.; Kavukcuoglu, Koray; Kohli, Pushmeet; Hassabis, Demis
    Nature (London, United Kingdom) (2021), 596 (7873), 583-589CODEN: NATUAS; ISSN:0028-0836. (Nature Portfolio)
    Proteins are essential to life, and understanding their structure can facilitate a mechanistic understanding of their function. Through an enormous exptl. effort, the structures of around 100,000 unique proteins have been detd., but this represents a small fraction of the billions of known protein sequences. Structural coverage is bottlenecked by the months to years of painstaking effort required to det. a single protein structure. Accurate computational approaches are needed to address this gap and to enable large-scale structural bioinformatics. Predicting the three-dimensional structure that a protein will adopt based solely on its amino acid sequence-the structure prediction component of the 'protein folding problem'-has been an important open research problem for more than 50 years. Despite recent progress, existing methods fall far short of at. accuracy, esp. when no homologous structure is available. Here we provide the first computational method that can regularly predict protein structures with at. accuracy even in cases in which no similar structure is known. We validated an entirely redesigned version of our neural network-based model, AlphaFold, in the challenging 14th Crit. Assessment of protein Structure Prediction (CASP14), demonstrating accuracy competitive with exptl. structures in a majority of cases and greatly outperforming other methods. Underpinning the latest version of AlphaFold is a novel machine learning approach that incorporates phys. and biol. knowledge about protein structure, leveraging multi-sequence alignments, into the design of the deep learning algorithm.
  84. 84
    Landrum, G. RDKit: A software suite for cheminformatics, computational chemistry, and predictive modeling. http://www.rdkit.org/RDKit_Overview.pdf, 2013; Accessed: 2023–12–13.
    Google Scholar
    There is no corresponding record for this reference.
  85. 85
    Mukherjee, S.; Ghosh, M.; Basuchowdhuri, P. Proceedings of the 2022 SIAM International Conference on Data Mining (SDM); Proceedings; Society for Industrial and Applied Mathematics, 2022; pp 729– 737.
    Google Scholar
    There is no corresponding record for this reference.
  86. 86
    Chen, L.; Tan, X.; Wang, D.; Zhong, F.; Liu, X.; Yang, T.; Luo, X.; Chen, K.; Jiang, H.; Zheng, M. TransformerCPI: improving compound–protein interaction prediction by sequence-based deep learning with self-attention mechanism and label reversal experiments. Bioinformatics 2020, 36, 4406– 4414,  DOI: 10.1093/bioinformatics/btaa524
    Google Scholar
    86
    TransformerCPI: improving compound-protein interaction prediction by sequence-based deep learning with self-attention mechanism and label reversal experiments
    Chen, Lifan; Tan, Xiaoqin; Wang, Dingyan; Zhong, Feisheng; Liu, Xiaohong; Yang, Tianbiao; Luo, Xiaomin; Chen, Kaixian; Jiang, Hualiang; Zheng, Mingyue
    Bioinformatics (2020), 36 (16), 4406-4414CODEN: BOINFP; ISSN:1367-4811. (Oxford University Press)
    Motivation: Identifying compd.-protein interaction (CPI) is a crucial task in drug discovery and chemogenomics studies, and proteins without three-dimensional structure account for a large part of potential biol. targets, which requires developing methods using only protein sequence information to predict CPI. However, sequence-based CPI models may face some specific pitfalls, including using inappropriate datasets, hidden ligand bias and splitting datasets inappropriately, resulting in overestimation of their prediction performance. Results: To address these issues, we here constructed new datasets specific for CPI prediction, proposed a novel transformer neural network named TransformerCPI, and introduced a more rigorous label reversal expt. to test whether a model learns true interaction features. TransformerCPI achieved much improved performance on the new expts., and it can be deconvolved to highlight important interacting regions of protein sequences and compd. atoms, which may contribute chem. biol. studies with useful guidance for further ligand structural optimization.
  87. 87
    Aly Abdelkader, G.; Ngnamsie Njimbouom, S.; Oh, T.-J.; Kim, J.-D. ResBiGAAT: Residual Bi-GRU with attention for protein-ligand binding affinity prediction. Comput. Biol. Chem. 2023, 107, 107969,  DOI: 10.1016/j.compbiolchem.2023.107969
    Google Scholar
    There is no corresponding record for this reference.
  88. 88
    Li, Q.; Zhang, X.; Wu, L.; Bo, X.; He, S.; Wang, S. PLA-MoRe: AProtein–Ligand Binding Affinity Prediction Model via Comprehensive Molecular Representations. J. Chem. Inf. Model. 2022, 62, 4380– 4390,  DOI: 10.1021/acs.jcim.2c00960
    Google Scholar
    There is no corresponding record for this reference.
  89. 89
    Abramson, J. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 2024, 636, E4,  DOI: 10.1038/s41586-024-08416-7
    Google Scholar
    There is no corresponding record for this reference.
  90. 90
    Gaulton, A.; Bellis, L. J.; Bento, A. P.; Chambers, J.; Davies, M.; Hersey, A.; Light, Y.; McGlinchey, S.; Michalovich, D.; Al-Lazikani, B.; Overington, J. P. ChEMBL: a large-scale bioactivity database for drug discovery. Nucleic Acids Res. 2012, 40, D1100– 7,  DOI: 10.1093/nar/gkr777
    Google Scholar
    90
    ChEMBL: a large-scale bioactivity database for drug discovery
    Gaulton, Anna; Bellis, Louisa J.; Bento, A. Patricia; Chambers, Jon; Davies, Mark; Hersey, Anne; Light, Yvonne; McGlinchey, Shaun; Michalovich, David; Al-Lazikani, Bissan; Overington, John P.
    Nucleic Acids Research (2012), 40 (D1), D1100-D1107CODEN: NARHAD; ISSN:0305-1048. (Oxford University Press)
    ChEMBL is an Open Data database contg. binding, functional and ADMET information for a large no. of drug-like bioactive compds. These data are manually abstracted from the primary published literature on a regular basis, then further curated and standardized to maximize their quality and utility across a wide range of chem. biol. and drug-discovery research problems. Currently, the database contains 5.4 million bioactivity measurements for more than 1 million compds. and 5200 protein targets. Access is available through a web-based interface, data downloads and web services at: https://www.ebi.ac.uk/chembldb.
  91. 91
    Wishart, D. S.; Knox, C.; Guo, A. C.; Shrivastava, S.; Hassanali, M.; Stothard, P.; Chang, Z.; Woolsey, J. DrugBank: a comprehensive resource for in silico drug discovery and exploration. Nucleic Acids Res. 2006, 34, D668– 72,  DOI: 10.1093/nar/gkj067
    Google Scholar
    91
    DrugBank: a comprehensive resource for in silico drug discovery and exploration
    Wishart, David S.; Knox, Craig; Guo, An Chi; Shrivastava, Savita; Hassanali, Murtaza; Stothard, Paul; Chang, Zhan; Woolsey, Jennifer
    Nucleic Acids Research (2006), 34 (Database), D668-D672CODEN: NARHAD; ISSN:0305-1048. (Oxford University Press)
    DrugBank is a unique bioinformatics/cheminformatics resource that combines detailed drug (i.e. chem.) data with comprehensive drug target (i.e. protein) information. The database contains >4100 drug entries including >800 FDA approved small mol. and biotech drugs as well as >3200 exptl. drugs. Addnl., >14 000 protein or drug target sequences are linked to these drug entries. Each DrugCard entry contains >80 data fields with half of the information being devoted to drug/chem. data and the other half devoted to drug target or protein data. Many data fields are hyperlinked to other databases (KEGG, PubChem, ChEBI, PDB, Swiss-Prot and GenBank) and a variety of structure viewing applets. The database is fully searchable supporting extensive text, sequence, chem. structure and relational query searches. Potential applications of DrugBank include in silico drug target discovery, drug design, drug docking or screening, drug metab. prediction, drug interaction prediction and general pharmaceutical education. DrugBank is available at http://redpoll.pharmacy.ualberta.ca/drugbank/.
  92. 92
    Berman, H. M.; Westbrook, J.; Feng, Z.; Gilliland, G.; Bhat, T. N.; Weissig, H.; Shindyalov, I. N.; Bourne, P. E. The Protein Data Bank. Nucleic Acids Res. 2000, 28, 235– 242,  DOI: 10.1093/nar/28.1.235
    Google Scholar
    92
    The Protein Data Bank
    Berman, Helen M.; Westbrook, John; Feng, Zukang; Gilliland, Gary; Bhat, T. N.; Weissig, Helge; Shindyalov, Ilya N.; Bourne, Philip E.
    Nucleic Acids Research (2000), 28 (1), 235-242CODEN: NARHAD; ISSN:0305-1048. (Oxford University Press)
    The Protein Data Bank (PDB; http://www.rcsb.org/pdb/)is the single worldwide archive of structural data of biol. macromols. This paper describes the goals of the PDB, the systems in place for data deposition and access, how to obtain further information, and near-term plans for the future development of the resource.
  93. 93
    Acids research, N. 2017 UniProt: the universal protein knowledgebase. Nucleic Acids Res. 2017, 45, D158– D169,  DOI: 10.1093/nar/gkw1099
    Google Scholar
    There is no corresponding record for this reference.
  94. 94
    Davis, M. I.; Hunt, J. P.; Herrgard, S.; Ciceri, P.; Wodicka, L. M.; Pallares, G.; Hocker, M.; Treiber, D. K.; Zarrinkar, P. P. Comprehensive analysis of kinase inhibitor selectivity. Nat. Biotechnol. 2011, 29, 1046– 1051,  DOI: 10.1038/nbt.1990
    Google Scholar
    94
    Comprehensive analysis of kinase inhibitor selectivity
    Davis, Mindy I.; Hunt, Jeremy P.; Herrgard, Sanna; Ciceri, Pietro; Wodicka, Lisa M.; Pallares, Gabriel; Hocker, Michael; Treiber, Daniel K.; Zarrinkar, Patrick P.
    Nature Biotechnology (2011), 29 (11), 1046-1051CODEN: NABIF9; ISSN:1087-0156. (Nature Publishing Group)
    We tested the interaction of 72 kinase inhibitors with 442 kinases covering >80% of the human catalytic protein kinome. Our data show that, as a class, type II inhibitors are more selective than type I inhibitors, but that there are important exceptions to this trend. The data further illustrate that selective inhibitors have been developed against the majority of kinases targeted by the compds. tested. Anal. of the interaction patterns reveals a class of 'group-selective' inhibitors broadly active against a single subfamily of kinases, but selective outside that subfamily. The data set suggests compds. to use as tools to study kinases for which no dedicated inhibitors exist. It also provides a foundation for further exploring kinase inhibitor biol. and toxicity, as well as for studying the structural basis of the obsd. interaction patterns. Our findings will help to realize the direct enabling potential of genomics for drug development and basic research about cellular signaling.
  95. 95
    Tang, J.; Szwajda, A.; Shakyawar, S.; Xu, T.; Hintsanen, P.; Wennerberg, K.; Aittokallio, T. Making sense of large-scale kinase inhibitor bioactivity data sets: a comparative and integrative analysis. J. Chem. Inf. Model. 2014, 54, 735– 743,  DOI: 10.1021/ci400709d
    Google Scholar
    95
    Making Sense of Large-Scale Kinase Inhibitor Bioactivity Data Sets: A Comparative and Integrative Analysis
    Tang, Jing; Szwajda, Agnieszka; Shakyawar, Sushil; Xu, Tao; Hintsanen, Petteri; Wennerberg, Krister; Aittokallio, Tero
    Journal of Chemical Information and Modeling (2014), 54 (3), 735-743CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    We carried out a systematic evaluation of target selectivity profiles across three recent large-scale biochem. assays of kinase inhibitors and further compared these standardized bioactivity assays with data reported in the widely used databases ChEMBL and STITCH. Our comparative evaluation revealed relative benefits and potential limitations among the bioactivity types, as well as pinpointed biases in the database curation processes. Ignoring such issues in data heterogeneity and representation may lead to biased modeling of drugs' polypharmacol. effects as well as to unrealistic evaluation of computational strategies for the prediction of drug-target interaction networks. Toward making use of the complementary information captured by the various bioactivity types, including IC50, Ki, and Kd, we also introduce a model-based integration approach, termed KIBA, and demonstrate here how it can be used to classify kinase inhibitor targets and to pinpoint potential errors in database-reported drug-target interactions. An integrated drug-target bioactivity matrix across 52 498 chem. compds. and 467 kinase targets, including a total of 246 088 KIBA scores, has been made freely available.
  96. 96
    Wang, R.; Fang, X.; Lu, Y.; Wang, S. The PDBbind database: collection of binding affinities for protein-ligand complexes with known three-dimensional structures. J. Med. Chem. 2004, 47, 2977– 2980,  DOI: 10.1021/jm030580l
    Google Scholar
    96
    The PDBbind database: Collection of binding affinities for protein-ligand complexes with known three-dimensional structures
    Wang, Renxiao; Fang, Xueliang; Lu, Yipin; Wang, Shaomeng
    Journal of Medicinal Chemistry (2004), 47 (12), 2977-2980CODEN: JMCMAR; ISSN:0022-2623. (American Chemical Society)
    We have screened the entire Protein Data Bank (Release No. 103, Jan. 2003) and identified 5671 protein-ligand complexes out of 19 621 exptl. structures. A systematic examn. of the primary refs. of these entries has led to a collection of binding affinity data (Kd, Ki, and IC50) for a total of 1359 complexes. The outcomes of this project have been organized into a Web-accessible database named the PDBbind database.
  97. 97
    Chen, S.; Zhang, S.; Fang, X.; Lin, L.; Zhao, H.; Yang, Y. Protein complex structure modeling by cross-modal alignment between cryo-EM maps and protein sequences. Nat. Commun. 2024, 15, 8808,  DOI: 10.1038/s41467-024-53116-5
    Google Scholar
    There is no corresponding record for this reference.
  98. 98
    Bishop, M.C. Pattern Recognition and Machine Learning, 1st ed.; Information Science and Statistics; Springer: New York, NY, 2006.
    Google Scholar
    There is no corresponding record for this reference.
  99. 99
    Yang, J.; Shen, C.; Huang, N. Predicting or Pretending: Artificial Intelligence for Protein-Ligand Interactions Lack of Sufficiently Large and Unbiased Datasets. Front. Pharmacol. 2020, 11, 69,  DOI: 10.3389/fphar.2020.00069
    Google Scholar
    99
    Predicting or pretending: artificial intelligence for protein-ligand interactions lack of sufficiently large and unbiased datasets
    Yang, Jincai; Shen, Cheng; Huang, Niu
    Frontiers in Pharmacology (2020), 11 (), 69CODEN: FPRHAU; ISSN:1663-9812. (Frontiers Media S.A.)
    Predicting protein-ligand interactions using artificial intelligence (AI) models has attracted great interest in recent years. However, data-driven AI models unequivocally suffer from a lack of sufficiently large and unbiased datasets. Here, we systematically investigated the data biases on the PDBbind and DUD-E datasets. We examd. the model performance of at. convolutional neural network (ACNN) on the PDBbind core set and achieved a Pearson R2 of 0.73 between exptl. and predicted binding affinities. Strikingly, the ACNN models did not require learning the essential protein-ligand interactions in complex structures and achieved similar performance even on datasets contg. only ligand structures or only protein structures, while data splitting based on similarity clustering (protein sequence or ligand scaffold) significantly reduced the model performance. We also identified the property and topol. biases in the DUD-E dataset which led to the artificially increased enrichment performance of virtual screening. The property bias in DUD-E was reduced by enforcing the more stringent ligand property matching rules, while the topol. bias still exists due to the use of mol. fingerprint similarity as a decoy selection criterion. Therefore, we believe that sufficiently large and unbiased datasets are desirable for training robust AI models to accurately predict protein-ligand interactions.
  100. 100
    Kryshtafovych, A.; Schwede, T.; Topf, M.; Fidelis, K.; Moult, J. Critical assessment of methods of protein structure prediction (CASP)–Round XIII. Proteins 2019, 87, 1011– 1020,  DOI: 10.1002/prot.25823
    Google Scholar
    100
    Critical assessment of methods of protein structure prediction (CASP)-Round XIII
    Kryshtafovych, Andriy; Schwede, Torsten; Topf, Maya; Fidelis, Krzysztof; Moult, John
    Proteins: Structure, Function, and Bioinformatics (2019), 87 (12), 1011-1020CODEN: PSFBAF; ISSN:1097-0134. (Wiley-Blackwell)
    A review. CASP (crit. assessment of structure prediction) assesses the state of the art in modeling protein structure from amino acid sequence. The most recent expt. (CASP13 held in 2018) saw dramatic progress in structure modeling without use of structural templates (historically "ab initio" modeling). Progress was driven by the successful application of deep learning techniques to predict inter-residue distances. In turn, these results drove dramatic improvements in three-dimensional structure accuracy: With the proviso that there are an adequate no. of sequences known for the protein family, the new methods essentially solve the long-standing problem of predicting the fold topol. of monomeric proteins. Further, the no. of sequences required in the alignment has fallen substantially. There is also substantial improvement in the accuracy of template-based models. Other areas-model refinement, accuracy estn., and the structure of protein assemblies-have again yielded interesting results. CASP13 placed increased emphasis on the use of sparse data together with modeling and chem. crosslinking, SAXS, and NMR all yielded more mature results. This paper summarizes the key outcomes of CASP13. The special issue of PROTEINS contains papers describing the CASP13 assessments in each modeling category and contributions from the participants.
  101. 101
    Janin, J.; Henrick, K.; Moult, J.; Eyck, L. T.; Sternberg, M. J. E.; Vajda, S.; Vakser, I.; Wodak, S. J. CAPRI: A Critical Assessment of PRedicted Interactions. Proteins 2003, 52, 2– 9,  DOI: 10.1002/prot.10381
    Google Scholar
    101
    CAPRI: A critical assessment of predicted interactions
    Janin, Joel; Henrick, Kim; Moult, John; Ten Eyck, Lynn; Sternberg, Michael J. E.; Vajda, Sandor; Vakser, Ilya; Wodak, Shoshana J.
    Proteins: Structure, Function, and Genetics (2003), 52 (1), 2-9CODEN: PSFGEY; ISSN:0887-3585. (Wiley-Liss, Inc.)
    A review. CAPRI is a community wide expt. to assess the capacity of protein-docking methods to predict protein-protein interactions. Nineteen groups participated in rounds 1 and 2 of CAPRI and submitted blind structure predictions for seven protein-protein complexes based on the known structure of the component proteins. The predictions were compared to the unpublished X-ray structures of the complexes. We describe here the motivations for launching CAPRI, the rules that we applied to select targets and run the expt., and some conclusions that can already be drawn. The results stress the need for new scoring functions and for methods handling the conformation changes that were obsd. in some of the target systems. CAPRI has already been a powerful drive for the community of computational biologists who development docking algorithms. We hope that this issue of Proteins will also be of interest to the community of structural biologists, which we call upon to provide new targets for future rounds of CAPRI, and to all mol. biologists who view protein-protein recognition as an essential process.
  102. 102
    Lensink, M. F.; Nadzirin, N.; Velankar, S.; Wodak, S. J. Modeling protein-protein, protein-peptide, and protein-oligosaccharide complexes: CAPRI 7th edition. Proteins 2020, 88, 916– 938,  DOI: 10.1002/prot.25870
    Google Scholar
    102
    Modeling protein-protein, protein-peptide, and protein-oligosaccharide complexes: CAPRI 7th edition
    Lensink, Marc F.; Nadzirin, Nurul; Velankar, Sameer; Wodak, Shoshana J.
    Proteins: Structure, Function, and Bioinformatics (2020), 88 (8), 916-938CODEN: PSFBAF; ISSN:1097-0134. (Wiley-Blackwell)
    The authors present the seventh report on the performance of methods for predicting the at. resoln. structures of protein complexes offered as targets to the community-wide initiative on the Crit. Assessment of Predicted Interactions. Performance was evaluated on the basis of 36,114 models of protein complexes submitted by 57 groups-including 13 automatic servers-in prediction rounds held during the years 2016 to 2019 for eight protein-protein, three protein-peptide, and five protein-oligosaccharide targets with different length ligands. Six of the protein-protein targets represented challenging hetero-complexes, due to factors such as availability of distantly related templates for the individual subunits, or for the full complex, inter-domain flexibility, conformational adjustments at the binding region, or the multi-component nature of the complex. The main challenge for the protein-peptide and protein-oligosaccharide complexes was to accurately model the ligand conformation and its interactions at the interface. Encouragingly, models of acceptable quality, or better, were obtained for a total of six protein-protein complexes, which included four of the challenging hetero-complexes and a homo-decamer. But fewer of these targets were predicted with medium or higher accuracy. High accuracy models were obtained for two of the three protein-peptide targets, and for one of the protein-oligosaccharide targets. The remaining protein-sugar targets were predicted with medium accuracy. The authors' anal. indicates that progress in predicting increasingly challenging and diverse types of targets is due to closer integration of template-based modeling techniques with docking, scoring, and model refinement procedures, and to significant incremental improvements in the underlying methodologies.
  103. 103
    Schomburg, I.; Chang, A.; Schomburg, D. BRENDA, enzyme data and metabolic information. Nucleic Acids Res. 2002, 30, 47– 49,  DOI: 10.1093/nar/30.1.47
    Google Scholar
    103
    BRENDA, enzyme data and metabolic information
    Schomburg, Ida; Chang, Antje; Schomburg, Dietmar
    Nucleic Acids Research (2002), 30 (1), 47-49CODEN: NARHAD; ISSN:0305-1048. (Oxford University Press)
    A review with 6 refs. BRENDA is a comprehensive relational database on functional and mol. information of enzymes, based on primary literature. The database contains information extd. and evaluated from ∼46,000 refs., holding data of at least 40,000 different enzymes from >6900 different organisms, classified in ∼3900 EC nos. BRENDA is an important tool for biochem. and medical research covering information on properties of all classified enzymes, including data on the occurrence, catalyzed reaction, kinetics, substrates/products, inhibitors, cofactors, activators, structure and stability. All data are connected to literature refs. which in turn are linked to PubMed. The data and information provide a fundamental tool for research of enzyme mechanisms, metabolic pathways, the evolution of metab. and, furthermore, for medicinal diagnostics and pharmaceutical research. The database is a resource for data of enzymes, classified according to the EC system of the IUBMB Enzyme Nomenclature Committee, and the entries are cross-referenced to other databases, i.e., organism classification, protein sequence, protein structure, and literature refs. BRENDA provides an academic web access at http://www.brenda.uni-koeln.de.
  104. 104
    Liu, T.; Lin, Y.; Wen, X.; Jorissen, R. N.; Gilson, M. K. BindingDB: a web-accessible database of experimentally determined protein-ligand binding affinities. Nucleic Acids Res. 2007, 35, D198– 201,  DOI: 10.1093/nar/gkl999
    Google Scholar
    104
    BindingDB: a web-accessible database of experimentally determined protein-ligand binding affinities
    Liu, Tiqing; Lin, Yuhmei; Wen, Xin; Jorissen, Robert N.; Gilson, Michael K.
    Nucleic Acids Research (2007), 35 (Database Iss), D198-D201CODEN: NARHAD; ISSN:0305-1048. (Oxford University Press)
    BindingDB is a publicly accessible database currently contg. ∼20 000 exptl. detd. binding affinities of protein-ligand complexes, for 110 protein targets including isoforms and mutational variants, and ∼11 000 small mol. ligands. The data are extd. from the scientific literature, data collection focusing on proteins that are drug-targets or candidate drug-targets and for which structural data are present in the Protein Data Bank. The BindingDB website supports a range of query types, including searches by chem. structure, substructure and similarity; protein sequence; ligand and protein names; affinity ranges and mol. wt. Data sets generated by BindingDB queries can be downloaded in the form of annotated SDfiles for further anal., or used as the basis for virtual screening of a compd. database uploaded by the user. The data in BindingDB are linked both to structural data in the PDB via PDB IDs and chem. and sequence searches, and to the literature in PubMed via PubMed IDs.
  105. 105
    Amemiya, T.; Koike, R.; Kidera, A.; Ota, M. PSCDB: a database for protein structural change upon ligand binding. Nucleic Acids Res. 2012, 40, D554– 8,  DOI: 10.1093/nar/gkr966
    Google Scholar
    There is no corresponding record for this reference.
  106. 106
    Mysinger, M. M.; Carchia, M.; Irwin, J. J.; Shoichet, B. K. Directory of useful decoys, enhanced (DUD-E): better ligands and decoys for better benchmarking. J. Med. Chem. 2012, 55, 6582– 6594,  DOI: 10.1021/jm300687e
    Google Scholar
    106
    Directory of Useful Decoys, Enhanced (DUD-E): Better Ligands and Decoys for Better Benchmarking
    Mysinger, Michael M.; Carchia, Michael; Irwin, John. J.; Shoichet, Brian K.
    Journal of Medicinal Chemistry (2012), 55 (14), 6582-6594CODEN: JMCMAR; ISSN:0022-2623. (American Chemical Society)
    A key metric to assess mol. docking remains ligand enrichment against challenging decoys. Whereas the directory of useful decoys (DUD) has been widely used, clear areas for optimization have emerged. Here we describe an improved benchmarking set that includes more diverse targets such as GPCRs and ion channels, totaling 102 proteins with 22886 clustered ligands drawn from ChEMBL, each with 50 property-matched decoys drawn from ZINC. To ensure chemotype diversity, we cluster each target's ligands by their Bemis-Murcko at. frameworks. We add net charge to the matched physicochem. properties and include only the most dissimilar decoys, by topol., from the ligands. An online automated tool (http://decoys.docking.org) generates these improved matched decoys for user-supplied ligands. We test this data set by docking all 102 targets, using the results to improve the balance between ligand desolvation and electrostatics in DOCK 3.6. The complete DUD-E benchmarking set is freely available at http://dude.docking.org.
  107. 107
    Warren, G. L.; Do, T. D.; Kelley, B. P.; Nicholls, A.; Warren, S. D. Essential considerations for using protein-ligand structures in drug discovery. Drug Discovery Today 2012, 17, 1270– 1281,  DOI: 10.1016/j.drudis.2012.06.011
    Google Scholar
    107
    Essential considerations for using protein-ligand structures in drug discovery
    Warren, Gregory L.; Do, Thanh D.; Kelley, Brian P.; Nicholls, Anthony; Warren, Stephen D.
    Drug Discovery Today (2012), 17 (23-24), 1270-1281CODEN: DDTOFS; ISSN:1359-6446. (Elsevier Ltd.)
    A review. Protein-ligand structures are the core data required for structure-based drug design (SBDD). Understanding the error present in this data is essential to the successful development of SBDD tools. Methods for assessing protein-ligand structure quality and a new set of identification criteria are presented here. When these criteria were applied to a set of 728 structures previously used to validate mol. docking software, only 17% were found to be acceptable. Structures were re-refined to maintain internal consistency in the comparison and assessment of the quality criteria. This process resulted in Iridium, a highly trustworthy protein-ligand structure database to be used for development and validation of structure-based design tools for drug discovery.
  108. 108
    Puvanendrampillai, D.; Mitchell, J. B. O. L/D Protein Ligand Database (PLD): additional understanding of the nature and specificity of protein-ligand complexes. Bioinformatics 2003, 19, 1856– 1857,  DOI: 10.1093/bioinformatics/btg243
    Google Scholar
    There is no corresponding record for this reference.
  109. 109
    Wang, C.; Hu, G.; Wang, K.; Brylinski, M.; Xie, L.; Kurgan, L. PDID: database of molecular-level putative protein-drug interactions in the structural human proteome. Bioinformatics 2016, 32, 579– 586,  DOI: 10.1093/bioinformatics/btv597
    Google Scholar
    There is no corresponding record for this reference.
  110. 110
    Zhu, M.; Song, X.; Chen, P.; Wang, W.; Wang, B. dbHDPLS: A database of human disease-related protein-ligand structures. Comput. Biol. Chem. 2019, 78, 353– 358,  DOI: 10.1016/j.compbiolchem.2018.12.023
    Google Scholar
    110
    dbHDPLS: A database of human disease-related protein-ligand structures
    Zhu, Muchun; Song, Xiaoping; Chen, Peng; Wang, Wenyan; Wang, Bing
    Computational Biology and Chemistry (2019), 78 (), 353-358CODEN: CBCOCH; ISSN:1476-9271. (Elsevier B.V.)
    Protein-ligand complexes perform specific functions, most of which are related to human diseases. The database, called as human disease-related protein-ligand structures (dbHDPLS), collected 8833 structures which were extd. from protein data bank (PDB) and other related databases. The database is annotated with comprehensive information involving ligands and drugs, related human diseases and protein-ligand interaction information, with the information of protein structures. The database may be a reliable resource for structure-based drug target discoveries and druggability predictions of protein-ligand binding sites, drug-disease relationships based on protein-ligand complex structures.
  111. 111
    Gao, M.; Moumbock, A. F. A.; Qaseem, A.; Xu, Q.; Günther, S. CovPDB: a high-resolution coverage of the covalent protein-ligand interactome. Nucleic Acids Res. 2022, 50, D445– D450,  DOI: 10.1093/nar/gkab868
    Google Scholar
    111
    CovPDB: a high-resolution coverage of the covalent protein-ligand interactome
    Gao, Mingjie; Moumbock, Aurelien F. A.; Qaseem, Ammar; Xu, Qianqing; Guenther, Stefan
    Nucleic Acids Research (2022), 50 (D1), D445-D450CODEN: NARHAD; ISSN:1362-4962. (Oxford University Press)
    In recent years, the drug discovery paradigm has shifted toward compds. that covalently modify disease-assocd. target proteins, because they tend to possess high potency, selectivity, and duration of action. The rational design of novel targeted covalent inhibitors (TCIs) typically starts from resolved macromol. structures of target proteins in their apo or holo forms. However, the existing TCI databases contain only a paucity of covalent protein-ligand (cP-L) complexes. Herein, we report CovPDB, the first database solely dedicated to highresoln. cocrystal structures of biol. relevant cP-L complexes, curated from the Protein Data Bank. For these curated complexes, the chem. structures and warheads of pre-reactive electrophilic ligands as well as the covalent bonding mechanisms to their target proteins were expertly manually annotated. Totally, CovPDB contains 733 proteins and 1,501 ligands, relating to 2,294 cP-L complexes, 93 reactive warheads, 14 targetable residues, and 21 covalent mechanisms. Users are provided with an intuitive and interactive web interface that allows multiple search and browsing options to explore the covalent interactome at a mol. level in order to develop novel TCIs. CovPDB is freely accessible and its contents are available for download as flat files of various formats.
  112. 112
    Ammar, A.; Cavill, R.; Evelo, C.; Willighagen, E. PSnpBind: a database of mutated binding site protein-ligand complexes constructed using a multithreaded virtual screening workflow. J. Cheminform. 2022, 14, 8,  DOI: 10.1186/s13321-021-00573-5
    Google Scholar
    There is no corresponding record for this reference.
  113. 113
    Lingė, D. PLBD: protein-ligand binding database of thermodynamic and kinetic intrinsic parameters. Database 2023,  DOI: 10.1093/database/baad040
    Google Scholar
    There is no corresponding record for this reference.
  114. 114
    Wei, H.; Wang, W.; Peng, Z.; Yang, J. Q-BioLiP: A Comprehensive Resource for Quaternary Structure-based Protein–ligand Interactions. bioRxiv , 2023, 2023.06.23.546351.
    Google Scholar
    There is no corresponding record for this reference.
  115. 115
    Korlepara, D. B. PLAS-20k: Extended dataset of protein-ligand affinities from MD simulations for machine learning applications. Sci. Data 2024,  DOI: 10.1038/s41597-023-02872-y
    Google Scholar
    There is no corresponding record for this reference.
  116. 116
    Xenarios, I.; Rice, D. W.; Salwinski, L.; Baron, M. K.; Marcotte, E. M.; Eisenberg, D. DIP: the database of interacting proteins. Nucleic Acids Res. 2000, 28, 289– 291,  DOI: 10.1093/nar/28.1.289
    Google Scholar
    116
    DIP: the Database of Interacting Proteins
    Xenarios, Ioannis; Rice, Danny W.; Salwinski, Lukasz; Baron, Marisa K.; Marcotte, Edward M.; Eisenberg, David
    Nucleic Acids Research (2000), 28 (1), 289-291CODEN: NARHAD; ISSN:0305-1048. (Oxford University Press)
    The Database of Interacting Proteins is a database that documents exptl. detd. protein-protein interactions. This database is intended to provide the scientific community with a comprehensive and integrated tool for browsing and efficiently extg. information about protein interactions and interaction networks in biol. processes. Beyond cataloging details of protein-protein interactions, the DIP is useful for understanding protein function and protein-protein relationships, studying the properties of networks of interacting proteins, benchmarking predictions of protein-protein interactions, and studying the evolution of protein-protein interactions.
  117. 117
    Wallach, I.; Lilien, R. The protein-small-molecule database, a non-redundant structural resource for the analysis of protein-ligand binding. Bioinformatics 2009, 25, 615– 620,  DOI: 10.1093/bioinformatics/btp035
    Google Scholar
    There is no corresponding record for this reference.
  118. 118
    Wang, S.; Lin, H.; Huang, Z.; He, Y.; Deng, X.; Xu, Y.; Pei, J.; Lai, L. CavitySpace: A Database of Potential Ligand Binding Sites in the Human Proteome. Biomolecules 2022, 12, 967,  DOI: 10.3390/biom12070967
    Google Scholar
    118
    CavitySpace: A Database of Potential Ligand Binding Sites in the Human Proteome
    Wang, Shiwei; Lin, Haoyu; Huang, Zhixian; He, Yufeng; Deng, Xiaobing; Xu, Youjun; Pei, Jianfeng; Lai, Luhua
    Biomolecules (2022), 12 (7), 967CODEN: BIOMHC; ISSN:2218-273X. (MDPI AG)
    Location and properties of ligand binding sites provide important information to uncover protein functions and to direct structure-based drug design approaches. However, as binding site detection depends on the three-dimensional (3D) structural data of proteins, functional anal. based on protein ligand binding sites is formidable for proteins without structural information. Recent developments in protein structure prediction and the 3D structures built by AlphaFold provide an unprecedented opportunity for analyzing ligand binding sites in human proteins. Here, we constructed the CavitySpace database, the first pocket library for all the proteins in the human proteome, using a widely-applied ligand binding site detection program CAVITY. Our anal. showed that known ligand binding sites could be well recovered. We grouped the predicted binding sites according to their similarity which can be used in protein function prediction and drug repurposing studies. Novel binding sites in highly reliable predicted structure regions provide new opportunities for drug discovery. Our CavitySpace is freely available and provides a valuable tool for drug discovery and protein function studies.
  119. 119
    Otter, D. W.; Medina, J. R.; Kalita, J. K. A survey of the usages of deep learning for natural language processing. IEEE Trans. Neural Netw. Learn. Syst. 2021, 32, 604– 624,  DOI: 10.1109/TNNLS.2020.2979670
    Google Scholar
    119
    A Survey of the Usages of Deep Learning for Natural Language Processing
    Otter Daniel W; Medina Julian R; Kalita Jugal K
    IEEE transactions on neural networks and learning systems (2021), 32 (2), 604-624 ISSN:.
    Over the last several years, the field of natural language processing has been propelled forward by an explosion in the use of deep learning models. This article provides a brief introduction to the field and a quick overview of deep learning architectures and methods. It then sifts through the plethora of recent studies and summarizes a large assortment of relevant contributions. Analyzed research areas include several core linguistic processing issues in addition to many applications of computational linguistics. A discussion of the current state of the art is then provided along with recommendations for future research in the field.
  120. 120
    Wang, Y.; You, Z.-H.; Yang, S.; Li, X.; Jiang, T.-H.; Zhou, X. A high efficient biological language model for predicting Protein-Protein interactions. Cells 2019, 8, 122,  DOI: 10.3390/cells8020122
    Google Scholar
    120
    A high efficient biological language model for predicting protein-protein interactions
    Wang, Yanbin; You, Zhu-Hong; Yang, Shan; Li, Xiao; Jiang, Tong-Hai; Zhou, Xi
    Cells (2019), 8 (2), 122CODEN: CELLC6; ISSN:2073-4409. (MDPI AG)
    : Many life activities and key functions in organisms are maintained by different types of protein-protein interactions (PPIs). In order to accelerate the discovery of PPIs for different species, many computational methods have been developed. Unfortunately, even though computational methods are constantly evolving, efficient methods for predicting PPIs from protein sequence information have not been found for many years due to limiting factors including both methodol. and technol. Inspired by the similarity of biol. sequences and languages, developing a biol. language processing technol. may provide a brand new theor. perspective and feasible method for the study of biol. sequences. In this paper, a pure biol. language processing model is proposed for predicting protein-protein interactions only using a protein sequence. The model was constructed based on a feature representation method for biol. sequences called bio-to-vector (Bio2Vec) and a convolution neural network (CNN). The Bio2Vec obtains protein sequence features by using a "bio-word" segmentation system and a word representation model used for learning the distributed representation for each "bio-word". The Bio2Vec supplies a frame that allows researchers to consider the context information and implicit semantic information of a bio sequence. A remarkable improvement in PPIs prediction performance has been obsd. by using the proposed model compared with state-of-the-art methods. The presentation of this approach marks the start of "bio language processing technol.," which could cause a technol. revolution and could be applied to improve the quality of predictions in other problems.
  121. 121
    Abbasi, K.; Razzaghi, P.; Poso, A.; Amanlou, M.; Ghasemi, J. B.; Masoudi-Nejad, A. DeepCDA: deep cross-domain compound–protein affinity prediction through LSTM and convolutional neural networks. Bioinformatics 2020, 36, 4633– 4642,  DOI: 10.1093/bioinformatics/btaa544
    Google Scholar
    121
    DeepCDA: deep cross-domain compound-protein affinity prediction through LSTM and convolutional neural networks
    Abbasi, Karim; Razzaghi, Parvin; Poso, Antti; Amanlou, Massoud; Ghasemi, Jahan B.; Masoudi-Nejad, Ali
    Bioinformatics (2020), 36 (17), 4633-4642CODEN: BOINFP; ISSN:1367-4811. (Oxford University Press)
    Motivation: An essential part of drug discovery is the accurate prediction of the binding affinity of new compd.-protein pairs. Most of the std. computational methods assume that compds. or proteins of the test data are obsd. during the training phase. However, in real-world situations, the test and training data are sampled from different domains with different distributions. To cope with this challenge, we propose a deep learning-based approach that consists of three steps. In the first step, the training encoder network learns a novel representation of compds. and proteins. To this end, we combine convolutional layers and long-short-term memory layers so that the occurrence patterns of local substructures through a protein and a compd. sequence are learned. Also, to encode the interaction strength of the protein and compd. substructures, we propose a two-sided attention mechanism. In the second phase, to deal with the different distributions of the training and test domains, a feature encoder network is learned for the test domain by utilizing an adversarial domain adaptation approach. In the third phase, the learned test encoder network is applied to new compd.-protein pairs to predict their binding affinity. Results: To evaluate the proposed approach, we applied it to KIBA, Davis and BindingDB datasets. The results show that the proposed method learns a more reliable model for the test domain in more challenging situations.
  122. 122
    Zhou, G.; Gao, Z.; Ding, Q.; Zheng, H.; Xu, H.; Wei, Z.; Zhang, L.; Ke, G. Uni-Mol: AUniversal 3D Molecular Representation Learning Framework. ChemRxiv , 2023.
    Google Scholar
    There is no corresponding record for this reference.
  123. 123
    Zhou, D.; Xu, Z.; Li, W.; Xie, X.; Peng, S. MultiDTI: drug–target interaction prediction based on multi-modal representation learning to bridge the gap between new chemical entities and known heterogeneous network. Bioinformatics 2021, 37, 4485– 4492,  DOI: 10.1093/bioinformatics/btab473
    Google Scholar
    123
    MultiDTI: drug-target interaction prediction based on multi-modal representation learning to bridge the gap between new chemical entities and known heterogeneous network
    Zhou, Deshan; Xu, Zhijian; Li, WenTao; Xie, Xiaolan; Peng, Shaoliang
    Bioinformatics (2021), 37 (23), 4485-4492CODEN: BOINFP; ISSN:1367-4811. (Oxford University Press)
    Motivation: Predicting new drug-target interactions is an important step in new drug development, understanding of its side effects and drug repositioning. Heterogeneous data sources can provide comprehensive information and different perspectives for drug-target interaction prediction. Thus, there have been many calcn. methods relying on heterogeneous networks. Most of them use graph-related algorithms to characterize nodes in heterogeneous networks for predicting new drug-target interactions (DTI). However, these methods can only make predictions in known heterogeneous network datasets, and cannot support the prediction of new chem. entities outside the heterogeneous network, which hinder further drug discovery and development. Results: To solve this problem, we proposed a multi-modal DTI prediction model named 'MultiDTI' which uses our proposed joint learning framework based on heterogeneous networks. It combines the interaction or assocn. information of the heterogeneous network and the drug/target sequence information, and maps the drugs, targets, side effects and disease nodes in the heterogeneous network into a common space. In this way, 'MultiDTI' can map the new chem. entity to this learned common space based on the chem. structure of the new entity. That is, bridging the gap between new chem. entities and known heterogeneous network. Our model has strong predictive performance, and the area under the receiver operating characteristic curve of the model is 0.961 and the area under the precision recall curve is 0.947 with 10-fold cross validation. In addn., some predicted new DTIs have been confirmed by ChEMBL database. Our results indicate that 'MultiDTI' is a powerful and practical tool for predicting new DTI, which can promote the development of drug discovery or drug repositioning.
  124. 124
    Özçelik, R.; Öztürk, H.; Özgür, A.; Ozkirimli, E. ChemBoost: A chemical language based approach for protein - ligand binding affinity prediction. Mol. Inform. 2021, 40, e2000212,  DOI: 10.1002/minf.202000212
    Google Scholar
    There is no corresponding record for this reference.
  125. 125
    Gaspar, H. A.; Ahmed, M.; Edlich, T.; Fabian, B.; Varszegi, Z.; Segler, M.; Meyers, J.; Fiscato, M. Proteochemometric Models Using Multiple Sequence Alignments and a Subword Segmented Masked Language Model. ChemRxiv , 2021.
    Google Scholar
    There is no corresponding record for this reference.
  126. 126
    Arseniev-Koehler, A. Theoretical foundations and limits of word embeddings: What types of meaning can they capture. Sociol. Methods Res. 2022, 004912412211401
    Google Scholar
    There is no corresponding record for this reference.
  127. 127
    Lake, B. M.; Murphy, G. L. Word meaning in minds and machines. Psychol. Rev. 2023, 130, 401– 431,  DOI: 10.1037/rev0000297
    Google Scholar
    There is no corresponding record for this reference.
  128. 128
    Winchester, S. A Verb for Our Frantic Times. https://www.nytimes.com/2011/05/29/opinion/29winchester.html, 2011; Accessed: 2024–9-15.
    Google Scholar
    There is no corresponding record for this reference.
  129. 129
    Panapitiya, G.; Girard, M.; Hollas, A.; Sepulveda, J.; Murugesan, V.; Wang, W.; Saldanha, E. Evaluation of deep learning architectures for aqueous solubility prediction. ACS Omega 2022, 7, 15695– 15710,  DOI: 10.1021/acsomega.2c00642
    Google Scholar
    129
    Evaluation of Deep Learning Architectures for Aqueous Solubility Prediction
    Panapitiya, Gihan; Girard, Michael; Hollas, Aaron; Sepulveda, Jonathan; Murugesan, Vijayakumar; Wang, Wei; Saldanha, Emily
    ACS Omega (2022), 7 (18), 15695-15710CODEN: ACSODF; ISSN:2470-1343. (American Chemical Society)
    Detg. the aq. soly. of mols. is a vital step in many pharmaceutical, environmental, and energy storage applications. Despite efforts made over decades, there are still challenges assocd. with developing a soly. prediction model with satisfactory accuracy for many of these applications. The goals of this study are to assess current deep learning methods for soly. prediction, develop a general model capable of predicting the soly. of a broad range of org. mols., and to understand the impact of data properties, mol. representation, and modeling architecture on predictive performance. Using the largest currently available soly. data set, we implement deep learning-based models to predict soly. from the mol. structure and explore several different mol. representations including mol. descriptors, simplified mol.-input line-entry system strings, mol. graphs, and three-dimensional at. coordinates using four different neural network architectures-fully connected neural networks, recurrent neural networks, graph neural networks (GNNs), and SchNet. We find that models using mol. descriptors achieve the best performance, with GNN models also achieving good performance. We perform extensive error anal. to understand the mol. properties that influence model performance, perform feature anal. to understand which information about the mol. structure is most valuable for prediction, and perform a transfer learning and data size study to understand the impact of data availability on model performance.
  130. 130
    Wu, X.; Yu, L. EPSOL: sequence-based protein solubility prediction using multidimensional embedding. Bioinformatics 2021, 37, 4314– 4320,  DOI: 10.1093/bioinformatics/btab463
    Google Scholar
    There is no corresponding record for this reference.
  131. 131
    Krogh, A. What are artificial neural networks?. Nat. Biotechnol. 2008, 26, 195– 197,  DOI: 10.1038/nbt1386
    Google Scholar
    131
    What are artificial neural networks?
    Krogh, Anders
    Nature Biotechnology (2008), 26 (2), 195-197CODEN: NABIF9; ISSN:1087-0156. (Nature Publishing Group)
    A review. Artificial neural networks have been applied to problems ranging from speech recognition to prediction of protein secondary structure, classification of cancers and gene prediction. How do they work and what might they be good for.
  132. 132
    Rumelhart, D.; Hinton, G. E.; Williams, R. J. Learning internal representations by error propagation. cmapspublic2.ihmc.us 1986, 673– 695
    Google Scholar
    There is no corresponding record for this reference.
  133. 133
    Salehinejad, H.; Sankar, S.; Barfett, J.; Colak, E.; Valaee, S. Recent Advances in Recurrent Neural Networks. arXiv , 2017.
    Google Scholar
    There is no corresponding record for this reference.
  134. 134
    Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, L.; Polosukhin, I. Attention is all you need. Adv. Neural Inf. Process. Syst. 2017, 30
    Google Scholar
    There is no corresponding record for this reference.
  135. 135
    Sherstinsky, A. Fundamentals of recurrent neural network (RNN) and long short-term memory (LSTM) network. arXiv , 2018.
    Google Scholar
    There is no corresponding record for this reference.
  136. 136
    Chen, G. A gentle tutorial of recurrent neural network with error backpropagation. arXiv , 2016.
    Google Scholar
    There is no corresponding record for this reference.
  137. 137
    Cho, K.; van Merrienboer, B.; Gulcehre, C.; Bahdanau, D.; Bougares, F.; Schwenk, H.; Bengio, Y. Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation. arXiv , 2014.
    Google Scholar
    There is no corresponding record for this reference.
  138. 138
    Hochreiter, S.; Schmidhuber, J. Long short-term memory. Neural Comput. 1997, 9, 1735– 1780,  DOI: 10.1162/neco.1997.9.8.1735
    Google Scholar
    138
    Long short-term memory
    Hochreiter S; Schmidhuber J
    Neural computation (1997), 9 (8), 1735-80 ISSN:0899-7667.
    Learning to store information over extended time intervals by recurrent backpropagation takes a very long time, mostly because of insufficient, decaying error backflow. We briefly review Hochreiter's (1991) analysis of this problem, then address it by introducing a novel, efficient, gradient-based method called long short-term memory (LSTM). Truncating the gradient where this does not do harm, LSTM can learn to bridge minimal time lags in excess of 1000 discrete-time steps by enforcing constant error flow through constant error carousels within special units. Multiplicative gate units learn to open and close access to the constant error flow. LSTM is local in space and time; its computational complexity per time step and weight is O(1). Our experiments with artificial data involve local, distributed, real-valued, and noisy pattern representations. In comparisons with real-time recurrent learning, back propagation through time, recurrent cascade correlation, Elman nets, and neural sequence chunking, LSTM leads to many more successful runs, and learns much faster. LSTM also solves complex, artificial long-time-lag tasks that have never been solved by previous recurrent network algorithms.
  139. 139
    Graves, A.; Schmidhuber, J. Framewise phoneme classification with bidirectional LSTM and other neural network architectures. Neural Netw. 2005, 18, 602– 610,  DOI: 10.1016/j.neunet.2005.06.042
    Google Scholar
    139
    Framewise phoneme classification with bidirectional LSTM and other neural network architectures
    Graves Alex; Schmidhuber Jurgen
    Neural networks : the official journal of the International Neural Network Society (2005), 18 (5-6), 602-10 ISSN:0893-6080.
    In this paper, we present bidirectional Long Short Term Memory (LSTM) networks, and a modified, full gradient version of the LSTM learning algorithm. We evaluate Bidirectional LSTM (BLSTM) and several other network architectures on the benchmark task of framewise phoneme classification, using the TIMIT database. Our main findings are that bidirectional networks outperform unidirectional ones, and Long Short Term Memory (LSTM) is much faster and also more accurate than both standard Recurrent Neural Nets (RNNs) and time-windowed Multilayer Perceptrons (MLPs). Our results support the view that contextual information is crucial to speech processing, and suggest that BLSTM is an effective architecture with which to exploit it.
  140. 140
    Thafar, M. A.; Alshahrani, M.; Albaradei, S.; Gojobori, T.; Essack, M.; Gao, X. Affinity2Vec: drug-target binding affinity prediction through representation learning, graph mining, and machine learning. Sci. Rep. 2022, 12, 4751,  DOI: 10.1038/s41598-022-08787-9
    Google Scholar
    140
    Affinity2Vec: drug-target binding affinity prediction through representation learning, graph mining, and machine learning
    Thafar, Maha A.; Alshahrani, Mona; Albaradei, Somayah; Gojobori, Takashi; Essack, Magbubah; Gao, Xin
    Scientific Reports (2022), 12 (1), 4751CODEN: SRCEC3; ISSN:2045-2322. (Nature Portfolio)
    Abstr.: Drug-target interaction (DTI) prediction plays a crucial role in drug repositioning and virtual drug screening. Most DTI prediction methods cast the problem as a binary classification task to predict if interactions exist or as a regression task to predict continuous values that indicate a drug's ability to bind to a specific target. The regression-based methods provide insight beyond the binary relationship. However, most of these methods require the three-dimensional (3D) structural information of targets which are still not generally available to the targets. Despite this bottleneck, only a few methods address the drug-target binding affinity (DTBA) problem from a non-structure-based approach to avoid the 3D structure limitations. Here we propose Affinity2Vec, as a novel regression-based method that formulates the entire task as a graph-based problem. To develop this method, we constructed a weighted heterogeneous graph that integrates data from several sources, including drug-drug similarity, target-target similarity, and drug-target binding affinities. Affinity2Vec further combines several computational techniques from feature representation learning, graph mining, and machine learning to generate or ext. features, build the model, and predict the binding affinity between the drug and the target with no 3D structural data. We conducted extensive expts. to evaluate and demonstrate the robustness and efficiency of the proposed method on benchmark datasets used in state-of-the-art non-structured-based drug-target binding affinity studies. Affinity2Vec showed superior and competitive results compared to the state-of-the-art methods based on several evaluation metrics, including mean squared error, rm2, concordance index, and area under the precision-recall curve.
  141. 141
    Wei, B.; Zhang, Y.; Gong, X. 519. DeepLPI: A Novel Drug Repurposing Model based on Ligand-Protein Interaction Using Deep Learning. Open Forum Infect. Dis. 2022, 9, ofac492.574,  DOI: 10.1093/ofid/ofac492.574
    Google Scholar
    There is no corresponding record for this reference.
  142. 142
    Yuan, W.; Chen, G.; Chen, C. Y.-C. FusionDTA: attention-based feature polymerizer and knowledge distillation for drug-target binding affinity prediction. Brief. Bioinform. 2022,  DOI: 10.1093/bib/bbab506
    Google Scholar
    There is no corresponding record for this reference.
  143. 143
    West-Roberts, J.; Valentin-Alvarado, L.; Mullen, S.; Sachdeva, R.; Smith, J.; Hug, L. A.; Gregoire, D. S.; Liu, W.; Lin, T.-Y.; Husain, G.; Amano, Y.; Ly, L.; Banfield, J. F. Giant genes are rare but implicated in cell wall degradation by predatory bacteria. bioRxiv , 2023.
    Google Scholar
    There is no corresponding record for this reference.
  144. 144
    Hernández, A.; Amigó, J. Attention mechanisms and their applications to complex systems. Entropy (Basel) 2021, 23, 283,  DOI: 10.3390/e23030283
    Google Scholar
    There is no corresponding record for this reference.
  145. 145
    Yang, X. An overview of the attention mechanisms in computer vision. 2020.
    Google Scholar
    There is no corresponding record for this reference.
  146. 146
    Hu, D. An introductory survey on attention mechanisms in NLP problems. arXiv , 2018.
    Google Scholar
    There is no corresponding record for this reference.
  147. 147
    Vig, J., Madani, A., Varshney, L. R., Xiong, C., Socher, R., Rajani, N. F. BERTology Meets Biology: Interpreting Attention in Protein Language Models. arXiv , 2020.
    Google Scholar
    There is no corresponding record for this reference.
  148. 148
    Wu, H.; Liu, J.; Jiang, T.; Zou, Q.; Qi, S.; Cui, Z.; Tiwari, P.; Ding, Y. AttentionMGT-DTA: A multi-modal drug-target affinity prediction using graph transformer and attention mechanism. Neural Netw. 2024, 169, 623– 636,  DOI: 10.1016/j.neunet.2023.11.018
    Google Scholar
    There is no corresponding record for this reference.
  149. 149
    Koyama, K.; Kamiya, K.; Shimada, K. Cross attention dti: Drug-target interaction prediction with cross attention module in the blind evaluation setup. BIOKDD2020 2020.
    Google Scholar
    There is no corresponding record for this reference.
  150. 150
    Kurata, H.; Tsukiyama, S. ICAN: Interpretable cross-attention network for identifying drug and target protein interactions. PLoS One 2022, 17, e0276609,  DOI: 10.1371/journal.pone.0276609
    Google Scholar
    There is no corresponding record for this reference.
  151. 151
    Zhao, Q.; Zhao, H.; Zheng, K.; Wang, J. HyperAttentionDTI: improving drug–protein interaction prediction by sequence-based deep learning with attention mechanism. Bioinformatics 2022, 38, 655– 662,  DOI: 10.1093/bioinformatics/btab715
    Google Scholar
    151
    HyperAttentionDTI: improving drug-protein interaction prediction by sequence-based deep learning with attention mechanism
    Zhao, Qichang; Zhao, Haochen; Zheng, Kai; Wang, Jianxin
    Bioinformatics (2022), 38 (3), 655-662CODEN: BOINFP; ISSN:1367-4811. (Oxford University Press)
    Motivation: Identifying drug-target interactions (DTIs) is a crucial step in drug repurposing and drug discovery. Accurately identifying DTIs in silico can significantly shorten development time and reduce costs. Recently, many sequence-based methods are proposed for DTI prediction and improve performance by introducing the attention mechanism. However, these methods only model single non-covalent inter-mol. interactions among drugs and proteins and ignore the complex interaction between atoms and amino acids. Results: In this article, we propose an end-to-end bio-inspired model based on the convolutional neural network (CNN) and attention mechanism, named HyperAttentionDTI, for predicting DTIs. We use deep CNNs to learn the feature matrixes of drugs and proteins. To model complex non-covalent inter-mol. interactions among atoms and amino acids, we utilize the attention mechanism on the feature matrixes and assign an attention vector to each atom or amino acid. We evaluate HpyerAttentionDTI on three benchmark datasets and the results show that our model achieves significantly improved performance compared with the state-of-the-art baselines. Moreover, a case study on the human Gamma-aminobutyric acid receptors confirm that our model can be used as a powerful tool to predict DTIs.
  152. 152
    Jiang, M.; Li, Z.; Zhang, S.; Wang, S.; Wang, X.; Yuan, Q. Drug-target affinity prediction using graph neural network and contact maps. RSC Adv. 2020, 10, 20701,  DOI: 10.1039/D0RA02297G
    Google Scholar
    152
    Drug-target affinity prediction using graph neural network and contact maps
    Jiang, Mingjian; Li, Zhen; Zhang, Shugang; Wang, Shuang; Wang, Xiaofeng; Yuan, Qing; Wei, Zhiqiang
    RSC Advances (2020), 10 (35), 20701-20712CODEN: RSCACL; ISSN:2046-2069. (Royal Society of Chemistry)
    Computer-aided drug design uses high-performance computers to simulate the tasks in drug design, which is a promising research area. Drug-target affinity (DTA) prediction is the most important step of computer-aided drug design, which could speed up drug development and reduce resource consumption. With the development of deep learning, the introduction of deep learning to DTA prediction and improving the accuracy have become a focus of research. In this paper, utilizing the structural information of mols. and proteins, two graphs of drug mols. and proteins are built up resp. Graph neural networks are introduced to obtain their representations, and a method called DGraphDTA is proposed for DTA prediction. Specifically, the protein graph is constructed based on the contact map output from the prediction method, which could predict the structural characteristics of the protein according to its sequence. It can be seen from the test of various metrics on benchmark datasets that the method proposed in this paper has strong robustness and generalizability.
  153. 153
    Nguyen, T. M.; Nguyen, T.; Le, T. M.; Tran, T. GEFA: Early Fusion Approach in Drug-Target Affinity Prediction. IEEE/ACM Trans. Comput. Biol. Bioinform. 2022, 19, 718– 728,  DOI: 10.1109/TCBB.2021.3094217
    Google Scholar
    153
    GEFA: early fusion approach in drug-target affinity prediction
    Nguyen, Tri Minh; Nguyen, Thin; Le, Thao Minh; Tran, Truyen
    IEEE/ACM Transactions on Computational Biology and Bioinformatics (2022), 19 (2), 718-728CODEN: ITCBCY; ISSN:1557-9964. (Institute of Electrical and Electronics Engineers)
    Predicting the interaction between a compd. and a target is crucial for rapid drug repurposing. Deep learning has been successfully applied in drug-target affinity (DTA) problem. However, previous deep learning-based methods ignore modeling the direct interactions between drug and protein residues. This would lead to inaccurate learning of target representation which may change due to the drug binding effects. In addn., previous DTA methods learn protein representation solely based on a small no. of protein sequences in DTA datasets while neglecting the use of proteins outside of the DTA datasets. We propose GEFA (Graph Early Fusion Affinity), a novel graph-in-graph neural network with attention mechanism to address the changes in target representation because of the binding effects. Specifically, a drug is modeled as a graph of atoms, which then serves as a node in a larger graph of residues-drug complex. The resulting model is an expressive deep nested graph neural network. We also use pre-trained protein representation powered by the recent effort of learning contextualized protein representation. The expts. are conducted under different settings to evaluate scenarios such as novel drugs or targets. The results demonstrate the effectiveness of the pre-trained protein embedding and the advantages our GEFA in modeling the nested graph for drug-target interaction.
  154. 154
    Yu, J.; Li, Z.; Chen, G.; Kong, X.; Hu, J.; Wang, D.; Cao, D.; Li, Y.; Huo, R.; Wang, G.; Liu, X.; Jiang, H.; Li, X.; Luo, X.; Zheng, M. Computing the relative binding affinity of ligands based on a pairwise binding comparison network. Nature Computational Science 2023, 3, 860– 872,  DOI: 10.1038/s43588-023-00529-9
    Google Scholar
    There is no corresponding record for this reference.
  155. 155
    Knutson, C.; Bontha, M.; Bilbrey, J. A.; Kumar, N. Decoding the protein–ligand interactions using parallel graph neural networks. Sci. Rep. 2022, 12, 1– 14,  DOI: 10.1038/s41598-022-10418-2
    Google Scholar
    There is no corresponding record for this reference.
  156. 156
    Kyro, G. W.; Brent, R. I.; Batista, V. S. HAC-Net: A Hybrid Attention-Based Convolutional Neural Network for Highly Accurate Protein–Ligand Binding Affinity Prediction. J. Chem. Inf. Model. 2023, 63, 1947– 1960,  DOI: 10.1021/acs.jcim.3c00251
    Google Scholar
    There is no corresponding record for this reference.
  157. 157
    Yousefi, N.; Yazdani-Jahromi, M.; Tayebi, A.; Kolanthai, E.; Neal, C. J.; Banerjee, T.; Gosai, A.; Balasubramanian, G.; Seal, S.; Ozmen Garibay, O. BindingSite-AugmentedDTA: enabling a next-generation pipeline for interpretable prediction models in drug repurposing. Brief. Bioinform. 2023,  DOI: 10.1093/bib/bbad136
    Google Scholar
    There is no corresponding record for this reference.
  158. 158
    Yazdani-Jahromi, M.; Yousefi, N.; Tayebi, A.; Kolanthai, E.; Neal, C. J.; Seal, S.; Garibay, O. O. AttentionSiteDTI: an interpretable graph-based model for drug-target interaction prediction using NLP sentence-level relation classification. Brief. Bioinform. 2022,  DOI: 10.1093/bib/bbac272
    Google Scholar
    There is no corresponding record for this reference.
  159. 159
    Bronstein, M. M.; Bruna, J.; Cohen, T.; Veličković, P. Geometric deep learning: Grids, groups, graphs, geodesics, and gauges. arXiv , 2021.
    Google Scholar
    There is no corresponding record for this reference.
  160. 160
    Lim, J.; Ryu, S.; Park, K.; Choe, Y. J.; Ham, J.; Kim, W. Y. Predicting Drug–Target Interaction Using a Novel Graph Neural Network with 3D Structure-Embedded Graph Representation. J. Chem. Inf. Model. 2019, 59, 3981– 3988,  DOI: 10.1021/acs.jcim.9b00387
    Google Scholar
    160
    Predicting Drug-Target Interaction Using a Novel Graph Neural Network with 3D Structure-Embedded Graph Representation
    Lim, Jaechang; Ryu, Seongok; Park, Kyubyong; Choe, Yo Joong; Ham, Jiyeon; Kim, Woo Youn
    Journal of Chemical Information and Modeling (2019), 59 (9), 3981-3988CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    We propose a novel deep learning approach for predicting drug-target interaction using a graph neural network. We introduce a distance-aware graph attention algorithm to differentiate various types of intermol. interactions. Furthermore, we ext. the graph feature of intermol. interactions directly from the 3D structural information on the protein-ligand binding pose. Thus, the model can learn key features for accurate predictions of drug-target interaction rather than just memorize certain patterns of ligand mols. As a result, our model shows better performance than docking and other deep learning methods for both virtual screening (AUROC of 0.968 for the DUD-E test set) and pose prediction (AUROC of 0.935 for the PDBbind test set). In addn., it can reproduce the natural population distribution of active mols. and inactive mols.
  161. 161
    Jin, Z.; Wu, T.; Chen, T.; Pan, D.; Wang, X.; Xie, J.; Quan, L.; Lyu, Q. CAPLA: improved prediction of protein–ligand binding affinity by a deep learning approach based on a cross-attention mechanism. Bioinformatics 2023, 39, btad049,  DOI: 10.1093/bioinformatics/btad049
    Google Scholar
    There is no corresponding record for this reference.
  162. 162
    Lin, Z.; Akin, H.; Rao, R.; Hie, B.; Zhu, Z.; Lu, W.; Smetanin, N.; Verkuil, R.; Kabeli, O.; Shmueli, Y.; Dos Santos Costa, A.; Fazel-Zarandi, M.; Sercu, T.; Candido, S.; Rives, A. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science 2023, 379, 1123– 1130,  DOI: 10.1126/science.ade2574
    Google Scholar
    162
    Evolutionary-scale prediction of atomic-level protein structure with a language model
    Lin, Zeming; Akin, Halil; Rao, Roshan; Hie, Brian; Zhu, Zhongkai; Lu, Wenting; Smetanin, Nikita; Verkuil, Robert; Kabeli, Ori; Shmueli, Yaniv; dos Santos Costa, Allan; Fazel-Zarandi, Maryam; Sercu, Tom; Candido, Salvatore; Rives, Alexander
    Science (Washington, DC, United States) (2023), 379 (6637), 1123-1130CODEN: SCIEAS; ISSN:1095-9203. (American Association for the Advancement of Science)
    Recent advances in machine learning have leveraged evolutionary information in multiple sequence alignments to predict protein structure. We demonstrate direct inference of full at.-level protein structure from primary sequence using a large language model. As language models of protein sequences are scaled up to 15 billion parameters, an at.-resoln. picture of protein structure emerges in the learned representations. This results in an order-of-magnitude acceleration of high-resoln. structure prediction, which enables large-scale structural characterization of metagenomic proteins. We apply this capability to construct the ESM Metagenomic Atlas by predicting structures for >617 million metagenomic protein sequences, including >225 million that are predicted with high confidence, which gives a view into the vast breadth and diversity of natural proteins.
  163. 163
    Zhang, S.; Fan, R.; Liu, Y.; Chen, S.; Liu, Q.; Zeng, W. Applications of transformer-based language models in bioinformatics: a survey. Bioinform. Adv. 2023, 3, vbad001,  DOI: 10.1093/bioadv/vbad001
    Google Scholar
    There is no corresponding record for this reference.
  164. 164
    Bahdanau, D.; Cho, K.; Bengio, Y. Neural Machine Translation by Jointly Learning to Align and Translate. arXiv , 2014.
    Google Scholar
    There is no corresponding record for this reference.
  165. 165
    Vinyals, O.; Toshev, A.; Bengio, S.; Erhan, D. Show and tell: A neural image caption generator. Pattern Recognition (CVPR) 2015, 3156– 3164
    Google Scholar
    There is no corresponding record for this reference.
  166. 166
    Sutskever, I.; Vinyals, O.; Le, Q. V. Sequence to sequence learning with Neural Networks. arXiv , 2014;.
    Google Scholar
    There is no corresponding record for this reference.
  167. 167
    Cho, K.; van Merrienboer, B.; Bahdanau, D.; Bengio, Y. On the properties of neural machine translation: Encoder-decoder approaches. arXiv , 2014.
    Google Scholar
    There is no corresponding record for this reference.
  168. 168
    Devlin, J.; Chang, M.-W.; Lee, K.; Toutanova, K. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. arXiv , 2018.
    Google Scholar
    There is no corresponding record for this reference.
  169. 169
    Zeyer, A.; Bahar, P.; Irie, K.; Schlüter, R.; Ney, H. A Comparison of Transformer and LSTM Encoder Decoder Models for ASR. 2019 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU) 2019, 8– 15
    Google Scholar
    There is no corresponding record for this reference.
  170. 170
    Irie, K.; Zeyer, A.; Schlüter, R.; Ney, H. Language Modeling with Deep Transformers. arXiv , 2019.
    Google Scholar
    There is no corresponding record for this reference.
  171. 171
    Zouitni, C.; Sabri, M. A.; Aarab, A. A Comparison Between LSTM and Transformers for Image Captioning. Digital Technologies and Applications 2023, 669, 492– 500,  DOI: 10.1007/978-3-031-29860-8_50
    Google Scholar
    There is no corresponding record for this reference.
  172. 172
    Parisotto, E.; Song, F.; Rae, J.; Pascanu, R.; Gulcehre, C.; Jayakumar, S.; Jaderberg, M.; Kaufman, R. L.; Clark, A.; Noury, S.; Botvinick, M.; Heess, N.; Hadsell, R. Stabilizing Transformers for Reinforcement Learning. Proceedings of the 37th International Conference on Machine Learning 2020, 7487– 7498
    Google Scholar
    There is no corresponding record for this reference.
  173. 173
    Bilokon, P.; Qiu, Y. Transformers versus LSTMs for electronic trading. arXiv , 2023.
    Google Scholar
    There is no corresponding record for this reference.
  174. 174
    Merity, S. Single Headed Attention RNN: Stop Thinking With Your Head. arXiv , 2019.
    Google Scholar
    There is no corresponding record for this reference.
  175. 175
    Ezen-Can, A. A Comparison of LSTM and BERT for Small Corpus. arXiv , 2020.
    Google Scholar
    There is no corresponding record for this reference.
  176. 176
    Unsal, S.; Atas, H.; Albayrak, M.; Turhan, K.; Acar, A. C.; Doğan, T. Learning functional properties of proteins with language models. Nature Machine Intelligence 2022, 4, 227– 245,  DOI: 10.1038/s42256-022-00457-9
    Google Scholar
    There is no corresponding record for this reference.
  177. 177
    Brandes, N.; Ofer, D.; Peleg, Y.; Rappoport, N.; Linial, M. ProteinBERT: a universal deep-learning model of protein sequence and function. Bioinformatics 2022, 38, 2102– 2110,  DOI: 10.1093/bioinformatics/btac020
    Google Scholar
    177
    ProteinBERT: a universal deep-learning model of protein sequence and function
    Brandes, Nadav; Ofer, Dan; Peleg, Yam; Rappoport, Nadav; Linial, Michal
    Bioinformatics (2022), 38 (8), 2102-2110CODEN: BOINFP; ISSN:1367-4811. (Oxford University Press)
    Self-supervised deep language modeling has shown unprecedented success across natural language tasks, and has recently been repurposed to biol. sequences. However, existing models and pretraining methods are designed and optimized for text anal. We introduce ProteinBERT, a deep language model specifically designed for proteins. Our pretraining scheme combines language modeling with a novel task of Gene Ontol. (GO) annotation prediction. We introduce novel architectural elements that make the model highly efficient and flexible to long sequences. The architecture of ProteinBERT consists of both local and global representations, allowing end-to-end processing of these types of inputs and outputs. ProteinBERT obtains near state-of-the-art performance, and sometimes exceeds it, on multiple benchmarks covering diverse protein properties (including protein structure, post-translational modifications and biophys. attributes), despite using a far smaller and faster model than competing deep-learning methods. Overall, ProteinBERT provides an efficient framework for rapidly training protein predictors, even with limited labeled data.
  178. 178
    Luo, S.; Chen, T.; Xu, Y.; Zheng, S.; Liu, T.-Y.; Wang, L.; He, D. One Transformer Can Understand Both 2D & 3D Molecular Data. arXiv , 2022.
    Google Scholar
    There is no corresponding record for this reference.
  179. 179
    Clark, K.; Luong, M.-T.; Le, Q. V.; Manning, C. D. ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators. arXiv , 2020.
    Google Scholar
    There is no corresponding record for this reference.
  180. 180
    Wang, J.; Wen, N.; Wang, C.; Zhao, L.; Cheng, L. ELECTRA-DTA: a new compound-protein binding affinity prediction model based on the contextualized sequence encoding. J. Cheminform. 2022, 14, 14,  DOI: 10.1186/s13321-022-00591-x
    Google Scholar
    There is no corresponding record for this reference.
  181. 181
    Shin, B.; Park, S.; Kang, K.; Ho, J. C. Self-Attention Based Molecule Representation for Predicting Drug-Target Interaction. Proceedings of the 4th Machine Learning for Healthcare Conference 2019, 230– 248
    Google Scholar
    There is no corresponding record for this reference.
  182. 182
    Huang, K.; Xiao, C.; Glass, L. M.; Sun, J. MolTrans: Molecular Interaction Transformer for drug–target interaction prediction. Bioinformatics 2021, 37, 830– 836,  DOI: 10.1093/bioinformatics/btaa880
    Google Scholar
    182
    MolTrans: molecular interaction transformer for drug-target interaction prediction
    Huang, Kexin; Xiao, Cao; Glass, Lucas M.; Sun, Jimeng
    Bioinformatics (2021), 37 (6), 830-836CODEN: BOINFP; ISSN:1367-4811. (Oxford University Press)
    Motivation: Drug-target interaction (DTI) prediction is a foundational task for in-silico drug discovery, which is costly and time-consuming due to the need of exptl. search over large drug compd. space. Recent years have witnessed promising progress for deep learning in DTI predictions. However, the following challenges are still open: (i) existing mol. representation learning approaches ignore the sub-structural nature of DTI, thus produce results that are less accurate and difficult to explain and (ii) existing methods focus on limited labeled data while ignoring the value of massive unlabeled mol. data. Results: We propose a Mol. Interaction Transformer (MolTrans) to address these limitations via: (i) knowledge inspired sub-structural pattern mining algorithm and interaction modeling module for more accurate and interpretable DTI prediction and (ii) an augmented transformer encoder to better ext. and capture the semantic relations among sub-structures extd. from massive unlabeled biomedical data. We evaluate MolTrans on real-world data and show it improved DTI prediction performance compared to state-of-the-art baselines.
  183. 183
    Shen, L.; Feng, H.; Qiu, Y.; Wei, G.-W. SVSBI: sequence-based virtual screening of biomolecular interactions. Commun. Biol. 2023, 6, 536,  DOI: 10.1038/s42003-023-04866-3
    Google Scholar
    183
    SVSBI: sequence-based virtual screening of biomolecular interactions
    Shen Li; Feng Hongsong; Qiu Yuchi; Wei Guo-Wei; Wei Guo-Wei; Wei Guo-Wei
    Communications biology (2023), 6 (1), 536 ISSN:.
    Virtual screening (VS) is a critical technique in understanding biomolecular interactions, particularly in drug design and discovery. However, the accuracy of current VS models heavily relies on three-dimensional (3D) structures obtained through molecular docking, which is often unreliable due to the low accuracy. To address this issue, we introduce a sequence-based virtual screening (SVS) as another generation of VS models that utilize advanced natural language processing (NLP) algorithms and optimized deep K-embedding strategies to encode biomolecular interactions without relying on 3D structure-based docking. We demonstrate that SVS outperforms state-of-the-art performance for four regression datasets involving protein-ligand binding, protein-protein, protein-nucleic acid binding, and ligand inhibition of protein-protein interactions and five classification datasets for protein-protein interactions in five biological species. SVS has the potential to transform current practices in drug discovery and protein engineering.
  184. 184
    Wang, J.; Hu, J.; Sun, H.; Xu, M.; Yu, Y.; Liu, Y.; Cheng, L. MGPLI: exploring multigranular representations for protein–ligand interaction prediction. Bioinformatics 2022, 38, 4859– 4867,  DOI: 10.1093/bioinformatics/btac597
    Google Scholar
    There is no corresponding record for this reference.
  185. 185
    Qian, Y.; Wu, J.; Zhang, Q. CAT-CPI: Combining CNN and transformer to learn compound image features for predicting compound-protein interactions. Front Mol. Biosci 2022, 9, 963912,  DOI: 10.3389/fmolb.2022.963912
    Google Scholar
    There is no corresponding record for this reference.
  186. 186
    Cang, Z.; Mu, L.; Wei, G.-W. Representability of algebraic topology for biomolecules in machine learning based scoring and virtual screening. PLoS Comput. Biol. 2018, 14, e1005929,  DOI: 10.1371/journal.pcbi.1005929
    Google Scholar
    186
    Representability of algebraic topology for biomolecules in machine learning based scoring and virtual screening
    Cang, Zixuan; Mu, Lin; Wei, Guo-Wei
    PLoS Computational Biology (2018), 14 (1), e1005929/1-e1005929/44CODEN: PCBLBG; ISSN:1553-7358. (Public Library of Science)
    This work introduces a no. of algebraic topol. approaches, including multi-component persistent homol., multi-level persistent homol., and electrostatic persistence for the representation, characterization, and description of small mols. and biomol. complexes. In contrast to the conventional persistent homol., multi-component persistent homol. retains crit. chem. and biol. information during the topol. simplification of biomol. geometric complexity. Multi-level persistent homol. enables a tailored topol. description of inter- and/or intra-mol. interactions of interest. Electrostatic persistence incorporates partial charge information into topol. invariants. These topol. methods are paired with Wasserstein distance to characterize similarities between mols. and are further integrated with a variety of machine learning algorithms, including k-nearest neighbors, ensemble of trees, and deep convolutional neural networks, to manifest their descriptive and predictive powers for protein-ligand binding anal. and virtual screening of small mols. Extensive numerical expts. involving 4,414 protein- ligand complexes from the PDBBind database and 128,374 ligand-target and decoytarget pairs in the DUD database are performed to test resp. the scoring power and the discriminatory power of the proposed topol. learning strategies. It is demonstrated that the present topol. learning outperforms other existing methods in protein-ligand binding affinity prediction and ligand-decoy discrimination.
  187. 187
    Chen, D.; Liu, J.; Wei, G.-W. Multiscale topology-enabled structure-to-sequence transformer for protein-ligand interaction predictions. Nat. Mac. Intell. 2024, 6, 799– 810,  DOI: 10.1038/s42256-024-00855-1
    Google Scholar
    There is no corresponding record for this reference.
  188. 188
    Bubeck, S.; Chandrasekaran, V.; Eldan, R.; Gehrke, J.; Horvitz, E.; Kamar, E.; Lee, P.; Lee, Y. T.; Li, Y.; Lundberg, S.; Nori, H.; Palangi, H.; Ribeiro, M. T.; Zhang, Y. Sparks of Artificial General Intelligence: Early experiments with GPT-4. arXiv , 2023.
    Google Scholar
    There is no corresponding record for this reference.
  189. 189
    Elnaggar, A.; Essam, H.; Salah-Eldin, W.; Moustafa, W.; Elkerdawy, M.; Rochereau, C.; Rost, B. Ankh: Optimized protein language model unlocks general-purpose modelling. bioRxiv , 2023.
    Google Scholar
    There is no corresponding record for this reference.
  190. 190
    Hwang, Y.; Cornman, A. L.; Kellogg, E. H.; Ovchinnikov, S.; Girguis, P. R. Genomic language model predicts protein co-regulation and function. Nat. Commun. 2024, 15, 2880,  DOI: 10.1038/s41467-024-46947-9
    Google Scholar
    There is no corresponding record for this reference.
  191. 191
    Vu, M. H.; Akbar, R.; Robert, P. A.; Swiatczak, B.; Greiff, V.; Sandve, G. K.; Haug, D. T. T. Linguistically inspired roadmap for building biologically reliable protein language models. arXiv , 2022.
    Google Scholar
    There is no corresponding record for this reference.
  192. 192
    Xu, M.; Zhang, Z.; Lu, J.; Zhu, Z.; Zhang, Y.; Ma, C.; Liu, R.; Tang, J. PEER: A comprehensive and multi-task benchmark for Protein sEquence undERstanding. arXiv 2022, 35156– 35173.
    Google Scholar
    There is no corresponding record for this reference.
  193. 193
    Schmirler, R.; Heinzinger, M.; Rost, B. Fine-tuning protein language models boosts predictions across diverse tasks. Nat. Commun. 2024, 15, 7407,  DOI: 10.1038/s41467-024-51844-2
    Google Scholar
    There is no corresponding record for this reference.
  194. 194
    Heinzinger, M.; Elnaggar, A.; Wang, Y.; Dallago, C.; Nechaev, D.; Matthes, F.; Rost, B. Modeling aspects of the language of life through transfer-learning protein sequences. BMC Bioinformatics 2019, 20, 723,  DOI: 10.1186/s12859-019-3220-8
    Google Scholar
    There is no corresponding record for this reference.
  195. 195
    Manfredi, M.; Savojardo, C.; Martelli, P. L.; Casadio, R. E-SNPs&GO: embedding of protein sequence and function improves the annotation of human pathogenic variants. Bioinformatics 2022, 38, 5168– 5174,  DOI: 10.1093/bioinformatics/btac678
    Google Scholar
    There is no corresponding record for this reference.
  196. 196
    Anteghini, M.; Martins Dos Santos, V.; Saccenti, E. In-Pero: Exploiting deep learning embeddings of protein sequences to predict the localisation of peroxisomal proteins. Int. J. Mol. Sci. 2021, 22, 6409,  DOI: 10.3390/ijms22126409
    Google Scholar
    There is no corresponding record for this reference.
  197. 197
    Nguyen, T.; Le, H.; Quinn, T. P.; Nguyen, T.; Le, T. D.; Venkatesh, S. GraphDTA: predicting drug–target binding affinity with graph neural networks. Bioinformatics 2021, 37, 1140– 1147,  DOI: 10.1093/bioinformatics/btaa921
    Google Scholar
    197
    GraphDTA: predicting drug-target binding affinity with graph neural networks
    Nguyen, Thin; Le, Hang; Quinn, Thomas P.; Nguyen, Tri; Le, Thuc Duy; Venkatesh, Svetha
    Bioinformatics (2021), 37 (8), 1140-1147CODEN: BOINFP; ISSN:1367-4811. (Oxford University Press)
    Summary: The development of new drugs is costly, time consuming, often accompanied with safety issues. Drug repurposing can avoid the expensive, lengthy process of drug development by finding new uses for already approved drugs. In order to repurpose drugs effectively, it is useful to know which proteins are targeted by which drugs. Computational models that est. the interaction strength of new drug-target pairs have the potential to expedite drug repurposing. Several models have been proposed for this task. However, these models represent the drugs as strings, which is not a natural way to represent mols. We propose a new model called GraphDTA that represents drugs as graphs, uses graph neural networks to predict drug-target affinity. We show that graph neural networks not only predict drug-target affinity better than non-deep learning models, but also outperform competing deep learning methods. Our results confirm that deep learning models are appropriate for drug- target binding affinity prediction, that representing drugs as graphs can lead to further improvements.
  198. 198
    Nam, H.; Ha, J.-W.; Kim, J. Dual attention networks for multimodal reasoning and matching. Pattern Recognition (CVPR) 2017, 299– 307
    Google Scholar
    There is no corresponding record for this reference.
  199. 199
    Wang, X.; Liu, D.; Zhu, J.; Rodriguez-Paton, A.; Song, T. CSConv2d: A 2-D Structural Convolution Neural Network with a Channel and Spatial Attention Mechanism for Protein-Ligand Binding Affinity Prediction. Biomolecules 2021,  DOI: 10.3390/biom11050643
    Google Scholar
    There is no corresponding record for this reference.
  200. 200
    Anteghini, M.; Santos, V. A. M. D.; Saccenti, E. PortPred: Exploiting deep learning embeddings of amino acid sequences for the identification of transporter proteins and their substrates. J. Cell. Biochem. 2023, 124, 1803,  DOI: 10.1002/jcb.30490
    Google Scholar
    There is no corresponding record for this reference.
  201. 201
    Huang, K.; Fu, T.; Glass, L. M.; Zitnik, M.; Xiao, C.; Sun, J. DeepPurpose: a deep learning library for drug–target interaction prediction. Bioinformatics 2021, 36, 5545– 5547,  DOI: 10.1093/bioinformatics/btaa1005
    Google Scholar
    201
    DeepPurpose: a deep learning library for drug-target interaction prediction
    Huang Kexin; Zitnik Marinka; Fu Tianfan; Glass Lucas M; Xiao Cao; Sun Jimeng
    Bioinformatics (Oxford, England) (2021), 36 (22-23), 5545-5547 ISSN:.
    SUMMARY: Accurate prediction of drug-target interactions (DTI) is crucial for drug discovery. Recently, deep learning (DL) models for show promising performance for DTI prediction. However, these models can be difficult to use for both computer scientists entering the biomedical field and bioinformaticians with limited DL experience. We present DeepPurpose, a comprehensive and easy-to-use DL library for DTI prediction. DeepPurpose supports training of customized DTI prediction models by implementing 15 compound and protein encoders and over 50 neural architectures, along with providing many other useful features. We demonstrate state-of-the-art performance of DeepPurpose on several benchmark datasets. AVAILABILITY AND IMPLEMENTATION: https://github.com/kexinhuang12345/DeepPurpose. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.
  202. 202
    Zhao, L.; Wang, J.; Pang, L.; Liu, Y.; Zhang, J. GANsDTA: Predicting Drug-Target Binding Affinity Using GANs. Front. Genet. 2020, 10, 1243,  DOI: 10.3389/fgene.2019.01243
    Google Scholar
    There is no corresponding record for this reference.
  203. 203
    Hu, F.; Jiang, J.; Wang, D.; Zhu, M.; Yin, P. Multi-PLI: interpretable multi-task deep learning model for unifying protein–ligand interaction datasets. J. Cheminform. 2021, 13, 30,  DOI: 10.1186/s13321-021-00510-6
    Google Scholar
    There is no corresponding record for this reference.
  204. 204
    Zheng, S.; Li, Y.; Chen, S.; Xu, J.; Yang, Y. Predicting Drug Protein Interaction using Quasi-Visual Question Answering System. bioRxiv 2019, 588178
    Google Scholar
    There is no corresponding record for this reference.
  205. 205
    Tsubaki, M.; Tomii, K.; Sese, J. Compound–protein interaction prediction with end-to-end learning of neural networks for graphs and sequences. Bioinformatics 2019, 35, 309– 318,  DOI: 10.1093/bioinformatics/bty535
    Google Scholar
    205
    Compound-protein interaction prediction with end-to-end learning of neural networks for graphs and sequences
    Tsubaki, Masashi; Tomii, Kentaro; Sese, Jun
    Bioinformatics (2019), 35 (2), 309-318CODEN: BOINFP; ISSN:1367-4811. (Oxford University Press)
    Motivation: In bioinformatics, machine learning-based methods that predict the compd.-protein interactions (CPIs) play an important role in the virtual screening for drug discovery. Recently, end-to-end representation learning for discrete symbolic data (e.g. words in natural language processing) using deep neural networks has demonstrated excellent performance on various difficult problems. For the CPI problem, data are provided as discrete symbolic data, i.e. compds. are represented as graphs where the vertices are atoms, the edges are chem. bonds, and proteins are sequences in which the characters are amino acids. In this study, we investigate the use of end-to-end representation learning for compds. and proteins, integrate the representations, and develop a new CPI prediction approach by combining a graph neural network (GNN) for compds. and a convolutional neural network (CNN) for proteins. Results: Our expts. using three CPI datasets demonstrated that the proposed end-to-end approach achieves competitive or higher performance as compared to various existing CPI prediction methods. In addn., the proposed approach significantly outperformed existing methods on an unbalanced dataset. This suggests that data-driven representations of compds. and proteins obtained by end-to-end GNNs and CNNs are more robust than traditional chem. and biol. features obtained from databases. Although analyzing deep learning models is difficult due to their black-box nature, we address this issue using a neural attention mechanism, which allows us to consider which subsequences in a protein are more important for a drug compd. when predicting its interaction. The neural attention mechanism also provides effective visualization, which makes it easier to analyze a model even when modeling is performed using real-valued representations instead of discrete features.
  206. 206
    Karimi, M.; Wu, D.; Wang, Z.; Shen, Y. DeepAffinity: interpretable deep learning of compound–protein affinity through unified recurrent and convolutional neural networks. Bioinformatics 2019, 35, 3329– 3338,  DOI: 10.1093/bioinformatics/btz111
    Google Scholar
    206
    DeepAffinity: interpretable deep learning of compound-protein affinity through unified recurrent and convolutional neural networks
    Karimi, Mostafa; Wu, Di; Wang, Zhangyang; Shen, Yang
    Bioinformatics (2019), 35 (18), 3329-3338CODEN: BOINFP; ISSN:1367-4811. (Oxford University Press)
    Motivation: Drug discovery demands rapid quantification of compd.-protein interaction (CPI). However, there is a lack of methods that can predict compd.-protein affinity from sequences alone with high applicability, accuracy and interpretability. Results: We present a seamless integration of domain knowledges and learning-based approaches. Under novel representations of structurally annotated protein sequences, a semisupervised deep learning model that unifies recurrent and convolutional neural networks has been proposed to exploit both unlabeled and labeled data, for jointly encoding mol. representations and predicting affinities. Our representations and models outperform conventional options in achieving relative error in IC50 within 5-fold for test cases and 20-fold for protein classes not included for training. Performances for new protein classes with few labeled data are further improved by transfer learning. Furthermore, sep. and joint attention mechanisms are developed and embedded to our model to add to its interpretability, as illustrated in case studies for predicting and explaining selective drug-target interactions. Lastly, alternative representations using protein sequences or compd. graphs and a unified RNN/GCNN-CNN model using graph CNN (GCNN) are also explored to reveal algorithmic challenges ahead.
  207. 207
    Li, S.; Wan, F.; Shu, H.; Jiang, T.; Zhao, D.; Zeng, J. MONN: a multi-objective neural network for predicting compound-protein interactions and affinities. Cell Systems 2020, 10, 308– 322,  DOI: 10.1016/j.cels.2020.03.002
    Google Scholar
    207
    MONN: A Multi-objective Neural Network for Predicting Compound-Protein Interactions and Affinities
    Li, Shuya; Wan, Fangping; Shu, Hantao; Jiang, Tao; Zhao, Dan; Zeng, Jianyang
    Cell Systems (2020), 10 (4), 308-322.e11CODEN: CSEYA4; ISSN:2405-4712. (Cell Press)
    Computational approaches for understanding compd.-protein interactions (CPIs) can greatly facilitate drug development. Recently, a no. of deep-learning-based methods have been proposed to predict binding affinities and attempt to capture local interaction sites in compds. and proteins through neural attentions (i.e., neural network architectures that enable the interpretation of feature importance). Here, we compiled a benchmark dataset contg. the inter-mol. non-covalent interactions for more than 10,000 compd.-protein pairs and systematically evaluated the interpretability of neural attentions in existing models. We also developed a multi-objective neural network, called MONN, to predict both non-covalent interactions and binding affinities between compds. and proteins. Comprehensive evaluation demonstrated that MONN can successfully predict the non-covalent interactions between compds. and proteins that cannot be effectively captured by neural attentions in previous prediction methods. Moreover, MONN outperforms other state-of-the-art methods in predicting binding affinities.
  208. 208
    Zhao, M.; Yuan, M.; Yang, Y.; Xu, S. X. CPGL: Prediction of Compound-Protein Interaction by Integrating Graph Attention Network With Long Short-Term Memory Neural Network. IEEE/ACM Trans. Comput. Biol. Bioinform. 2023, 20, 1935– 1942,  DOI: 10.1109/TCBB.2022.3225296
    Google Scholar
    There is no corresponding record for this reference.
  209. 209
    Yu, L.; Qiu, W.; Lin, W.; Cheng, X.; Xiao, X.; Dai, J. HGDTI: predicting drug–target interaction by using information aggregation based on heterogeneous graph neural network. BMC Bioinformatics 2022, 23, 126,  DOI: 10.1186/s12859-022-04655-5
    Google Scholar
    There is no corresponding record for this reference.
  210. 210
    Lee, I.; Nam, H. Sequence-based prediction of protein binding regions and drug-target interactions. J. Cheminform. 2022, 14, 5,  DOI: 10.1186/s13321-022-00584-w
    Google Scholar
    There is no corresponding record for this reference.
  211. 211
    Gönen, M.; Heller, G. Concordance probability and discriminatory power in proportional hazards regression. Biometrika 2005, 92, 965– 970,  DOI: 10.1093/biomet/92.4.965
    Google Scholar
    There is no corresponding record for this reference.
  212. 212
    Deller, M. C.; Rupp, B. Models of protein-ligand crystal structures: trust, but verify. J. Comput. Aided Mol. Des. 2015, 29, 817– 836,  DOI: 10.1007/s10822-015-9833-8
    Google Scholar
    212
    Models of protein-ligand crystal structures: trust, but verify
    Deller, Marc C.; Rupp, Bernhard
    Journal of Computer-Aided Molecular Design (2015), 29 (9), 817-836CODEN: JCADEQ; ISSN:0920-654X. (Springer)
    X-ray crystallog. provides the most accurate models of protein-ligand structures. These models serve as the foundation of many computational methods including structure prediction, mol. modeling, and structure-based drug design. The success of these computational methods ultimately depends on the quality of the underlying protein-ligand models. X-ray crystallog. offers the unparalleled advantage of a clear math. formalism relating the exptl. data to the protein-ligand model. In the case of X-ray crystallog., the primary exptl. evidence is the electron d. of the mols. forming the crystal. The first step in the generation of an accurate and precise crystallog. model is the interpretation of the electron d. of the crystal, typically carried out by construction of an at. model. The at. model must then be validated for fit to the exptl. electron d. and also for agreement with prior expectations of stereochem. Stringent validation of protein-ligand models has become possible as a result of the mandatory deposition of primary diffraction data, and many computational tools are now available to aid in the validation process. Validation of protein-ligand complexes has revealed some instances of overenthusiastic interpretation of ligand d. Fundamental concepts and metrics of protein-ligand quality validation are discussed and we highlight software tools to assist in this process. It is essential that end users select high quality protein-ligand models for their computational and biol. studies, and we provide an overview of how this can be achieved.
  213. 213
    Kalakoti, Y.; Yadav, S.; Sundar, D. TransDTI: Transformer-based language models for estimating DTIs and building a drug recommendation workflow. ACS Omega 2022, 7, 2706– 2717,  DOI: 10.1021/acsomega.1c05203
    Google Scholar
    213
    TransDTI: Transformer-Based Language Models for Estimating DTIs and Building a Drug Recommendation Workflow
    Kalakoti, Yogesh; Yadav, Shashank; Sundar, Durai
    ACS Omega (2022), 7 (3), 2706-2717CODEN: ACSODF; ISSN:2470-1343. (American Chemical Society)
    The identification of novel drug-target interactions is a labor-intensive and low-throughput process. In silico alternatives have proved to be of immense importance in assisting the drug discovery process. Here, we present TransDTI, a multiclass classification and regression workflow employing transformer-based language models to segregate interactions between drug-target pairs as active, inactive, and intermediate. The models were trained with large-scale drug-target interaction (DTI) data sets, which reported an improvement in performance in terms of the area under receiver operating characteristic (auROC), the area under precision recall (auPR), Matthew's correlation coeff. (MCC), and R2 over baseline methods. The results showed that models based on transformer-based language models effectively predict novel drug-target interactions from sequence data. The proposed models significantly outperformed existing methods like DeepConvDTI, DeepDTA, and DeepDTI on a test data set. Further, the validity of novel interactions predicted by TransDTI was found to be backed by mol. docking and simulation anal., where the model prediction had similar or better interaction potential for MAP2k and transforming growth factor-β (TGFβ) and their known inhibitors. Proposed approaches can have a significant impact on the development of personalized therapy and clin. decision making.
  214. 214
    Chatterjee, A.; Walters, R.; Shafi, Z.; Ahmed, O. S.; Sebek, M.; Gysi, D.; Yu, R.; Eliassi-Rad, T.; Barabási, A.-L.; Menichetti, G. Improving the generalizability of protein-ligand binding predictions with AI-Bind. Nat. Commun. 2023, 14, 1989,  DOI: 10.1038/s41467-023-37572-z
    Google Scholar
    There is no corresponding record for this reference.
  215. 215
    Gudivada, V.; Apon, A.; Ding, J. Data quality considerations for big data and machine learning: Going beyond data cleaning and transformations. Int. J. Adv. Softw. 2017, 10, 1– 20
    Google Scholar
    There is no corresponding record for this reference.
  216. 216
    Nasteski, V. An overview of the supervised machine learning methods. Horizons 2017, 4, 51– 62,  DOI: 10.20544/HORIZONS.B.04.1.17.P05
    Google Scholar
    There is no corresponding record for this reference.
  217. 217
    Kozlov, M. So you got a null result. Will anyone publish it?. Nature 2024, 631, 728– 730,  DOI: 10.1038/d41586-024-02383-9
    Google Scholar
    There is no corresponding record for this reference.
  218. 218
    Edfeldt, K. A data science roadmap for open science organizations engaged in early-stage drug discovery. Nat. Commun. 2024, 15, 5640,  DOI: 10.1038/s41467-024-49777-x
    Google Scholar
    There is no corresponding record for this reference.
  219. 219
    Mlinarić, A.; Horvat, M.; Šupak Smolčić, V. Dealing with the positive publication bias: Why you should really publish your negative results. Biochem. Med. 2017, 27, 030201,  DOI: 10.11613/BM.2017.030201
    Google Scholar
    There is no corresponding record for this reference.
  220. 220
    Fanelli, D. Negative results are disappearing from most disciplines and countries. Scientometrics 2012, 90, 891– 904,  DOI: 10.1007/s11192-011-0494-7
    Google Scholar
    There is no corresponding record for this reference.
  221. 221
    Albalate, A.; Minker, W. Semi-supervised and unervised machine learning: Novel strategies; Wiley-ISTE, 2013.
    Google Scholar
    There is no corresponding record for this reference.
  222. 222
    Sajadi, S. Z.; Zare Chahooki, M. A.; Gharaghani, S.; Abbasi, K. AutoDTI++: deep unsupervised learning for DTI prediction by autoencoders. BMC Bioinformatics 2021, 22, 204,  DOI: 10.1186/s12859-021-04127-2
    Google Scholar
    There is no corresponding record for this reference.
  223. 223
    Najm, M.; Azencott, C.-A.; Playe, B.; Stoven, V. Drug Target Identification with Machine Learning: How to Choose Negative Examples. Int. J. Mol. Sci. 2021, 22, 5118,  DOI: 10.3390/ijms22105118
    Google Scholar
    There is no corresponding record for this reference.
  224. 224
    Sieg, J.; Flachsenberg, F.; Rarey, M. In Need of Bias Control: Evaluating Chemical Data for Machine Learning in Structure-Based Virtual Screening. J. Chem. Inf. Model. 2019, 59, 947– 961,  DOI: 10.1021/acs.jcim.8b00712
    Google Scholar
    224
    In Need of Bias Control: Evaluating Chemical Data for Machine Learning in Structure-Based Virtual Screening
    Sieg, Jochen; Flachsenberg, Florian; Rarey, Matthias
    Journal of Chemical Information and Modeling (2019), 59 (3), 947-961CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    A review. Reports of successful applications of machine learning (ML) methods in structure-based virtual screening (SBVS) are increasing. ML methods such as convolutional neural networks show promising results and often outperform traditional methods such as empirical scoring functions in retrospective validation. However, trained ML models are often treated as black boxes and are not straightforwardly interpretable. In most cases, it is unknown which features in the data are decisive and whether a model's predictions are right for the right reason. Hence, the authors reevaluated three widely used benchmark data sets in the context of ML methods and came to the conclusion that not every benchmark data set is suitable. Moreover, the authors demonstrate on two examples from current literature that bias is learned implicitly and unnoticed from std. benchmarks. On the basis of these results, the authors conclude that there is a need for eligible validation expts. and benchmark data sets suited to ML for more bias-controlled validation in ML-based SBVS. Therefore, the authors provide guidelines for setting up validation expts. and give a perspective on how new data sets could be generated.
  225. 225
    Volkov, M.; Turk, J.-A.; Drizard, N.; Martin, N.; Hoffmann, B.; Gaston-Mathé, Y.; Rognan, D. On the Frustration to Predict Binding Affinities from Protein–Ligand Structures with Deep Neural Networks. J. Med. Chem. 2022, 65, 7946– 7958,  DOI: 10.1021/acs.jmedchem.2c00487
    Google Scholar
    225
    On the Frustration to Predict Binding Affinities from Protein-Ligand Structures with Deep Neural Networks
    Volkov, Mikhail; Turk, Joseph-Andre; Drizard, Nicolas; Martin, Nicolas; Hoffmann, Brice; Gaston-Mathe, Yann; Rognan, Didier
    Journal of Medicinal Chemistry (2022), 65 (11), 7946-7958CODEN: JMCMAR; ISSN:0022-2623. (American Chemical Society)
    Accurate prediction of binding affinities from protein-ligand at. coordinates remains a major challenge in early stages of drug discovery. Using modular message passing graph neural networks describing both the ligand and the protein in their free and bound states, we unambiguously evidence that an explicit description of protein-ligand noncovalent interactions does not provide any advantage with respect to ligand or protein descriptors. Simple models, inferring binding affinities of test samples from that of the closest ligands or proteins in the training set, already exhibit good performances, suggesting that memorization largely dominates true learning in the deep neural networks. The current study suggests considering only noncovalent interactions while omitting their protein and ligand at. environments. Removing all hidden biases probably requires much denser protein-ligand training matrixes and a coordinated effort of the drug design community to solve the necessary protein-ligand structures.
  226. 226
    Shivakumar, D.; Williams, J.; Wu, Y.; Damm, W.; Shelley, J.; Sherman, W. Prediction of absolute solvation free energies using molecular dynamics free energy perturbation and the OPLS force field. J. Chem. Theory Comput. 2010, 6, 1509– 1519,  DOI: 10.1021/ct900587b
    Google Scholar
    226
    Prediction of Absolute Solvation Free Energies using Molecular Dynamics Free Energy Perturbation and the OPLS Force Field
    Shivakumar, Devleena; Williams, Joshua; Wu, Yujie; Damm, Wolfgang; Shelley, John; Sherman, Woody
    Journal of Chemical Theory and Computation (2010), 6 (5), 1509-1519CODEN: JCTCCE; ISSN:1549-9618. (American Chemical Society)
    The accurate prediction of protein-ligand binding free energies is a primary objective in computer-aided drug design. The solvation free energy of a small mol. provides a surrogate to the desolvation of the ligand in the thermodn. process of protein-ligand binding. Here, we use explicit solvent mol. dynamics free energy perturbation to predict the abs. solvation free energies of a set of 239 small mols., spanning diverse chem. functional groups commonly found in drugs and drug-like mols. We also compare the performance of abs. solvation free energies obtained using the OPLS_2005 force field with two other commonly used small mol. force fields - general AMBER force field (GAFF) with AM1-BCC charges and CHARMm-MSI with CHelpG charges. Using the OPLS_2005 force field, we obtain high correlation with exptl. solvation free energies (R2 = 0.94) and low av. unsigned errors for a majority of the functional groups compared to AM1-BCC/GAFF or CHelpG/CHARMm-MSI. However, OPLS_2005 has errors of over 1.3 kcal/mol for certain classes of polar compds. We show that predictions on these compd. classes can be improved by using a semiempirical charge assignment method with an implicit bond charge correction.
  227. 227
    El Hage, K.; Mondal, P.; Meuwly, M. Free energy simulations for protein ligand binding and stability. Mol. Simul. 2018, 44, 1044– 1061,  DOI: 10.1080/08927022.2017.1416115
    Google Scholar
    227
    Free energy simulations for protein ligand binding and stability
    El Hage, Krystel; Mondal, Padmabati; Meuwly, Markus
    Molecular Simulation (2018), 44 (13-14), 1044-1061CODEN: MOSIEA; ISSN:0892-7022. (Taylor & Francis Ltd.)
    We summarize several computational techniques to det. relative free energies for condensed-phase systems. The focus is on practical considerations which are capable of making direct contact with expts. Particular applications include the thermodn. stability of apo- and holo-myoglobin, insulin dimerization free energy, ligand binding in lysozyme, and ligand diffusion in globular proteins. In addn. to provide differential free energies between neighboring states, converged umbrella sampling simulations provide insight into migration barriers and ligand dissocn. barriers and anal. of the trajectories yield addnl. insight into the structural dynamics of fundamental processes. Also, such simulations are useful tools to quantify relative stability changes for situations where expts. are difficult. This is illustrated for NO-bound myoglobin. For the dissocn. of benzonitrile from lysozyme it is found that long umbrella sampling simulations are required to approx. converge the free energy profile. Then, however, the resulting differential free energy between the bound and unbound state is in good agreement with ests. from mol. mechanics with generalized Born surface area simulations. Furthermore, comparing the barrier height for ligand escape suggests that ligand dissocn. contains a non-equil. component.
  228. 228
    Ngo, S. T.; Pham, M. Q. Umbrella sampling-based method to compute ligand-binding affinity. Methods Mol. Biol. 2022, 2385, 313– 323,  DOI: 10.1007/978-1-0716-1767-0_14
    Google Scholar
    There is no corresponding record for this reference.
  229. 229
    Pandey, M.; Fernandez, M.; Gentile, F.; Isayev, O.; Tropsha, A.; Stern, A. C.; Cherkasov, A. The transformational role of GPU computing and deep learning in drug discovery. Nat. Mach. Intell. 2022, 4, 211– 221,  DOI: 10.1038/s42256-022-00463-x
    Google Scholar
    There is no corresponding record for this reference.
  230. 230
    Bibal, A.; Cardon, R.; Alfter, D.; Wilkens, R.; Wang, X.; François, T.; Watrin, P. Is Attention Explanation? An Introduction to the Debate. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Dublin, Ireland, 2022, 3889– 3900
    Google Scholar
    There is no corresponding record for this reference.
  231. 231
    Wiegreffe, S.; Pinter, Y. Attention is not not Explanation. arXiv , 2019.
    Google Scholar
    There is no corresponding record for this reference.
  232. 232
    Jain, S.; Wallace, B. C. Attention is not Explanation. arXiv , 2019.
    Google Scholar
    There is no corresponding record for this reference.
  233. 233
    Lundberg, S. M.; Lee, S.-I. A unified approach to interpreting model predictions. Neural Inf. Process. Syst. 2017, 30, 4765– 4774
    Google Scholar
    There is no corresponding record for this reference.
  234. 234
    Gu, Y.; Zhang, X.; Xu, A.; Chen, W.; Liu, K.; Wu, L.; Mo, S.; Hu, Y.; Liu, M.; Luo, Q. Protein-ligand binding affinity prediction with edge awareness and supervised attention. iScience 2023, 26, 105892,  DOI: 10.1016/j.isci.2022.105892
    Google Scholar
    There is no corresponding record for this reference.
  235. 235
    Rodis, N.; Sardianos, C.; Papadopoulos, G. T.; Radoglou-Grammatikis, P.; Sarigiannidis, P.; Varlamis, I. Multimodal Explainable Artificial Intelligence: A Comprehensive Review of Methodological Advances and Future Research Directions. arXiv [cs.AI] 2023.
    Google Scholar
    There is no corresponding record for this reference.
  236. 236
    Gilpin, L. H.; Bau, D.; Yuan, B. Z.; Bajwa, A.; Specter, M.; Kagal, L. Explaining explanations: An overview of interpretability of machine learning. 2018 IEEE 5th International Conference on Data Science and Advanced Analytics (DSAA) 2018, 80– 89
    Google Scholar
    There is no corresponding record for this reference.
  237. 237
    Luo, D.; Liu, D.; Qu, X.; Dong, L.; Wang, B. Enhancing generalizability in protein-ligand binding affinity prediction with multimodal contrastive learning. J. Chem. Inf. Model. 2024, 64, 1892– 1906,  DOI: 10.1021/acs.jcim.3c01961
    Google Scholar
    There is no corresponding record for this reference.
  238. 238
    Rao, R.; Bhattacharya, N.; Thomas, N.; Duan, Y.; Chen, X.; Canny, J.; Abbeel, P.; Song, Y. S. Evaluating protein transfer learning with TAPE. bioRxiv , 2019.
    Google Scholar
    There is no corresponding record for this reference.
  239. 239
    Suzek, B. E.; Wang, Y.; Huang, H.; McGarvey, P. B.; Wu, C. H. UniProt Consortium UniRef clusters: a comprehensive and scalable alternative for improving sequence similarity searches. Bioinformatics 2015, 31, 926– 932,  DOI: 10.1093/bioinformatics/btu739
    Google Scholar
    There is no corresponding record for this reference.
  240. 240
    Eguida, M.; Rognan, D. A Computer Vision Approach to Align and Compare Protein Cavities: Application to Fragment-Based Drug Design. J. Med. Chem. 2020, 63, 7127– 7142,  DOI: 10.1021/acs.jmedchem.0c00422
    Google Scholar
    240
    A Computer Vision Approach to Align and Compare Protein Cavities: Application to Fragment-Based Drug Design
    Eguida, Merveille; Rognan, Didier
    Journal of Medicinal Chemistry (2020), 63 (13), 7127-7142CODEN: JMCMAR; ISSN:0022-2623. (American Chemical Society)
    Identifying local similarities in binding sites from distant proteins is a major hurdle to rational drug design. We herewith present a novel method, borrowed from computer vision, adapted to mine fragment subpockets and compare them to whole ligand-binding sites. Pockets are represented by pharmacophore-annotated point clouds mimicking ideal ligands or fragments. Point cloud registration is used to find the transformation enabling an optimal overlap of points sharing similar topol. and pharmacophoric neighborhoods. The method (ProCare) was calibrated on a large set of druggable cavities and applied to the comparison of fragment subpockets to entire cavities. A collection of 33,953 subpockets annotated with their bound fragments was screened for local similarity to cavities from recently described protein X-ray structures. ProCare was able to detect local similarities between remote pockets and transfer the corresponding fragments to the query cavity space, thereby proposing a first step to fragment-based design approaches targeting orphan cavities.
  241. 241
    Öztürk, H.; Özgür, A.; Ozkirimli, E. DeepDTA: deep drug-target binding affinity prediction. Bioinformatics 2018, 34, i821– i829,  DOI: 10.1093/bioinformatics/bty593
    Google Scholar
    241
    DeepDTA: deep drug-target binding affinity prediction
    Ozturk, Hakime; Ozgur, Arzucan; Ozkirimli, Elif
    Bioinformatics (2018), 34 (17), i821-i829CODEN: BOINFP; ISSN:1367-4811. (Oxford University Press)
    Motivation: The identification of novel drug-target (DT) interactions is a substantial part of the drug discovery process. Most of the computational methods that have been proposed to predict DT interactions have focused on binary classification, where the goal is to det. whether a DT pair interacts or not. However, protein-ligand interactions assume a continuum of binding strength values, also called binding affinity and predicting this value still remains a challenge. The increase in the affinity data available in DT knowledge-bases allows the use of advanced learning techniques such as deep learning architectures in the prediction of binding affinities. In this study, we propose a deep-learning based model that uses only sequence information of both targets and drugs to predict DT interaction binding affinities. The few studies that focus on DT binding affinity prediction use either 3D structures of protein-ligand complexes or 2D features of compds. One novel approach used in this work is the modeling of protein sequences and compd. 1D representations with convolutional neural networks (CNNs). Results: The results show that the proposed deep learning based model that uses the 1D representations of targets and drugs is an effective approach for drug target binding affinity prediction. The model in which high-level representations of a drug and a target are constructed via CNNs achieved the best Concordance Index (CI) performance in one of our larger benchmark datasets, outperforming the KronRLS algorithm and SimBoost, a state-of-the-art method for DT binding affinity prediction.
  242. 242
    Evans, R. Protein complex prediction with AlphaFold-Multimer. bioRxiv , 2021.
    Google Scholar
    There is no corresponding record for this reference.
  243. 243
    Omidi, A.; Møller, M. H.; Malhis, N.; Bui, J. M.; Gsponer, J. AlphaFold-Multimer accurately captures interactions and dynamics of intrinsically disordered protein regions. Proc. Natl. Acad. Sci. U. S. A. 2024, 121, e2406407121,  DOI: 10.1073/pnas.2406407121
    Google Scholar
    There is no corresponding record for this reference.
  244. 244
    Zhu, W.; Shenoy, A.; Kundrotas, P.; Elofsson, A. Evaluation of AlphaFold-Multimer prediction on multi-chain protein complexes. Bioinformatics 2023, 39, btad424,  DOI: 10.1093/bioinformatics/btad424
    Google Scholar
    There is no corresponding record for this reference.
  245. 245
    Rombach, R.; Blattmann, A.; Lorenz, D.; Esser, P.; Ommer, B. High-resolution image synthesis with latent diffusion models. Pattern Recognition (CVPR) 2022, 10684– 10695
    Google Scholar
    There is no corresponding record for this reference.
  246. 246
    Dhariwal, P.; Nichol, A. Diffusion models beat GANs on image synthesis. Neural Inf. Process. Syst. 2021, 8780– 8794
    Google Scholar
    There is no corresponding record for this reference.
  247. 247
    Karras, T.; Aittala, M.; Aila, T.; Laine, S. Elucidating the Design Space of Diffusion-Based Generative Models. Adv. Neural Inf. Process. Syst. 2022, 26565– 26577
    Google Scholar
    There is no corresponding record for this reference.
  248. 248
    Buttenschoen, M.; Morris, G.; Deane, C. PoseBusters: AI-based docking methods fail to generate physically valid poses or generalise to novel sequences. Chem. Sci. 2024, 15, 3130– 3139,  DOI: 10.1039/D3SC04185A
    Google Scholar
    There is no corresponding record for this reference.
  249. 249
    Wee, J.; Wei, G.-W. Benchmarking AlphaFold3’s protein-protein complex accuracy and machine learning prediction reliability for binding free energy changes upon mutation. arXiv , 2024.
    Google Scholar
    There is no corresponding record for this reference.
  250. 250
    Bernard, C.; Postic, G.; Ghannay, S.; Tahi, F. Has AlphaFold 3 reached its success for RNAs? bioRxiv , 2024.
    Google Scholar
    There is no corresponding record for this reference.
  251. 251
    Zonta, F.; Pantano, S. From sequence to mechanobiology? Promises and challenges for AlphaFold 3. Mechanobiology in Medicine 2024, 2, 100083,  DOI: 10.1016/j.mbm.2024.100083
    Google Scholar
    There is no corresponding record for this reference.
  252. 252
    He, X.-H.; Li, J.-R.; Shen, S.-Y.; Xu, H. E. AlphaFold3 versus experimental structures: assessment of the accuracy in ligand-bound G protein-coupled receptors. Acta Pharmacol. Sin. 2024, 1– 12,  DOI: 10.1038/s41401-024-01429-y
    Google Scholar
    There is no corresponding record for this reference.
  253. 253
    Desai, D.; Kantliwala, S. V.; Vybhavi, J.; Ravi, R.; Patel, H.; Patel, J. Review of AlphaFold 3: Transformative advances in drug design and therapeutics. Cureus 2024, 16, e63646,  DOI: 10.7759/cureus.63646
    Google Scholar
    There is no corresponding record for this reference.
  254. 254
    Baek, M. Accurate prediction of protein structures and interactions using a three-track neural network. Science 2021, 373, 871– 876,  DOI: 10.1126/science.abj8754
    Google Scholar
    254
    Accurate prediction of protein structures and interactions using a three-track neural network
    Baek, Minkyung; DiMaio, Frank; Anishchenko, Ivan; Dauparas, Justas; Ovchinnikov, Sergey; Lee, Gyu Rie; Wang, Jue; Cong, Qian; Kinch, Lisa N.; Schaeffer, R. Dustin; Millan, Claudia; Park, Hahnbeom; Adams, Carson; Glassman, Caleb R.; DeGiovanni, Andy; Pereira, Jose H.; Rodrigues, Andria V.; van Dijk, Alberdina A.; Ebrecht, Ana C.; Opperman, Diederik J.; Sagmeister, Theo; Buhlheller, Christoph; Pavkov-Keller, Tea; Rathinaswamy, Manoj K.; Dalwadi, Udit; Yip, Calvin K.; Burke, John E.; Garcia, K. Christopher; Grishin, Nick V.; Adams, Paul D.; Read, Randy J.; Baker, David
    Science (Washington, DC, United States) (2021), 373 (6557), 871-876CODEN: SCIEAS; ISSN:1095-9203. (American Association for the Advancement of Science)
    DeepMind presented notably accurate predictions at the recent 14th Crit. Assessment of Structure Prediction (CASP14) conference. We explored network architectures that incorporate related ideas and obtained the best performance with a three-track network in which information at the one-dimensional (1D) sequence level, the 2D distance map level, and the 3D coordinate level is successively transformed and integrated. The three-track network produces structure predictions with accuracies approaching those of DeepMind in CASP14, enables the rapid soln. of challenging x-ray crystallog. and cryo-electron microscopy structure modeling problems, and provides insights into the functions of proteins of currently unknown structure. The network also enables rapid generation of accurate protein-protein complex models from sequence information alone, short-circuiting traditional approaches that require modeling of individual subunits followed by docking. We make the method available to the scientific community to speed biol. research.
  255. 255
    Ahdritz, G. OpenFold: retraining AlphaFold2 yields new insights into its learning mechanisms and capacity for generalization. Nat. Methods 2024, 21, 1514– 1524,  DOI: 10.1038/s41592-024-02272-z
    Google Scholar
    There is no corresponding record for this reference.
  256. 256
    Liao, C.; Yu, Y.; Mei, Y.; Wei, Y. From words to molecules: A survey of Large Language Models in chemistry. arXiv , 2024.
    Google Scholar
    There is no corresponding record for this reference.
  257. 257
    Bagal, V.; Aggarwal, R.; Vinod, P. K.; Priyakumar, U. D. MolGPT: Molecular generation using a transformer-decoder model. J. Chem. Inf. Model. 2022, 62, 2064– 2076,  DOI: 10.1021/acs.jcim.1c00600
    Google Scholar
    257
    MolGPT: Molecular Generation Using a Transformer-Decoder Model
    Bagal, Viraj; Aggarwal, Rishal; Vinod, P. K.; Priyakumar, U. Deva
    Journal of Chemical Information and Modeling (2022), 62 (9), 2064-2076CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    Application of deep learning techniques for de novo generation of mols., termed as inverse mol. design, has been gaining enormous traction in drug design. The representation of mols. in SMILES notation as a string of characters enables the usage of state of the art models in natural language processing, such as Transformers, for mol. design in general. Inspired by generative pre-training (GPT) models that have been shown to be successful in generating meaningful text, we train a transformer-decoder on the next token prediction task using masked self-attention for the generation of druglike mols. in this study. We show that our model, MolGPT, performs on par with other previously proposed modern machine learning frameworks for mol. generation in terms of generating valid, unique, and novel mols. Furthermore, we demonstrate that the model can be trained conditionally to control multiple properties of the generated mols. We also show that the model can be used to generate mols. with desired scaffolds as well as desired mol. properties by conditioning the generation on scaffold SMILES strings of desired scaffolds and property values. Using saliency maps, we highlight the interpretability of the generative process of the model.
  258. 258
    Janakarajan, N.; Erdmann, T.; Swaminathan, S.; Laino, T.; Born, J. Language models in molecular discovery. arXiv , 2023.
    Google Scholar
    There is no corresponding record for this reference.
  259. 259
    Park, Y.; Metzger, B. P. H.; Thornton, J. W. The simplicity of protein sequence-function relationships. Nat. Commun. 2024, 15, 7953,  DOI: 10.1038/s41467-024-51895-5
    Google Scholar
    There is no corresponding record for this reference.
  260. 260
    Stahl, K.; Warneke, R.; Demann, L.; Bremenkamp, R.; Hormes, B.; Brock, O.; Stülke, J.; Rappsilber, J. Modelling protein complexes with crosslinking mass spectrometry and deep learning. Nat. Commun. 2024, 15, 7866,  DOI: 10.1038/s41467-024-51771-2
    Google Scholar
    There is no corresponding record for this reference.
  261. 261
    Senior, A. W. Improved protein structure prediction using potentials from deep learning. Nature 2020, 577, 706– 710,  DOI: 10.1038/s41586-019-1923-7
    Google Scholar
    261
    Improved protein structure prediction using potentials from deep learning
    Senior, Andrew W.; Evans, Richard; Jumper, John; Kirkpatrick, James; Sifre, Laurent; Green, Tim; Qin, Chongli; Zidek, Augustin; Nelson, Alexander W. R.; Bridgland, Alex; Penedones, Hugo; Petersen, Stig; Simonyan, Karen; Crossan, Steve; Kohli, Pushmeet; Jones, David T.; Silver, David; Kavukcuoglu, Koray; Hassabis, Demis
    Nature (London, United Kingdom) (2020), 577 (7792), 706-710CODEN: NATUAS; ISSN:0028-0836. (Nature Research)
    Protein structure prediction can be used to det. the three-dimensional shape of a protein from its amino acid sequence1. This problem is of fundamental importance as the structure of a protein largely dets. its function2; however, protein structures can be difficult to det. exptl. Considerable progress has recently been made by leveraging genetic information. It is possible to infer which amino acid residues are in contact by analyzing covariation in homologous sequences, which aids in the prediction of protein structures3. The authors can train a neural network to make accurate predictions of the distances between pairs of residues, which convey more information about the structure than contact predictions. Using this information, the authors construct a potential of mean force4 that can accurately describe the shape of a protein. The resulting potential can be optimized by a simple gradient descent algorithm to generate structures without complex sampling procedures. The resulting system, named AlphaFold, achieves high accuracy, even for sequences with fewer homologous sequences. In the recent Crit. Assessment of Protein Structure Prediction5 (CASP13)-a blind assessment of the state of the field-AlphaFold created high-accuracy structures (with template modeling (TM) scores6 of 0.7 or higher) for 24 out of 43 free modeling domains, whereas the next best method, which used sampling and contact information, achieved such accuracy for only 14 out of 43 domains. AlphaFold represents a considerable advance in protein-structure prediction. The authors expect this increased accuracy to enable insights into the function and malfunction of proteins, esp. in cases for which no structures for homologous proteins have been exptl. detd.7.
  262. 262
    Kramer, M. A. Nonlinear principal component analysis using autoassociative neural networks. AIChE J. 1991, 37, 233– 243,  DOI: 10.1002/aic.690370209
    Google Scholar
    262
    Nonlinear principal component analysis using autoassociative neural networks
    Kramer, Mark A.
    AIChE Journal (1991), 37 (2), 233-43CODEN: AICEAC; ISSN:0001-1541.
    Nonlinear-principal-component anal. (NLPCA), is a novel technique for multivariate data anal., similar to the method of principal-component anal. (PCA). NLPCA like PCA, is used to identify and remove correlations among problem variables as an aid to dimensionality redn., visualization, and exploratory data anal. While PCA identifies only linear correlations between variables, NLPCA uncovers both linear and nonlinear correlations, without restriction on the character of the nonlinearities present in the data. NLPCA operates by training a feedforward neural network to perform the identity mapping, where the network inputs are reproduced at the output layer. The network contains an internal bottleneck layer (contg. fewer nodes than input or output layers), which forces the network to develop a compact representation of the input data and 2 addnl. hidden layers. The NLPCA method is demonstrated by using time-dependent, simulated batch-reaction data. NLPCA can reduce dimensionality and produce a feature space map resembling the actual distribution of the underlying system parameters.
  263. 263
    Rogers, D.; Hahn, M. Extended-connectivity fingerprints. J. Chem. Inf. Model. 2010, 50, 742– 754,  DOI: 10.1021/ci100050t
    Google Scholar
    263
    Extended-Connectivity Fingerprints
    Rogers, David; Hahn, Mathew
    Journal of Chemical Information and Modeling (2010), 50 (5), 742-754CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    Extended-connectivity fingerprints (ECFPs) are a novel class of topol. fingerprints for mol. characterization. Historically, topol. fingerprints were developed for substructure and similarity searching. ECFPs were developed specifically for structure-activity modeling. ECFPs are circular fingerprints with a no. of useful qualities: they can be very rapidly calcd.; they are not predefined and can represent an essentially infinite no. of different mol. features (including stereochem. information); their features represent the presence of particular substructures, allowing easier interpretation of anal. results; and the ECFP algorithm can be tailored to generate different types of circular fingerprints, optimized for different uses. While the use of ECFPs has been widely adopted and validated, a description of their implementation has not previously been presented in the literature.
  264. 264
    Veličković, P.; Cucurull, G.; Casanova, A.; Romero, A.; Liò, P.; Bengio, Y. Graph Attention Networks. arXiv , 2017.
    Google Scholar
    There is no corresponding record for this reference.
  265. 265
    Kipf, T. N.; Welling, M. Semi-Supervised Classification with Graph Convolutional Networks. arXiv , 2016.
    Google Scholar
    There is no corresponding record for this reference.
  266. 266
    Xu, Z.; Wang, S.; Zhu, F.; Huang, J. Seq2seq Fingerprint: An Unsupervised Deep Molecular Embedding for Drug Discovery. Proceedings of the 8th ACM International Conference on Bioinformatics, Computational Biology, and Health Informatics. New York, NY, USA 2017, 285– 294
    Google Scholar
    There is no corresponding record for this reference.
  267. 267
    Gilmer, J.; Schoenholz, S.; Riley, P. F.; Vinyals, O.; Dahl, G. E. Neural Message Passing for Quantum Chemistry. ICML 2017, 1263– 1272
    Google Scholar
    There is no corresponding record for this reference.
  268. 268
    Asgari, E.; Mofrad, M. R. K. Continuous Distributed Representation of Biological Sequences for Deep Proteomics and Genomics. PLoS One 2015, 10, e0141287,  DOI: 10.1371/journal.pone.0141287
    Google Scholar
    268
    Continuous distributed representation of biological sequences for deep proteomics and genomics
    Asgari, Ehsaneddin; Mofrad, Mohammad R. K.
    PLoS One (2015), 10 (11), e0141287/1-e0141287/15CODEN: POLNCL; ISSN:1932-6203. (Public Library of Science)
    We introduce a new representation and feature extn. method for biol. sequences. Named bio-vectors (BioVec) to refer to biol. sequences in general with protein-vectors (ProtVec) for proteins (amino-acid sequences) and gene-vectors (GeneVec) for gene sequences, this representation can be widely used in applications of deep learning in proteomics and genomics. In the present paper, we focus on protein-vectors that can be utilized in a wide array of bioinformatics investigations such as family classification, protein visualization, structure prediction, disordered protein identification, and protein-protein interaction prediction. In this method, we adopt artificial neural network approaches and represent a protein sequence with a single dense n-dimensional vector. To evaluate this method, we apply it in classification of 324,018 protein sequences obtained from Swiss-Prot belonging to 7,027 protein families, where an av. family classification accuracy of 93% ± 0.06%is obtained, outperforming existing family classification methods. In addn., we use ProtVec representation to predict disordered proteins from structured proteins. Two databases of disordered sequences are used: the DisProt database as well as a database featuring the disordered regions of nucleoporins rich with phenylalanine-glycine repeats (FG-Nups). Using support vector machine classifiers, FG-Nup sequences are distinguished from structured protein sequences found in Protein Data Bank (PDB) with a 99.8% accuracy, and unstructured DisProt sequences are differentiated from structured DisProt sequences with 100.0% accuracy. These results indicate that by only providing sequence data for various proteins into this model, accurate information about protein structure can be detd. Importantly, this model needs to be trained only once and can then be applied to ext. a comprehensive set of information regarding proteins of interest. Moreover, this representation can be considered as pre-training for various applications of deep learning in bioinformatics.
  269. 269
    He, K.; Zhang, X.; Ren, S.; Sun, J. Deep residual learning for image recognition. Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recognit. 2015, 770– 778
    Google Scholar
    There is no corresponding record for this reference.
  270. 270
    Öztürk, H.; Ozkirimli, E.; Özgür, A. A novel methodology on distributed representations of proteins using their interacting ligands. Bioinformatics 2018, 34, i295– i303,  DOI: 10.1093/bioinformatics/bty287
    Google Scholar
    There is no corresponding record for this reference.
  271. 271
    Hu, J.; Shen, L.; Sun, G. Squeeze-and-excitation networks. Proc. IEEE Conf. Comput. Vis. Pattern Recognition 2018, 7132– 7141
    Google Scholar
    There is no corresponding record for this reference.

Cited By

Click to copy section linkSection link copied!

This article has not yet been cited by other publications.

Download PDF
Go to Journal of Chemical Information and Modeling
Get e-Alerts
Get e-Alerts

Journal of Chemical Information and Modeling

Cite this: J. Chem. Inf. Model. 2025, 65, 5, 2191–2213
Click to copy citationCitation copied!
https://doi.org/10.1021/acs.jcim.4c01907
Published February 24, 2025

Copyright © 2025 The Authors. Published by American Chemical Society. This publication is licensed under

CC-BY-NC-ND 4.0 .

Article Views

1803

Altmetric

-

Citations

-
Learn about these metrics

Article Views are the COUNTER-compliant sum of full text article downloads since November 2008 (both PDF and HTML) across all institutions and individuals. These metrics are regularly updated to reflect usage leading up to the last few days.

Citations are the number of other articles citing this article, calculated by Crossref and updated daily. Find more information about Crossref citation counts.

The Altmetric Attention Score is a quantitative measure of the attention that a research article has received online. Clicking on the donut icon will load a page at altmetric.com with additional details about the score and the social media presence for the given article. Find more information on the Altmetric Attention Score and how the score is calculated.

  • Abstract

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 1

    Figure 1. Language of protein sequences and the ligand SMILES representation: NLP methods can be applied to text representations to infer local and global properties of human language, proteins, and molecules alike. Local properties are inferred from subsequences in text: (left) for human language, this includes a part of speech or role a word serves; (middle) for protein sequences, this includes motifs, functional sites, and domains; and (right) for SMILES strings, this can include functional groups and special characters used in SMILES syntax to indicate chemical attributes. Similarly, global properties can theoretically be inferred from a text in its entirety.

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 2

    Figure 2. Summary of the data preparation, model creation, and model evaluation workflow. Model Creation for PLI studies follows an Extract-Fuse-Predict Framework: input protein and ligand data are extracted and embedded, combined, and passed into a machine learning model to generate predictions.

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 3

    Figure 3. Framework diagrams for RNN (and its variant LSTM), transformer, and attention with arrows representing a flow of information. (A) The "unrolled" structure of an RNN and the recurrent units, where hidden states propagate across time steps. The recurrent unit takes the current token Xt as input, combines it with the value of the current hidden state ht, and computes their weighted sum before generating the response Ot and an updated hidden state ht+1. Weighted sums depend upon the associated network weights Wxh, Whh, or Woh, which connect input to hidden state, hidden state to hidden state, and hidden state to output, respectively. LSTM differs in that a memory state is updated during each iteration, facilitating long-term dependency learning. (B) A simplified framework of a transformer's encoder-decoder architecture, and associated attention mechanism. A scaled product of the Query and Key vectors yields attention weights that can provide interpretability, with the new embedding vector (or the output vector) updated based on this specific key.

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 4

    Figure 4. Sample attention weights for relating protein and ligand. The heatmaps on the left help visualize the weighted importance of select protein residues and ligand atoms in a PLI. Structural views of the protein–ligand binding pocket are shown in the middle, with insets of the 2D ligand structures on the right. The colored residues and red color highlights indicate AAs in the protein binding pocket and ligand atoms with high attention scores. Reproduced with permission from Figure 7 of Wu et al. (148) Used with permission under license CC BY 4.0. Copyright 2023 The Author(s). Published by Elsevier Ltd.

    High Resolution Image
    Download MS PowerPoint Slide

  • References

    This article references 271 other publications.

    1. 1
      Songyang, Z.; Cantley, L. C. Recognition and specificity in protein tyrosine kinase-mediated signalling. Trends Biochem. Sci. 1995, 20, 470– 475,  DOI: 10.1016/S0968-0004(00)89103-3
      1
      Recognition and specificity in protein tyrosine kinase-mediated signaling
      Songyang, Zhou; Cantley, Lewis C.
      Trends in Biochemical Sciences (1995), 20 (11), 470-5CODEN: TBSCDB; ISSN:0968-0004. (Elsevier Trends Journals)
      A review, with 46 refs. There are several factors that contribute to the specificities of protein tyrosine kinases (PTKs) in signal transduction pathways. While protein-protein interaction domains, such as the Src homol. (SH2 and SH3) domains, regulate the cellular localization of PTKs and their substrates, the specificities of PTKs are ultimately detd. by their catalytic domains. The use of peptide libraries has revealed the substrate specificities of SH2 domains and PTK catalytic domains, and has suggested cross-talk between these domains.
    2. 2
      Johnson, L. N.; Lowe, E. D.; Noble, M. E.; Owen, D. J. The Eleventh Datta Lecture. The structural basis for substrate recognition and control by protein kinases. FEBS Lett. 1998, 430, 1– 11,  DOI: 10.1016/S0014-5793(98)00606-1
      2
      The structural basis for substrate recognition and control by protein kinases
      Johnson, Louise N.; Lowe, Edward D.; Noble, Martin E. M.; Owen, David J.
      FEBS Letters (1998), 430 (1,2), 1-11CODEN: FEBLAL; ISSN:0014-5793. (Elsevier Science B.V.)
      A review with 49 refs. Protein kinases catalyze phospho transfer reactions from ATP to serine, threonine or tyrosine residues in target substrates and provide key mechanisms for control of cellular signaling processes. The crystal structures of 12 protein kinases are now known. These include structures of kinases in the active state in ternary complexes with ATP (or analogs) and inhibitor or peptide substrates (e.g. cAMP dependent protein kinase, phosphorylase kinase and insulin receptor tyrosine kinase); kinases in both active and inactive states (e.g., CDK2/cyclin A, insulin receptor tyrosine kinase and MAPK); kinases in the active state (e.g. casein kinase 1, Lck); and kinases in inactive states (e.g. twitchin kinase, calcium calmodulin kinase 1, FGF receptor kinase, c-Src and Hck). This paper summarizes the detailed information obtained with active phosphorylase kinase ternary complex and reviews the results with ref. to other kinase structures for insights into mechanisms for substrate recognition and control.
    3. 3
      Kristiansen, K. Molecular mechanisms of ligand binding, signaling, and regulation within the superfamily of G-protein-coupled receptors: molecular modeling and mutagenesis approaches to receptor structure and function. Pharmacol. Ther. 2004, 103, 21– 80,  DOI: 10.1016/j.pharmthera.2004.05.002
      3
      Molecular mechanisms of ligand binding, signaling, and regulation within the superfamily of G-protein-coupled receptors: molecular modeling and mutagenesis approaches to receptor structure and function
      Kristiansen, Kurt
      Pharmacology & Therapeutics (2004), 103 (1), 21-80CODEN: PHTHDT; ISSN:0163-7258. (Elsevier Science B.V.)
      A review. The superfamily of G-protein-coupled receptors (GPCRs) could be subclassified into 7 families (A, B, large N-terminal family B-7 transmembrane helix, C, Frizzled/Smoothened, taste 2, and vomeronasal 1 receptors) among mammalian species. Cloning and functional studies of GPCRs have revealed that the superfamily of GPCRs comprises receptors for chem. diverse native ligands including endogenous compds. like amines, peptides, and Wnt proteins (i.e., secreted proteins activating Frizzled receptors); endogenous cell surface adhesion mols.; and photons and exogenous compds. like odorants. The combined use of site-directed mutagenesis and mol. modeling approaches have provided detailed insight into mol. mechanisms of ligand binding, receptor folding, receptor activation, G-protein coupling, and regulation of GPCRs. The vast majority of family A, B, C, vomeronasal 1, and taste 2 receptors are able to transduce signals into cells through G-protein coupling. However, G-protein-independent signaling mechanisms have also been reported for many GPCRs. Specific interaction motifs in the intracellular parts of these receptors allow them to interact with scaffold proteins. Protein engineering techniques have provided information on mol. mechanisms of GPCR-accessory protein, GPCR-GPCR, and GPCR-scaffold protein interactions. Site-directed mutagenesis and mol. dynamics simulations have revealed that the inactive state conformations are stabilized by specific interhelical and intrahelical salt bridge interactions and hydrophobic-type interactions. Constitutively activating mutations or agonist binding disrupts such constraining interactions leading to receptor conformations that assocs. with and activate G-proteins.
    4. 4
      West, I. C. What determines the substrate specificity of the multi-drug-resistance pump?. Trends Biochem. Sci. 1990, 15, 42– 46,  DOI: 10.1016/0968-0004(90)90171-7
      There is no corresponding record for this reference.
    5. 5
      Vivier, E.; Malissen, B. Innate and adaptive immunity: specificities and signaling hierarchies revisited. Nat. Immunol. 2005, 6, 17– 21,  DOI: 10.1038/ni1153
      5
      Innate and adaptive immunity: specificities and signaling hierarchies revisited
      Vivier, Eric; Malissen, Bernard
      Nature Immunology (2005), 6 (1), 17-21CODEN: NIAMCZ; ISSN:1529-2908. (Nature Publishing Group)
      A review. The conventional classification of known immune responses by specificity may need re-evaluation. The immune system can be classified into two subsystems: the innate and adaptive immune systems. In general, innate immunity is considered a nonspecific response, whereas the adaptive immune system is thought of as being very specific. In addn., the antigen receptors of the adaptive immune response are commonly viewed as 'master sensors' whose engagement dictates lymphocyte function. Here the authors propose that these ideas do not genuinely reflect the organization of immune responses and that they bias the authors' view of immunity as well as the authors' teaching of immunol. Indeed, the level of specificity and mode of signaling integration used by the main cellular participants in the adaptive and innate immune systems are more similar than previously appreciated.
    6. 6
      Desvergne, B.; Michalik, L.; Wahli, W. Transcriptional regulation of metabolism. Physiol. Rev. 2006, 86, 465– 514,  DOI: 10.1152/physrev.00025.2005
      6
      Transcriptional regulation of metabolism
      Desvergne, Beatrice; Michalik, Liliane; Wahli, Walter
      Physiological Reviews (2006), 86 (2), 465-514CODEN: PHREA7; ISSN:0031-9333. (American Physiological Society)
      A review. Our understanding of metab. is undergoing a dramatic shift. Indeed, the efforts made towards elucidating the mechanisms controlling the major regulatory pathways are now being rewarded. At the mol. level, the crucial role of transcription factors is particularly well-illustrated by the link between alterations of their functions and the occurrence of major metabolic diseases. In addn., the possibility of manipulating the ligand-dependent activity of some of these transcription factors makes them attractive as therapeutic targets. The aim of this review is to summarize recent knowledge on the transcriptional control of metabolic homeostasis. We first review data on the transcriptional regulation of the intermediary metab., i.e., glucose, amino acid, lipid, and cholesterol metab. Then, we analyze how transcription factors integrate signals from various pathways to ensure homeostasis. One example of this coordination is the daily adaptation to the circadian fasting and feeding rhythm. This section also discusses the dysregulations causing the metabolic syndrome, which reveals the intricate nature of glucose and lipid metab. and the role of the transcription factor PPARγ in orchestrating this assocn. Finally, we discuss the mol. mechanisms underlying metabolic regulations, which provide new opportunities for treating complex metabolic disorders.
    7. 7
      Atkinson, D. E. Biological feedback control at the molecular level: Interaction between metabolite-modulated enzymes seems to be a major factor in metabolic regulation. Science 1965, 150, 851– 857,  DOI: 10.1126/science.150.3698.851
      There is no corresponding record for this reference.
    8. 8
      Huang, S.-Y.; Zou, X. Advances and challenges in protein-ligand docking. Int. J. Mol. Sci. 2010, 11, 3016– 3034,  DOI: 10.3390/ijms11083016
      8
      Advances and challenges in protein-ligand docking
      Huang, Sheng-You; Zou, Xiaoqin
      International Journal of Molecular Sciences (2010), 11 (), 3016-3034CODEN: IJMCFK; ISSN:1422-0067. (Molecular Diversity Preservation International)
      A review. Mol. docking is a widely-used computational tool for the study of mol. recognition, which aims to predict the binding mode and binding affinity of a complex formed by two or more constituent mols. with known structures. An important type of mol. docking is protein-ligand docking because of its therapeutic applications in modern structure-based drug design. Here, we review the recent advances of protein flexibility, ligand sampling, and scoring functions - the three important aspects in protein-ligand docking. Challenges and possible future directions are discussed in the conclusion.
    9. 9
      Chaires, J. B. Calorimetry and thermodynamics in drug design. Annu. Rev. Biophys. 2008, 37, 135– 151,  DOI: 10.1146/annurev.biophys.36.040306.132812
      9
      Calorimetry and thermodynamics in drug design
      Chaires, Jonathan B.
      Annual Review of Biophysics (2008), 37 (), 135-151CODEN: ARBNCV ISSN:. (Annual Reviews Inc.)
      A review. Modern instrumentation for calorimetry permits direct detn. of enthalpy values for binding reactions and conformational transitions in biomols. Complete thermodn. profiles consisting of free energy, enthalpy, and entropy may be obtained for reactions of interest in a relatively straightforward manner. Such profiles are of enormous value in drug design because they provide information about the balance of driving forces that cannot be obtained from structural or computational methods alone. This perspective shows several examples of the insight provided by thermodn. data in drug design.
    10. 10
      Serhan, C. N. Signalling the fat controller. Nature 1996, 384, 23– 24,  DOI: 10.1038/384023a0
      There is no corresponding record for this reference.
    11. 11
      McAllister, C. H.; Beatty, P. H.; Good, A. G. Engineering nitrogen use efficient crop plants: the current status: Engineering nitrogen use efficient crop plants. Plant Biotechnol. J. 2012, 10, 1011– 1025,  DOI: 10.1111/j.1467-7652.2012.00700.x
      There is no corresponding record for this reference.
    12. 12
      Goldsmith, M.; Tawfik, D. S. Enzyme engineering: reaching the maximal catalytic efficiency peak. Curr. Opin. Struct. Biol. 2017, 47, 140– 150,  DOI: 10.1016/j.sbi.2017.09.002
      12
      Enzyme engineering: reaching the maximal catalytic efficiency peak
      Goldsmith, Moshe; Tawfik, Dan S.
      Current Opinion in Structural Biology (2017), 47 (), 140-150CODEN: COSBEF; ISSN:0959-440X. (Elsevier Ltd.)
      A review. The practical need for highly efficient enzymes presents new challenges in enzyme engineering, in particular, the need to improve catalytic turnover (kcat) or efficiency (kcat/KM) by several orders of magnitude. However, optimizing catalysis demands navigation through complex and rugged fitness landscapes, with optimization trajectories often leading to strong diminishing returns and dead-ends. When no further improvements are obsd. in library screens or selections, it remains unclear whether the maximal catalytic efficiency of the enzyme (the catalytic 'fitness peak') has been reached; or perhaps, an alternative combination of mutations exists that could yield addnl. improvements. Here, we discuss fundamental aspects of the process of catalytic optimization, and offer practical solns. with respect to overcoming optimization plateaus.
    13. 13
      Vajda, S.; Guarnieri, F. Characterization of protein-ligand interaction sites using experimental and computational methods. Curr. Opin. Drug Discovery Devel. 2006, 9, 354– 362
      13
      Characterization of protein-ligand interaction sites using experimental and computational methods
      Vajda, Sandor; Guarnieri, Frank
      Current Opinion in Drug Discovery & Development (2006), 9 (3), 354-362CODEN: CODDFF; ISSN:1367-6733. (Thomson Scientific)
      A review. The ability to identify the sites of a protein that can bind with high affinity to small, drug-like compds. has been an important goal in drug design. Accurate prediction of druggable sites and the identification of small compds. binding in those sites have provided the input for fragment-based combinatorial approaches that allow for a more thorough exploration of the chem. space, and that have the potential to yield mols. that are more lead-like than those found using traditional high-throughput screening. Current progress in exptl. and computational methods for identifying and characterizing druggable ligand binding sites on protein targets is reviewed herein, including a discussion of successful NMR, x-ray crystallog. and tethering technologies. Classical geometric and energy-based computational methods are also discussed, with particular focus on two powerful technologies, i.e., computational solvent mapping and grand canonical Monte Carlo simulations (as used by Locus Pharmaceuticals Inc). Both methods can be used to reliably identify druggable sites on proteins and to facilitate the design of novel, low-nanomolar-affinity ligands.
    14. 14
      Du, X.; Li, Y.; Xia, Y.-L.; Ai, S.-M.; Liang, J.; Sang, P.; Ji, X.-L.; Liu, S.-Q. Insights into protein-ligand interactions: Mechanisms, models, and methods. Int. J. Mol. Sci. 2016, 17, 144,  DOI: 10.3390/ijms17020144
      14
      Insights into protein-ligand interactions: mechanisms, models, and methods
      Du, Xing; Li, Yi; Xia, Yuan-Ling; Ai, Shi-Meng; Liang, Jing; Sang, Peng; Ji, Xing-Lai; Liu, Shu-Qun
      International Journal of Molecular Sciences (2016), 17 (2), 144/1-144/34CODEN: IJMCFK; ISSN:1422-0067. (MDPI AG)
      Mol. recognition, which is the process of biol. macromols. interacting with each other or various small mols. with a high specificity and affinity to form a specific complex, constitutes the basis of all processes in living organisms. Proteins, an important class of biol. macromols., realize their functions through binding to themselves or other mols. A detailed understanding of the protein-ligand interactions is therefore central to understanding biol. at the mol. level. Moreover, knowledge of the mechanisms responsible for the protein-ligand recognition and binding will also facilitate the discovery, design, and development of drugs. In the present review, first, the physicochem. mechanisms underlying protein-ligand binding, including the binding kinetics, thermodn. concepts and relationships, and binding driving forces, are introduced and rationalized. Next, three currently existing protein-ligand binding models-the "lock-and-key", "induced fit", and "conformational selection"-are described and their underlying thermodn. mechanisms are discussed. Finally, the methods available for investigating protein-ligand binding affinity, including exptl. and theor./computational approaches, are introduced, and their advantages, disadvantages, and challenges are discussed.
    15. 15
      Fan, F. J.; Shi, Y. Effects of data quality and quantity on deep learning for protein-ligand binding affinity prediction. Bioorg. Med. Chem. 2022, 72, 117003,  DOI: 10.1016/j.bmc.2022.117003
      15
      Effects of data quality and quantity on deep learning for protein-ligand binding affinity prediction
      Fan, Frankie J.; Shi, Yun
      Bioorganic & Medicinal Chemistry (2022), 72 (), 117003CODEN: BMECEP; ISSN:0968-0896. (Elsevier B.V.)
      Prediction of protein-ligand binding affinities is crucial for computational drug discovery. A no. of deep learning approaches have been developed in recent years to improve the accuracy of such affinity prediction. While the predicting power of these systems have advanced to some degrees depending on the dataset used for model training and testing, the effects of the quality and quantity of the underlying data have not been thoroughly examd. In this study, we employed erroneous datasets and data subsets of different sizes, created from one of the largest databases of exptl. binding affinities, to train and evaluate a deep learning system based on convolutional neural networks. Our results show that data quality and quantity do have significant impacts on the prediction performance of trained models. Depending on the variations in data quality and quantity, the performance discrepancies could be comparable to or even larger than those obsd. among different deep learning approaches. In particular, the presence of proteins in the training data leads to a dramatic increase in prediction accuracy. This implies that continued accumulation of high-quality affinity data, esp. for new protein targets, is indispensable for improving deep learning models to better predict protein-ligand binding affinities.
    16. 16
      Sousa, S. F.; Ribeiro, A. J. M.; Coimbra, J. T. S.; Neves, R. P. P.; Martins, S. A.; Moorthy, N. S. H. N.; Fernandes, P. A.; Ramos, M. J. Protein-Ligand Docking in the New Millennium A Retrospective of 10 Years in the Field. Curr. Med. Chem. 2013, 20, 2296– 2314,  DOI: 10.2174/0929867311320180002
      There is no corresponding record for this reference.
    17. 17
      Morris, C. J.; Corte, D. D. Using molecular docking and molecular dynamics to investigate protein-ligand interactions. Mod. Phys. Lett. B 2021, 35, 2130002,  DOI: 10.1142/S0217984921300027
      17
      Using molecular docking and molecular dynamics to investigate protein-ligand interactions
      Morris, Connor J.; Corte, Dennis Della
      Modern Physics Letters B (2021), 35 (8), 2130002CODEN: MPLBET; ISSN:0217-9849. (World Scientific Publishing Co. Pte. Ltd.)
      A review. Mol. docking and mol. dynamics (MD) are powerful tools used to investigate protein-ligand interactions. Mol. docking programs predict the binding pose and affinity of a protein-ligand complex, while MD can be used to incorporate flexibility into docking calcns. and gain further information on the kinetics and stability of the protein-ligand bond. This review covers state-of-the-art methods of using mol. docking and MD to explore protein-ligand interactions, with emphasis on application to drug discovery. We also call for further research on combining common mol. docking and MD methods.
    18. 18
      Lecina, D.; Gilabert, J. F.; Guallar, V. Adaptive simulations, towards interactive protein-ligand modeling. Sci. Rep. 2017, 7, 8466,  DOI: 10.1038/s41598-017-08445-5
      18
      Adaptive simulations, towards interactive protein-ligand modeling
      Lecina Daniel; Gilabert Joan F; Guallar Victor; Guallar Victor
      Scientific reports (2017), 7 (1), 8466 ISSN:.
      Modeling the dynamic nature of protein-ligand binding with atomistic simulations is one of the main challenges in computational biophysics, with important implications in the drug design process. Although in the past few years hardware and software advances have significantly revamped the use of molecular simulations, we still lack a fast and accurate ab initio description of the binding mechanism in complex systems, available only for up-to-date techniques and requiring several hours or days of heavy computation. Such delay is one of the main limiting factors for a larger penetration of protein dynamics modeling in the pharmaceutical industry. Here we present a game-changing technology, opening up the way for fast reliable simulations of protein dynamics by combining an adaptive reinforcement learning procedure with Monte Carlo sampling in the frame of modern multi-core computational resources. We show remarkable performance in mapping the protein-ligand energy landscape, being able to reproduce the full binding mechanism in less than half an hour, or the active site induced fit in less than 5 minutes. We exemplify our method by studying diverse complex targets, including nuclear hormone receptors and GPCRs, demonstrating the potential of using the new adaptive technique in screening and lead optimization studies.
    19. 19
      Ikebata, H.; Hongo, K.; Isomura, T.; Maezono, R.; Yoshida, R. Bayesian molecular design with a chemical language model. J. Comput. Aided Mol. Des. 2017, 31, 379– 391,  DOI: 10.1007/s10822-016-0008-z
      There is no corresponding record for this reference.
    20. 20
      Rives, A.; Meier, J.; Sercu, T.; Goyal, S.; Lin, Z.; Liu, J.; Guo, D.; Ott, M.; Zitnick, C. L.; Ma, J.; Fergus, R. Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences. Proc. Natl. Acad. Sci. U. S. A. 2021,  DOI: 10.1073/pnas.2016239118
      There is no corresponding record for this reference.
    21. 21
      Cao, Y.; Shen, Y. TALE: Transformer-based protein function Annotation with joint sequence–Label Embedding. Bioinformatics 2021, 37, 2825– 2833,  DOI: 10.1093/bioinformatics/btab198
      There is no corresponding record for this reference.
    22. 22
      Wang, S.; Guo, Y.; Wang, Y.; Sun, H.; Huang, J. SMILES-BERT: Large Scale Unsupervised Pre-Training for Molecular Property Prediction. Proceedings of the 10th ACM International Conference on Bioinformatics, Computational Biology and Health Informatics. New York, NY, USA 2019, 429– 436
      There is no corresponding record for this reference.
    23. 23
      Chithrananda, S.; Grand, G.; Ramsundar, B. ChemBERTa: Large-Scale Self-Supervised Pretraining for Molecular Property Prediction. arXiv , 2020.
      There is no corresponding record for this reference.
    24. 24
      Kumar, N.; Acharya, V. Machine intelligence-driven framework for optimized hit selection in virtual screening. J. Cheminform. 2022, 14, 48,  DOI: 10.1186/s13321-022-00630-7
      There is no corresponding record for this reference.
    25. 25
      Erikawa, D.; Yasuo, N.; Sekijima, M. MERMAID: an open source automated hit-to-lead method based on deep reinforcement learning. J. Cheminform. 2021, 13, 94,  DOI: 10.1186/s13321-021-00572-6
      There is no corresponding record for this reference.
    26. 26
      Zhou, M.; Duan, N.; Liu, S.; Shum, H.-Y. Progress in neural NLP: Modeling, learning, and reasoning. Engineering (Beijing) 2020, 6, 275– 290,  DOI: 10.1016/j.eng.2019.12.014
      There is no corresponding record for this reference.
    27. 27
      Patwardhan, N.; Marrone, S.; Sansone, C. Transformers in the real world: A survey on NLP applications. Inf. 2023, 14, 242,  DOI: 10.3390/info14040242
      There is no corresponding record for this reference.
    28. 28
      Bijral, R. K.; Singh, I.; Manhas, J.; Sharma, V. Exploring Artificial Intelligence in Drug Discovery: A Comprehensive Review. Arch. Comput. Methods Eng. 2022, 29, 2513– 2529,  DOI: 10.1007/s11831-021-09661-z
      There is no corresponding record for this reference.
    29. 29
      Ray, P. P. ChatGPT: A comprehensive review on background, applications, key challenges, bias, ethics, limitations and future scope. Internet of Things and Cyber-Physical Systems 2023, 3, 121– 154,  DOI: 10.1016/j.iotcps.2023.04.003
      There is no corresponding record for this reference.
    30. 30
      Radford, A.; Narasimhan, K.; Salimans, T.; Sutskever, I. Improving language understanding by generative pre-training. https://www.mikecaptain.com/resources/pdf/GPT-1.pdf, Accessed: 2023–10–27.
      There is no corresponding record for this reference.
    31. 31
      Goodside, R, Papay, Meet Claude: Anthropic’s Rival to ChatGPT. https://scale.com/blog/chatgpt-vs-claude, 2023.
      There is no corresponding record for this reference.
    32. 32
      Bing Copilot. Bing Copilot; https://copilot.microsoft.com/.
      There is no corresponding record for this reference.
    33. 33
      Rahul; Adhikari, S.; Monika NLP based Machine Learning Approaches for Text Summarization. 2020 Fourth International Conference on Computing Methodologies and Communication (ICCMC) 2020, 535– 538
      There is no corresponding record for this reference.
    34. 34
      Nasukawa, T.; Yi, J. Sentiment analysis: capturing favorability using natural language processing. Proceedings of the 2nd international conference on Knowledge capture. New York, NY, USA 2003, 70– 77
      There is no corresponding record for this reference.
    35. 35
      Lample, G.; Charton, F. Deep Learning for Symbolic Mathematics. arXiv , 2019.
      There is no corresponding record for this reference.
    36. 36
      Feng, Z.; Guo, D.; Tang, D.; Duan, N.; Feng, X.; Gong, M.; Shou, L.; Qin, B.; Liu, T.; Jiang, D.; Zhou, M. CodeBERT: APre-Trained Model for Programming and Natural Languages. arXiv , 2020.
      There is no corresponding record for this reference.
    37. 37
      Mielke, S. J.; Alyafeai, Z.; Salesky, E.; Raffel, C.; Dey, M.; Gallé, M.; Raja, A.; Si, C.; Lee, W. Y.; Sagot, B.; Tan, S. Between words and characters: A Brief History of Open-Vocabulary Modeling and Tokenization in NLP. arXiv , 2021.
      There is no corresponding record for this reference.
    38. 38
      Camacho-Collados, J.; Pilehvar, M. T. From word to sense embeddings: A survey on vector representations of meaning. J. Artif. Intell. Res. 2018, 63, 743– 788,  DOI: 10.1613/jair.1.11259
      There is no corresponding record for this reference.
    39. 39
      Ashok, V. G.; Feng, S.; Choi, Y. Success with style: Using writing style to predict the success of novelsd.
      There is no corresponding record for this reference.
    40. 40
      Barberá, P.; Boydstun, A. E.; Linn, S.; McMahon, R.; Nagler, J. Automated text classification of news articles: A practical guide. Polit. Anal. 2021, 29, 19– 42,  DOI: 10.1017/pan.2020.8
      There is no corresponding record for this reference.
    41. 41
      Wang, H.; Wu, H.; He, Z.; Huang, L.; Church, K. W. Progress in machine translation. Engineering (Beijing) 2022, 18, 143– 153,  DOI: 10.1016/j.eng.2021.03.023
      There is no corresponding record for this reference.
    42. 42
      Sønderby, S. K.; Winther, O. Protein Secondary Structure Prediction with Long Short Term Memory Networks. arXiv , 2014.
      There is no corresponding record for this reference.
    43. 43
      Guo, Y.; Li, W.; Wang, B.; Liu, H.; Zhou, D. DeepACLSTM: deep asymmetric convolutional long short-term memory neural models for protein secondary structure prediction. BMC Bioinformatics 2019, 20, 341,  DOI: 10.1186/s12859-019-2940-0
      There is no corresponding record for this reference.
    44. 44
      Bhasuran, B.; Natarajan, J. Automatic extraction of gene-disease associations from literature using joint ensemble learning. PLoS One 2018, 13, e0200699,  DOI: 10.1371/journal.pone.0200699
      There is no corresponding record for this reference.
    45. 45
      Pang, M.; Su, K.; Li, M. Leveraging information in spatial transcriptomics to predict super-resolution gene expression from histology images in tumors. bioRxiv , 2021, 2021.11.28.470212.
      There is no corresponding record for this reference.
    46. 46
      Bouatta, N.; Sorger, P.; AlQuraishi, M. Protein structure prediction by AlphaFold2: are attention and symmetries all you need?. Acta Crystallogr. D Struct Biol. 2021, 77, 982– 991,  DOI: 10.1107/S2059798321007531
      There is no corresponding record for this reference.
    47. 47
      Skolnick, J.; Gao, M.; Zhou, H.; Singh, S. AlphaFold 2: Why It Works and Its Implications for Understanding the Relationships of Protein Sequence, Structure, and Function. J. Chem. Inf. Model. 2021, 61, 4827– 4831,  DOI: 10.1021/acs.jcim.1c01114
      47
      AlphaFold 2: Why It Works and Its Implications for Understanding the Relationships of Protein Sequence, Structure, and Function
      Skolnick, Jeffrey; Gao, Mu; Zhou, Hongyi; Singh, Suresh
      Journal of Chemical Information and Modeling (2021), 61 (10), 4827-4831CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      AlphaFold 2 (AF2) was the star of CASP14, the last biannual structure prediction expt. Using novel deep learning, AF2 predicted the structures of many difficult protein targets at or near exptl. resoln. Here, the authors present the authors' perspective of why AF2 works and show that it is a very sophisticated fold recognition algorithm that exploits the completeness of the library of single domain PDB structures. It also learned local side chain packing rearrangements that enable it to refine proteins to high resoln. The benefits and limitations of its ability to predict the structures of many more proteins at or close to at. detail are discussed.
    48. 48
      Adadi, A.; Berrada, M. Peeking inside the black-box: a survey on explainable artificial intelligence (XAI). IEEE Access 2018, 6, 52138,  DOI: 10.1109/ACCESS.2018.2870052
      There is no corresponding record for this reference.
    49. 49
      Box, G. E. P. Science and Statistics. J. Am. Stat. Assoc. 1976, 71, 791– 799,  DOI: 10.1080/01621459.1976.10480949
      There is no corresponding record for this reference.
    50. 50
      Geirhos, R.; Jacobsen, J.-H.; Michaelis, C.; Zemel, R.; Brendel, W.; Bethge, M.; Wichmann, F. A. Shortcut learning in deep neural networks. Nature Machine Intelligence 2020, 2, 665– 673,  DOI: 10.1038/s42256-020-00257-z
      There is no corresponding record for this reference.
    51. 51
      Outeiral, C.; Nissley, D. A.; Deane, C. M. Current structure predictors are not learning the physics of protein folding. Bioinformatics 2022, 38, 1881– 1887,  DOI: 10.1093/bioinformatics/btab881
      51
      Current structure predictors are not learning the physics of protein folding
      Outeiral, Carlos; Nissley, Daniel A.; Deane, Charlotte M.
      Bioinformatics (2022), 38 (7), 1881-1887CODEN: BOINFP; ISSN:1367-4811. (Oxford University Press)
      Predicting the native state of a protein has long been considered a gateway problem for understanding protein folding. Recent advances in structural modeling driven by deep learning have achieved unprecedented success at predicting a protein's crystal structure, but it is not clear if these models are learning the physics of how proteins dynamically fold into their equil. structure or are just accurate knowledge-based predictors of the final state. In this work, we compare the pathways generated by state-of-the-art protein structure prediction methods to exptl. data about protein folding pathways. The methods considered were AlphaFold 2, RoseTTAFold, trRosetta, RaptorX, DMPfold, EVfold, SAINT2 and Rosetta. We find evidence that their simulated dynamics capture some information about the folding pathway, but their predictive ability is worse than a trivial classifier using sequence-agnostic features like chain length. The folding trajectories produced are also uncorrelated with exptl. observables such as intermediate structures and the folding rate const. These results suggest that recent advances in structure prediction do not yet provide an enhanced understanding of protein folding.
    52. 52
      Steels, L. Modeling the cultural evolution of language. Phys. Life Rev. 2011, 8, 339– 356,  DOI: 10.1016/j.plrev.2011.10.014
      There is no corresponding record for this reference.
    53. 53
      Maurya, H. C.; Gupta, P.; Choudhary, N. Natural language ambiguity and its effect on machine learning. Int. J. Modern Eng. Res. 2015, 5, 25– 30
      There is no corresponding record for this reference.
    54. 54
      Tenney, I.; Xia, P.; Chen, B.; Wang, A.; Poliak, A.; McCoy, R. T.; Kim, N.; Van Durme, B.; Bowman, S. R.; Das, D.; Pavlick, E. What do you learn from context? Probing for sentence structure in contextualized word representations. arXiv , 2019.
      There is no corresponding record for this reference.
    55. 55
      Miyagawa, S.; Berwick, R. C.; Okanoya, K. The emergence of hierarchical structure in human language. Front. Psychol. 2013, 4, 71,  DOI: 10.3389/fpsyg.2013.00071
      There is no corresponding record for this reference.
    56. 56
      Liu, H.; Xu, C.; Liang, J. Dependency distance: A new perspective on syntactic patterns in natural languages. Phys. Life Rev. 2017, 21, 171– 193,  DOI: 10.1016/j.plrev.2017.03.002
      There is no corresponding record for this reference.
    57. 57
      Frank, S. L.; Bod, R.; Christiansen, M. H. How hierarchical is language use?. Proc. Biol. Sci. 2012, 279, 4522– 4531,  DOI: 10.1098/rspb.2012.1741
      There is no corresponding record for this reference.
    58. 58
      Oesch, N.; Dunbar, R. I. M. The emergence of recursion in human language: Mentalising predicts recursive syntax task performance. J. Neurolinguistics 2017, 43, 95– 106,  DOI: 10.1016/j.jneuroling.2016.09.008
      There is no corresponding record for this reference.
    59. 59
      Ferruz, N.; Höcker, B. Controllable protein design with language models. Nature Machine Intelligence 2022, 4, 521– 532,  DOI: 10.1038/s42256-022-00499-z
      There is no corresponding record for this reference.
    60. 60
      Ofer, D.; Brandes, N.; Linial, M. The language of proteins: NLP, machine learning & protein sequences. Comput. Struct. Biotechnol. J. 2021, 19, 1750– 1758,  DOI: 10.1016/j.csbj.2021.03.022
      60
      The language of proteins: NLP, machine learning & protein sequences
      Ofer, Dan; Brandes, Nadav; Linial, Michal
      Computational and Structural Biotechnology Journal (2021), 19 (), 1750-1758CODEN: CSBJAC; ISSN:2001-0370. (Elsevier B.V.)
      A review. Natural language processing (NLP) is a field of computer science concerned with automated text and language anal. In recent years, following a series of breakthroughs in deep and machine learning, NLP methods have shown overwhelming progress. Here, we review the success, promise and pitfalls of applying NLP algorithms to the study of proteins. Proteins, which can be represented as strings of amino-acid letters, are a natural fit to many NLP methods. We explore the conceptual similarities and differences between proteins and language, and review a range of protein-related tasks amenable to machine learning. We present methods for encoding the information of proteins as text and analyzing it with NLP methods, reviewing classic concepts such as bag-of-words, k-mers/n-grams and text search, as well as modern techniques such as word embedding, contextualized embedding, deep learning and neural language models. In particular, we focus on recent innovations such as masked language modeling, self-supervised learning and attention-based models. Finally, we discuss trends and challenges in the intersection of NLP and protein research.
    61. 61
      Ptitsyn, O. B. How does protein synthesis give rise to the 3D-structure?. FEBS Lett. 1991, 285, 176– 181,  DOI: 10.1016/0014-5793(91)80799-9
      There is no corresponding record for this reference.
    62. 62
      Yu, L.; Tanwar, D. K.; Penha, E. D. S.; Wolf, Y. I.; Koonin, E. V.; Basu, M. K. Grammar of protein domain architectures 2019, 116, 3636– 3645,  DOI: 10.1073/pnas.1814684116
      There is no corresponding record for this reference.
    63. 63
      Petsko, G. A.; Ringe, D. Protein Structure and Function; Primers in Biology; Blackwell Publishing: London, England, 2003.
      There is no corresponding record for this reference.
    64. 64
      Shenoy, S. R.; Jayaram, B. Proteins: sequence to structure and function-current status. Curr. Protein Pept. Sci. 2010, 11, 498– 514,  DOI: 10.2174/138920310794109094
      64
      Proteins: sequence to structure and function - current status
      Shenoy, Sandhya R.; Jayaram, B.
      Current Protein and Peptide Science (2010), 11 (7), 498-514CODEN: CPPSCM; ISSN:1389-2037. (Bentham Science Publishers Ltd.)
      A review. In an era that has been dominated by structural biol. for the last 30-40 yr, a dramatic change of focus toward sequence anal. has spurred the advent of the genome projects and the resultant diverging sequence/structure deficit. The central challenge of computational structural biol. is therefore to rationalize the mass of sequence information into biochem. and biophys. knowledge and to decipher the structural, functional, and evolutionary clues encoded in the language of biol. sequences. In investigating the meaning of sequences, 2 distinct anal. themes have emerged: (1) in the 1st approach, pattern recognition techniques are used to detect similarity between sequences and hence to infer related structures and functions; (2) in the 2nd, ab initio prediction methods are used to deduce 3-dimensional structure, and ultimately to infer function, directly from the linear sequence. Here, the authors attempt to provide a crit. assessment of what one may and may not expect from the biol. sequences and to identify major issues yet to be resolved. The presentation is organized under several subtitles such as protein sequences, pattern recognition techniques, protein tertiary structure prediction, membrane protein bioinformatics, human proteome, protein-protein interactions, metabolic networks, potential drug targets based on simple sequence properties, disordered proteins, the sequence-structure relation, and chem. logic of protein sequences.
    65. 65
      Takahashi, M.; Maraboeuf, F.; Nordén, B. Locations of functional domains in the RecA protein. Overlap of domains and regulation of activities. Eur. J. Biochem. 1996, 242, 20– 28,  DOI: 10.1111/j.1432-1033.1996.0020r.x
      There is no corresponding record for this reference.
    66. 66
      Liang, W.; KaiYong, Z. Detecting “protein words” through unsupervised word segmentation. arXiv , 2014.
      There is no corresponding record for this reference.
    67. 67
      Kuntz, I. D.; Crippen, G. M.; Kollman, P. A.; Kimelman, D. Calculation of protein tertiary structure. J. Mol. Biol. 1976, 106, 983– 994,  DOI: 10.1016/0022-2836(76)90347-8
      There is no corresponding record for this reference.
    68. 68
      Rodrigue, N.; Lartillot, N.; Bryant, D.; Philippe, H. Site interdependence attributed to tertiary structure in amino acid sequence evolution. Gene 2005, 347, 207– 217,  DOI: 10.1016/j.gene.2004.12.011
      There is no corresponding record for this reference.
    69. 69
      Eisenhaber, F.; Persson, B.; Argos, P. Protein structure prediction: recognition of primary, secondary, and tertiary structural features from amino acid sequence. Crit. Rev. Biochem. Mol. Biol. 1995, 30, 1– 94,  DOI: 10.3109/10409239509085139
      69
      Protein structure prediction: recognition of primary, secondary, and tertiary structural features from amino acid sequence
      Eisenhaber F; Persson B; Argos P
      Critical reviews in biochemistry and molecular biology (1995), 30 (1), 1-94 ISSN:1040-9238.
      This review attempts a critical stock-taking of the current state of the science aimed at predicting structural features of proteins from their amino acid sequences. At the primary structure level, methods are considered for detection of remotely related sequences and for recognizing amino acid patterns to predict posttranslational modifications and binding sites. The techniques involving secondary structural features include prediction of secondary structure, membrane-spanning regions, and secondary structural class. At the tertiary structural level, methods for threading a sequence into a mainchain fold, homology modeling and assigning sequences to protein families with similar folds are discussed. A literature analysis suggests that, to date, threading techniques are not able to show their superiority over sequence pattern recognition methods. Recent progress in the state of ab initio structure calculation is reviewed in detail. The analysis shows that many structural features can be predicted from the amino acid sequence much better than just a few years ago and with attendant utility in experimental research. Best prediction can be achieved for new protein sequences that can be assigned to well-studied protein families. For single sequences without homologues, the folding problem has not yet been solved.
    70. 70
      Garfield, E. Chemico-linguistics: computer translation of chemical nomenclature. Nature 1961, 192, 192,  DOI: 10.1038/192192a0
      There is no corresponding record for this reference.
    71. 71
      Wigh, D. S.; Goodman, J. M.; Lapkin, A. A. A review of molecular representation in the age of machine learning. Wiley Interdiscip. Rev. Comput. Mol. Sci. 2022,  DOI: 10.1002/wcms.1603
      There is no corresponding record for this reference.
    72. 72
      Weininger, D. SMILES, a chemical language and information system. 1 Introduction to methodology and encoding rules. J. Chem. Inf. Comput. Sci. 1988, 28, 31– 36,  DOI: 10.1021/ci00057a005
      72
      SMILES, a chemical language and information system. 1. Introduction to methodology and encoding rules
      Weininger, David
      Journal of Chemical Information and Computer Sciences (1988), 28 (1), 31-6CODEN: JCISD8; ISSN:0095-2338.
      The SMILES (simplified mol. input line entry system) chem. notation system is described for information processing. The system is based on principles of mol. graph theory and it allows structure specification by use of a very small and natural grammar well suited for high-speed machine processing. The system is easy to use, has high machine compatibility, and allows many computer applications, including notation generation, const. speed database retrieval, substructure searching, and property prediction models.
    73. 73
      Wang, Y.; Xiao, J.; Suzek, T. O.; Zhang, J.; Wang, J.; Bryant, S. H. PubChem: a public information system for analyzing bioactivities of small molecules. Nucleic Acids Res. 2009, 37, W623– 33,  DOI: 10.1093/nar/gkp456
      73
      PubChem: a public information system for analyzing bioactivities of small molecules
      Wang, Yanli; Xiao, Jewen; Suzek, Tugba O.; Zhang, Jian; Wang, Jiyao; Bryant, Stephen H.
      Nucleic Acids Research (2009), 37 (Web Server), W623-W633CODEN: NARHAD; ISSN:0305-1048. (Oxford University Press)
      PubChem (http://pubchem.ncbi.nlm.nih.gov) is a public repository for biol. properties of small mols. hosted by the US National Institutes of Health (NIH). PubChem BioAssay database currently contains biol. test results for more than 700 000 compds. The goal of PubChem is to make this information easily accessible to biomedical researchers. In this work, we present a set of web servers to facilitate and optimize the utility of biol. activity information within PubChem. These web-based services provide tools for rapid data retrieval, integration and comparison of biol. screening results, exploratory structure-activity anal., and target selectivity examn. This article reviews these bioactivity anal. tools and discusses their uses. Most of the tools described in this work can be directly accessed at http://pubchem.ncbi.nlm.nih.gov/assay/. URLs for accessing other tools described in this work are specified individually.
    74. 74
      Degtyarenko, K.; de Matos, P.; Ennis, M.; Hastings, J.; Zbinden, M.; McNaught, A.; Alcántara, R.; Darsow, M.; Guedj, M.; Ashburner, M. ChEBI: a database and ontology for chemical entities of biological interest. Nucleic Acids Res. 2007, 36, D344– 50,  DOI: 10.1093/nar/gkm791
      There is no corresponding record for this reference.
    75. 75
      Wishart, D. S.; Knox, C.; Guo, A. C.; Cheng, D.; Shrivastava, S.; Tzur, D.; Gautam, B.; Hassanali, M. DrugBank: a knowledgebase for drugs, drug actions and drug targets. Nucleic Acids Res. 2008, 36, D901– 6,  DOI: 10.1093/nar/gkm958
      75
      DrugBank: a knowledgebase for drugs, drug actions and drug targets
      Wishart, David S.; Knox, Craig; Guo, An Chi; Cheng, Dean; Shrivastava, Savita; Tzur, Dan; Gautam, Bijaya; Hassanali, Murtaza
      Nucleic Acids Research (2008), 36 (Database Iss), D901-D906CODEN: NARHAD; ISSN:0305-1048. (Oxford University Press)
      DrugBank is a richly annotated resource that combines detailed drug data with comprehensive drug target and drug action information. Since its first release in 2006, DrugBank has been widely used to facilitate in silico drug target discovery, drug design, drug docking or screening, drug metab. prediction, drug interaction prediction and general pharmaceutical education. The latest version of DrugBank (release 2.0) has been expanded significantly over the previous release. With ∼4900 drug entries, it now contains 60% more FDA-approved small mol. and biotech drugs including 10% more exptl.' drugs. Significantly, more protein target data has also been added to the database, with the latest version of DrugBank contg. three times as many non-redundant protein or drug target sequences as before (1565 vs. 524). Each DrugCard entry now contains more than 100 data fields with half of the information being devoted to drug/chem. data and the other half devoted to pharmacol., pharmacogenomic and mol. biol. data. A no. of new data fields, including food-drug interactions, drug-drug interactions and exptl. ADME data have been added in response to numerous user requests. DrugBank has also significantly improved the power and simplicity of its structure query and text query searches. DrugBank is available at http://www.drugbank.ca.
    76. 76
      Wang, X.; Hao, J.; Yang, Y.; He, K. Natural language adversarial defense through synonym encoding. Proceedings of the Thirty-Seventh Conference on Uncertainty in Artificial Intelligence 2021, 823– 833
      There is no corresponding record for this reference.
    77. 77
      Bjerrum, E. J. SMILES Enumeration as Data Augmentation for Neural Network Modeling of Molecules. arXiv , 2017.
      There is no corresponding record for this reference.
    78. 78
      Lee, I.; Nam, H. Infusing Linguistic Knowledge of SMILES into Chemical Language Models. arXiv , 2022.
      There is no corresponding record for this reference.
    79. 79
      Skinnider, M. A. Invalid SMILES are beneficial rather than detrimental to chemical language models. Nature Mach. Intell. 2024, 6, 437,  DOI: 10.1038/s42256-024-00821-x
      There is no corresponding record for this reference.
    80. 80
      O’Boyle, N.; Dalke, A. DeepSMILES: An Adaptation of SMILES for Use in Machine-Learning of Chemical Structures. ChemRxiv , 2018.
      There is no corresponding record for this reference.
    81. 81
      Krenn, M.; Häse, F.; Nigam, A.; Friederich, P.; Aspuru-Guzik, A. Self-referencing embedded strings (SELFIES): A 100% robust molecular string representation. Mach. Learn.: Sci. Technol. 2020, 1, 045024,  DOI: 10.1088/2632-2153/aba947
      There is no corresponding record for this reference.
    82. 82
      Gohlke, H.; Mannhold, R.; Kubinyi, H.; Folkers, G. In Protein-Ligand Interactions; Gohlke, H., Ed.; Methods and Principles in Medicinal Chemistry; Wiley-VCH Verlag: Weinheim, Germany, 2012.
      There is no corresponding record for this reference.
    83. 83
      Jumper, J. Highly accurate protein structure prediction with AlphaFold. Nature 2021, 596, 583– 589,  DOI: 10.1038/s41586-021-03819-2
      83
      Highly accurate protein structure prediction with AlphaFold
      Jumper, John; Evans, Richard; Pritzel, Alexander; Green, Tim; Figurnov, Michael; Ronneberger, Olaf; Tunyasuvunakool, Kathryn; Bates, Russ; Zidek, Augustin; Potapenko, Anna; Bridgland, Alex; Meyer, Clemens; Kohl, Simon A. A.; Ballard, Andrew J.; Cowie, Andrew; Romera-Paredes, Bernardino; Nikolov, Stanislav; Jain, Rishub; Adler, Jonas; Back, Trevor; Petersen, Stig; Reiman, David; Clancy, Ellen; Zielinski, Michal; Steinegger, Martin; Pacholska, Michalina; Berghammer, Tamas; Bodenstein, Sebastian; Silver, David; Vinyals, Oriol; Senior, Andrew W.; Kavukcuoglu, Koray; Kohli, Pushmeet; Hassabis, Demis
      Nature (London, United Kingdom) (2021), 596 (7873), 583-589CODEN: NATUAS; ISSN:0028-0836. (Nature Portfolio)
      Proteins are essential to life, and understanding their structure can facilitate a mechanistic understanding of their function. Through an enormous exptl. effort, the structures of around 100,000 unique proteins have been detd., but this represents a small fraction of the billions of known protein sequences. Structural coverage is bottlenecked by the months to years of painstaking effort required to det. a single protein structure. Accurate computational approaches are needed to address this gap and to enable large-scale structural bioinformatics. Predicting the three-dimensional structure that a protein will adopt based solely on its amino acid sequence-the structure prediction component of the 'protein folding problem'-has been an important open research problem for more than 50 years. Despite recent progress, existing methods fall far short of at. accuracy, esp. when no homologous structure is available. Here we provide the first computational method that can regularly predict protein structures with at. accuracy even in cases in which no similar structure is known. We validated an entirely redesigned version of our neural network-based model, AlphaFold, in the challenging 14th Crit. Assessment of protein Structure Prediction (CASP14), demonstrating accuracy competitive with exptl. structures in a majority of cases and greatly outperforming other methods. Underpinning the latest version of AlphaFold is a novel machine learning approach that incorporates phys. and biol. knowledge about protein structure, leveraging multi-sequence alignments, into the design of the deep learning algorithm.
    84. 84
      Landrum, G. RDKit: A software suite for cheminformatics, computational chemistry, and predictive modeling. http://www.rdkit.org/RDKit_Overview.pdf, 2013; Accessed: 2023–12–13.
      There is no corresponding record for this reference.
    85. 85
      Mukherjee, S.; Ghosh, M.; Basuchowdhuri, P. Proceedings of the 2022 SIAM International Conference on Data Mining (SDM); Proceedings; Society for Industrial and Applied Mathematics, 2022; pp 729– 737.
      There is no corresponding record for this reference.
    86. 86
      Chen, L.; Tan, X.; Wang, D.; Zhong, F.; Liu, X.; Yang, T.; Luo, X.; Chen, K.; Jiang, H.; Zheng, M. TransformerCPI: improving compound–protein interaction prediction by sequence-based deep learning with self-attention mechanism and label reversal experiments. Bioinformatics 2020, 36, 4406– 4414,  DOI: 10.1093/bioinformatics/btaa524
      86
      TransformerCPI: improving compound-protein interaction prediction by sequence-based deep learning with self-attention mechanism and label reversal experiments
      Chen, Lifan; Tan, Xiaoqin; Wang, Dingyan; Zhong, Feisheng; Liu, Xiaohong; Yang, Tianbiao; Luo, Xiaomin; Chen, Kaixian; Jiang, Hualiang; Zheng, Mingyue
      Bioinformatics (2020), 36 (16), 4406-4414CODEN: BOINFP; ISSN:1367-4811. (Oxford University Press)
      Motivation: Identifying compd.-protein interaction (CPI) is a crucial task in drug discovery and chemogenomics studies, and proteins without three-dimensional structure account for a large part of potential biol. targets, which requires developing methods using only protein sequence information to predict CPI. However, sequence-based CPI models may face some specific pitfalls, including using inappropriate datasets, hidden ligand bias and splitting datasets inappropriately, resulting in overestimation of their prediction performance. Results: To address these issues, we here constructed new datasets specific for CPI prediction, proposed a novel transformer neural network named TransformerCPI, and introduced a more rigorous label reversal expt. to test whether a model learns true interaction features. TransformerCPI achieved much improved performance on the new expts., and it can be deconvolved to highlight important interacting regions of protein sequences and compd. atoms, which may contribute chem. biol. studies with useful guidance for further ligand structural optimization.
    87. 87
      Aly Abdelkader, G.; Ngnamsie Njimbouom, S.; Oh, T.-J.; Kim, J.-D. ResBiGAAT: Residual Bi-GRU with attention for protein-ligand binding affinity prediction. Comput. Biol. Chem. 2023, 107, 107969,  DOI: 10.1016/j.compbiolchem.2023.107969
      There is no corresponding record for this reference.
    88. 88
      Li, Q.; Zhang, X.; Wu, L.; Bo, X.; He, S.; Wang, S. PLA-MoRe: AProtein–Ligand Binding Affinity Prediction Model via Comprehensive Molecular Representations. J. Chem. Inf. Model. 2022, 62, 4380– 4390,  DOI: 10.1021/acs.jcim.2c00960
      There is no corresponding record for this reference.
    89. 89
      Abramson, J. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 2024, 636, E4,  DOI: 10.1038/s41586-024-08416-7
      There is no corresponding record for this reference.
    90. 90
      Gaulton, A.; Bellis, L. J.; Bento, A. P.; Chambers, J.; Davies, M.; Hersey, A.; Light, Y.; McGlinchey, S.; Michalovich, D.; Al-Lazikani, B.; Overington, J. P. ChEMBL: a large-scale bioactivity database for drug discovery. Nucleic Acids Res. 2012, 40, D1100– 7,  DOI: 10.1093/nar/gkr777
      90
      ChEMBL: a large-scale bioactivity database for drug discovery
      Gaulton, Anna; Bellis, Louisa J.; Bento, A. Patricia; Chambers, Jon; Davies, Mark; Hersey, Anne; Light, Yvonne; McGlinchey, Shaun; Michalovich, David; Al-Lazikani, Bissan; Overington, John P.
      Nucleic Acids Research (2012), 40 (D1), D1100-D1107CODEN: NARHAD; ISSN:0305-1048. (Oxford University Press)
      ChEMBL is an Open Data database contg. binding, functional and ADMET information for a large no. of drug-like bioactive compds. These data are manually abstracted from the primary published literature on a regular basis, then further curated and standardized to maximize their quality and utility across a wide range of chem. biol. and drug-discovery research problems. Currently, the database contains 5.4 million bioactivity measurements for more than 1 million compds. and 5200 protein targets. Access is available through a web-based interface, data downloads and web services at: https://www.ebi.ac.uk/chembldb.
    91. 91
      Wishart, D. S.; Knox, C.; Guo, A. C.; Shrivastava, S.; Hassanali, M.; Stothard, P.; Chang, Z.; Woolsey, J. DrugBank: a comprehensive resource for in silico drug discovery and exploration. Nucleic Acids Res. 2006, 34, D668– 72,  DOI: 10.1093/nar/gkj067
      91
      DrugBank: a comprehensive resource for in silico drug discovery and exploration
      Wishart, David S.; Knox, Craig; Guo, An Chi; Shrivastava, Savita; Hassanali, Murtaza; Stothard, Paul; Chang, Zhan; Woolsey, Jennifer
      Nucleic Acids Research (2006), 34 (Database), D668-D672CODEN: NARHAD; ISSN:0305-1048. (Oxford University Press)
      DrugBank is a unique bioinformatics/cheminformatics resource that combines detailed drug (i.e. chem.) data with comprehensive drug target (i.e. protein) information. The database contains >4100 drug entries including >800 FDA approved small mol. and biotech drugs as well as >3200 exptl. drugs. Addnl., >14 000 protein or drug target sequences are linked to these drug entries. Each DrugCard entry contains >80 data fields with half of the information being devoted to drug/chem. data and the other half devoted to drug target or protein data. Many data fields are hyperlinked to other databases (KEGG, PubChem, ChEBI, PDB, Swiss-Prot and GenBank) and a variety of structure viewing applets. The database is fully searchable supporting extensive text, sequence, chem. structure and relational query searches. Potential applications of DrugBank include in silico drug target discovery, drug design, drug docking or screening, drug metab. prediction, drug interaction prediction and general pharmaceutical education. DrugBank is available at http://redpoll.pharmacy.ualberta.ca/drugbank/.
    92. 92
      Berman, H. M.; Westbrook, J.; Feng, Z.; Gilliland, G.; Bhat, T. N.; Weissig, H.; Shindyalov, I. N.; Bourne, P. E. The Protein Data Bank. Nucleic Acids Res. 2000, 28, 235– 242,  DOI: 10.1093/nar/28.1.235
      92
      The Protein Data Bank
      Berman, Helen M.; Westbrook, John; Feng, Zukang; Gilliland, Gary; Bhat, T. N.; Weissig, Helge; Shindyalov, Ilya N.; Bourne, Philip E.
      Nucleic Acids Research (2000), 28 (1), 235-242CODEN: NARHAD; ISSN:0305-1048. (Oxford University Press)
      The Protein Data Bank (PDB; http://www.rcsb.org/pdb/)is the single worldwide archive of structural data of biol. macromols. This paper describes the goals of the PDB, the systems in place for data deposition and access, how to obtain further information, and near-term plans for the future development of the resource.
    93. 93
      Acids research, N. 2017 UniProt: the universal protein knowledgebase. Nucleic Acids Res. 2017, 45, D158– D169,  DOI: 10.1093/nar/gkw1099
      There is no corresponding record for this reference.
    94. 94
      Davis, M. I.; Hunt, J. P.; Herrgard, S.; Ciceri, P.; Wodicka, L. M.; Pallares, G.; Hocker, M.; Treiber, D. K.; Zarrinkar, P. P. Comprehensive analysis of kinase inhibitor selectivity. Nat. Biotechnol. 2011, 29, 1046– 1051,  DOI: 10.1038/nbt.1990
      94
      Comprehensive analysis of kinase inhibitor selectivity
      Davis, Mindy I.; Hunt, Jeremy P.; Herrgard, Sanna; Ciceri, Pietro; Wodicka, Lisa M.; Pallares, Gabriel; Hocker, Michael; Treiber, Daniel K.; Zarrinkar, Patrick P.
      Nature Biotechnology (2011), 29 (11), 1046-1051CODEN: NABIF9; ISSN:1087-0156. (Nature Publishing Group)
      We tested the interaction of 72 kinase inhibitors with 442 kinases covering >80% of the human catalytic protein kinome. Our data show that, as a class, type II inhibitors are more selective than type I inhibitors, but that there are important exceptions to this trend. The data further illustrate that selective inhibitors have been developed against the majority of kinases targeted by the compds. tested. Anal. of the interaction patterns reveals a class of 'group-selective' inhibitors broadly active against a single subfamily of kinases, but selective outside that subfamily. The data set suggests compds. to use as tools to study kinases for which no dedicated inhibitors exist. It also provides a foundation for further exploring kinase inhibitor biol. and toxicity, as well as for studying the structural basis of the obsd. interaction patterns. Our findings will help to realize the direct enabling potential of genomics for drug development and basic research about cellular signaling.
    95. 95
      Tang, J.; Szwajda, A.; Shakyawar, S.; Xu, T.; Hintsanen, P.; Wennerberg, K.; Aittokallio, T. Making sense of large-scale kinase inhibitor bioactivity data sets: a comparative and integrative analysis. J. Chem. Inf. Model. 2014, 54, 735– 743,  DOI: 10.1021/ci400709d
      95
      Making Sense of Large-Scale Kinase Inhibitor Bioactivity Data Sets: A Comparative and Integrative Analysis
      Tang, Jing; Szwajda, Agnieszka; Shakyawar, Sushil; Xu, Tao; Hintsanen, Petteri; Wennerberg, Krister; Aittokallio, Tero
      Journal of Chemical Information and Modeling (2014), 54 (3), 735-743CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      We carried out a systematic evaluation of target selectivity profiles across three recent large-scale biochem. assays of kinase inhibitors and further compared these standardized bioactivity assays with data reported in the widely used databases ChEMBL and STITCH. Our comparative evaluation revealed relative benefits and potential limitations among the bioactivity types, as well as pinpointed biases in the database curation processes. Ignoring such issues in data heterogeneity and representation may lead to biased modeling of drugs' polypharmacol. effects as well as to unrealistic evaluation of computational strategies for the prediction of drug-target interaction networks. Toward making use of the complementary information captured by the various bioactivity types, including IC50, Ki, and Kd, we also introduce a model-based integration approach, termed KIBA, and demonstrate here how it can be used to classify kinase inhibitor targets and to pinpoint potential errors in database-reported drug-target interactions. An integrated drug-target bioactivity matrix across 52 498 chem. compds. and 467 kinase targets, including a total of 246 088 KIBA scores, has been made freely available.
    96. 96
      Wang, R.; Fang, X.; Lu, Y.; Wang, S. The PDBbind database: collection of binding affinities for protein-ligand complexes with known three-dimensional structures. J. Med. Chem. 2004, 47, 2977– 2980,  DOI: 10.1021/jm030580l
      96
      The PDBbind database: Collection of binding affinities for protein-ligand complexes with known three-dimensional structures
      Wang, Renxiao; Fang, Xueliang; Lu, Yipin; Wang, Shaomeng
      Journal of Medicinal Chemistry (2004), 47 (12), 2977-2980CODEN: JMCMAR; ISSN:0022-2623. (American Chemical Society)
      We have screened the entire Protein Data Bank (Release No. 103, Jan. 2003) and identified 5671 protein-ligand complexes out of 19 621 exptl. structures. A systematic examn. of the primary refs. of these entries has led to a collection of binding affinity data (Kd, Ki, and IC50) for a total of 1359 complexes. The outcomes of this project have been organized into a Web-accessible database named the PDBbind database.
    97. 97
      Chen, S.; Zhang, S.; Fang, X.; Lin, L.; Zhao, H.; Yang, Y. Protein complex structure modeling by cross-modal alignment between cryo-EM maps and protein sequences. Nat. Commun. 2024, 15, 8808,  DOI: 10.1038/s41467-024-53116-5
      There is no corresponding record for this reference.
    98. 98
      Bishop, M.C. Pattern Recognition and Machine Learning, 1st ed.; Information Science and Statistics; Springer: New York, NY, 2006.
      There is no corresponding record for this reference.
    99. 99
      Yang, J.; Shen, C.; Huang, N. Predicting or Pretending: Artificial Intelligence for Protein-Ligand Interactions Lack of Sufficiently Large and Unbiased Datasets. Front. Pharmacol. 2020, 11, 69,  DOI: 10.3389/fphar.2020.00069
      99
      Predicting or pretending: artificial intelligence for protein-ligand interactions lack of sufficiently large and unbiased datasets
      Yang, Jincai; Shen, Cheng; Huang, Niu
      Frontiers in Pharmacology (2020), 11 (), 69CODEN: FPRHAU; ISSN:1663-9812. (Frontiers Media S.A.)
      Predicting protein-ligand interactions using artificial intelligence (AI) models has attracted great interest in recent years. However, data-driven AI models unequivocally suffer from a lack of sufficiently large and unbiased datasets. Here, we systematically investigated the data biases on the PDBbind and DUD-E datasets. We examd. the model performance of at. convolutional neural network (ACNN) on the PDBbind core set and achieved a Pearson R2 of 0.73 between exptl. and predicted binding affinities. Strikingly, the ACNN models did not require learning the essential protein-ligand interactions in complex structures and achieved similar performance even on datasets contg. only ligand structures or only protein structures, while data splitting based on similarity clustering (protein sequence or ligand scaffold) significantly reduced the model performance. We also identified the property and topol. biases in the DUD-E dataset which led to the artificially increased enrichment performance of virtual screening. The property bias in DUD-E was reduced by enforcing the more stringent ligand property matching rules, while the topol. bias still exists due to the use of mol. fingerprint similarity as a decoy selection criterion. Therefore, we believe that sufficiently large and unbiased datasets are desirable for training robust AI models to accurately predict protein-ligand interactions.
    100. 100
      Kryshtafovych, A.; Schwede, T.; Topf, M.; Fidelis, K.; Moult, J. Critical assessment of methods of protein structure prediction (CASP)–Round XIII. Proteins 2019, 87, 1011– 1020,  DOI: 10.1002/prot.25823
      100
      Critical assessment of methods of protein structure prediction (CASP)-Round XIII
      Kryshtafovych, Andriy; Schwede, Torsten; Topf, Maya; Fidelis, Krzysztof; Moult, John
      Proteins: Structure, Function, and Bioinformatics (2019), 87 (12), 1011-1020CODEN: PSFBAF; ISSN:1097-0134. (Wiley-Blackwell)
      A review. CASP (crit. assessment of structure prediction) assesses the state of the art in modeling protein structure from amino acid sequence. The most recent expt. (CASP13 held in 2018) saw dramatic progress in structure modeling without use of structural templates (historically "ab initio" modeling). Progress was driven by the successful application of deep learning techniques to predict inter-residue distances. In turn, these results drove dramatic improvements in three-dimensional structure accuracy: With the proviso that there are an adequate no. of sequences known for the protein family, the new methods essentially solve the long-standing problem of predicting the fold topol. of monomeric proteins. Further, the no. of sequences required in the alignment has fallen substantially. There is also substantial improvement in the accuracy of template-based models. Other areas-model refinement, accuracy estn., and the structure of protein assemblies-have again yielded interesting results. CASP13 placed increased emphasis on the use of sparse data together with modeling and chem. crosslinking, SAXS, and NMR all yielded more mature results. This paper summarizes the key outcomes of CASP13. The special issue of PROTEINS contains papers describing the CASP13 assessments in each modeling category and contributions from the participants.
    101. 101
      Janin, J.; Henrick, K.; Moult, J.; Eyck, L. T.; Sternberg, M. J. E.; Vajda, S.; Vakser, I.; Wodak, S. J. CAPRI: A Critical Assessment of PRedicted Interactions. Proteins 2003, 52, 2– 9,  DOI: 10.1002/prot.10381
      101
      CAPRI: A critical assessment of predicted interactions
      Janin, Joel; Henrick, Kim; Moult, John; Ten Eyck, Lynn; Sternberg, Michael J. E.; Vajda, Sandor; Vakser, Ilya; Wodak, Shoshana J.
      Proteins: Structure, Function, and Genetics (2003), 52 (1), 2-9CODEN: PSFGEY; ISSN:0887-3585. (Wiley-Liss, Inc.)
      A review. CAPRI is a community wide expt. to assess the capacity of protein-docking methods to predict protein-protein interactions. Nineteen groups participated in rounds 1 and 2 of CAPRI and submitted blind structure predictions for seven protein-protein complexes based on the known structure of the component proteins. The predictions were compared to the unpublished X-ray structures of the complexes. We describe here the motivations for launching CAPRI, the rules that we applied to select targets and run the expt., and some conclusions that can already be drawn. The results stress the need for new scoring functions and for methods handling the conformation changes that were obsd. in some of the target systems. CAPRI has already been a powerful drive for the community of computational biologists who development docking algorithms. We hope that this issue of Proteins will also be of interest to the community of structural biologists, which we call upon to provide new targets for future rounds of CAPRI, and to all mol. biologists who view protein-protein recognition as an essential process.
    102. 102
      Lensink, M. F.; Nadzirin, N.; Velankar, S.; Wodak, S. J. Modeling protein-protein, protein-peptide, and protein-oligosaccharide complexes: CAPRI 7th edition. Proteins 2020, 88, 916– 938,  DOI: 10.1002/prot.25870
      102
      Modeling protein-protein, protein-peptide, and protein-oligosaccharide complexes: CAPRI 7th edition
      Lensink, Marc F.; Nadzirin, Nurul; Velankar, Sameer; Wodak, Shoshana J.
      Proteins: Structure, Function, and Bioinformatics (2020), 88 (8), 916-938CODEN: PSFBAF; ISSN:1097-0134. (Wiley-Blackwell)
      The authors present the seventh report on the performance of methods for predicting the at. resoln. structures of protein complexes offered as targets to the community-wide initiative on the Crit. Assessment of Predicted Interactions. Performance was evaluated on the basis of 36,114 models of protein complexes submitted by 57 groups-including 13 automatic servers-in prediction rounds held during the years 2016 to 2019 for eight protein-protein, three protein-peptide, and five protein-oligosaccharide targets with different length ligands. Six of the protein-protein targets represented challenging hetero-complexes, due to factors such as availability of distantly related templates for the individual subunits, or for the full complex, inter-domain flexibility, conformational adjustments at the binding region, or the multi-component nature of the complex. The main challenge for the protein-peptide and protein-oligosaccharide complexes was to accurately model the ligand conformation and its interactions at the interface. Encouragingly, models of acceptable quality, or better, were obtained for a total of six protein-protein complexes, which included four of the challenging hetero-complexes and a homo-decamer. But fewer of these targets were predicted with medium or higher accuracy. High accuracy models were obtained for two of the three protein-peptide targets, and for one of the protein-oligosaccharide targets. The remaining protein-sugar targets were predicted with medium accuracy. The authors' anal. indicates that progress in predicting increasingly challenging and diverse types of targets is due to closer integration of template-based modeling techniques with docking, scoring, and model refinement procedures, and to significant incremental improvements in the underlying methodologies.
    103. 103
      Schomburg, I.; Chang, A.; Schomburg, D. BRENDA, enzyme data and metabolic information. Nucleic Acids Res. 2002, 30, 47– 49,  DOI: 10.1093/nar/30.1.47
      103
      BRENDA, enzyme data and metabolic information
      Schomburg, Ida; Chang, Antje; Schomburg, Dietmar
      Nucleic Acids Research (2002), 30 (1), 47-49CODEN: NARHAD; ISSN:0305-1048. (Oxford University Press)
      A review with 6 refs. BRENDA is a comprehensive relational database on functional and mol. information of enzymes, based on primary literature. The database contains information extd. and evaluated from ∼46,000 refs., holding data of at least 40,000 different enzymes from >6900 different organisms, classified in ∼3900 EC nos. BRENDA is an important tool for biochem. and medical research covering information on properties of all classified enzymes, including data on the occurrence, catalyzed reaction, kinetics, substrates/products, inhibitors, cofactors, activators, structure and stability. All data are connected to literature refs. which in turn are linked to PubMed. The data and information provide a fundamental tool for research of enzyme mechanisms, metabolic pathways, the evolution of metab. and, furthermore, for medicinal diagnostics and pharmaceutical research. The database is a resource for data of enzymes, classified according to the EC system of the IUBMB Enzyme Nomenclature Committee, and the entries are cross-referenced to other databases, i.e., organism classification, protein sequence, protein structure, and literature refs. BRENDA provides an academic web access at http://www.brenda.uni-koeln.de.
    104. 104
      Liu, T.; Lin, Y.; Wen, X.; Jorissen, R. N.; Gilson, M. K. BindingDB: a web-accessible database of experimentally determined protein-ligand binding affinities. Nucleic Acids Res. 2007, 35, D198– 201,  DOI: 10.1093/nar/gkl999
      104
      BindingDB: a web-accessible database of experimentally determined protein-ligand binding affinities
      Liu, Tiqing; Lin, Yuhmei; Wen, Xin; Jorissen, Robert N.; Gilson, Michael K.
      Nucleic Acids Research (2007), 35 (Database Iss), D198-D201CODEN: NARHAD; ISSN:0305-1048. (Oxford University Press)
      BindingDB is a publicly accessible database currently contg. ∼20 000 exptl. detd. binding affinities of protein-ligand complexes, for 110 protein targets including isoforms and mutational variants, and ∼11 000 small mol. ligands. The data are extd. from the scientific literature, data collection focusing on proteins that are drug-targets or candidate drug-targets and for which structural data are present in the Protein Data Bank. The BindingDB website supports a range of query types, including searches by chem. structure, substructure and similarity; protein sequence; ligand and protein names; affinity ranges and mol. wt. Data sets generated by BindingDB queries can be downloaded in the form of annotated SDfiles for further anal., or used as the basis for virtual screening of a compd. database uploaded by the user. The data in BindingDB are linked both to structural data in the PDB via PDB IDs and chem. and sequence searches, and to the literature in PubMed via PubMed IDs.
    105. 105
      Amemiya, T.; Koike, R.; Kidera, A.; Ota, M. PSCDB: a database for protein structural change upon ligand binding. Nucleic Acids Res. 2012, 40, D554– 8,  DOI: 10.1093/nar/gkr966
      There is no corresponding record for this reference.
    106. 106
      Mysinger, M. M.; Carchia, M.; Irwin, J. J.; Shoichet, B. K. Directory of useful decoys, enhanced (DUD-E): better ligands and decoys for better benchmarking. J. Med. Chem. 2012, 55, 6582– 6594,  DOI: 10.1021/jm300687e
      106
      Directory of Useful Decoys, Enhanced (DUD-E): Better Ligands and Decoys for Better Benchmarking
      Mysinger, Michael M.; Carchia, Michael; Irwin, John. J.; Shoichet, Brian K.
      Journal of Medicinal Chemistry (2012), 55 (14), 6582-6594CODEN: JMCMAR; ISSN:0022-2623. (American Chemical Society)
      A key metric to assess mol. docking remains ligand enrichment against challenging decoys. Whereas the directory of useful decoys (DUD) has been widely used, clear areas for optimization have emerged. Here we describe an improved benchmarking set that includes more diverse targets such as GPCRs and ion channels, totaling 102 proteins with 22886 clustered ligands drawn from ChEMBL, each with 50 property-matched decoys drawn from ZINC. To ensure chemotype diversity, we cluster each target's ligands by their Bemis-Murcko at. frameworks. We add net charge to the matched physicochem. properties and include only the most dissimilar decoys, by topol., from the ligands. An online automated tool (http://decoys.docking.org) generates these improved matched decoys for user-supplied ligands. We test this data set by docking all 102 targets, using the results to improve the balance between ligand desolvation and electrostatics in DOCK 3.6. The complete DUD-E benchmarking set is freely available at http://dude.docking.org.
    107. 107
      Warren, G. L.; Do, T. D.; Kelley, B. P.; Nicholls, A.; Warren, S. D. Essential considerations for using protein-ligand structures in drug discovery. Drug Discovery Today 2012, 17, 1270– 1281,  DOI: 10.1016/j.drudis.2012.06.011
      107
      Essential considerations for using protein-ligand structures in drug discovery
      Warren, Gregory L.; Do, Thanh D.; Kelley, Brian P.; Nicholls, Anthony; Warren, Stephen D.
      Drug Discovery Today (2012), 17 (23-24), 1270-1281CODEN: DDTOFS; ISSN:1359-6446. (Elsevier Ltd.)
      A review. Protein-ligand structures are the core data required for structure-based drug design (SBDD). Understanding the error present in this data is essential to the successful development of SBDD tools. Methods for assessing protein-ligand structure quality and a new set of identification criteria are presented here. When these criteria were applied to a set of 728 structures previously used to validate mol. docking software, only 17% were found to be acceptable. Structures were re-refined to maintain internal consistency in the comparison and assessment of the quality criteria. This process resulted in Iridium, a highly trustworthy protein-ligand structure database to be used for development and validation of structure-based design tools for drug discovery.
    108. 108
      Puvanendrampillai, D.; Mitchell, J. B. O. L/D Protein Ligand Database (PLD): additional understanding of the nature and specificity of protein-ligand complexes. Bioinformatics 2003, 19, 1856– 1857,  DOI: 10.1093/bioinformatics/btg243
      There is no corresponding record for this reference.
    109. 109
      Wang, C.; Hu, G.; Wang, K.; Brylinski, M.; Xie, L.; Kurgan, L. PDID: database of molecular-level putative protein-drug interactions in the structural human proteome. Bioinformatics 2016, 32, 579– 586,  DOI: 10.1093/bioinformatics/btv597
      There is no corresponding record for this reference.
    110. 110
      Zhu, M.; Song, X.; Chen, P.; Wang, W.; Wang, B. dbHDPLS: A database of human disease-related protein-ligand structures. Comput. Biol. Chem. 2019, 78, 353– 358,  DOI: 10.1016/j.compbiolchem.2018.12.023
      110
      dbHDPLS: A database of human disease-related protein-ligand structures
      Zhu, Muchun; Song, Xiaoping; Chen, Peng; Wang, Wenyan; Wang, Bing
      Computational Biology and Chemistry (2019), 78 (), 353-358CODEN: CBCOCH; ISSN:1476-9271. (Elsevier B.V.)
      Protein-ligand complexes perform specific functions, most of which are related to human diseases. The database, called as human disease-related protein-ligand structures (dbHDPLS), collected 8833 structures which were extd. from protein data bank (PDB) and other related databases. The database is annotated with comprehensive information involving ligands and drugs, related human diseases and protein-ligand interaction information, with the information of protein structures. The database may be a reliable resource for structure-based drug target discoveries and druggability predictions of protein-ligand binding sites, drug-disease relationships based on protein-ligand complex structures.
    111. 111
      Gao, M.; Moumbock, A. F. A.; Qaseem, A.; Xu, Q.; Günther, S. CovPDB: a high-resolution coverage of the covalent protein-ligand interactome. Nucleic Acids Res. 2022, 50, D445– D450,  DOI: 10.1093/nar/gkab868
      111
      CovPDB: a high-resolution coverage of the covalent protein-ligand interactome
      Gao, Mingjie; Moumbock, Aurelien F. A.; Qaseem, Ammar; Xu, Qianqing; Guenther, Stefan
      Nucleic Acids Research (2022), 50 (D1), D445-D450CODEN: NARHAD; ISSN:1362-4962. (Oxford University Press)
      In recent years, the drug discovery paradigm has shifted toward compds. that covalently modify disease-assocd. target proteins, because they tend to possess high potency, selectivity, and duration of action. The rational design of novel targeted covalent inhibitors (TCIs) typically starts from resolved macromol. structures of target proteins in their apo or holo forms. However, the existing TCI databases contain only a paucity of covalent protein-ligand (cP-L) complexes. Herein, we report CovPDB, the first database solely dedicated to highresoln. cocrystal structures of biol. relevant cP-L complexes, curated from the Protein Data Bank. For these curated complexes, the chem. structures and warheads of pre-reactive electrophilic ligands as well as the covalent bonding mechanisms to their target proteins were expertly manually annotated. Totally, CovPDB contains 733 proteins and 1,501 ligands, relating to 2,294 cP-L complexes, 93 reactive warheads, 14 targetable residues, and 21 covalent mechanisms. Users are provided with an intuitive and interactive web interface that allows multiple search and browsing options to explore the covalent interactome at a mol. level in order to develop novel TCIs. CovPDB is freely accessible and its contents are available for download as flat files of various formats.
    112. 112
      Ammar, A.; Cavill, R.; Evelo, C.; Willighagen, E. PSnpBind: a database of mutated binding site protein-ligand complexes constructed using a multithreaded virtual screening workflow. J. Cheminform. 2022, 14, 8,  DOI: 10.1186/s13321-021-00573-5
      There is no corresponding record for this reference.
    113. 113
      Lingė, D. PLBD: protein-ligand binding database of thermodynamic and kinetic intrinsic parameters. Database 2023,  DOI: 10.1093/database/baad040
      There is no corresponding record for this reference.
    114. 114
      Wei, H.; Wang, W.; Peng, Z.; Yang, J. Q-BioLiP: A Comprehensive Resource for Quaternary Structure-based Protein–ligand Interactions. bioRxiv , 2023, 2023.06.23.546351.
      There is no corresponding record for this reference.
    115. 115
      Korlepara, D. B. PLAS-20k: Extended dataset of protein-ligand affinities from MD simulations for machine learning applications. Sci. Data 2024,  DOI: 10.1038/s41597-023-02872-y
      There is no corresponding record for this reference.
    116. 116
      Xenarios, I.; Rice, D. W.; Salwinski, L.; Baron, M. K.; Marcotte, E. M.; Eisenberg, D. DIP: the database of interacting proteins. Nucleic Acids Res. 2000, 28, 289– 291,  DOI: 10.1093/nar/28.1.289
      116
      DIP: the Database of Interacting Proteins
      Xenarios, Ioannis; Rice, Danny W.; Salwinski, Lukasz; Baron, Marisa K.; Marcotte, Edward M.; Eisenberg, David
      Nucleic Acids Research (2000), 28 (1), 289-291CODEN: NARHAD; ISSN:0305-1048. (Oxford University Press)
      The Database of Interacting Proteins is a database that documents exptl. detd. protein-protein interactions. This database is intended to provide the scientific community with a comprehensive and integrated tool for browsing and efficiently extg. information about protein interactions and interaction networks in biol. processes. Beyond cataloging details of protein-protein interactions, the DIP is useful for understanding protein function and protein-protein relationships, studying the properties of networks of interacting proteins, benchmarking predictions of protein-protein interactions, and studying the evolution of protein-protein interactions.
    117. 117
      Wallach, I.; Lilien, R. The protein-small-molecule database, a non-redundant structural resource for the analysis of protein-ligand binding. Bioinformatics 2009, 25, 615– 620,  DOI: 10.1093/bioinformatics/btp035
      There is no corresponding record for this reference.
    118. 118
      Wang, S.; Lin, H.; Huang, Z.; He, Y.; Deng, X.; Xu, Y.; Pei, J.; Lai, L. CavitySpace: A Database of Potential Ligand Binding Sites in the Human Proteome. Biomolecules 2022, 12, 967,  DOI: 10.3390/biom12070967
      118
      CavitySpace: A Database of Potential Ligand Binding Sites in the Human Proteome
      Wang, Shiwei; Lin, Haoyu; Huang, Zhixian; He, Yufeng; Deng, Xiaobing; Xu, Youjun; Pei, Jianfeng; Lai, Luhua
      Biomolecules (2022), 12 (7), 967CODEN: BIOMHC; ISSN:2218-273X. (MDPI AG)
      Location and properties of ligand binding sites provide important information to uncover protein functions and to direct structure-based drug design approaches. However, as binding site detection depends on the three-dimensional (3D) structural data of proteins, functional anal. based on protein ligand binding sites is formidable for proteins without structural information. Recent developments in protein structure prediction and the 3D structures built by AlphaFold provide an unprecedented opportunity for analyzing ligand binding sites in human proteins. Here, we constructed the CavitySpace database, the first pocket library for all the proteins in the human proteome, using a widely-applied ligand binding site detection program CAVITY. Our anal. showed that known ligand binding sites could be well recovered. We grouped the predicted binding sites according to their similarity which can be used in protein function prediction and drug repurposing studies. Novel binding sites in highly reliable predicted structure regions provide new opportunities for drug discovery. Our CavitySpace is freely available and provides a valuable tool for drug discovery and protein function studies.
    119. 119
      Otter, D. W.; Medina, J. R.; Kalita, J. K. A survey of the usages of deep learning for natural language processing. IEEE Trans. Neural Netw. Learn. Syst. 2021, 32, 604– 624,  DOI: 10.1109/TNNLS.2020.2979670
      119
      A Survey of the Usages of Deep Learning for Natural Language Processing
      Otter Daniel W; Medina Julian R; Kalita Jugal K
      IEEE transactions on neural networks and learning systems (2021), 32 (2), 604-624 ISSN:.
      Over the last several years, the field of natural language processing has been propelled forward by an explosion in the use of deep learning models. This article provides a brief introduction to the field and a quick overview of deep learning architectures and methods. It then sifts through the plethora of recent studies and summarizes a large assortment of relevant contributions. Analyzed research areas include several core linguistic processing issues in addition to many applications of computational linguistics. A discussion of the current state of the art is then provided along with recommendations for future research in the field.
    120. 120
      Wang, Y.; You, Z.-H.; Yang, S.; Li, X.; Jiang, T.-H.; Zhou, X. A high efficient biological language model for predicting Protein-Protein interactions. Cells 2019, 8, 122,  DOI: 10.3390/cells8020122
      120
      A high efficient biological language model for predicting protein-protein interactions
      Wang, Yanbin; You, Zhu-Hong; Yang, Shan; Li, Xiao; Jiang, Tong-Hai; Zhou, Xi
      Cells (2019), 8 (2), 122CODEN: CELLC6; ISSN:2073-4409. (MDPI AG)
      : Many life activities and key functions in organisms are maintained by different types of protein-protein interactions (PPIs). In order to accelerate the discovery of PPIs for different species, many computational methods have been developed. Unfortunately, even though computational methods are constantly evolving, efficient methods for predicting PPIs from protein sequence information have not been found for many years due to limiting factors including both methodol. and technol. Inspired by the similarity of biol. sequences and languages, developing a biol. language processing technol. may provide a brand new theor. perspective and feasible method for the study of biol. sequences. In this paper, a pure biol. language processing model is proposed for predicting protein-protein interactions only using a protein sequence. The model was constructed based on a feature representation method for biol. sequences called bio-to-vector (Bio2Vec) and a convolution neural network (CNN). The Bio2Vec obtains protein sequence features by using a "bio-word" segmentation system and a word representation model used for learning the distributed representation for each "bio-word". The Bio2Vec supplies a frame that allows researchers to consider the context information and implicit semantic information of a bio sequence. A remarkable improvement in PPIs prediction performance has been obsd. by using the proposed model compared with state-of-the-art methods. The presentation of this approach marks the start of "bio language processing technol.," which could cause a technol. revolution and could be applied to improve the quality of predictions in other problems.
    121. 121
      Abbasi, K.; Razzaghi, P.; Poso, A.; Amanlou, M.; Ghasemi, J. B.; Masoudi-Nejad, A. DeepCDA: deep cross-domain compound–protein affinity prediction through LSTM and convolutional neural networks. Bioinformatics 2020, 36, 4633– 4642,  DOI: 10.1093/bioinformatics/btaa544
      121
      DeepCDA: deep cross-domain compound-protein affinity prediction through LSTM and convolutional neural networks
      Abbasi, Karim; Razzaghi, Parvin; Poso, Antti; Amanlou, Massoud; Ghasemi, Jahan B.; Masoudi-Nejad, Ali
      Bioinformatics (2020), 36 (17), 4633-4642CODEN: BOINFP; ISSN:1367-4811. (Oxford University Press)
      Motivation: An essential part of drug discovery is the accurate prediction of the binding affinity of new compd.-protein pairs. Most of the std. computational methods assume that compds. or proteins of the test data are obsd. during the training phase. However, in real-world situations, the test and training data are sampled from different domains with different distributions. To cope with this challenge, we propose a deep learning-based approach that consists of three steps. In the first step, the training encoder network learns a novel representation of compds. and proteins. To this end, we combine convolutional layers and long-short-term memory layers so that the occurrence patterns of local substructures through a protein and a compd. sequence are learned. Also, to encode the interaction strength of the protein and compd. substructures, we propose a two-sided attention mechanism. In the second phase, to deal with the different distributions of the training and test domains, a feature encoder network is learned for the test domain by utilizing an adversarial domain adaptation approach. In the third phase, the learned test encoder network is applied to new compd.-protein pairs to predict their binding affinity. Results: To evaluate the proposed approach, we applied it to KIBA, Davis and BindingDB datasets. The results show that the proposed method learns a more reliable model for the test domain in more challenging situations.
    122. 122
      Zhou, G.; Gao, Z.; Ding, Q.; Zheng, H.; Xu, H.; Wei, Z.; Zhang, L.; Ke, G. Uni-Mol: AUniversal 3D Molecular Representation Learning Framework. ChemRxiv , 2023.
      There is no corresponding record for this reference.
    123. 123
      Zhou, D.; Xu, Z.; Li, W.; Xie, X.; Peng, S. MultiDTI: drug–target interaction prediction based on multi-modal representation learning to bridge the gap between new chemical entities and known heterogeneous network. Bioinformatics 2021, 37, 4485– 4492,  DOI: 10.1093/bioinformatics/btab473
      123
      MultiDTI: drug-target interaction prediction based on multi-modal representation learning to bridge the gap between new chemical entities and known heterogeneous network
      Zhou, Deshan; Xu, Zhijian; Li, WenTao; Xie, Xiaolan; Peng, Shaoliang
      Bioinformatics (2021), 37 (23), 4485-4492CODEN: BOINFP; ISSN:1367-4811. (Oxford University Press)
      Motivation: Predicting new drug-target interactions is an important step in new drug development, understanding of its side effects and drug repositioning. Heterogeneous data sources can provide comprehensive information and different perspectives for drug-target interaction prediction. Thus, there have been many calcn. methods relying on heterogeneous networks. Most of them use graph-related algorithms to characterize nodes in heterogeneous networks for predicting new drug-target interactions (DTI). However, these methods can only make predictions in known heterogeneous network datasets, and cannot support the prediction of new chem. entities outside the heterogeneous network, which hinder further drug discovery and development. Results: To solve this problem, we proposed a multi-modal DTI prediction model named 'MultiDTI' which uses our proposed joint learning framework based on heterogeneous networks. It combines the interaction or assocn. information of the heterogeneous network and the drug/target sequence information, and maps the drugs, targets, side effects and disease nodes in the heterogeneous network into a common space. In this way, 'MultiDTI' can map the new chem. entity to this learned common space based on the chem. structure of the new entity. That is, bridging the gap between new chem. entities and known heterogeneous network. Our model has strong predictive performance, and the area under the receiver operating characteristic curve of the model is 0.961 and the area under the precision recall curve is 0.947 with 10-fold cross validation. In addn., some predicted new DTIs have been confirmed by ChEMBL database. Our results indicate that 'MultiDTI' is a powerful and practical tool for predicting new DTI, which can promote the development of drug discovery or drug repositioning.
    124. 124
      Özçelik, R.; Öztürk, H.; Özgür, A.; Ozkirimli, E. ChemBoost: A chemical language based approach for protein - ligand binding affinity prediction. Mol. Inform. 2021, 40, e2000212,  DOI: 10.1002/minf.202000212
      There is no corresponding record for this reference.
    125. 125
      Gaspar, H. A.; Ahmed, M.; Edlich, T.; Fabian, B.; Varszegi, Z.; Segler, M.; Meyers, J.; Fiscato, M. Proteochemometric Models Using Multiple Sequence Alignments and a Subword Segmented Masked Language Model. ChemRxiv , 2021.
      There is no corresponding record for this reference.
    126. 126
      Arseniev-Koehler, A. Theoretical foundations and limits of word embeddings: What types of meaning can they capture. Sociol. Methods Res. 2022, 004912412211401
      There is no corresponding record for this reference.
    127. 127
      Lake, B. M.; Murphy, G. L. Word meaning in minds and machines. Psychol. Rev. 2023, 130, 401– 431,  DOI: 10.1037/rev0000297
      There is no corresponding record for this reference.
    128. 128
      Winchester, S. A Verb for Our Frantic Times. https://www.nytimes.com/2011/05/29/opinion/29winchester.html, 2011; Accessed: 2024–9-15.
      There is no corresponding record for this reference.
    129. 129
      Panapitiya, G.; Girard, M.; Hollas, A.; Sepulveda, J.; Murugesan, V.; Wang, W.; Saldanha, E. Evaluation of deep learning architectures for aqueous solubility prediction. ACS Omega 2022, 7, 15695– 15710,  DOI: 10.1021/acsomega.2c00642
      129
      Evaluation of Deep Learning Architectures for Aqueous Solubility Prediction
      Panapitiya, Gihan; Girard, Michael; Hollas, Aaron; Sepulveda, Jonathan; Murugesan, Vijayakumar; Wang, Wei; Saldanha, Emily
      ACS Omega (2022), 7 (18), 15695-15710CODEN: ACSODF; ISSN:2470-1343. (American Chemical Society)
      Detg. the aq. soly. of mols. is a vital step in many pharmaceutical, environmental, and energy storage applications. Despite efforts made over decades, there are still challenges assocd. with developing a soly. prediction model with satisfactory accuracy for many of these applications. The goals of this study are to assess current deep learning methods for soly. prediction, develop a general model capable of predicting the soly. of a broad range of org. mols., and to understand the impact of data properties, mol. representation, and modeling architecture on predictive performance. Using the largest currently available soly. data set, we implement deep learning-based models to predict soly. from the mol. structure and explore several different mol. representations including mol. descriptors, simplified mol.-input line-entry system strings, mol. graphs, and three-dimensional at. coordinates using four different neural network architectures-fully connected neural networks, recurrent neural networks, graph neural networks (GNNs), and SchNet. We find that models using mol. descriptors achieve the best performance, with GNN models also achieving good performance. We perform extensive error anal. to understand the mol. properties that influence model performance, perform feature anal. to understand which information about the mol. structure is most valuable for prediction, and perform a transfer learning and data size study to understand the impact of data availability on model performance.
    130. 130
      Wu, X.; Yu, L. EPSOL: sequence-based protein solubility prediction using multidimensional embedding. Bioinformatics 2021, 37, 4314– 4320,  DOI: 10.1093/bioinformatics/btab463
      There is no corresponding record for this reference.
    131. 131
      Krogh, A. What are artificial neural networks?. Nat. Biotechnol. 2008, 26, 195– 197,  DOI: 10.1038/nbt1386
      131
      What are artificial neural networks?
      Krogh, Anders
      Nature Biotechnology (2008), 26 (2), 195-197CODEN: NABIF9; ISSN:1087-0156. (Nature Publishing Group)
      A review. Artificial neural networks have been applied to problems ranging from speech recognition to prediction of protein secondary structure, classification of cancers and gene prediction. How do they work and what might they be good for.
    132. 132
      Rumelhart, D.; Hinton, G. E.; Williams, R. J. Learning internal representations by error propagation. cmapspublic2.ihmc.us 1986, 673– 695
      There is no corresponding record for this reference.
    133. 133
      Salehinejad, H.; Sankar, S.; Barfett, J.; Colak, E.; Valaee, S. Recent Advances in Recurrent Neural Networks. arXiv , 2017.
      There is no corresponding record for this reference.
    134. 134
      Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, L.; Polosukhin, I. Attention is all you need. Adv. Neural Inf. Process. Syst. 2017, 30
      There is no corresponding record for this reference.
    135. 135
      Sherstinsky, A. Fundamentals of recurrent neural network (RNN) and long short-term memory (LSTM) network. arXiv , 2018.
      There is no corresponding record for this reference.
    136. 136
      Chen, G. A gentle tutorial of recurrent neural network with error backpropagation. arXiv , 2016.
      There is no corresponding record for this reference.
    137. 137
      Cho, K.; van Merrienboer, B.; Gulcehre, C.; Bahdanau, D.; Bougares, F.; Schwenk, H.; Bengio, Y. Learning Phrase Representations using RNN Encoder-Decoder for Statistical Machine Translation. arXiv , 2014.
      There is no corresponding record for this reference.
    138. 138
      Hochreiter, S.; Schmidhuber, J. Long short-term memory. Neural Comput. 1997, 9, 1735– 1780,  DOI: 10.1162/neco.1997.9.8.1735
      138
      Long short-term memory
      Hochreiter S; Schmidhuber J
      Neural computation (1997), 9 (8), 1735-80 ISSN:0899-7667.
      Learning to store information over extended time intervals by recurrent backpropagation takes a very long time, mostly because of insufficient, decaying error backflow. We briefly review Hochreiter's (1991) analysis of this problem, then address it by introducing a novel, efficient, gradient-based method called long short-term memory (LSTM). Truncating the gradient where this does not do harm, LSTM can learn to bridge minimal time lags in excess of 1000 discrete-time steps by enforcing constant error flow through constant error carousels within special units. Multiplicative gate units learn to open and close access to the constant error flow. LSTM is local in space and time; its computational complexity per time step and weight is O(1). Our experiments with artificial data involve local, distributed, real-valued, and noisy pattern representations. In comparisons with real-time recurrent learning, back propagation through time, recurrent cascade correlation, Elman nets, and neural sequence chunking, LSTM leads to many more successful runs, and learns much faster. LSTM also solves complex, artificial long-time-lag tasks that have never been solved by previous recurrent network algorithms.
    139. 139
      Graves, A.; Schmidhuber, J. Framewise phoneme classification with bidirectional LSTM and other neural network architectures. Neural Netw. 2005, 18, 602– 610,  DOI: 10.1016/j.neunet.2005.06.042
      139
      Framewise phoneme classification with bidirectional LSTM and other neural network architectures
      Graves Alex; Schmidhuber Jurgen
      Neural networks : the official journal of the International Neural Network Society (2005), 18 (5-6), 602-10 ISSN:0893-6080.
      In this paper, we present bidirectional Long Short Term Memory (LSTM) networks, and a modified, full gradient version of the LSTM learning algorithm. We evaluate Bidirectional LSTM (BLSTM) and several other network architectures on the benchmark task of framewise phoneme classification, using the TIMIT database. Our main findings are that bidirectional networks outperform unidirectional ones, and Long Short Term Memory (LSTM) is much faster and also more accurate than both standard Recurrent Neural Nets (RNNs) and time-windowed Multilayer Perceptrons (MLPs). Our results support the view that contextual information is crucial to speech processing, and suggest that BLSTM is an effective architecture with which to exploit it.
    140. 140
      Thafar, M. A.; Alshahrani, M.; Albaradei, S.; Gojobori, T.; Essack, M.; Gao, X. Affinity2Vec: drug-target binding affinity prediction through representation learning, graph mining, and machine learning. Sci. Rep. 2022, 12, 4751,  DOI: 10.1038/s41598-022-08787-9
      140
      Affinity2Vec: drug-target binding affinity prediction through representation learning, graph mining, and machine learning
      Thafar, Maha A.; Alshahrani, Mona; Albaradei, Somayah; Gojobori, Takashi; Essack, Magbubah; Gao, Xin
      Scientific Reports (2022), 12 (1), 4751CODEN: SRCEC3; ISSN:2045-2322. (Nature Portfolio)
      Abstr.: Drug-target interaction (DTI) prediction plays a crucial role in drug repositioning and virtual drug screening. Most DTI prediction methods cast the problem as a binary classification task to predict if interactions exist or as a regression task to predict continuous values that indicate a drug's ability to bind to a specific target. The regression-based methods provide insight beyond the binary relationship. However, most of these methods require the three-dimensional (3D) structural information of targets which are still not generally available to the targets. Despite this bottleneck, only a few methods address the drug-target binding affinity (DTBA) problem from a non-structure-based approach to avoid the 3D structure limitations. Here we propose Affinity2Vec, as a novel regression-based method that formulates the entire task as a graph-based problem. To develop this method, we constructed a weighted heterogeneous graph that integrates data from several sources, including drug-drug similarity, target-target similarity, and drug-target binding affinities. Affinity2Vec further combines several computational techniques from feature representation learning, graph mining, and machine learning to generate or ext. features, build the model, and predict the binding affinity between the drug and the target with no 3D structural data. We conducted extensive expts. to evaluate and demonstrate the robustness and efficiency of the proposed method on benchmark datasets used in state-of-the-art non-structured-based drug-target binding affinity studies. Affinity2Vec showed superior and competitive results compared to the state-of-the-art methods based on several evaluation metrics, including mean squared error, rm2, concordance index, and area under the precision-recall curve.
    141. 141
      Wei, B.; Zhang, Y.; Gong, X. 519. DeepLPI: A Novel Drug Repurposing Model based on Ligand-Protein Interaction Using Deep Learning. Open Forum Infect. Dis. 2022, 9, ofac492.574,  DOI: 10.1093/ofid/ofac492.574
      There is no corresponding record for this reference.
    142. 142
      Yuan, W.; Chen, G.; Chen, C. Y.-C. FusionDTA: attention-based feature polymerizer and knowledge distillation for drug-target binding affinity prediction. Brief. Bioinform. 2022,  DOI: 10.1093/bib/bbab506
      There is no corresponding record for this reference.
    143. 143
      West-Roberts, J.; Valentin-Alvarado, L.; Mullen, S.; Sachdeva, R.; Smith, J.; Hug, L. A.; Gregoire, D. S.; Liu, W.; Lin, T.-Y.; Husain, G.; Amano, Y.; Ly, L.; Banfield, J. F. Giant genes are rare but implicated in cell wall degradation by predatory bacteria. bioRxiv , 2023.
      There is no corresponding record for this reference.
    144. 144
      Hernández, A.; Amigó, J. Attention mechanisms and their applications to complex systems. Entropy (Basel) 2021, 23, 283,  DOI: 10.3390/e23030283
      There is no corresponding record for this reference.
    145. 145
      Yang, X. An overview of the attention mechanisms in computer vision. 2020.
      There is no corresponding record for this reference.
    146. 146
      Hu, D. An introductory survey on attention mechanisms in NLP problems. arXiv , 2018.
      There is no corresponding record for this reference.
    147. 147
      Vig, J., Madani, A., Varshney, L. R., Xiong, C., Socher, R., Rajani, N. F. BERTology Meets Biology: Interpreting Attention in Protein Language Models. arXiv , 2020.
      There is no corresponding record for this reference.
    148. 148
      Wu, H.; Liu, J.; Jiang, T.; Zou, Q.; Qi, S.; Cui, Z.; Tiwari, P.; Ding, Y. AttentionMGT-DTA: A multi-modal drug-target affinity prediction using graph transformer and attention mechanism. Neural Netw. 2024, 169, 623– 636,  DOI: 10.1016/j.neunet.2023.11.018
      There is no corresponding record for this reference.
    149. 149
      Koyama, K.; Kamiya, K.; Shimada, K. Cross attention dti: Drug-target interaction prediction with cross attention module in the blind evaluation setup. BIOKDD2020 2020.
      There is no corresponding record for this reference.
    150. 150
      Kurata, H.; Tsukiyama, S. ICAN: Interpretable cross-attention network for identifying drug and target protein interactions. PLoS One 2022, 17, e0276609,  DOI: 10.1371/journal.pone.0276609
      There is no corresponding record for this reference.
    151. 151
      Zhao, Q.; Zhao, H.; Zheng, K.; Wang, J. HyperAttentionDTI: improving drug–protein interaction prediction by sequence-based deep learning with attention mechanism. Bioinformatics 2022, 38, 655– 662,  DOI: 10.1093/bioinformatics/btab715
      151
      HyperAttentionDTI: improving drug-protein interaction prediction by sequence-based deep learning with attention mechanism
      Zhao, Qichang; Zhao, Haochen; Zheng, Kai; Wang, Jianxin
      Bioinformatics (2022), 38 (3), 655-662CODEN: BOINFP; ISSN:1367-4811. (Oxford University Press)
      Motivation: Identifying drug-target interactions (DTIs) is a crucial step in drug repurposing and drug discovery. Accurately identifying DTIs in silico can significantly shorten development time and reduce costs. Recently, many sequence-based methods are proposed for DTI prediction and improve performance by introducing the attention mechanism. However, these methods only model single non-covalent inter-mol. interactions among drugs and proteins and ignore the complex interaction between atoms and amino acids. Results: In this article, we propose an end-to-end bio-inspired model based on the convolutional neural network (CNN) and attention mechanism, named HyperAttentionDTI, for predicting DTIs. We use deep CNNs to learn the feature matrixes of drugs and proteins. To model complex non-covalent inter-mol. interactions among atoms and amino acids, we utilize the attention mechanism on the feature matrixes and assign an attention vector to each atom or amino acid. We evaluate HpyerAttentionDTI on three benchmark datasets and the results show that our model achieves significantly improved performance compared with the state-of-the-art baselines. Moreover, a case study on the human Gamma-aminobutyric acid receptors confirm that our model can be used as a powerful tool to predict DTIs.
    152. 152
      Jiang, M.; Li, Z.; Zhang, S.; Wang, S.; Wang, X.; Yuan, Q. Drug-target affinity prediction using graph neural network and contact maps. RSC Adv. 2020, 10, 20701,  DOI: 10.1039/D0RA02297G
      152
      Drug-target affinity prediction using graph neural network and contact maps
      Jiang, Mingjian; Li, Zhen; Zhang, Shugang; Wang, Shuang; Wang, Xiaofeng; Yuan, Qing; Wei, Zhiqiang
      RSC Advances (2020), 10 (35), 20701-20712CODEN: RSCACL; ISSN:2046-2069. (Royal Society of Chemistry)
      Computer-aided drug design uses high-performance computers to simulate the tasks in drug design, which is a promising research area. Drug-target affinity (DTA) prediction is the most important step of computer-aided drug design, which could speed up drug development and reduce resource consumption. With the development of deep learning, the introduction of deep learning to DTA prediction and improving the accuracy have become a focus of research. In this paper, utilizing the structural information of mols. and proteins, two graphs of drug mols. and proteins are built up resp. Graph neural networks are introduced to obtain their representations, and a method called DGraphDTA is proposed for DTA prediction. Specifically, the protein graph is constructed based on the contact map output from the prediction method, which could predict the structural characteristics of the protein according to its sequence. It can be seen from the test of various metrics on benchmark datasets that the method proposed in this paper has strong robustness and generalizability.
    153. 153
      Nguyen, T. M.; Nguyen, T.; Le, T. M.; Tran, T. GEFA: Early Fusion Approach in Drug-Target Affinity Prediction. IEEE/ACM Trans. Comput. Biol. Bioinform. 2022, 19, 718– 728,  DOI: 10.1109/TCBB.2021.3094217
      153
      GEFA: early fusion approach in drug-target affinity prediction
      Nguyen, Tri Minh; Nguyen, Thin; Le, Thao Minh; Tran, Truyen
      IEEE/ACM Transactions on Computational Biology and Bioinformatics (2022), 19 (2), 718-728CODEN: ITCBCY; ISSN:1557-9964. (Institute of Electrical and Electronics Engineers)
      Predicting the interaction between a compd. and a target is crucial for rapid drug repurposing. Deep learning has been successfully applied in drug-target affinity (DTA) problem. However, previous deep learning-based methods ignore modeling the direct interactions between drug and protein residues. This would lead to inaccurate learning of target representation which may change due to the drug binding effects. In addn., previous DTA methods learn protein representation solely based on a small no. of protein sequences in DTA datasets while neglecting the use of proteins outside of the DTA datasets. We propose GEFA (Graph Early Fusion Affinity), a novel graph-in-graph neural network with attention mechanism to address the changes in target representation because of the binding effects. Specifically, a drug is modeled as a graph of atoms, which then serves as a node in a larger graph of residues-drug complex. The resulting model is an expressive deep nested graph neural network. We also use pre-trained protein representation powered by the recent effort of learning contextualized protein representation. The expts. are conducted under different settings to evaluate scenarios such as novel drugs or targets. The results demonstrate the effectiveness of the pre-trained protein embedding and the advantages our GEFA in modeling the nested graph for drug-target interaction.
    154. 154
      Yu, J.; Li, Z.; Chen, G.; Kong, X.; Hu, J.; Wang, D.; Cao, D.; Li, Y.; Huo, R.; Wang, G.; Liu, X.; Jiang, H.; Li, X.; Luo, X.; Zheng, M. Computing the relative binding affinity of ligands based on a pairwise binding comparison network. Nature Computational Science 2023, 3, 860– 872,  DOI: 10.1038/s43588-023-00529-9
      There is no corresponding record for this reference.
    155. 155
      Knutson, C.; Bontha, M.; Bilbrey, J. A.; Kumar, N. Decoding the protein–ligand interactions using parallel graph neural networks. Sci. Rep. 2022, 12, 1– 14,  DOI: 10.1038/s41598-022-10418-2
      There is no corresponding record for this reference.
    156. 156
      Kyro, G. W.; Brent, R. I.; Batista, V. S. HAC-Net: A Hybrid Attention-Based Convolutional Neural Network for Highly Accurate Protein–Ligand Binding Affinity Prediction. J. Chem. Inf. Model. 2023, 63, 1947– 1960,  DOI: 10.1021/acs.jcim.3c00251
      There is no corresponding record for this reference.
    157. 157
      Yousefi, N.; Yazdani-Jahromi, M.; Tayebi, A.; Kolanthai, E.; Neal, C. J.; Banerjee, T.; Gosai, A.; Balasubramanian, G.; Seal, S.; Ozmen Garibay, O. BindingSite-AugmentedDTA: enabling a next-generation pipeline for interpretable prediction models in drug repurposing. Brief. Bioinform. 2023,  DOI: 10.1093/bib/bbad136
      There is no corresponding record for this reference.
    158. 158
      Yazdani-Jahromi, M.; Yousefi, N.; Tayebi, A.; Kolanthai, E.; Neal, C. J.; Seal, S.; Garibay, O. O. AttentionSiteDTI: an interpretable graph-based model for drug-target interaction prediction using NLP sentence-level relation classification. Brief. Bioinform. 2022,  DOI: 10.1093/bib/bbac272
      There is no corresponding record for this reference.
    159. 159
      Bronstein, M. M.; Bruna, J.; Cohen, T.; Veličković, P. Geometric deep learning: Grids, groups, graphs, geodesics, and gauges. arXiv , 2021.
      There is no corresponding record for this reference.
    160. 160
      Lim, J.; Ryu, S.; Park, K.; Choe, Y. J.; Ham, J.; Kim, W. Y. Predicting Drug–Target Interaction Using a Novel Graph Neural Network with 3D Structure-Embedded Graph Representation. J. Chem. Inf. Model. 2019, 59, 3981– 3988,  DOI: 10.1021/acs.jcim.9b00387
      160
      Predicting Drug-Target Interaction Using a Novel Graph Neural Network with 3D Structure-Embedded Graph Representation
      Lim, Jaechang; Ryu, Seongok; Park, Kyubyong; Choe, Yo Joong; Ham, Jiyeon; Kim, Woo Youn
      Journal of Chemical Information and Modeling (2019), 59 (9), 3981-3988CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      We propose a novel deep learning approach for predicting drug-target interaction using a graph neural network. We introduce a distance-aware graph attention algorithm to differentiate various types of intermol. interactions. Furthermore, we ext. the graph feature of intermol. interactions directly from the 3D structural information on the protein-ligand binding pose. Thus, the model can learn key features for accurate predictions of drug-target interaction rather than just memorize certain patterns of ligand mols. As a result, our model shows better performance than docking and other deep learning methods for both virtual screening (AUROC of 0.968 for the DUD-E test set) and pose prediction (AUROC of 0.935 for the PDBbind test set). In addn., it can reproduce the natural population distribution of active mols. and inactive mols.
    161. 161
      Jin, Z.; Wu, T.; Chen, T.; Pan, D.; Wang, X.; Xie, J.; Quan, L.; Lyu, Q. CAPLA: improved prediction of protein–ligand binding affinity by a deep learning approach based on a cross-attention mechanism. Bioinformatics 2023, 39, btad049,  DOI: 10.1093/bioinformatics/btad049
      There is no corresponding record for this reference.
    162. 162
      Lin, Z.; Akin, H.; Rao, R.; Hie, B.; Zhu, Z.; Lu, W.; Smetanin, N.; Verkuil, R.; Kabeli, O.; Shmueli, Y.; Dos Santos Costa, A.; Fazel-Zarandi, M.; Sercu, T.; Candido, S.; Rives, A. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science 2023, 379, 1123– 1130,  DOI: 10.1126/science.ade2574
      162
      Evolutionary-scale prediction of atomic-level protein structure with a language model
      Lin, Zeming; Akin, Halil; Rao, Roshan; Hie, Brian; Zhu, Zhongkai; Lu, Wenting; Smetanin, Nikita; Verkuil, Robert; Kabeli, Ori; Shmueli, Yaniv; dos Santos Costa, Allan; Fazel-Zarandi, Maryam; Sercu, Tom; Candido, Salvatore; Rives, Alexander
      Science (Washington, DC, United States) (2023), 379 (6637), 1123-1130CODEN: SCIEAS; ISSN:1095-9203. (American Association for the Advancement of Science)
      Recent advances in machine learning have leveraged evolutionary information in multiple sequence alignments to predict protein structure. We demonstrate direct inference of full at.-level protein structure from primary sequence using a large language model. As language models of protein sequences are scaled up to 15 billion parameters, an at.-resoln. picture of protein structure emerges in the learned representations. This results in an order-of-magnitude acceleration of high-resoln. structure prediction, which enables large-scale structural characterization of metagenomic proteins. We apply this capability to construct the ESM Metagenomic Atlas by predicting structures for >617 million metagenomic protein sequences, including >225 million that are predicted with high confidence, which gives a view into the vast breadth and diversity of natural proteins.
    163. 163
      Zhang, S.; Fan, R.; Liu, Y.; Chen, S.; Liu, Q.; Zeng, W. Applications of transformer-based language models in bioinformatics: a survey. Bioinform. Adv. 2023, 3, vbad001,  DOI: 10.1093/bioadv/vbad001
      There is no corresponding record for this reference.
    164. 164
      Bahdanau, D.; Cho, K.; Bengio, Y. Neural Machine Translation by Jointly Learning to Align and Translate. arXiv , 2014.
      There is no corresponding record for this reference.
    165. 165
      Vinyals, O.; Toshev, A.; Bengio, S.; Erhan, D. Show and tell: A neural image caption generator. Pattern Recognition (CVPR) 2015, 3156– 3164
      There is no corresponding record for this reference.
    166. 166
      Sutskever, I.; Vinyals, O.; Le, Q. V. Sequence to sequence learning with Neural Networks. arXiv , 2014;.
      There is no corresponding record for this reference.
    167. 167
      Cho, K.; van Merrienboer, B.; Bahdanau, D.; Bengio, Y. On the properties of neural machine translation: Encoder-decoder approaches. arXiv , 2014.
      There is no corresponding record for this reference.
    168. 168
      Devlin, J.; Chang, M.-W.; Lee, K.; Toutanova, K. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. arXiv , 2018.
      There is no corresponding record for this reference.
    169. 169
      Zeyer, A.; Bahar, P.; Irie, K.; Schlüter, R.; Ney, H. A Comparison of Transformer and LSTM Encoder Decoder Models for ASR. 2019 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU) 2019, 8– 15
      There is no corresponding record for this reference.
    170. 170
      Irie, K.; Zeyer, A.; Schlüter, R.; Ney, H. Language Modeling with Deep Transformers. arXiv , 2019.
      There is no corresponding record for this reference.
    171. 171
      Zouitni, C.; Sabri, M. A.; Aarab, A. A Comparison Between LSTM and Transformers for Image Captioning. Digital Technologies and Applications 2023, 669, 492– 500,  DOI: 10.1007/978-3-031-29860-8_50
      There is no corresponding record for this reference.
    172. 172
      Parisotto, E.; Song, F.; Rae, J.; Pascanu, R.; Gulcehre, C.; Jayakumar, S.; Jaderberg, M.; Kaufman, R. L.; Clark, A.; Noury, S.; Botvinick, M.; Heess, N.; Hadsell, R. Stabilizing Transformers for Reinforcement Learning. Proceedings of the 37th International Conference on Machine Learning 2020, 7487– 7498
      There is no corresponding record for this reference.
    173. 173
      Bilokon, P.; Qiu, Y. Transformers versus LSTMs for electronic trading. arXiv , 2023.
      There is no corresponding record for this reference.
    174. 174
      Merity, S. Single Headed Attention RNN: Stop Thinking With Your Head. arXiv , 2019.
      There is no corresponding record for this reference.
    175. 175
      Ezen-Can, A. A Comparison of LSTM and BERT for Small Corpus. arXiv , 2020.
      There is no corresponding record for this reference.
    176. 176
      Unsal, S.; Atas, H.; Albayrak, M.; Turhan, K.; Acar, A. C.; Doğan, T. Learning functional properties of proteins with language models. Nature Machine Intelligence 2022, 4, 227– 245,  DOI: 10.1038/s42256-022-00457-9
      There is no corresponding record for this reference.
    177. 177
      Brandes, N.; Ofer, D.; Peleg, Y.; Rappoport, N.; Linial, M. ProteinBERT: a universal deep-learning model of protein sequence and function. Bioinformatics 2022, 38, 2102– 2110,  DOI: 10.1093/bioinformatics/btac020
      177
      ProteinBERT: a universal deep-learning model of protein sequence and function
      Brandes, Nadav; Ofer, Dan; Peleg, Yam; Rappoport, Nadav; Linial, Michal
      Bioinformatics (2022), 38 (8), 2102-2110CODEN: BOINFP; ISSN:1367-4811. (Oxford University Press)
      Self-supervised deep language modeling has shown unprecedented success across natural language tasks, and has recently been repurposed to biol. sequences. However, existing models and pretraining methods are designed and optimized for text anal. We introduce ProteinBERT, a deep language model specifically designed for proteins. Our pretraining scheme combines language modeling with a novel task of Gene Ontol. (GO) annotation prediction. We introduce novel architectural elements that make the model highly efficient and flexible to long sequences. The architecture of ProteinBERT consists of both local and global representations, allowing end-to-end processing of these types of inputs and outputs. ProteinBERT obtains near state-of-the-art performance, and sometimes exceeds it, on multiple benchmarks covering diverse protein properties (including protein structure, post-translational modifications and biophys. attributes), despite using a far smaller and faster model than competing deep-learning methods. Overall, ProteinBERT provides an efficient framework for rapidly training protein predictors, even with limited labeled data.
    178. 178
      Luo, S.; Chen, T.; Xu, Y.; Zheng, S.; Liu, T.-Y.; Wang, L.; He, D. One Transformer Can Understand Both 2D & 3D Molecular Data. arXiv , 2022.
      There is no corresponding record for this reference.
    179. 179
      Clark, K.; Luong, M.-T.; Le, Q. V.; Manning, C. D. ELECTRA: Pre-training Text Encoders as Discriminators Rather Than Generators. arXiv , 2020.
      There is no corresponding record for this reference.
    180. 180
      Wang, J.; Wen, N.; Wang, C.; Zhao, L.; Cheng, L. ELECTRA-DTA: a new compound-protein binding affinity prediction model based on the contextualized sequence encoding. J. Cheminform. 2022, 14, 14,  DOI: 10.1186/s13321-022-00591-x
      There is no corresponding record for this reference.
    181. 181
      Shin, B.; Park, S.; Kang, K.; Ho, J. C. Self-Attention Based Molecule Representation for Predicting Drug-Target Interaction. Proceedings of the 4th Machine Learning for Healthcare Conference 2019, 230– 248
      There is no corresponding record for this reference.
    182. 182
      Huang, K.; Xiao, C.; Glass, L. M.; Sun, J. MolTrans: Molecular Interaction Transformer for drug–target interaction prediction. Bioinformatics 2021, 37, 830– 836,  DOI: 10.1093/bioinformatics/btaa880
      182
      MolTrans: molecular interaction transformer for drug-target interaction prediction
      Huang, Kexin; Xiao, Cao; Glass, Lucas M.; Sun, Jimeng
      Bioinformatics (2021), 37 (6), 830-836CODEN: BOINFP; ISSN:1367-4811. (Oxford University Press)
      Motivation: Drug-target interaction (DTI) prediction is a foundational task for in-silico drug discovery, which is costly and time-consuming due to the need of exptl. search over large drug compd. space. Recent years have witnessed promising progress for deep learning in DTI predictions. However, the following challenges are still open: (i) existing mol. representation learning approaches ignore the sub-structural nature of DTI, thus produce results that are less accurate and difficult to explain and (ii) existing methods focus on limited labeled data while ignoring the value of massive unlabeled mol. data. Results: We propose a Mol. Interaction Transformer (MolTrans) to address these limitations via: (i) knowledge inspired sub-structural pattern mining algorithm and interaction modeling module for more accurate and interpretable DTI prediction and (ii) an augmented transformer encoder to better ext. and capture the semantic relations among sub-structures extd. from massive unlabeled biomedical data. We evaluate MolTrans on real-world data and show it improved DTI prediction performance compared to state-of-the-art baselines.
    183. 183
      Shen, L.; Feng, H.; Qiu, Y.; Wei, G.-W. SVSBI: sequence-based virtual screening of biomolecular interactions. Commun. Biol. 2023, 6, 536,  DOI: 10.1038/s42003-023-04866-3
      183
      SVSBI: sequence-based virtual screening of biomolecular interactions
      Shen Li; Feng Hongsong; Qiu Yuchi; Wei Guo-Wei; Wei Guo-Wei; Wei Guo-Wei
      Communications biology (2023), 6 (1), 536 ISSN:.
      Virtual screening (VS) is a critical technique in understanding biomolecular interactions, particularly in drug design and discovery. However, the accuracy of current VS models heavily relies on three-dimensional (3D) structures obtained through molecular docking, which is often unreliable due to the low accuracy. To address this issue, we introduce a sequence-based virtual screening (SVS) as another generation of VS models that utilize advanced natural language processing (NLP) algorithms and optimized deep K-embedding strategies to encode biomolecular interactions without relying on 3D structure-based docking. We demonstrate that SVS outperforms state-of-the-art performance for four regression datasets involving protein-ligand binding, protein-protein, protein-nucleic acid binding, and ligand inhibition of protein-protein interactions and five classification datasets for protein-protein interactions in five biological species. SVS has the potential to transform current practices in drug discovery and protein engineering.
    184. 184
      Wang, J.; Hu, J.; Sun, H.; Xu, M.; Yu, Y.; Liu, Y.; Cheng, L. MGPLI: exploring multigranular representations for protein–ligand interaction prediction. Bioinformatics 2022, 38, 4859– 4867,  DOI: 10.1093/bioinformatics/btac597
      There is no corresponding record for this reference.
    185. 185
      Qian, Y.; Wu, J.; Zhang, Q. CAT-CPI: Combining CNN and transformer to learn compound image features for predicting compound-protein interactions. Front Mol. Biosci 2022, 9, 963912,  DOI: 10.3389/fmolb.2022.963912
      There is no corresponding record for this reference.
    186. 186
      Cang, Z.; Mu, L.; Wei, G.-W. Representability of algebraic topology for biomolecules in machine learning based scoring and virtual screening. PLoS Comput. Biol. 2018, 14, e1005929,  DOI: 10.1371/journal.pcbi.1005929
      186
      Representability of algebraic topology for biomolecules in machine learning based scoring and virtual screening
      Cang, Zixuan; Mu, Lin; Wei, Guo-Wei
      PLoS Computational Biology (2018), 14 (1), e1005929/1-e1005929/44CODEN: PCBLBG; ISSN:1553-7358. (Public Library of Science)
      This work introduces a no. of algebraic topol. approaches, including multi-component persistent homol., multi-level persistent homol., and electrostatic persistence for the representation, characterization, and description of small mols. and biomol. complexes. In contrast to the conventional persistent homol., multi-component persistent homol. retains crit. chem. and biol. information during the topol. simplification of biomol. geometric complexity. Multi-level persistent homol. enables a tailored topol. description of inter- and/or intra-mol. interactions of interest. Electrostatic persistence incorporates partial charge information into topol. invariants. These topol. methods are paired with Wasserstein distance to characterize similarities between mols. and are further integrated with a variety of machine learning algorithms, including k-nearest neighbors, ensemble of trees, and deep convolutional neural networks, to manifest their descriptive and predictive powers for protein-ligand binding anal. and virtual screening of small mols. Extensive numerical expts. involving 4,414 protein- ligand complexes from the PDBBind database and 128,374 ligand-target and decoytarget pairs in the DUD database are performed to test resp. the scoring power and the discriminatory power of the proposed topol. learning strategies. It is demonstrated that the present topol. learning outperforms other existing methods in protein-ligand binding affinity prediction and ligand-decoy discrimination.
    187. 187
      Chen, D.; Liu, J.; Wei, G.-W. Multiscale topology-enabled structure-to-sequence transformer for protein-ligand interaction predictions. Nat. Mac. Intell. 2024, 6, 799– 810,  DOI: 10.1038/s42256-024-00855-1
      There is no corresponding record for this reference.
    188. 188
      Bubeck, S.; Chandrasekaran, V.; Eldan, R.; Gehrke, J.; Horvitz, E.; Kamar, E.; Lee, P.; Lee, Y. T.; Li, Y.; Lundberg, S.; Nori, H.; Palangi, H.; Ribeiro, M. T.; Zhang, Y. Sparks of Artificial General Intelligence: Early experiments with GPT-4. arXiv , 2023.
      There is no corresponding record for this reference.
    189. 189
      Elnaggar, A.; Essam, H.; Salah-Eldin, W.; Moustafa, W.; Elkerdawy, M.; Rochereau, C.; Rost, B. Ankh: Optimized protein language model unlocks general-purpose modelling. bioRxiv , 2023.
      There is no corresponding record for this reference.
    190. 190
      Hwang, Y.; Cornman, A. L.; Kellogg, E. H.; Ovchinnikov, S.; Girguis, P. R. Genomic language model predicts protein co-regulation and function. Nat. Commun. 2024, 15, 2880,  DOI: 10.1038/s41467-024-46947-9
      There is no corresponding record for this reference.
    191. 191
      Vu, M. H.; Akbar, R.; Robert, P. A.; Swiatczak, B.; Greiff, V.; Sandve, G. K.; Haug, D. T. T. Linguistically inspired roadmap for building biologically reliable protein language models. arXiv , 2022.
      There is no corresponding record for this reference.
    192. 192
      Xu, M.; Zhang, Z.; Lu, J.; Zhu, Z.; Zhang, Y.; Ma, C.; Liu, R.; Tang, J. PEER: A comprehensive and multi-task benchmark for Protein sEquence undERstanding. arXiv 2022, 35156– 35173.
      There is no corresponding record for this reference.
    193. 193
      Schmirler, R.; Heinzinger, M.; Rost, B. Fine-tuning protein language models boosts predictions across diverse tasks. Nat. Commun. 2024, 15, 7407,  DOI: 10.1038/s41467-024-51844-2
      There is no corresponding record for this reference.
    194. 194
      Heinzinger, M.; Elnaggar, A.; Wang, Y.; Dallago, C.; Nechaev, D.; Matthes, F.; Rost, B. Modeling aspects of the language of life through transfer-learning protein sequences. BMC Bioinformatics 2019, 20, 723,  DOI: 10.1186/s12859-019-3220-8
      There is no corresponding record for this reference.
    195. 195
      Manfredi, M.; Savojardo, C.; Martelli, P. L.; Casadio, R. E-SNPs&GO: embedding of protein sequence and function improves the annotation of human pathogenic variants. Bioinformatics 2022, 38, 5168– 5174,  DOI: 10.1093/bioinformatics/btac678
      There is no corresponding record for this reference.
    196. 196
      Anteghini, M.; Martins Dos Santos, V.; Saccenti, E. In-Pero: Exploiting deep learning embeddings of protein sequences to predict the localisation of peroxisomal proteins. Int. J. Mol. Sci. 2021, 22, 6409,  DOI: 10.3390/ijms22126409
      There is no corresponding record for this reference.
    197. 197
      Nguyen, T.; Le, H.; Quinn, T. P.; Nguyen, T.; Le, T. D.; Venkatesh, S. GraphDTA: predicting drug–target binding affinity with graph neural networks. Bioinformatics 2021, 37, 1140– 1147,  DOI: 10.1093/bioinformatics/btaa921
      197
      GraphDTA: predicting drug-target binding affinity with graph neural networks
      Nguyen, Thin; Le, Hang; Quinn, Thomas P.; Nguyen, Tri; Le, Thuc Duy; Venkatesh, Svetha
      Bioinformatics (2021), 37 (8), 1140-1147CODEN: BOINFP; ISSN:1367-4811. (Oxford University Press)
      Summary: The development of new drugs is costly, time consuming, often accompanied with safety issues. Drug repurposing can avoid the expensive, lengthy process of drug development by finding new uses for already approved drugs. In order to repurpose drugs effectively, it is useful to know which proteins are targeted by which drugs. Computational models that est. the interaction strength of new drug-target pairs have the potential to expedite drug repurposing. Several models have been proposed for this task. However, these models represent the drugs as strings, which is not a natural way to represent mols. We propose a new model called GraphDTA that represents drugs as graphs, uses graph neural networks to predict drug-target affinity. We show that graph neural networks not only predict drug-target affinity better than non-deep learning models, but also outperform competing deep learning methods. Our results confirm that deep learning models are appropriate for drug- target binding affinity prediction, that representing drugs as graphs can lead to further improvements.
    198. 198
      Nam, H.; Ha, J.-W.; Kim, J. Dual attention networks for multimodal reasoning and matching. Pattern Recognition (CVPR) 2017, 299– 307
      There is no corresponding record for this reference.
    199. 199
      Wang, X.; Liu, D.; Zhu, J.; Rodriguez-Paton, A.; Song, T. CSConv2d: A 2-D Structural Convolution Neural Network with a Channel and Spatial Attention Mechanism for Protein-Ligand Binding Affinity Prediction. Biomolecules 2021,  DOI: 10.3390/biom11050643
      There is no corresponding record for this reference.
    200. 200
      Anteghini, M.; Santos, V. A. M. D.; Saccenti, E. PortPred: Exploiting deep learning embeddings of amino acid sequences for the identification of transporter proteins and their substrates. J. Cell. Biochem. 2023, 124, 1803,  DOI: 10.1002/jcb.30490
      There is no corresponding record for this reference.
    201. 201
      Huang, K.; Fu, T.; Glass, L. M.; Zitnik, M.; Xiao, C.; Sun, J. DeepPurpose: a deep learning library for drug–target interaction prediction. Bioinformatics 2021, 36, 5545– 5547,  DOI: 10.1093/bioinformatics/btaa1005
      201
      DeepPurpose: a deep learning library for drug-target interaction prediction
      Huang Kexin; Zitnik Marinka; Fu Tianfan; Glass Lucas M; Xiao Cao; Sun Jimeng
      Bioinformatics (Oxford, England) (2021), 36 (22-23), 5545-5547 ISSN:.
      SUMMARY: Accurate prediction of drug-target interactions (DTI) is crucial for drug discovery. Recently, deep learning (DL) models for show promising performance for DTI prediction. However, these models can be difficult to use for both computer scientists entering the biomedical field and bioinformaticians with limited DL experience. We present DeepPurpose, a comprehensive and easy-to-use DL library for DTI prediction. DeepPurpose supports training of customized DTI prediction models by implementing 15 compound and protein encoders and over 50 neural architectures, along with providing many other useful features. We demonstrate state-of-the-art performance of DeepPurpose on several benchmark datasets. AVAILABILITY AND IMPLEMENTATION: https://github.com/kexinhuang12345/DeepPurpose. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.
    202. 202
      Zhao, L.; Wang, J.; Pang, L.; Liu, Y.; Zhang, J. GANsDTA: Predicting Drug-Target Binding Affinity Using GANs. Front. Genet. 2020, 10, 1243,  DOI: 10.3389/fgene.2019.01243
      There is no corresponding record for this reference.
    203. 203
      Hu, F.; Jiang, J.; Wang, D.; Zhu, M.; Yin, P. Multi-PLI: interpretable multi-task deep learning model for unifying protein–ligand interaction datasets. J. Cheminform. 2021, 13, 30,  DOI: 10.1186/s13321-021-00510-6
      There is no corresponding record for this reference.
    204. 204
      Zheng, S.; Li, Y.; Chen, S.; Xu, J.; Yang, Y. Predicting Drug Protein Interaction using Quasi-Visual Question Answering System. bioRxiv 2019, 588178
      There is no corresponding record for this reference.
    205. 205
      Tsubaki, M.; Tomii, K.; Sese, J. Compound–protein interaction prediction with end-to-end learning of neural networks for graphs and sequences. Bioinformatics 2019, 35, 309– 318,  DOI: 10.1093/bioinformatics/bty535
      205
      Compound-protein interaction prediction with end-to-end learning of neural networks for graphs and sequences
      Tsubaki, Masashi; Tomii, Kentaro; Sese, Jun
      Bioinformatics (2019), 35 (2), 309-318CODEN: BOINFP; ISSN:1367-4811. (Oxford University Press)
      Motivation: In bioinformatics, machine learning-based methods that predict the compd.-protein interactions (CPIs) play an important role in the virtual screening for drug discovery. Recently, end-to-end representation learning for discrete symbolic data (e.g. words in natural language processing) using deep neural networks has demonstrated excellent performance on various difficult problems. For the CPI problem, data are provided as discrete symbolic data, i.e. compds. are represented as graphs where the vertices are atoms, the edges are chem. bonds, and proteins are sequences in which the characters are amino acids. In this study, we investigate the use of end-to-end representation learning for compds. and proteins, integrate the representations, and develop a new CPI prediction approach by combining a graph neural network (GNN) for compds. and a convolutional neural network (CNN) for proteins. Results: Our expts. using three CPI datasets demonstrated that the proposed end-to-end approach achieves competitive or higher performance as compared to various existing CPI prediction methods. In addn., the proposed approach significantly outperformed existing methods on an unbalanced dataset. This suggests that data-driven representations of compds. and proteins obtained by end-to-end GNNs and CNNs are more robust than traditional chem. and biol. features obtained from databases. Although analyzing deep learning models is difficult due to their black-box nature, we address this issue using a neural attention mechanism, which allows us to consider which subsequences in a protein are more important for a drug compd. when predicting its interaction. The neural attention mechanism also provides effective visualization, which makes it easier to analyze a model even when modeling is performed using real-valued representations instead of discrete features.
    206. 206
      Karimi, M.; Wu, D.; Wang, Z.; Shen, Y. DeepAffinity: interpretable deep learning of compound–protein affinity through unified recurrent and convolutional neural networks. Bioinformatics 2019, 35, 3329– 3338,  DOI: 10.1093/bioinformatics/btz111
      206
      DeepAffinity: interpretable deep learning of compound-protein affinity through unified recurrent and convolutional neural networks
      Karimi, Mostafa; Wu, Di; Wang, Zhangyang; Shen, Yang
      Bioinformatics (2019), 35 (18), 3329-3338CODEN: BOINFP; ISSN:1367-4811. (Oxford University Press)
      Motivation: Drug discovery demands rapid quantification of compd.-protein interaction (CPI). However, there is a lack of methods that can predict compd.-protein affinity from sequences alone with high applicability, accuracy and interpretability. Results: We present a seamless integration of domain knowledges and learning-based approaches. Under novel representations of structurally annotated protein sequences, a semisupervised deep learning model that unifies recurrent and convolutional neural networks has been proposed to exploit both unlabeled and labeled data, for jointly encoding mol. representations and predicting affinities. Our representations and models outperform conventional options in achieving relative error in IC50 within 5-fold for test cases and 20-fold for protein classes not included for training. Performances for new protein classes with few labeled data are further improved by transfer learning. Furthermore, sep. and joint attention mechanisms are developed and embedded to our model to add to its interpretability, as illustrated in case studies for predicting and explaining selective drug-target interactions. Lastly, alternative representations using protein sequences or compd. graphs and a unified RNN/GCNN-CNN model using graph CNN (GCNN) are also explored to reveal algorithmic challenges ahead.
    207. 207
      Li, S.; Wan, F.; Shu, H.; Jiang, T.; Zhao, D.; Zeng, J. MONN: a multi-objective neural network for predicting compound-protein interactions and affinities. Cell Systems 2020, 10, 308– 322,  DOI: 10.1016/j.cels.2020.03.002
      207
      MONN: A Multi-objective Neural Network for Predicting Compound-Protein Interactions and Affinities
      Li, Shuya; Wan, Fangping; Shu, Hantao; Jiang, Tao; Zhao, Dan; Zeng, Jianyang
      Cell Systems (2020), 10 (4), 308-322.e11CODEN: CSEYA4; ISSN:2405-4712. (Cell Press)
      Computational approaches for understanding compd.-protein interactions (CPIs) can greatly facilitate drug development. Recently, a no. of deep-learning-based methods have been proposed to predict binding affinities and attempt to capture local interaction sites in compds. and proteins through neural attentions (i.e., neural network architectures that enable the interpretation of feature importance). Here, we compiled a benchmark dataset contg. the inter-mol. non-covalent interactions for more than 10,000 compd.-protein pairs and systematically evaluated the interpretability of neural attentions in existing models. We also developed a multi-objective neural network, called MONN, to predict both non-covalent interactions and binding affinities between compds. and proteins. Comprehensive evaluation demonstrated that MONN can successfully predict the non-covalent interactions between compds. and proteins that cannot be effectively captured by neural attentions in previous prediction methods. Moreover, MONN outperforms other state-of-the-art methods in predicting binding affinities.
    208. 208
      Zhao, M.; Yuan, M.; Yang, Y.; Xu, S. X. CPGL: Prediction of Compound-Protein Interaction by Integrating Graph Attention Network With Long Short-Term Memory Neural Network. IEEE/ACM Trans. Comput. Biol. Bioinform. 2023, 20, 1935– 1942,  DOI: 10.1109/TCBB.2022.3225296
      There is no corresponding record for this reference.
    209. 209
      Yu, L.; Qiu, W.; Lin, W.; Cheng, X.; Xiao, X.; Dai, J. HGDTI: predicting drug–target interaction by using information aggregation based on heterogeneous graph neural network. BMC Bioinformatics 2022, 23, 126,  DOI: 10.1186/s12859-022-04655-5
      There is no corresponding record for this reference.
    210. 210
      Lee, I.; Nam, H. Sequence-based prediction of protein binding regions and drug-target interactions. J. Cheminform. 2022, 14, 5,  DOI: 10.1186/s13321-022-00584-w
      There is no corresponding record for this reference.
    211. 211
      Gönen, M.; Heller, G. Concordance probability and discriminatory power in proportional hazards regression. Biometrika 2005, 92, 965– 970,  DOI: 10.1093/biomet/92.4.965
      There is no corresponding record for this reference.
    212. 212
      Deller, M. C.; Rupp, B. Models of protein-ligand crystal structures: trust, but verify. J. Comput. Aided Mol. Des. 2015, 29, 817– 836,  DOI: 10.1007/s10822-015-9833-8
      212
      Models of protein-ligand crystal structures: trust, but verify
      Deller, Marc C.; Rupp, Bernhard
      Journal of Computer-Aided Molecular Design (2015), 29 (9), 817-836CODEN: JCADEQ; ISSN:0920-654X. (Springer)
      X-ray crystallog. provides the most accurate models of protein-ligand structures. These models serve as the foundation of many computational methods including structure prediction, mol. modeling, and structure-based drug design. The success of these computational methods ultimately depends on the quality of the underlying protein-ligand models. X-ray crystallog. offers the unparalleled advantage of a clear math. formalism relating the exptl. data to the protein-ligand model. In the case of X-ray crystallog., the primary exptl. evidence is the electron d. of the mols. forming the crystal. The first step in the generation of an accurate and precise crystallog. model is the interpretation of the electron d. of the crystal, typically carried out by construction of an at. model. The at. model must then be validated for fit to the exptl. electron d. and also for agreement with prior expectations of stereochem. Stringent validation of protein-ligand models has become possible as a result of the mandatory deposition of primary diffraction data, and many computational tools are now available to aid in the validation process. Validation of protein-ligand complexes has revealed some instances of overenthusiastic interpretation of ligand d. Fundamental concepts and metrics of protein-ligand quality validation are discussed and we highlight software tools to assist in this process. It is essential that end users select high quality protein-ligand models for their computational and biol. studies, and we provide an overview of how this can be achieved.
    213. 213
      Kalakoti, Y.; Yadav, S.; Sundar, D. TransDTI: Transformer-based language models for estimating DTIs and building a drug recommendation workflow. ACS Omega 2022, 7, 2706– 2717,  DOI: 10.1021/acsomega.1c05203
      213
      TransDTI: Transformer-Based Language Models for Estimating DTIs and Building a Drug Recommendation Workflow
      Kalakoti, Yogesh; Yadav, Shashank; Sundar, Durai
      ACS Omega (2022), 7 (3), 2706-2717CODEN: ACSODF; ISSN:2470-1343. (American Chemical Society)
      The identification of novel drug-target interactions is a labor-intensive and low-throughput process. In silico alternatives have proved to be of immense importance in assisting the drug discovery process. Here, we present TransDTI, a multiclass classification and regression workflow employing transformer-based language models to segregate interactions between drug-target pairs as active, inactive, and intermediate. The models were trained with large-scale drug-target interaction (DTI) data sets, which reported an improvement in performance in terms of the area under receiver operating characteristic (auROC), the area under precision recall (auPR), Matthew's correlation coeff. (MCC), and R2 over baseline methods. The results showed that models based on transformer-based language models effectively predict novel drug-target interactions from sequence data. The proposed models significantly outperformed existing methods like DeepConvDTI, DeepDTA, and DeepDTI on a test data set. Further, the validity of novel interactions predicted by TransDTI was found to be backed by mol. docking and simulation anal., where the model prediction had similar or better interaction potential for MAP2k and transforming growth factor-β (TGFβ) and their known inhibitors. Proposed approaches can have a significant impact on the development of personalized therapy and clin. decision making.
    214. 214
      Chatterjee, A.; Walters, R.; Shafi, Z.; Ahmed, O. S.; Sebek, M.; Gysi, D.; Yu, R.; Eliassi-Rad, T.; Barabási, A.-L.; Menichetti, G. Improving the generalizability of protein-ligand binding predictions with AI-Bind. Nat. Commun. 2023, 14, 1989,  DOI: 10.1038/s41467-023-37572-z
      There is no corresponding record for this reference.
    215. 215
      Gudivada, V.; Apon, A.; Ding, J. Data quality considerations for big data and machine learning: Going beyond data cleaning and transformations. Int. J. Adv. Softw. 2017, 10, 1– 20
      There is no corresponding record for this reference.
    216. 216
      Nasteski, V. An overview of the supervised machine learning methods. Horizons 2017, 4, 51– 62,  DOI: 10.20544/HORIZONS.B.04.1.17.P05
      There is no corresponding record for this reference.
    217. 217
      Kozlov, M. So you got a null result. Will anyone publish it?. Nature 2024, 631, 728– 730,  DOI: 10.1038/d41586-024-02383-9
      There is no corresponding record for this reference.
    218. 218
      Edfeldt, K. A data science roadmap for open science organizations engaged in early-stage drug discovery. Nat. Commun. 2024, 15, 5640,  DOI: 10.1038/s41467-024-49777-x
      There is no corresponding record for this reference.
    219. 219
      Mlinarić, A.; Horvat, M.; Šupak Smolčić, V. Dealing with the positive publication bias: Why you should really publish your negative results. Biochem. Med. 2017, 27, 030201,  DOI: 10.11613/BM.2017.030201
      There is no corresponding record for this reference.
    220. 220
      Fanelli, D. Negative results are disappearing from most disciplines and countries. Scientometrics 2012, 90, 891– 904,  DOI: 10.1007/s11192-011-0494-7
      There is no corresponding record for this reference.
    221. 221
      Albalate, A.; Minker, W. Semi-supervised and unervised machine learning: Novel strategies; Wiley-ISTE, 2013.
      There is no corresponding record for this reference.
    222. 222
      Sajadi, S. Z.; Zare Chahooki, M. A.; Gharaghani, S.; Abbasi, K. AutoDTI++: deep unsupervised learning for DTI prediction by autoencoders. BMC Bioinformatics 2021, 22, 204,  DOI: 10.1186/s12859-021-04127-2
      There is no corresponding record for this reference.
    223. 223
      Najm, M.; Azencott, C.-A.; Playe, B.; Stoven, V. Drug Target Identification with Machine Learning: How to Choose Negative Examples. Int. J. Mol. Sci. 2021, 22, 5118,  DOI: 10.3390/ijms22105118
      There is no corresponding record for this reference.
    224. 224
      Sieg, J.; Flachsenberg, F.; Rarey, M. In Need of Bias Control: Evaluating Chemical Data for Machine Learning in Structure-Based Virtual Screening. J. Chem. Inf. Model. 2019, 59, 947– 961,  DOI: 10.1021/acs.jcim.8b00712
      224
      In Need of Bias Control: Evaluating Chemical Data for Machine Learning in Structure-Based Virtual Screening
      Sieg, Jochen; Flachsenberg, Florian; Rarey, Matthias
      Journal of Chemical Information and Modeling (2019), 59 (3), 947-961CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      A review. Reports of successful applications of machine learning (ML) methods in structure-based virtual screening (SBVS) are increasing. ML methods such as convolutional neural networks show promising results and often outperform traditional methods such as empirical scoring functions in retrospective validation. However, trained ML models are often treated as black boxes and are not straightforwardly interpretable. In most cases, it is unknown which features in the data are decisive and whether a model's predictions are right for the right reason. Hence, the authors reevaluated three widely used benchmark data sets in the context of ML methods and came to the conclusion that not every benchmark data set is suitable. Moreover, the authors demonstrate on two examples from current literature that bias is learned implicitly and unnoticed from std. benchmarks. On the basis of these results, the authors conclude that there is a need for eligible validation expts. and benchmark data sets suited to ML for more bias-controlled validation in ML-based SBVS. Therefore, the authors provide guidelines for setting up validation expts. and give a perspective on how new data sets could be generated.
    225. 225
      Volkov, M.; Turk, J.-A.; Drizard, N.; Martin, N.; Hoffmann, B.; Gaston-Mathé, Y.; Rognan, D. On the Frustration to Predict Binding Affinities from Protein–Ligand Structures with Deep Neural Networks. J. Med. Chem. 2022, 65, 7946– 7958,  DOI: 10.1021/acs.jmedchem.2c00487
      225
      On the Frustration to Predict Binding Affinities from Protein-Ligand Structures with Deep Neural Networks
      Volkov, Mikhail; Turk, Joseph-Andre; Drizard, Nicolas; Martin, Nicolas; Hoffmann, Brice; Gaston-Mathe, Yann; Rognan, Didier
      Journal of Medicinal Chemistry (2022), 65 (11), 7946-7958CODEN: JMCMAR; ISSN:0022-2623. (American Chemical Society)
      Accurate prediction of binding affinities from protein-ligand at. coordinates remains a major challenge in early stages of drug discovery. Using modular message passing graph neural networks describing both the ligand and the protein in their free and bound states, we unambiguously evidence that an explicit description of protein-ligand noncovalent interactions does not provide any advantage with respect to ligand or protein descriptors. Simple models, inferring binding affinities of test samples from that of the closest ligands or proteins in the training set, already exhibit good performances, suggesting that memorization largely dominates true learning in the deep neural networks. The current study suggests considering only noncovalent interactions while omitting their protein and ligand at. environments. Removing all hidden biases probably requires much denser protein-ligand training matrixes and a coordinated effort of the drug design community to solve the necessary protein-ligand structures.
    226. 226
      Shivakumar, D.; Williams, J.; Wu, Y.; Damm, W.; Shelley, J.; Sherman, W. Prediction of absolute solvation free energies using molecular dynamics free energy perturbation and the OPLS force field. J. Chem. Theory Comput. 2010, 6, 1509– 1519,  DOI: 10.1021/ct900587b
      226
      Prediction of Absolute Solvation Free Energies using Molecular Dynamics Free Energy Perturbation and the OPLS Force Field
      Shivakumar, Devleena; Williams, Joshua; Wu, Yujie; Damm, Wolfgang; Shelley, John; Sherman, Woody
      Journal of Chemical Theory and Computation (2010), 6 (5), 1509-1519CODEN: JCTCCE; ISSN:1549-9618. (American Chemical Society)
      The accurate prediction of protein-ligand binding free energies is a primary objective in computer-aided drug design. The solvation free energy of a small mol. provides a surrogate to the desolvation of the ligand in the thermodn. process of protein-ligand binding. Here, we use explicit solvent mol. dynamics free energy perturbation to predict the abs. solvation free energies of a set of 239 small mols., spanning diverse chem. functional groups commonly found in drugs and drug-like mols. We also compare the performance of abs. solvation free energies obtained using the OPLS_2005 force field with two other commonly used small mol. force fields - general AMBER force field (GAFF) with AM1-BCC charges and CHARMm-MSI with CHelpG charges. Using the OPLS_2005 force field, we obtain high correlation with exptl. solvation free energies (R2 = 0.94) and low av. unsigned errors for a majority of the functional groups compared to AM1-BCC/GAFF or CHelpG/CHARMm-MSI. However, OPLS_2005 has errors of over 1.3 kcal/mol for certain classes of polar compds. We show that predictions on these compd. classes can be improved by using a semiempirical charge assignment method with an implicit bond charge correction.
    227. 227
      El Hage, K.; Mondal, P.; Meuwly, M. Free energy simulations for protein ligand binding and stability. Mol. Simul. 2018, 44, 1044– 1061,  DOI: 10.1080/08927022.2017.1416115
      227
      Free energy simulations for protein ligand binding and stability
      El Hage, Krystel; Mondal, Padmabati; Meuwly, Markus
      Molecular Simulation (2018), 44 (13-14), 1044-1061CODEN: MOSIEA; ISSN:0892-7022. (Taylor & Francis Ltd.)
      We summarize several computational techniques to det. relative free energies for condensed-phase systems. The focus is on practical considerations which are capable of making direct contact with expts. Particular applications include the thermodn. stability of apo- and holo-myoglobin, insulin dimerization free energy, ligand binding in lysozyme, and ligand diffusion in globular proteins. In addn. to provide differential free energies between neighboring states, converged umbrella sampling simulations provide insight into migration barriers and ligand dissocn. barriers and anal. of the trajectories yield addnl. insight into the structural dynamics of fundamental processes. Also, such simulations are useful tools to quantify relative stability changes for situations where expts. are difficult. This is illustrated for NO-bound myoglobin. For the dissocn. of benzonitrile from lysozyme it is found that long umbrella sampling simulations are required to approx. converge the free energy profile. Then, however, the resulting differential free energy between the bound and unbound state is in good agreement with ests. from mol. mechanics with generalized Born surface area simulations. Furthermore, comparing the barrier height for ligand escape suggests that ligand dissocn. contains a non-equil. component.
    228. 228
      Ngo, S. T.; Pham, M. Q. Umbrella sampling-based method to compute ligand-binding affinity. Methods Mol. Biol. 2022, 2385, 313– 323,  DOI: 10.1007/978-1-0716-1767-0_14
      There is no corresponding record for this reference.
    229. 229
      Pandey, M.; Fernandez, M.; Gentile, F.; Isayev, O.; Tropsha, A.; Stern, A. C.; Cherkasov, A. The transformational role of GPU computing and deep learning in drug discovery. Nat. Mach. Intell. 2022, 4, 211– 221,  DOI: 10.1038/s42256-022-00463-x
      There is no corresponding record for this reference.
    230. 230
      Bibal, A.; Cardon, R.; Alfter, D.; Wilkens, R.; Wang, X.; François, T.; Watrin, P. Is Attention Explanation? An Introduction to the Debate. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). Dublin, Ireland, 2022, 3889– 3900
      There is no corresponding record for this reference.
    231. 231
      Wiegreffe, S.; Pinter, Y. Attention is not not Explanation. arXiv , 2019.
      There is no corresponding record for this reference.
    232. 232
      Jain, S.; Wallace, B. C. Attention is not Explanation. arXiv , 2019.
      There is no corresponding record for this reference.
    233. 233
      Lundberg, S. M.; Lee, S.-I. A unified approach to interpreting model predictions. Neural Inf. Process. Syst. 2017, 30, 4765– 4774
      There is no corresponding record for this reference.
    234. 234
      Gu, Y.; Zhang, X.; Xu, A.; Chen, W.; Liu, K.; Wu, L.; Mo, S.; Hu, Y.; Liu, M.; Luo, Q. Protein-ligand binding affinity prediction with edge awareness and supervised attention. iScience 2023, 26, 105892,  DOI: 10.1016/j.isci.2022.105892
      There is no corresponding record for this reference.
    235. 235
      Rodis, N.; Sardianos, C.; Papadopoulos, G. T.; Radoglou-Grammatikis, P.; Sarigiannidis, P.; Varlamis, I. Multimodal Explainable Artificial Intelligence: A Comprehensive Review of Methodological Advances and Future Research Directions. arXiv [cs.AI] 2023.
      There is no corresponding record for this reference.
    236. 236
      Gilpin, L. H.; Bau, D.; Yuan, B. Z.; Bajwa, A.; Specter, M.; Kagal, L. Explaining explanations: An overview of interpretability of machine learning. 2018 IEEE 5th International Conference on Data Science and Advanced Analytics (DSAA) 2018, 80– 89
      There is no corresponding record for this reference.
    237. 237
      Luo, D.; Liu, D.; Qu, X.; Dong, L.; Wang, B. Enhancing generalizability in protein-ligand binding affinity prediction with multimodal contrastive learning. J. Chem. Inf. Model. 2024, 64, 1892– 1906,  DOI: 10.1021/acs.jcim.3c01961
      There is no corresponding record for this reference.
    238. 238
      Rao, R.; Bhattacharya, N.; Thomas, N.; Duan, Y.; Chen, X.; Canny, J.; Abbeel, P.; Song, Y. S. Evaluating protein transfer learning with TAPE. bioRxiv , 2019.
      There is no corresponding record for this reference.
    239. 239
      Suzek, B. E.; Wang, Y.; Huang, H.; McGarvey, P. B.; Wu, C. H. UniProt Consortium UniRef clusters: a comprehensive and scalable alternative for improving sequence similarity searches. Bioinformatics 2015, 31, 926– 932,  DOI: 10.1093/bioinformatics/btu739
      There is no corresponding record for this reference.
    240. 240
      Eguida, M.; Rognan, D. A Computer Vision Approach to Align and Compare Protein Cavities: Application to Fragment-Based Drug Design. J. Med. Chem. 2020, 63, 7127– 7142,  DOI: 10.1021/acs.jmedchem.0c00422
      240
      A Computer Vision Approach to Align and Compare Protein Cavities: Application to Fragment-Based Drug Design
      Eguida, Merveille; Rognan, Didier
      Journal of Medicinal Chemistry (2020), 63 (13), 7127-7142CODEN: JMCMAR; ISSN:0022-2623. (American Chemical Society)
      Identifying local similarities in binding sites from distant proteins is a major hurdle to rational drug design. We herewith present a novel method, borrowed from computer vision, adapted to mine fragment subpockets and compare them to whole ligand-binding sites. Pockets are represented by pharmacophore-annotated point clouds mimicking ideal ligands or fragments. Point cloud registration is used to find the transformation enabling an optimal overlap of points sharing similar topol. and pharmacophoric neighborhoods. The method (ProCare) was calibrated on a large set of druggable cavities and applied to the comparison of fragment subpockets to entire cavities. A collection of 33,953 subpockets annotated with their bound fragments was screened for local similarity to cavities from recently described protein X-ray structures. ProCare was able to detect local similarities between remote pockets and transfer the corresponding fragments to the query cavity space, thereby proposing a first step to fragment-based design approaches targeting orphan cavities.
    241. 241
      Öztürk, H.; Özgür, A.; Ozkirimli, E. DeepDTA: deep drug-target binding affinity prediction. Bioinformatics 2018, 34, i821– i829,  DOI: 10.1093/bioinformatics/bty593
      241
      DeepDTA: deep drug-target binding affinity prediction
      Ozturk, Hakime; Ozgur, Arzucan; Ozkirimli, Elif
      Bioinformatics (2018), 34 (17), i821-i829CODEN: BOINFP; ISSN:1367-4811. (Oxford University Press)
      Motivation: The identification of novel drug-target (DT) interactions is a substantial part of the drug discovery process. Most of the computational methods that have been proposed to predict DT interactions have focused on binary classification, where the goal is to det. whether a DT pair interacts or not. However, protein-ligand interactions assume a continuum of binding strength values, also called binding affinity and predicting this value still remains a challenge. The increase in the affinity data available in DT knowledge-bases allows the use of advanced learning techniques such as deep learning architectures in the prediction of binding affinities. In this study, we propose a deep-learning based model that uses only sequence information of both targets and drugs to predict DT interaction binding affinities. The few studies that focus on DT binding affinity prediction use either 3D structures of protein-ligand complexes or 2D features of compds. One novel approach used in this work is the modeling of protein sequences and compd. 1D representations with convolutional neural networks (CNNs). Results: The results show that the proposed deep learning based model that uses the 1D representations of targets and drugs is an effective approach for drug target binding affinity prediction. The model in which high-level representations of a drug and a target are constructed via CNNs achieved the best Concordance Index (CI) performance in one of our larger benchmark datasets, outperforming the KronRLS algorithm and SimBoost, a state-of-the-art method for DT binding affinity prediction.
    242. 242
      Evans, R. Protein complex prediction with AlphaFold-Multimer. bioRxiv , 2021.
      There is no corresponding record for this reference.
    243. 243
      Omidi, A.; Møller, M. H.; Malhis, N.; Bui, J. M.; Gsponer, J. AlphaFold-Multimer accurately captures interactions and dynamics of intrinsically disordered protein regions. Proc. Natl. Acad. Sci. U. S. A. 2024, 121, e2406407121,  DOI: 10.1073/pnas.2406407121
      There is no corresponding record for this reference.
    244. 244
      Zhu, W.; Shenoy, A.; Kundrotas, P.; Elofsson, A. Evaluation of AlphaFold-Multimer prediction on multi-chain protein complexes. Bioinformatics 2023, 39, btad424,  DOI: 10.1093/bioinformatics/btad424
      There is no corresponding record for this reference.
    245. 245
      Rombach, R.; Blattmann, A.; Lorenz, D.; Esser, P.; Ommer, B. High-resolution image synthesis with latent diffusion models. Pattern Recognition (CVPR) 2022, 10684– 10695
      There is no corresponding record for this reference.
    246. 246
      Dhariwal, P.; Nichol, A. Diffusion models beat GANs on image synthesis. Neural Inf. Process. Syst. 2021, 8780– 8794
      There is no corresponding record for this reference.
    247. 247
      Karras, T.; Aittala, M.; Aila, T.; Laine, S. Elucidating the Design Space of Diffusion-Based Generative Models. Adv. Neural Inf. Process. Syst. 2022, 26565– 26577
      There is no corresponding record for this reference.
    248. 248
      Buttenschoen, M.; Morris, G.; Deane, C. PoseBusters: AI-based docking methods fail to generate physically valid poses or generalise to novel sequences. Chem. Sci. 2024, 15, 3130– 3139,  DOI: 10.1039/D3SC04185A
      There is no corresponding record for this reference.
    249. 249
      Wee, J.; Wei, G.-W. Benchmarking AlphaFold3’s protein-protein complex accuracy and machine learning prediction reliability for binding free energy changes upon mutation. arXiv , 2024.
      There is no corresponding record for this reference.
    250. 250
      Bernard, C.; Postic, G.; Ghannay, S.; Tahi, F. Has AlphaFold 3 reached its success for RNAs? bioRxiv , 2024.
      There is no corresponding record for this reference.
    251. 251
      Zonta, F.; Pantano, S. From sequence to mechanobiology? Promises and challenges for AlphaFold 3. Mechanobiology in Medicine 2024, 2, 100083,  DOI: 10.1016/j.mbm.2024.100083
      There is no corresponding record for this reference.
    252. 252
      He, X.-H.; Li, J.-R.; Shen, S.-Y.; Xu, H. E. AlphaFold3 versus experimental structures: assessment of the accuracy in ligand-bound G protein-coupled receptors. Acta Pharmacol. Sin. 2024, 1– 12,  DOI: 10.1038/s41401-024-01429-y
      There is no corresponding record for this reference.
    253. 253
      Desai, D.; Kantliwala, S. V.; Vybhavi, J.; Ravi, R.; Patel, H.; Patel, J. Review of AlphaFold 3: Transformative advances in drug design and therapeutics. Cureus 2024, 16, e63646,  DOI: 10.7759/cureus.63646
      There is no corresponding record for this reference.
    254. 254
      Baek, M. Accurate prediction of protein structures and interactions using a three-track neural network. Science 2021, 373, 871– 876,  DOI: 10.1126/science.abj8754
      254
      Accurate prediction of protein structures and interactions using a three-track neural network
      Baek, Minkyung; DiMaio, Frank; Anishchenko, Ivan; Dauparas, Justas; Ovchinnikov, Sergey; Lee, Gyu Rie; Wang, Jue; Cong, Qian; Kinch, Lisa N.; Schaeffer, R. Dustin; Millan, Claudia; Park, Hahnbeom; Adams, Carson; Glassman, Caleb R.; DeGiovanni, Andy; Pereira, Jose H.; Rodrigues, Andria V.; van Dijk, Alberdina A.; Ebrecht, Ana C.; Opperman, Diederik J.; Sagmeister, Theo; Buhlheller, Christoph; Pavkov-Keller, Tea; Rathinaswamy, Manoj K.; Dalwadi, Udit; Yip, Calvin K.; Burke, John E.; Garcia, K. Christopher; Grishin, Nick V.; Adams, Paul D.; Read, Randy J.; Baker, David
      Science (Washington, DC, United States) (2021), 373 (6557), 871-876CODEN: SCIEAS; ISSN:1095-9203. (American Association for the Advancement of Science)
      DeepMind presented notably accurate predictions at the recent 14th Crit. Assessment of Structure Prediction (CASP14) conference. We explored network architectures that incorporate related ideas and obtained the best performance with a three-track network in which information at the one-dimensional (1D) sequence level, the 2D distance map level, and the 3D coordinate level is successively transformed and integrated. The three-track network produces structure predictions with accuracies approaching those of DeepMind in CASP14, enables the rapid soln. of challenging x-ray crystallog. and cryo-electron microscopy structure modeling problems, and provides insights into the functions of proteins of currently unknown structure. The network also enables rapid generation of accurate protein-protein complex models from sequence information alone, short-circuiting traditional approaches that require modeling of individual subunits followed by docking. We make the method available to the scientific community to speed biol. research.
    255. 255
      Ahdritz, G. OpenFold: retraining AlphaFold2 yields new insights into its learning mechanisms and capacity for generalization. Nat. Methods 2024, 21, 1514– 1524,  DOI: 10.1038/s41592-024-02272-z
      There is no corresponding record for this reference.
    256. 256
      Liao, C.; Yu, Y.; Mei, Y.; Wei, Y. From words to molecules: A survey of Large Language Models in chemistry. arXiv , 2024.
      There is no corresponding record for this reference.
    257. 257
      Bagal, V.; Aggarwal, R.; Vinod, P. K.; Priyakumar, U. D. MolGPT: Molecular generation using a transformer-decoder model. J. Chem. Inf. Model. 2022, 62, 2064– 2076,  DOI: 10.1021/acs.jcim.1c00600
      257
      MolGPT: Molecular Generation Using a Transformer-Decoder Model
      Bagal, Viraj; Aggarwal, Rishal; Vinod, P. K.; Priyakumar, U. Deva
      Journal of Chemical Information and Modeling (2022), 62 (9), 2064-2076CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      Application of deep learning techniques for de novo generation of mols., termed as inverse mol. design, has been gaining enormous traction in drug design. The representation of mols. in SMILES notation as a string of characters enables the usage of state of the art models in natural language processing, such as Transformers, for mol. design in general. Inspired by generative pre-training (GPT) models that have been shown to be successful in generating meaningful text, we train a transformer-decoder on the next token prediction task using masked self-attention for the generation of druglike mols. in this study. We show that our model, MolGPT, performs on par with other previously proposed modern machine learning frameworks for mol. generation in terms of generating valid, unique, and novel mols. Furthermore, we demonstrate that the model can be trained conditionally to control multiple properties of the generated mols. We also show that the model can be used to generate mols. with desired scaffolds as well as desired mol. properties by conditioning the generation on scaffold SMILES strings of desired scaffolds and property values. Using saliency maps, we highlight the interpretability of the generative process of the model.
    258. 258
      Janakarajan, N.; Erdmann, T.; Swaminathan, S.; Laino, T.; Born, J. Language models in molecular discovery. arXiv , 2023.
      There is no corresponding record for this reference.
    259. 259
      Park, Y.; Metzger, B. P. H.; Thornton, J. W. The simplicity of protein sequence-function relationships. Nat. Commun. 2024, 15, 7953,  DOI: 10.1038/s41467-024-51895-5
      There is no corresponding record for this reference.
    260. 260
      Stahl, K.; Warneke, R.; Demann, L.; Bremenkamp, R.; Hormes, B.; Brock, O.; Stülke, J.; Rappsilber, J. Modelling protein complexes with crosslinking mass spectrometry and deep learning. Nat. Commun. 2024, 15, 7866,  DOI: 10.1038/s41467-024-51771-2
      There is no corresponding record for this reference.
    261. 261
      Senior, A. W. Improved protein structure prediction using potentials from deep learning. Nature 2020, 577, 706– 710,  DOI: 10.1038/s41586-019-1923-7
      261
      Improved protein structure prediction using potentials from deep learning
      Senior, Andrew W.; Evans, Richard; Jumper, John; Kirkpatrick, James; Sifre, Laurent; Green, Tim; Qin, Chongli; Zidek, Augustin; Nelson, Alexander W. R.; Bridgland, Alex; Penedones, Hugo; Petersen, Stig; Simonyan, Karen; Crossan, Steve; Kohli, Pushmeet; Jones, David T.; Silver, David; Kavukcuoglu, Koray; Hassabis, Demis
      Nature (London, United Kingdom) (2020), 577 (7792), 706-710CODEN: NATUAS; ISSN:0028-0836. (Nature Research)
      Protein structure prediction can be used to det. the three-dimensional shape of a protein from its amino acid sequence1. This problem is of fundamental importance as the structure of a protein largely dets. its function2; however, protein structures can be difficult to det. exptl. Considerable progress has recently been made by leveraging genetic information. It is possible to infer which amino acid residues are in contact by analyzing covariation in homologous sequences, which aids in the prediction of protein structures3. The authors can train a neural network to make accurate predictions of the distances between pairs of residues, which convey more information about the structure than contact predictions. Using this information, the authors construct a potential of mean force4 that can accurately describe the shape of a protein. The resulting potential can be optimized by a simple gradient descent algorithm to generate structures without complex sampling procedures. The resulting system, named AlphaFold, achieves high accuracy, even for sequences with fewer homologous sequences. In the recent Crit. Assessment of Protein Structure Prediction5 (CASP13)-a blind assessment of the state of the field-AlphaFold created high-accuracy structures (with template modeling (TM) scores6 of 0.7 or higher) for 24 out of 43 free modeling domains, whereas the next best method, which used sampling and contact information, achieved such accuracy for only 14 out of 43 domains. AlphaFold represents a considerable advance in protein-structure prediction. The authors expect this increased accuracy to enable insights into the function and malfunction of proteins, esp. in cases for which no structures for homologous proteins have been exptl. detd.7.
    262. 262
      Kramer, M. A. Nonlinear principal component analysis using autoassociative neural networks. AIChE J. 1991, 37, 233– 243,  DOI: 10.1002/aic.690370209
      262
      Nonlinear principal component analysis using autoassociative neural networks
      Kramer, Mark A.
      AIChE Journal (1991), 37 (2), 233-43CODEN: AICEAC; ISSN:0001-1541.
      Nonlinear-principal-component anal. (NLPCA), is a novel technique for multivariate data anal., similar to the method of principal-component anal. (PCA). NLPCA like PCA, is used to identify and remove correlations among problem variables as an aid to dimensionality redn., visualization, and exploratory data anal. While PCA identifies only linear correlations between variables, NLPCA uncovers both linear and nonlinear correlations, without restriction on the character of the nonlinearities present in the data. NLPCA operates by training a feedforward neural network to perform the identity mapping, where the network inputs are reproduced at the output layer. The network contains an internal bottleneck layer (contg. fewer nodes than input or output layers), which forces the network to develop a compact representation of the input data and 2 addnl. hidden layers. The NLPCA method is demonstrated by using time-dependent, simulated batch-reaction data. NLPCA can reduce dimensionality and produce a feature space map resembling the actual distribution of the underlying system parameters.
    263. 263
      Rogers, D.; Hahn, M. Extended-connectivity fingerprints. J. Chem. Inf. Model. 2010, 50, 742– 754,  DOI: 10.1021/ci100050t
      263
      Extended-Connectivity Fingerprints
      Rogers, David; Hahn, Mathew
      Journal of Chemical Information and Modeling (2010), 50 (5), 742-754CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      Extended-connectivity fingerprints (ECFPs) are a novel class of topol. fingerprints for mol. characterization. Historically, topol. fingerprints were developed for substructure and similarity searching. ECFPs were developed specifically for structure-activity modeling. ECFPs are circular fingerprints with a no. of useful qualities: they can be very rapidly calcd.; they are not predefined and can represent an essentially infinite no. of different mol. features (including stereochem. information); their features represent the presence of particular substructures, allowing easier interpretation of anal. results; and the ECFP algorithm can be tailored to generate different types of circular fingerprints, optimized for different uses. While the use of ECFPs has been widely adopted and validated, a description of their implementation has not previously been presented in the literature.
    264. 264
      Veličković, P.; Cucurull, G.; Casanova, A.; Romero, A.; Liò, P.; Bengio, Y. Graph Attention Networks. arXiv , 2017.
      There is no corresponding record for this reference.
    265. 265
      Kipf, T. N.; Welling, M. Semi-Supervised Classification with Graph Convolutional Networks. arXiv , 2016.
      There is no corresponding record for this reference.
    266. 266
      Xu, Z.; Wang, S.; Zhu, F.; Huang, J. Seq2seq Fingerprint: An Unsupervised Deep Molecular Embedding for Drug Discovery. Proceedings of the 8th ACM International Conference on Bioinformatics, Computational Biology, and Health Informatics. New York, NY, USA 2017, 285– 294
      There is no corresponding record for this reference.
    267. 267
      Gilmer, J.; Schoenholz, S.; Riley, P. F.; Vinyals, O.; Dahl, G. E. Neural Message Passing for Quantum Chemistry. ICML 2017, 1263– 1272
      There is no corresponding record for this reference.
    268. 268
      Asgari, E.; Mofrad, M. R. K. Continuous Distributed Representation of Biological Sequences for Deep Proteomics and Genomics. PLoS One 2015, 10, e0141287,  DOI: 10.1371/journal.pone.0141287
      268
      Continuous distributed representation of biological sequences for deep proteomics and genomics
      Asgari, Ehsaneddin; Mofrad, Mohammad R. K.
      PLoS One (2015), 10 (11), e0141287/1-e0141287/15CODEN: POLNCL; ISSN:1932-6203. (Public Library of Science)
      We introduce a new representation and feature extn. method for biol. sequences. Named bio-vectors (BioVec) to refer to biol. sequences in general with protein-vectors (ProtVec) for proteins (amino-acid sequences) and gene-vectors (GeneVec) for gene sequences, this representation can be widely used in applications of deep learning in proteomics and genomics. In the present paper, we focus on protein-vectors that can be utilized in a wide array of bioinformatics investigations such as family classification, protein visualization, structure prediction, disordered protein identification, and protein-protein interaction prediction. In this method, we adopt artificial neural network approaches and represent a protein sequence with a single dense n-dimensional vector. To evaluate this method, we apply it in classification of 324,018 protein sequences obtained from Swiss-Prot belonging to 7,027 protein families, where an av. family classification accuracy of 93% ± 0.06%is obtained, outperforming existing family classification methods. In addn., we use ProtVec representation to predict disordered proteins from structured proteins. Two databases of disordered sequences are used: the DisProt database as well as a database featuring the disordered regions of nucleoporins rich with phenylalanine-glycine repeats (FG-Nups). Using support vector machine classifiers, FG-Nup sequences are distinguished from structured protein sequences found in Protein Data Bank (PDB) with a 99.8% accuracy, and unstructured DisProt sequences are differentiated from structured DisProt sequences with 100.0% accuracy. These results indicate that by only providing sequence data for various proteins into this model, accurate information about protein structure can be detd. Importantly, this model needs to be trained only once and can then be applied to ext. a comprehensive set of information regarding proteins of interest. Moreover, this representation can be considered as pre-training for various applications of deep learning in bioinformatics.
    269. 269
      He, K.; Zhang, X.; Ren, S.; Sun, J. Deep residual learning for image recognition. Proc. IEEE Comput. Soc. Conf. Comput. Vis. Pattern Recognit. 2015, 770– 778
      There is no corresponding record for this reference.
    270. 270
      Öztürk, H.; Ozkirimli, E.; Özgür, A. A novel methodology on distributed representations of proteins using their interacting ligands. Bioinformatics 2018, 34, i295– i303,  DOI: 10.1093/bioinformatics/bty287
      There is no corresponding record for this reference.
    271. 271
      Hu, J.; Shen, L.; Sun, G. Squeeze-and-excitation networks. Proc. IEEE Conf. Comput. Vis. Pattern Recognition 2018, 7132– 7141
      There is no corresponding record for this reference.