ACS Publications. Most Trusted. Most Cited. Most Read
Quick View

OR SEARCH CITATIONS

My Activity
Publications
CONTENT TYPES

Figure 1Loading Img
Download Hi-Res ImageDownload to MS-PowerPointCite This:J. Chem. Inf. Model. 2021, 61, 11, 5343-5361
RETURN TO ISSUEPREVReviewNEXT

Generative Deep Learning for Targeted Compound Design

Cite this: J. Chem. Inf. Model. 2021, 61, 11, 5343–5361
Publication Date (Web):October 26, 2021
https://doi.org/10.1021/acs.jcim.0c01496

Copyright © 2022 The Authors. Published by American Chemical Society. This publication is licensed under

CC-BY-NC-ND 4.0.
  • Open Access

Article Views

14070

Altmetric

-

Citations

32
LEARN ABOUT THESE METRICS
PDF (4 MB)
Get e-Alerts

Abstract

High Resolution Image
Download MS PowerPoint Slide

In the past few years, de novo molecular design has increasingly been using generative models from the emergent field of Deep Learning, proposing novel compounds that are likely to possess desired properties or activities. De novo molecular design finds applications in different fields ranging from drug discovery and materials sciences to biotechnology. A panoply of deep generative models, including architectures as Recurrent Neural Networks, Autoencoders, and Generative Adversarial Networks, can be trained on existing data sets and provide for the generation of novel compounds. Typically, the new compounds follow the same underlying statistical distributions of properties exhibited on the training data set Additionally, different optimization strategies, including transfer learning, Bayesian optimization, reinforcement learning, and conditional generation, can direct the generation process toward desired aims, regarding their biological activities, synthesis processes or chemical features. Given the recent emergence of these technologies and their relevance, this work presents a systematic and critical review on deep generative models and related optimization methods for targeted compound design, and their applications.

This publication is licensed under

CC-BY-NC-ND 4.0.
  • cc licence
  • by licence
  • nc licence
  • nd licence
KEYWORDS:

Introduction

ARTICLE SECTIONS
Jump To
De novo molecular design aims to create new chemical entities with desired properties and/or activities. These properties may be easily quantifiable, such as molecular weight, or somewhat more abstract, as is the case of toxicity. This is an inherently difficult task owing to the immense search space of around 1033–1080 feasible molecules from which only a small fraction typically have the desired traits. (1) As such, de novo molecular design was, for many years, and mostly remains a process of almost exclusive trial and error, with human expert knowledge and intuition about chemistry playing a major role. (2)
Meanwhile, the high costs associated with developing new molecules, reaching $2.8 billion dollars for a single compound, have also led to the implementation of computational tools capable of assisting the process. These have proven valuable and have found wide usage in practical applications. (2,3) A forthright approach consists in enumerating all possible molecules that conform to valency rules and do not include chemically unstable functional groups. A notable example is the Chemical Space project, where this technique was employed to generate 166 billion molecules. (4,5) Another technique, reaction-based de novo design, uses a set of known chemical reactions to combine various readily available building blocks into new molecules. This process can be guided by a similarity criterion to a known molecule of interest, giving rise to a large number of new similar molecules while ensuring their synthetic plausibility. (5,6)
Evolutionary Algorithms (EAs) have also been successfully applied to de novo molecular design. As a recent example, AutoGrow4 (7) uses an EA to create new predicted ligands. At each iteration, new molecules are created using a mutation operator, that performs an in silico chemical reaction, or a crossover operator that merges two compounds into a new one by randomly combining their decorating moieties. Grammatical Evolution on string representations and evolving molecular graphs provide alternative approaches that enable EAs to generate novel compounds targeting desired properties. (8,9)
Although useful, these methods still leave room for improvement. For instance, enumeration often leads to molecules that are too difficult to synthesize, and reaction-based design is fundamentally restricted in its ability to explore the chemical space, both important aspects of molecular design. EAs, while computationally efficient and capable of performing on par with other recent approaches, rely on expertly encoded operations, possibly limiting the search space and not leveraging the large amounts of data currently available. (10)
More recently, advances in Deep Learning (DL) sparked a surge of novel approaches to these problems. (11) Over the past decade, DL has proven successful in multiple fields, such as computer vision, speech recognition, and translation, pushing the state-of-the-art forward and surpassing other Machine Learning approaches. (12,13) DL refers to the use of artificial neural networks with multiple hidden layers. (12) A common intuition is that successive layers can learn higher-level abstractions of the inputs.
Within DL, generative modeling aims at capturing an underlying data generation process, some unknown probabilistic distribution from which a data set was sampled, and has been successfully used for creative tasks, such as writing, composing music, and painting. (14) It usually deals with unlabeled data and attempts to create a model capable of generating new observations that closely resemble those from the training data and not simply producing copies. (14)
Improving on earlier approaches, which employed more traditional machine learning methods such as Gaussian Mixture Models, deep generative models have recently found use in the generation of novel molecular entities. (14−16) Starting around 2017, with works like that of Ǵomez-Bombarelli et al., (10) Yuan et al. (17) and Segler et al., (18) a large number of novel approaches have been put forward employing various neural network architectures and molecular representations. (16,19) Alongside these works, several reviews have also sought to condense the plethora of different approaches, shedding light into and discussing the different architectures, generation of various molecular representations and use of comparative metrics. (16,19−22)
Due to the vastness of chemical space and the costs associated with testing possible compounds, a rational exploration with regard to the desired properties is preferable. As such, a number of distinct approaches have been developed for directing and controlling the generating process of molecules toward compounds with defined chemical properties or desired activities.
In this arena, the evolution has been very fast, with novel methods appearing at an impressive rate. A systematic review of the main advances of DL in the generation of focused molecules seems particularly relevant for practitioners interested in understanding the main features of each method, their main advantages, and their limitations. Several reviews covering the recent explosion of interest in generating molecules leveraging DL have been published. These, however, have mainly focused on the various architectures and molecular representations while only briefly touching on the various methods for the targeted generation of compounds. Particularly, Schwalbe-Koda and Ǵomez-Bombarelli (20) mention some of these methods, but they primarily focused on the several molecular generation schemes. Likewise, several other reviews also make note of a couple of these methods while not delving into further discussion. (19,21−24) Closer to our work, Sanchez-Lengeling and Aspuru-Guzik (16) directly discussed the inverse molecular design problem, touching on some methods for controlling the properties of generated molecules.
Notwithstanding previous efforts on reviewing this field, we feel that a more rigorous approach to this subject, containing a more systematic coverage of the methods, can be important for researchers working on these topics. To that end, here we aim to provide a comprehensive review of DL methods for the targeted generation of novel compounds. As such, after an introduction to molecular representations, we present the most common deep generative models and the underlying neural network architectures. We, then, focus on the different optimization approaches that allow to focus the search on molecules with desired properties or activities, closing with a review of the main practical applications.

Representing Molecules

ARTICLE SECTIONS
Jump To
The use of Machine Learning (ML) for chemical applications requires the conversion of chemical compounds to a machine-readable format suitable for computer processing. Although trivial names such as ”benzene” and ”caffeine” are easy to remember, they carry little to no information on the structure and properties of the underlying compound. Systematic terminology, such as the International Union of Pure and Applied Chemistry (IUPAC) nomenclature of organic chemistry, can be very lengthy and not specify the full structure of a compound. As such, several notation systems have been developed to provide a suitable form of representing molecules, (25) each with a distinct impact on the performance and outcomes of chemical ML methods. These are reviewed next and are also illustrated by the example provided by Figure 1.

Figure 1

Figure 1. Acetaminophen (center) under various molecular representations. Top-left: Sequence based representations. Prior to being fed to the models, these sequences are also usually one-hot encoded. Top-right: Graph-based representations. While connection matrices are a suitable input for standard architectures, graphs can also be directly handled using graph neural networks. Bottom: Three dimensional representations, images from PubChem. (26) Graphs may be enhanced by including 3D information as node attributes, such as internal distances and angles, or based on a coordinate system such as Cartesian space. Molecular surfaces can be voxelized into a 3D grid for easier processing.

High Resolution Image
Download MS PowerPoint Slide

1D Sequences

Line notations, or sequences, represent chemical structures as human readable strings of characters. Although many line notations have been developed, such as the SYBYL (27) and the Wiswesser (28) line notations, some have achieved greater popularity. Next, we emphasize the linear notations more commonly used in DL models for compound generation.

SMILES

The Simplified Molecular Input Line Entry System (SMILES) notation allows the representation of molecules as sequences of “tokens”. It is built through a depth-first traversal of the molecular graph, encoding atoms and how they connect in a simple and machine-friendly form. (29) As the process of encoding a molecule into a string can start at different locations of the molecule, there exists a one-to-many relationship where a single molecule can be represented by several different SMILES. As such, canonicalization algorithms have been developed to ensure that a one-to-one relationship is possible. (25,30) The SMILES syntax also presents an added difficulty.
Rings and branches require symbols to occur in pairs which often leads to syntactically invalid SMILES strings. Interestingly, the one-to-many relationship between molecules and noncanonical SMILES can be leveraged to perform data augmentation in DL approaches, by allowing one to expand a given data set with the various possible SMILES of the molecules it contains. This was first suggested by Bjerrum (31) to improve on prediction tasks and later applied to molecular generation by Bjerrum and Sattarov (32) followed by several other teams. (33−36)

InChI

The International Chemical Identifier (InChI) system, proposed by the IUPAC, consists of a notation language that represents molecules as layered strings of characters and aims to be a machine-readable unique representation of a structure. (37) Here, molecules are encoded as predefined layers of information that are arranged in a specific order. InChI starts by specifying a core parent structure to which further information may be added. Each layer is separated by a delimiter “/” and prefixed by a lower case letter identifying the layer (except for the first layer).
Although very versatile in representing compounds, generating molecules represented as InChI in DL has not been successful. This was first observed by Ǵomez-Bombarelli et al. (10) who attributed this result to the added complexity of the InChI syntax, when compared with SMILES. A later work by Winter et al. (38) also noted inferior performance when translating from SMILES to InChI, reporting that the model failed to learn, identifying the same probable cause. Nonetheless, InChI provides a unique identifier for chemical structures which can be exploited to derive a canonical SMILES string.

DeepSMILES

O’Boyle and Dalke (39) proposed DeepSMILES, an adaptation of the SMILES syntax focusing on two of the major causes of generating invalid SMILES, unmatched ring and parentheses closures. This is addressed by using a single symbol to indicate rings and by denoting branches solely using closing parentheses.

SELFIES

Krenn et al. (40) introduced a new linear notation for constrained graphs termed SELFreferencIng Embedded Strings (SELFIES). It is capable of enforcing the generation of syntactically and semantically valid graphs and is readily translated to and from them. The team compared its performance against SMILES, DeepSMILES, and Kusner et al. Grammar-VAE, reporting improvements on validity, reconstruction accuracy, and diversity of generated molecules.

2D (Chemical) Structures

A popular solution is to store compounds as graphs, where nodes represent atoms and edges represent bonds. These molecular graphs are commonly implemented using an adjacency matrix that specifies which atoms are connected and the respective bond order/type. Furthermore, nodes and edges can have associated properties, such as, and respectively, relative spatial location and bond order/type. This format allows the encoding of detailed topological data in a readily processable form. (41)

3D Structures

Representing a molecule as a simple connection table of atoms overlooks its three-dimensional conformation and consequently disregards valuable information. Representing the arrangement of atoms in space can be accomplished by coupling a coordinate system to a connection table. A common solution is to represent the molecule inside a Cartesian space, assigning each atom in the connection table spatial coordinates (x, y, z). (25) An alternative is to use internal coordinates, such as bond length, bond angle, and torsion angle, to describe the position of each atom relative to its neighbors, foregoing the need for a fixed coordinate system. (25)
Although useful, these two solutions only depict a molecule as a three-dimensional graph without any volume. In reality, a molecule has an electron cloud surrounding its atoms from which many of its properties arise. In molecular surfaces, a molecule is represented as a closed surface that delimits the volume it occupies. Particularly, this surface outlines a threshold value of electron density in the electron cloud that surrounds the molecule. The molecular surface can also be associated with properties, such as the electrostatic or hydrophobicity potential at particular locations. (25)

Databases

There are currently various repositories offering vast collections of molecules. Some are more specialized, providing focused libraries of known active compounds, such as DrugBank, (42) while others store more diverse compounds as is the case of ChEMBL. (43) The assorted molecular representations can usually be directly obtained from a repository or derived from SMILES or InChI in a straightforward manner using a cheminformatics package, such as the open source toolkit RDKit (44) or the Chemistry Development Kit (CDK). (45) Several of commonly used databases in the de novo drug design are identified in Table 1.
Table 1. Databases of Interesta
databasemoleculesinformation
ChEMBL (43)2M compoundsbioactive drug-like small molecules
ExCAPE-DB (46)1M compoundsactive/inactive molecules by target
ZINC (47)750M compoundsdrug-like molecules, available for purchase
PubChem (26)111M compoundsmostly small molecules
DrugBank (42)13K drug entriesapproved and experimental drugs
GDB-17 (4)166B compoundscombinatorially generated molecules
REAL database (48)1.95B compoundsdatabase of enumerated structures
Tox21 (49)11K compoundstoxicity data for various assays
QM8 (50)22K compoundselectronic spectra and excited state energy
QM9 (51)134K compoundsgeometric, energetic, electronic, thermodynamic
PDBbind (52)17K compounds3D structures and binding affinity
a

Number of available molecules reported as of October 2020.

Deep Learning models for De Novo Molecular Design

ARTICLE SECTIONS
Jump To

Architectures

Recently, generative DL has emerged as a promising development for de novo molecular design, where deep neural networks are employed as generative models. This specific application has attracted considerable attention, with several novel architectures being proposed, that are briefly reviewed next, being also illustrated in Figure 2.

Figure 2

Figure 2. Top-left: Three layer Recurrent Neural Network (RNN) both rolled and unrolled. In each layer, the output of a step, besides flowing to the next layer, also flows to the next step of the layer itself. These recurrent connections are depicted in the unfolded view of the network as vertical arrows. Top-right: Variational Autoencoder (VAE) where the input is encoded to the parameters of a statistical distribution, namely, the means (μ) and standard deviation (σ). In practice, these correspond to two vectors which, on the sampling step, are interpreted as a set of means and standard deviations. Bottom-left: Generative Adversarial Network (GAN) composed by a generator and a discriminator. Training seeks not a minimum but a useful equilibrium between the generator and the discriminator. Bottom-right: Adversarial Autoencoder (AAE) where the attached discriminator must discern between encoded points and samples drawn from a prior statistical distribution.

High Resolution Image
Download MS PowerPoint Slide

Recurrent Neural Networks

RNNs assume a sequential structure in the data, one where a sample is composed of a set of steps. This assumption is implemented by processing an input consecutively and introducing a connection carrying the output from previous steps into the current step. However, as the number of steps increases, RNNs can suffer from vanishing or exploding gradients during backpropagation, impairing the training process and making the learning of long-term dependencies extremely difficult. In practice, this is handled by using specialized units such as gated recurrent units (GRUs) (53) or Long Short-Term Memory (LSTM), (54) which introduce gates, learnable parameters controlling the flow of information through the steps. (12,13)

Generative Adversarial Networks

GANs define a pair of networks, a generator, and a discriminator, trained in competition with each other. The generator is intended to transform random noise into real looking data and is trained to maximize the synthetic samples classified as real by the discriminator. Meanwhile, the discriminator is trained to better discern between generated and real data. The training framework resembles a competition, with both networks constantly improving and adapting to each other. (12,13,55)

Autoencoders

Autoencoders (AEs) are neural networks trained to copy their input into the output with restrictions imposed as to not simply learn the identity function. They are usually thought of as two separate parts, an encoder that transforms the input into a more compact latent state, and a decoder that reconstructs the input from this representation. Both are trained together to minimize the information lost from reconstructing. (12,13)
Variational Autoencoders (VAEs) are a special type of AE, which assume that the data was sampled from an arbitrary statistical distribution. The encoder transforms its input into the parameters of a multidimensional statistical distribution, that is, a set of means and standard deviations. A sampling then occurs, where a point is drawn from the encoded distribution and fed into the decoder that reconstructs it into the original input. The objective function used for training consists of a term penalizing reconstruction errors and a term restricting the parameters encoded to be close to a normal distribution. This stochastic process acts to regularize the network while constraining the encoded parameters close to those of a normal distribution helps in forming a useful latent space. (13,56)
Adversarial Autoencoders (AAEs) are an alternative to VAEs that employ adversarial training for structuring the latent space. In particular, the encoder transforms its input into a single point in the latent space. A discriminator network then attempts to discern between samples of a prior statistical distribution and encoded points. As such, the encoder can also be viewed as a generator engaged in a competition with the discriminator, ultimately balancing between the reconstruction and adversarial error. (57)

Generating Molecules

There have been several approaches to applying generative DL to molecular generation, mainly differing on the chosen molecular representation. As such, usually more than one method surfaced for generating each of the main representations discussed in section.
Borrowing from the natural language processing field, molecules can be generated as sequences, such as SMILES, by using RNNs. Specifically, when using RNNs as a generative model, each token in the string is encoded as a one-hot vector and the network is trained to predict the next character in the sequence. The generation of new data is achieved by running the network autoregressively, that is, using its output as the input for the next time-step. This process is usually seeded with a special start token and the generation of a molecule ends when a special stop token is sampled. These two tokens are also respectively prefixed and appended to each molecule during training, Figure 3 illustrates the generative process.

Figure 3

Figure 3. Three layer RNN, unfolded over four time-steps. In autoregressive sequence generation, the process is started with a special start token, here “G”. The model then predicts the next token, which is sampled and used as input for the next step. Generation ends when a stop token is predicted.

High Resolution Image
Download MS PowerPoint Slide
Several research groups have employed this method with a stacked RNN, usually with LSTM cells, leading to good rates of validity, novelty, and diversity. (18,35,58,59) More complex architectures such as VAEs and GANs have also been employed to generate molecules as strings; however, these also employ a RNN for the sequence generation process, either as the decoder or the generator. (10,60,61)
Despite some limitations of sequence-based approaches, such as the need to learn a complex syntax and the mismatch between the edit distance of two SMILES and the underlying molecular similarity, these methods have produced impressive results including some instances of experimental validation of the generated molecules. (17,62−64)
In an attempt to present molecules in a more natural and intuitive form, several methods have been proposed to directly generate molecular graphs. The generation process can be modeled as a sequence of decisions that progressively builds a graph. As Figure 4 describes, this approach usually employs multiple networks with specific functions, such as adding new nodes or adding edges between existing ones. In this paradigm, Li et al. (65) used two Graph Neural Networks (GNNs) to build graphs, one deciding whether to add a new node followed by another network deciding whether, and where, to add new edges between the existing nodes. Liu et al. (66) used a similar generation procedure, employing Gated Graph Neural Networks (GGNNs) to build a VAE. The decoding process starts by using a linear classifier to add attributes to a fixed number of unconnected nodes. Edges between these nodes are then progressively added with two densely connected networks, with one deciding the target node and the other deciding the type of edge. At each step, the attributes of the nodes are updated with a GGNN and the process ends when a special stop node is selected. Furthermore, invalid operations can be masked during the generation allowing, for example, the enforcement of valency rules. More recently, Mercado et al. (67) compared six GNN architectures coupled with a tiered Multi Layer Perceptron (MLP) for sequentially building molecular graphs, reporting that GGNNs showed the best performance for both speed and quality of the generated structures.

Figure 4

Figure 4. Left: In sequential graph generation, a graph is built by evaluating a current partial graph, adding a node/edge and repeating until the network outputs a stop signal. Right: In the one-shot generation of graphs, probabilities over the full adjacency matrix and node/edge attribute tensors are produced. The graph is then obtained by taking a sample or the argmax of these outputs.

High Resolution Image
Download MS PowerPoint Slide
An alternative is to generate graphs in a one-shot fashion by directly outputting an adjacency matrix and corresponding attribute tensors. De Cao and Kipf (68) applied this approach with MolGAN, a GAN whose generator outputs probabilities over the adjacency matrix and the annotation matrix of a molecular graph. Following a similar approach, Simonovsky and Komodakis (69) employed a VAE with a decoder that outputs three probability distributions, one over the adjacency matrix, one over the edge attribute tensor, and another over node attribute tensor. Furthermore, Ma et al. (70) proposed a regularization scheme capable of enforcing validity constraints to graphs generated in the same manner, significantly improving the validity of generated molecules.
Molecules are ultimately three-dimensional objects, with electron clouds surrounding their atoms and multiple possible spatial arrangements or stereoisomers. By generating molecules as sequences or graphs, important information is omitted, possibly hindering de novo design. Furthermore, determining the relevant conformations is not a trivial problem, as even small molecules can have many possible conformations. (71) As such, some attempts have been made toward generating molecules as three-dimensional entities. Skalic et al. (72) proposed to generate voxelized molecular shapes with a VAE and then caption them into SMILES with a separate network. A different approach was proposed by Gebauer et al. (73) where molecules are generated as point sets, iteratively built based on the pairwise distance to previously placed points/atoms. Figure 5 outlines the general process of these two last approaches. More recently, Ragoza et al. (74) proposed a method to generate molecules as 3D atomic density grids and then apply an optimization algorithm to find the best fitting 3D chemical structure.

Figure 5

Figure 5. Left: General procedure for the generation of 3D shapes as proposed by Skalic et al. (72) The convolutional decoder of a VAE is used to produce a 3D molecular shape which is converted to SMILES by a captioning network. Right: General process for generating molecules as 3D point sets, proposed by Gebauer et al. (73) It is conceptually similar to the sequential graph generation, operating on point sets with an internal coordinate system.

High Resolution Image
Download MS PowerPoint Slide

Evaluating Generative Models

Given the rapid growth this field is experiencing, the development of systematic and robust methods to assess the performance of novel approaches is essential to help guide future work. Aiming to help address this, Preuer et al. (75) introduced the Fŕechet ChemNet Distance (FCD), a metric for comparing the generated molecules against the training data set. This score measures the distance between the hidden representations of the two sets of molecules in ’ChemNet’, a recent multitask network for predicting biological activities. As this network was trained to predict the bioactivities of about 6000 assays, the team proposes that FCD combines into a single metric a multitude of important molecular features, which is therefore useful to evaluate generative models.
Arús-Pous et al. (76) proposed a method to evaluate how well a generative model learns to cover the relevant chemical space. According to the team, this can be accomplished by training the model on a fraction of a large enumerated data set, such as GDB-13, and then tracking the percentage of the total data set the model can recover, how uniform the coverage is, and also whether it generates molecules outside the data set. These results can then be compared to an ideal model, directly sampling the data set, that serves as an upper bound for performance. Furthermore, the team introduced a method for evaluating the quality of the training process by comparing the negative log-likelihood of the sampled, training and evaluation sets throughout the training process.
Brown et al. (77) introduced GuacaMol, a framework for benchmarking models for de novo molecular design. This framework was divided into two main sets of benchmarks, distribution-learning and goal-directed generation, aiming to emulate the two main use cases of these models. The first set measures how well the models learn to generate new molecules, using validity, uniqueness and novelty rates, and also whether they match the properties of the training data set, measuring the FCD and the divergence in the distribution of a variety of physicochemical descriptors. The second set evaluates the targeted generation performance, including benchmarks such as optimizing given molecular features or properties, generating molecules similar to a target compound or recover a specific target molecules. In addition, the team also included a benchmark for assessing molecular quality, leveraging rule sets for building high-throughput screening libraries. Lastly, a standardized data set alongside a number of baselines were also released with this framework to help compare novel approaches. Polykovskiy et al. (78) proposed the benchmarking framework MOSES, for evaluating the distribution learning performance of generative models. To assess whether the models can generate new molecules, it measures the validity, uniqueness and novelty rates along with the internal diversity and the fraction of generated molecules that pass a set of structural filters for molecular quality. The framework also provides a set of metrics meant to evaluate how well the model learned features of the training data set. For this purpose, it presents the FCD, the distance between the distribution of various physicochemical properties, the Tanimoto Similarity to a nearest neighbor, the cosine similarity of Bemis–Murcko scaffolds, and also the cosine similarity of BRICS fragments. These last two metrics help compare molecules at a substructure level, while the first three help to capture more abstract chemical and biological similarities. The team also released various useful baselines, as well as a standardized data set with a recommended train, test and scaffold test split.
Building on these early works, other approaches to evaluating molecular generative models have continued to be developed. Specifically, Renz et al. (79) highlighted some shortcomings of currently used evaluation metrics. Specifically, the team details how a trivial model can excel in distribution-learning benchmarks and also how goal-directed generation can exploit biases in the scoring functions, producing compounds with high scores but of little practical use. Cieplinski et al. (80) proposed to better represent real discovery problems by using docking as a benchmark of the different methods of goal-directed generation. Zhang et al. (81) improved on their earlier work, measuring the coverage of chemical space by generative models, (76) by also evaluating the coverage of functional groups and ring systems. Furthermore, the team provided results for various, recently introduced, generative model architectures allowing for their comparison as well as providing useful baselines for future works.

Generating Compounds of Interest

ARTICLE SECTIONS
Jump To
The automatic generation of novel molecules usually targets specific properties and characteristics, such as solubility or bioactivity. As such, the ability to create not just new, but also focused, molecules is of interest. Table 2 summarizes state-of-the-art approaches for targeted compound design, further described in this section.
Table 2. Methods for the Directed Generation of Molecules
methodrepresentationarchitectureref
Transfer LearningSMILESStacked RNN (18), (34), (59), (62), (63), (82)
  SMILES  (83)
  GraphGNN (84)
  3D point setsSchNet +2 MLP (73)
Reinforcement LearningPretrain + RLSMILESStacked RNN (18), (58), (82), (85)
  SMILESVAE (86)
  Graphtwo RNNs (87)
 Adversarial + RLSMILESGAN (60,88−90)
  GraphGNNs (91), (92)
  GraphGAN (68)
Latent Space NavigationBayesian OptimizationSMILESVAE (10), (93)
  SMILESAAE and VAE (94)
  SMILES (production rules)VAE (95), (96)
  Graph (junction trees)VAE (97)
  GraphVAE (98)
 Gradient AscentGraph (junction trees)VAE (97)
  GraphVAE (66), (99)
 DL ModelSMILESGAN + AE (32) (36)
  Graph (junction trees)CycleGAN + VAE (97) (100)
 GASMILESAE (101)
 PSOSMILESAE (102)
 CLaSSSMILESVAE (103)
ConditionedConditionedSMILESVAE (61)
  SMILESStacked RNN (104)
  SMILESTwo AAEs (105)
  SMILES (production rules)Two GANs (106)
  SELFIESVAE (107)
  GraphGNNs (65), (84)
  GraphVAE (69)
  Graph (junction trees)VAE (108)
  3D shapeVAE + RNN (72)
  3D shapeVAE + GAN (109)
 semisupervisedSMILESVAE (110)
  SMILESAAE (64)
  Graph (scaffold extension)VAE (111)

Screening

Testing large numbers of compounds to see if they show evidence of having desired properties is usually one of the first steps in the drug discovery pipeline. This usually entails a virtual screening for bioactivity either based on a target, as in docking, or on known ligands, as in ML classifiers and similarity searching. (112) This process can be used on its own or after other methods of biasing the generation process. Docking refers to a process of fitting a molecule to the binding site of a given target whose 3D structure is known. This is usually attained by scoring different poses (spatial orientations) of a molecule relative to its target. The score is calculated by a scoring function, usually hinged on predictive changes in Gibbs free energy. (113)
Yuan et al., (17) for example, docked generated molecules against VEGFR-2, a mediator of the VEGF angiogenesis pathway, to choose compounds to be synthesized. Similarly,
Polykovskiy et al. (64) docked molecules against Janus kinase 2 and 3 as part of the selection process for synthesis.
An alternative to predict how well a molecule may bind to a receptor is through the use of supervised machine learning classifiers trained to distinguish between known actives and not actives. An example of this approach is the work done by Olivecrona et al. (58) where a support vector machine was trained to predict activity toward the Dopamine type 2 Receptor (DRD2).
In a similarity search, the similarity between a molecule and a set of molecules that are known to be active is determined. A common approach is to compute the Tanimoto coefficient between the fingerprints of each molecule. Kadurin et al. (114) used this method to search PubChem for molecules similar to the fingerprints generated by their model.

Transfer Learning

Despite that a simple screening of unspecific model outputs may lead to finding molecules of interest, the process is somewhat inefficient as a large number of the generated molecules end up being discarded. A less wasteful approach would be to first bias the model toward producing (more) molecules meeting the desired properties. This can be achieved with transfer learning, a training procedure where a model first learns to perform a similar task, but for which larger data sets exist, later being fine-tuned on the intended data. (12) This approach assumes that several of the underlying attributes learned on the first set are transferable to the second set. In molecule generation, this is usually applied on sequence-based approaches where a large data set such as ZINC or ChEMBL first helps to learn the syntax of the string representation and then a smaller, targeted, data set biases the model toward particular attributes, such as a given biological activity, as it is illustrated by Figure 6.

Figure 6

Figure 6. In transfer learning, a general model is first trained on a large data set and then fine-tuned toward generating the desired properties with a smaller, focused, data set.

High Resolution Image
Download MS PowerPoint Slide
Transfer learning has been successfully applied to fine-tune stacked RNNs generating SMILES. In 2017, Segler et al. (18) applied this method to a RNN with three LSTM layers generating SMILES. The model was later fine-tuned on known ligands of specific receptors and reported to successfully recover molecules from a hold-out test set.
Merk et al. (62,63) employed the same method, where their fine-tuned SMILES-based RNN was used to generate compounds to be later synthesized. The team then performed in vitro activity testing, reporting that four out of five, and in a later work two out of four, were active. Gupta et al. (59) experimented with fine-tuning on small data sets, reporting that even just a set of five molecules can lead to a model capable of generating unseen actives. Moreover, Moret et al. (34) also looked into the applicability of transfer learning in low data regimes when combined with data augmentation. With just five dissimilar natural products, they were able to generate structurally diverse molecules covering a broad range of scaffolds. Departing from sequence-based representations, Gebauer et al. (73) used transfer learning with their point set based model to target a specific value range of HOMO–LUMO gap, a molecular property relevant for the development of organic semiconductors.
Transfer learning can also be used as part of a larger procedure, as a mean to accelerate training and improve results. Notably, Li et al. (84) fine-tuned their model before generating molecules conditioned for a specific bioactivity profile. In a similar vein, Blaschke et al. (82) used transfer learning to focus the model toward features relevant to their objective, facilitating their subsequent reinforcement learning procedure.

Reinforcement Learning

A different approach to bias models toward generating molecules of interest is reinforcement learning (RL). RL outlines a framework where an agent, or system, interacts with an environment through a sequence of actions that are dictated by a policy and evaluated by a reward signal. The agent must iteratively revise the policy to improve the cumulative rewards over the full sequence of actions. This framework aims at learning a system capable of adopting the best set of actions in a given environment. (115,116) In de novo generation of molecules, RL has been applied in both sequence and graph-based approaches.
One application is to first pretrain a model through maximum likelihood estimation and then optimize it in a RL framework toward generating molecules with desired properties. This concept is presented in Figure 7 (top). In this vein, Segler et al. (18) biased their stacked RNN by coupling it to a prediction model and iteratively fine-tuning on the generated active compounds. With just eight iterations, the team reported the successful recovery of active compounds from the test set. Olivecrona et al. (58) trained a stacked RNN to generate molecules as SMILES and, using RL, optimized it toward generating analogues to celocoxib and generating molecules predicted as active against DRD2. Popova et al. (87) applied this concept to the sequential generation of graphs. Their model, termed MolecularRNN, was first trained to generate diverse realistic samples and then optimized for either quantitative estimate of drug-likeness (QED), melting point, and octanol–water partition coefficient (log P) penalized by synthetic accessibility (SA) and large rings.

Figure 7

Figure 7. Top: The model is first pretrained through maximum likelihood estimation, learning the structure of the output space along with general chemical rules. Then, using RL, the model is optimized for specific properties such as binding affinity or solubility. While similar in concept to transfer learning, the use of RL allows one to bias the model toward a wider range of objectives. Bottom: Directed generation with RL and GAN. This method leverages adversarial training to produce feasible molecules and RL to bias the generation toward desired properties.

High Resolution Image
Download MS PowerPoint Slide
Blaschke et al. (82) developed a production ready generative method, termed REINVENT2.0, based on a stacked RNN leveraging randomized SMILES and reinforcement learning. Later, Blaschke et al. (85) proposed a method to improve diversity in the REINVENT framework, termed memory-assisted RL. This method creates ”buckets” grouping similar generated molecules, once a bucket reaches a set capacity, subsequent molecules falling in that cluster are penalized. This memory unit, therefore, helps lead the model to unexplored areas of chemical space. The framework was employed to target a specific range of log P and optimize the predicted activity for HTR1A and for DRD2. The team noted an increase in the generation of diverse scaffolds, while producing highly scored compounds.
Zhavoronkov et al. (86) used RL to optimize a SMILES based VAE toward generating selective DDR1 kinase inhibitors. The objective was based on the predictions of an ensemble of three Self-Organizing Maps (SOMs) predicting general activity toward kinases, selectivity for DDR1, and novelty of generated molecules. From the generated compounds, six were selected for synthesis and in vitro testing with two of those being reported as both active and stable. Further in vivo testing was also performed on one molecule, with reasonable pharmacokinetic properties being reported. Lastly, the authors also noted the substantial reduction in both time and costs of their DL based approach compared to traditional drug development pipelines.
GANs and RL can also be combined to generate realistic, but optimized, molecules. Figure 7 (bottom) outlines this method. More specifically, a GAN is trained in a RL framework combining the adversarial reward with other relevant objectives. This method was employed by Guimaraes et al. (60) in ORGAN to generate SMILES optimizing molecular properties such as log P, SA, and QED. Improving on the previous method, Sanchez-Lengeling et al. (88) proposed ORGANIC, optimizing for melting point, drug-likeness with QED, and Lipinski’s rule-of-five and finally for nonfullerene electron acceptors for use in organic solar cells. Putin et al. (89) proposed ATNC, improving ORGANIC with a differentiable neural computer as generator and a novel reward function to improve the diversity of generated structures. With this model, the team optimized for similarity to known kinase inhibitors and synthesized a molecule similar to a generated one.
With a similar method but departing from sequence-based generation, You et al. (91) proposed GCPN to sequentially generate molecular graphs with optimized properties. Reporting that it can be used to target specific ranges of log P and molecular weight and also optimize log P penalized by SA and large rings, while constrained by similarity to a starting molecule. Karimi et al. (92) employed a similar molecular generative process to generate new drug combinations. Specifically, the proposed method aimed to directly generate sets of novel molecules that could be useful as disease-specific drug combinations. To this end, the team employed a RL process to sequentially generate sets of molecular graphs guided by a chemical validity reward, an adversarial reward enforcing SA and drug-likeness, and a network-based reward to help target the desired disease by incorporating prior knowledge from gene–gene, gene-disease, and disease–disease networks.
Also combining GANs and RL, De Cao and Kipf (68) proposed the generation of molecular graphs in a one-shot fashion in their model MolGAN.

Exploration and Exploitation of Molecules Latent Space

Instead of optimizing models for the desired properties, models based on the AEs architecture provide a latent representation of molecules that can be used for property optimization or targeted generation of compounds. These methods, which are illustrated in Figure 8, make use of well structured latent spaces, for which VAE and AAE are common choices. Different approaches have been suggested to navigate and shape the latent space of models leveraging both sequence and graph-based representations.

Figure 8

Figure 8. Here, the latent space of an AE is used as a reversible and continuous molecular representation allowing for the application of various optimization algorithms.

High Resolution Image
Download MS PowerPoint Slide

Bayesian Optimization

Bayesian Optimization (BO) is a sequential model-based optimization method suitable for black-box problems. It has two main elements, a probabilistic surrogate model and an acquisition function. The surrogate serves to estimate the objective function given some currently known data, then the acquisition function leverages the model to determine the best point in the objective function to evaluate. These new data are then used to update the surrogate model and the process repeats for a set of iterations, ideally leading to the global maximum of the objective function. Common choices are a Gaussian Process for the surrogate and the Expected Improvement for the acquisition function. (117,118)
In the de novo design, BO is used to optimize the properties of molecules by operating on their latent representation, using the decoder to reconstruct molecules from the suggested points. In this application, approximate inference is often used, in the form of a sparse Gaussian Process, due to the large number of evaluations that are made.
BO has been often used to demonstrate that the latent space of a particular architecture can be effectively navigated. In the context of molecular generation, it was first suggested by Ǵomez-Bombarelli et al. (10) who, in an earlier version of their work, optimized the log P of molecules penalized by their SA and presence of large rings. This methodology, and objective function, was then adopted in subsequent approaches by Kusner et al. (95) with GrammarVAE and Dai et al. (96) with SD-VAE which leveraged SMILES production rules and by Jin et al. (97) in JT-VAE and Samanta et al. (98) in NEVAE which dealt with molecular graphs. A more practical objective was employed by Blaschke et al. (94) who optimized the predicted DRD2 activity, reporting that BO was capable of effectively navigating the latent space of their uniform AAE and find novel active molecules. Lastly, Griffiths and Herńandez-Lobato (93) applied constrained BO as a way to mitigate training set mismatch, where the BO would visit latent points far from the training data that the model struggles to reconstruct. More specifically, the acquisition function was altered to only consider latent points which decoded to valid molecules with a tangible molecular weight. This was reported to lead to the generation of higher quality molecules using three drug-likeness metrics.

Genetic Algorithms and Particle Swarms

Once a continuous latent space is obtained, other optimization algorithms can be employed to optimize for desired properties. For example, Sattarov et al. (101) used a Genetic Algorithm (GA) to explore the latent space of their SMILES based seq2seq AE. Setting as goal optimizing molecules for activity toward the adenosine A2A receptor, they reported the generation of libraries enriched with actives and novel scaffolds. Winter et al. (102) also explored the latent space of a SMILES based seq2seq AE using a different meta-heuristic, Particle Swarm Optimization (PSO). They experimented with optimizing different properties, such as QED, penalized log P, activity toward EGFR, and activity toward BACE1. Furthermore, multiobjective optimization was also attempted by minimizing and maximizing activity for each of the receptors, reporting molecules with the desired activity profile and favorable absorption, distribution, metabolism, excretion, and toxicity properties.

Gradient-Based Methods

Gradient-based methods are alternative optimization algorithms that additionally require the objective to be differentiable. Although fulfilling the differentiability requirement might not always be possible, training a secondary neural network to predict the desired properties, in parallel with the main model, allows to obtain the gradient of a latent encoding with regard to the so desired property. This process has been used to optimize chemical properties through gradient ascent in some approaches, mainly as a benchmark for the smoothness of latent space.
Jin et al. (97) optimized the penalized log P constrained to a set degree of similarity to the starting molecule. More specifically, the team used a feed-forward network as a predictor and constrained the optimization with the Tanimoto similarity to the original molecule. Liu et al. (66) employed a similar methodology, optimizing however for the QED based on the gradients of a gated regression network. Bresson and Laurent (99) used gradient ascent with a single multilayer perceptron to, and following the work of Jin et al., (97) optimize the penalized log P and the log P constrained by similarity.

Deep Learning for Latent Space Navigation

Deep Learning models can be trained to operate on molecular encodings of a separate generative model. Prykhodko et al. (36) applied such an approach to their LatentGAN, a GAN trained to generate latent points of a separate SMILES heteroencoder. The real data used for training is obtained by passing SMILES through the encoder of the AE, while during the generation the novel latent points are converted into molecules using the decoder part of the AE. By training their model to produce realist encodings of compounds with activity for either EGFR, HTR1A, and S1PR1 (separate model for each), they reported the generation of valid and novel SMILES with a large percentage predicted as active. In a similar vein, Maziarka et al. (100) proposed mol-cycleGAN, a cycleGAN operating on the latent space of the JT-VAE (97) to optimize molecules, while keeping the results similar to the starting compound. Specifically, the model learns to convert molecules (points in latent space) from one set into another and back again. For example, it can convert from a set with only three aromatic rings into a set with only two, all while keeping transformed points similar to the original. The team reported that the model was capable of removing halogen moieties, replace bioesters, alter the number of rings, and increase the predicted activity toward the DRD2.
More recently, Chenthamarakshan et al. (103) applied a VAE generating molecules as SMILES and achieved controlled generation with Conditional Latent Attribute Space Sampling (CLaSS). (119) Under CLaSS, a Gaussian mixture model is trained to match the posterior of the trained encoder and a binary classifier trained for predicting desired properties from latent encodings.
Then, samples are drawn from the mixture model and filtered using the latent classifiers with only those predicted to have the specified attributes being decoded back to sequences. This method was employed to generate possible leads targeting SARS-CoV-2, using latent attribute predictors for three molecular properties and the binding affinity to relevant protein targets.

Conditioned and Semisupervised Generation

An alternative to both biasing the model and latent space optimization is to include explicit inputs to the model for controlling the properties of the generated molecules. As illustrated in Figure 9 (top) with an AE, this is achieved by introducing a condition vector to the models’ input, effectively biasing/conditioning the generation process toward the specified values. During training, the condition vector corresponds to various properties of the encoded molecule, leading the model to infer a correlation between the two. Later, during sampling, this vector can be altered, controlling the properties of the generated molecules.

Figure 9

Figure 9. Top: In conditioned generation, the desired properties are introduced as explicit inputs to the model. These properties are precomputed for each compound of the training set and used during training to induce a correlation between the two. This correlation is then leveraged during the generation process to target specific property values. Bottom: In the semisupervised case of conditioned generation, only part of the training set has the desired properties available. To overcome this, a predictor network is trained on the labeled instances and used to predict the properties of unlabeled ones.

High Resolution Image
Download MS PowerPoint Slide
Li et al. (84) used a sequential graph generator based on GNNs and added a conditioning vector at each step of the graph building process. Their model was conditioned on molecular scaffolds, QED, SA and, after fine-tuning, the predicted activity toward JN3 and GSK3β. The team was able to successfully generate inhibitors for either receptor, as well as dual inhibitors. A similar process was employed by Li et al. (65) to condition a GNN for sequential graph generation by appending the condition vector solely to the initial node states. Conditioning on the number of atoms, bonds and aromatic rings, the model was able to extrapolate and successfully generate molecules when conditioned with values outside the training data.
Kotsias et al. (104) conditioned a SMILES based stacked RNN by setting the conditioning vector as the initial internal state of the network. Two different approaches were compared, either conditioning with molecular fingerprints or building the condition vector with molecular properties such as Topological Polar Surface Area (TPSA), molecular weight, and bioactivity. They reported that while both approaches could generate molecules satisfying the desired properties, the fingerprint-based model generated molecules with scaffolds similar to the seed compound, facilitating the encoding of structural restrictions. Meanwhile, the property-based model generated more dissimilar scaffolds, enabling a more versatile exploration of chemical space.
The conditioned generation methodology is also applicable to AEs, where the condition vector is generally appended to the input of both the encoder and decoder. For example,
Simonovsky and Komodakis (69) demonstrated conditioned generation with their GraphVAE by controlling the number of heavy atoms in generated molecules. Lim et al. (61) used a conditioned VAE to control molecular weight, TPSA and number of H+ donors and acceptors, reporting independent control of properties, as well as generating molecules with properties beyond those seen during training. Working with a 3D representation of molecules, Skalic et al. (72) proposed to condition their shape-based VAE with the location of pharmacophores. Specifically, the decoder received a 3D shape constructed by placing property points close to atoms with that property. The team noted that conditioning the reconstruction of random latent points often led to implausible output shapes, but conditioning the decoding of seed molecules improved the reconstruction of pharmacophore features.
When the desired properties are not readily determinable and no large labeled data sets are available, common with bioactivity data, semisupervised-AE can be used, Figure 9 (bottom). A semisupervised-AE consists of an AE with an added predictor network, which receives the molecule as input and outputs its properties. These are then appended to both the input and output of the encoder, ultimately conditioning the decoder. The architecture is termed semisupervised because the data set does not need to be fully labeled. Labeled instances are used for training the predictor, replacing its output, while unlabeled samples have their properties predicted by the predictor network. (120)
Kang and Cho (110) employed a semisupervised-VAE conditioned on molecular weight, log P, and QED and experimented with various fractions of labeled/unlabeled data used for training. Polykovskiy et al. (64) evaluated the application of different disentanglement techniques to a semisupervised-AAE. With their most successful method, termed semisupervised entangled AAE, the team was able to generate molecules conditioned on the activity toward the Janus kinase 2 and Janus kinase 3. By setting low activity for JK2 but high activity for JK3, they generated a set of selective inhibitors that were then filtered. A single molecule was synthesized and reported to have in vitro activity and selectivity for the Janus kinase 3.
Lim et al. (111) used a conditioned VAE to extend molecular scaffolds, producing molecules with predetermined scaffolds and desired properties. The model was reported to successfully condition the molecular weight, TPSA and log P. Furthermore, a semisupervised extension of the model was used to design EGFR inhibitors, reporting a significant enhancement of the predicted inhibition potency.
More recently, Ḿendez-Lucio et al. (106) used a VAE in conjunction with a GAN conditioned on gene expression data. More specifically, a two-stage GAN is trained to generate latent points of a GrammarVAE, (95) with its decoder serving to reconstruct the points generated by the GAN. Here, the conditioning vector is an input to the generator, alongside Gaussian noise. Also, two networks, one per stage, predict whether the generated latent points correspond to the gene expression profiles used for conditioning the generator. Using this method, the team conditioned the model on the gene expression of ten knockouts of pharmacological interest, reporting the generation of molecules similar to known active compounds.
Born et al. (107) proposed PaccMannRL, a framework that leverages gene expression data and combines reinforcement learning with conditioned VAEs. This method starts by training two separate VAEs, one to reconstruct molecules, represented as SMILES, and the other to reconstruct gene expression data. The two models are then combined, with the output of both encoders being summed together and used as input to the molecular decoder. This new architecture is then trained through reinforcement learning toward generating molecules targeting the specified gene expression profile. The framework was then employed to generate anticancer compounds, with the team reporting an improvement in predicted efficacy while maintaining similar validity scores.
Also exploiting omics data, Shayakhmetov et al. (105) proposed a SMILES based conditioned AAE to generate molecules capable of inducing a desired transcriptomic change. Particularly, the model is composed of two AAEs, one tasked with reconstructing molecules and the other with reconstructing gene expression profiles, with the conditioning vector being produced by the gene expression encoder as a separate latent vector. As such, this architecture produces a three-part latent space, with one part meant to encode molecule specific information, another one to encode expression specific information, and the last to encode features relevant to both. This specific arrangement was designed to aid the model in ignoring nonrelevant cellular processes that are included in the gene expression profile from the changes induced by the molecule.
Masuda et al. (109) proposed to condition their previously proposed model, generating molecules as 3D structures, with the 3D structure of the intended binding site. Specifically, a separate encoder module was added to encode the binding target structure into a latent representation, which was then concatenated to the output of the molecular encoder, conditioning it. With this approach, the team reported that, by sampling around a seed molecule, they often could generate new compounds with better binding affinities.
A different approach was taken by Jin et al. (108) who adapted JT-VAE to optimize input molecules by adding adversarial training and a conditioning vector to the latent encoding. That is, the model was trained to ”translate” molecules missing the desired properties into molecules with those qualities. The adversarial objective ensures that the new molecule has the desired properties, while the conditioning vector directs the generation process. With this method, the team reported success when optimizing for the penalized log P, QED and the predicted activity toward DRD2 while constraining by the similarity to the initial molecule.

Synthetic Accessibility

Successfully applying these methods to practical use cases will inevitably depend on synthesizing the novel compounds. However, ensuring the SA of the generated molecules is a major hurdle that often goes unnoticed.
Seeking to help address this, Gao and Coley (121) discussed and compared three approaches capable of guiding generative models toward synthesizable compounds. Specifically, they considered the use of synthesizability scores and retro-synthetic analysis to filter generated molecules, filter the training data set or to modify the objective function used for targeted generation. The team reported that, despite useful, the first two methods often proved insufficient. Furthermore, modifying the objective improved synthesizability, but at the cost of the main goals.
Horwood and Noutahi (122) directly addressed this issue in their RL based method by iteratively building novel molecules through a series of chemical reactions. With this approach, the team reported the successful optimization of multiple objectives, while maintaining good diversity among generated molecules and also providing a valid synthetic route for every novel molecule.
A similar approach was taken by Gottipati et al., (123) which also applied RL to generate molecules using a sequence of chemical reactions. Like the previous method, the team reported attaining high scoring generated compounds while ensuring SA.
Bradshaw et al. (124) also proposed to generate novel molecules by recursively combining simple building blocks through a series of chemical reactions. The team proposed two variants of their approach, one where RL is employed to perform targeted generation and, departing from the two previous works, a Wasserstein Autoencoder (WAE) based model. The later one, as discussed in the previous sections, should allow other methods of targeted generation to be used.

Current Applications

ARTICLE SECTIONS
Jump To
There are multiple proposed approaches to not only generate molecules, but also do so in a directed manner, controlling and optimizing for desired properties. The adopted objectives have more often been employed as a benchmark of the proposed methodologies than as goals in themselves. While some have been of limited real use, like maximizing log P penalized by SA and large rings, others such as optimizing for specific bioactivity profiles can have a more direct application in fields like drug discovery. For example, the swiftness of these methods has recently found use in the creation of therapeutic leads for SARS-CoV-2.

Drug Development

The large costs associated with drug development has often led to the implementation of various computational tools to assist and accelerate the process. As such, a large focus has been placed on applying deep generative models to various stages of early drug development. The most commonly suggested application has been to the de novo drug design. Indeed, the bulk of methods here described are capable of generating broad molecular libraries, with a large part capable of focused generation. These libraries can then be used for virtual screening or high-throughput screening, hopefully exploring previously unseen regions of the chemical space.
A different suggested application has been for molecule optimization. Although the usefulness of some objectives has been questioned, the ability to, for example, optimize solubility or improve SA while maintaining high similarity to an original structure can help inform lead optimization. Furthermore, easily and reliably finding close novel analogues to a molecule or introducing specific modifications like substituting bioesters can also be of practical use. Lastly, some attention has also been given to fragment-based drug development, with methods proposed for growing a molecule from a specified fragment (59) or linking two fragments together. (125,126)
Several different biological targets have also deserved attention. Table 3 summarizes these applications. It notes the various instances of experimental validation of de novo generated molecules that have already been performed. These are mostly successful in vitro activity testing, with two instances of in vivo validation. These successful practical realizations encourage further research into this blooming field.
Table 3. Experimental Validation of Molecules Generated with Generative DL
  activity 
targetdirected generationin silicoin vitroin vivoref
RXR, PPARtransfer learningSPiDER5 synthesized; reported: 4 active  (63)
RXRtransfer learningSPiDER WHALES4 synthesized; reported: 2 active  (62)
JK3 selectivereinforcement learningdocking1 synthesized; reported: active and selective for JK3  (64)
kinase inhibitorsreinforcement learning 50 purchased (similar); reported: 7 active  (89)
DRD2, 5-HT1A, 5-HT2Atransfer learningMT-DNN on ECFP41+6 analogues synthesized; reported: active for the 3 receptors1+6 analogues; 1 active and acceptable safety (127)
VEGFR-2train on activesdocking5 synthesized; reported: 3 active and noncytotoxic  (17)
DDR1reinforcement learningSOM pharmacophore6 synthesized; reported: 2 active and stable1 tested; half-life 3.5 h; 10 analogues tested (86)
p300/CBP inhibitorstransfer learningdocking1+26 analogues synthesized; reported: active, selective, and stablegood bioavailability, efficacy, safety (128)
LXR agoniststransfer learning 25 synthesized, 3 purchased; reported: 12 active  (129)

COVID-19

With the recent SARS-CoV-2 pandemic, generative DL methods became an attractive option for the de novo design of possible therapeutic leads. Indeed, in the span of a few months, several approaches were proposed and a large number of potential compounds were shared.
Bung et al. (83) biased a SMILES based stacked RNN with transfer learning toward generating possible binders of SARS-CoV-2 proteases. Specifically, a set of 1.6 million molecules from ChEMBL were used for pretraining and then around 2500 protease inhibitors were used for fine-tuning the model. Furthermore, reinforcement learning was used to control other molecular properties such as QED, log P, molecular weight and SA. After screening the generated molecules using docking simulations, the team shared 31 potential leads.
Shaker et al. (130) used their SMILES-based in-house model, named Rosalind, to generate molecules targeting the SARS-CoV-2 main protease Mpro. After applying filters for binding affinity predicted by docking, QED, molecular weight, structural alerts and predicted toxicity the team shared a list of 40 compounds.
Chenthamarakshan et al., (103) as described earlier, employed CLaSS on a SMILES based VAE to generate compounds with favorable binding to three relevant target proteins of SARS-CoV-2. The controlled sampling leveraged property predictors for QED, SA, log P, and binding affinity to a specified protein. Filters were then applied for toxicity, retrosynthesis prediction and binding affinity to the relevant target with docking and a set of 3.5K potential leads was shared by the team.
Zhavoronkov et al. (86) used an internal pipeline leveraging 28 different models and various molecular representations to generate molecules targeting the SARS-CoV-2 main protease Mpro. Docking was employed to rank the compounds by their affinity and a set of 10 molecules was shared by the team.
Born et al. (131) adapted their previously proposed framework, PaccMannRL, to generate compounds addressing 41 targets of SARS-CoV-2. Specifically, the molecular VAE was adapted to generate SELFIES (instead of SMILES) and the conditioning was performed based on a protein VAE (instead of gene expression data).

Organic Photovoltaics

The design of new organic photovoltaics has great potential to help reduce the dependence on fossil fuels by enabling cheaper and more efficient solar energy. (132) Deep generative methods can be employed to quickly design new molecules targeting desired properties such as a specific HOMO–LUMO gap and high Power Conversion Efficiency (PCE). Indeed, some attention has already been devoted to this application.
For instance, Sanchez-Lengeling et al. (88) applied ORGANIC toward generating nonfullerene electron acceptors for use in organic solar panels, reporting an increase in the average predicted PCE. Jørgensen et al. (133) used a GrammarVAE (95) for generating donoracceptor polymers optimized for a specific range of optical gap and LUMO energy. Griffiths and Hernández-Lobato (93) employed their proposed constrained BO method toward generating molecules optimized for PCE, reporting that the averaged score of the generated molecules lied above the 90th percentile of the training data. Gebauer et al. (73) leveraged transfer learning to fine-tune their model toward generating molecules targeting a specific HOMO–LUMO gap. Similarly, Yuan et al. (134) biased a RNN with transfer learning to generate donor–acceptor oligomers targeting a specific HOMO–LUMO gap.

Future Directions of Research

ARTICLE SECTIONS
Jump To
Over the last years, the field of deep generative learning for molecular design has seen an explosion of interest, with a great number of different and novel approaches being proposed. Within this blooming field, some trends emerged that have somewhat guided novel research and give clues for the direction of future developments.
Many of the early approaches to generating new molecules with DL borrowed from the NLP field, using RNNs for modeling and generating molecules as sequences. However, in the meantime, the state-of-the-art in language processing has evolved to leverage attention mechanisms within architectures like the transformer, BERT and GPT-3. (135−137) Indeed, these advances have begun to trickle into the field of generative molecular design with recent works, leveraging the transformer architecture to approach the generation of new bioactive compounds as a translation from the amino acid sequence of a target protein to an active SMILES, (138) to link fragments together (126) or to perform scaffold hopping. (139) Likewise, approaches leveraging these new methods, and others that may emerge from NLP, are likely to receive further attention in the future.
With regard to molecular representations, while early methods dealt mainly with SMILES, (18,58) several subsequent works sought a more meaningful representation by directly generating molecular graphs. (87,91) The pursuit of this goal has led to the proposal of novel graph generation procedures and will likely inspire interesting new research.
Nevertheless, graph-based representations still disregard the three-dimensional nature of molecules, possibly missing useful information by, for example, neglecting chirality. As such, recent interest has formed around the generation of 3D molecular structures leading to the proposal of a few methods (72−74) and, hopefully, motivating future works.
An interesting recent trend has been to leverage omics data to help direct the generation process. (105−107) Due to the large quantities of omics data already available and the information that it can convey, this strategy has the potential to be extremely useful. A handful of approaches have already been suggested and further work along these lines will likely be developed.
The synthetic accessibility of generated compounds, which is vital for the practical realization of these methods, still poses a significant hurdle to overcome. Some methods have recently emerged taking this into consideration, (121−124) while it is expected that future approaches will continue to address this issue.
With regard to new applications, methods targeting the biotechnological industry, such as novel artificial flavorings, dyes, catalysts, or pesticides, could be very relevant. Furthermore, employing methods leveraging omics data to aid in metabolic engineering tasks might also be a promising avenue for future work.
Lastly, some evidence has been reported in favor of using multiple generative models in parallel to cover different regions of the chemical space. (36) This further motivates new research and development in this area, as different architectures and methods could complement each other, learning specific chemical patterns and enabling a more diverse approach to the exploration of chemical space.

Author Information

ARTICLE SECTIONS
Jump To
  • Corresponding Author
  • Authors
    • Tiago Sousa - Centre of Biological Engineering, Campus Gualtar, University of Minho, 4710-057 Braga, Portugal;  Orcidhttps://orcid.org/0000-0003-4013-7012
    • João Correia - Centre of Biological Engineering, Campus Gualtar, University of Minho, 4710-057 Braga, Portugal
    • Vítor Pereira - Centre of Biological Engineering, Campus Gualtar, University of Minho, 4710-057 Braga, Portugal
  • Notes
    The authors declare no competing financial interest.

Acknowledgments

ARTICLE SECTIONS
Jump To

This project has received funding from the European Union’s Horizon 2020 research and innovation programme (Grant Agreement Number 814408).

References

ARTICLE SECTIONS
Jump To

This article references 139 other publications.

  1. 1
    Polishchuk, P. G.; Madzhidov, T. I.; Varnek, A. Estimation of the size of drug-like chemical space based on GDB-17 data. J. Comput.-Aided Mol. Des. 2013, 27, 675,  DOI: 10.1007/s10822-013-9672-4
    [Crossref], [PubMed], [CAS], Google Scholar
    1
    Estimation of the size of drug-like chemical space based on GDB-17 data
    Polishchuk, P. G.; Madzhidov, T. I.; Varnek, A.
    Journal of Computer-Aided Molecular Design (2013), 27 (8), 675-679CODEN: JCADEQ; ISSN:0920-654X. (Springer)
    The goal of this paper is to est. the no. of realistic drug-like mols. which could ever be synthesized. Unlike previous studies based on exhaustive enumeration of mol. graphs or on combinatorial enumeration preselected fragments, we used results of constrained graphs enumeration by Reymond to establish a correlation between the no. of generated structures (M) and the no. of heavy atoms (N): logM = 0.584 × N × logN + 0.356. The no. of atoms limiting drug-like chem. space of mols. which follow Lipinsky's rules (N = 36) has been obtained from the anal. of the PubChem database. This results in M ≈ 1033 which is in between the nos. estd. by Ertl (1023) and by Bohacek (1060).
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC3sXhtlWjtbjM&md5=80a95002031f04319351c7e342bc1a55
  2. 2
    Schneider, G. Automating drug discovery. Nat. Rev. Drug Discovery 2018, 17, 97– 113,  DOI: 10.1038/nrd.2017.232
    [Crossref], [PubMed], [CAS], Google Scholar
    2
    Automating drug discovery
    Schneider, Gisbert
    Nature Reviews Drug Discovery (2018), 17 (2), 97-113CODEN: NRDDAG; ISSN:1474-1776. (Nature Research)
    A review. Small mol. drug discovery can be viewed as a challenging multidimensional problem in which various characteristics of compds. including efficacy, pharmacokinetics and safety need to be optimized in parallel to provide drug candidates. Recent advances in areas such as microfluidics-assisted chem. synthesis and biol. testing, as well as artificial intelligence systems that improve a design hypothesis through feedback anal., are now providing a basis for the introduction of greater automation into aspects of this process. This could potentially accelerate time frames for compd. discovery and optimization and enable more effective searches of chem. space. However, such approaches also raise considerable conceptual, tech. and organizational challenges, as well as scepticism about the current hype around them. This article aims to identify the approaches and technologies that could be implemented robustly by medicinal chemists in the near future and to critically analyze the opportunities and challenges for their more widespread application.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC2sXhvFOntbbK&md5=4fc1e441a0a31d7421642eb91fbac8c7
  3. 3
    DiMasi, J. A.; Grabowski, H. G.; Hansen, R. W. Innovation in the pharmaceutical industry: New estimates of R&D costs. Journal of Health Economics 2016, 47, 20– 33,  DOI: 10.1016/j.jhealeco.2016.01.012
    [Crossref], [PubMed], [CAS], Google Scholar
    3
    Innovation in the pharmaceutical industry: New estimates of R&D costs
    DiMasi Joseph A; Grabowski Henry G; Hansen Ronald W
    Journal of health economics (2016), 47 (), 20-33 ISSN:.
    The research and development costs of 106 randomly selected new drugs were obtained from a survey of 10 pharmaceutical firms. These data were used to estimate the average pre-tax cost of new drug and biologics development. The costs of compounds abandoned during testing were linked to the costs of compounds that obtained marketing approval. The estimated average out-of-pocket cost per approved new compound is $1395 million (2013 dollars). Capitalizing out-of-pocket costs to the point of marketing approval at a real discount rate of 10.5% yields a total pre-approval cost estimate of $2558 million (2013 dollars). When compared to the results of the previous study in this series, total capitalized costs were shown to have increased at an annual rate of 8.5% above general price inflation. Adding an estimate of post-approval R&D costs increases the cost estimate to $2870 million (2013 dollars).
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A280%3ADC%252BC28jls1egsw%253D%253D&md5=a78649853406ddbfe54abb0052ca964b
  4. 4
    Ruddigkeit, L.; van Deursen, R.; Blum, L. C.; Reymond, J.-L. Enumeration of 166 Billion Organic Small Molecules in the Chemical Universe Database GDB-17. J. Chem. Inf. Model. 2012, 52, 2864– 2875,  DOI: 10.1021/ci300415d
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    4
    Enumeration of 166 Billion Organic Small Molecules in the Chemical Universe Database GDB-17
    Ruddigkeit, Lars; van Deursen, Ruud; Blum, Lorenz C.; Reymond, Jean-Louis
    Journal of Chemical Information and Modeling (2012), 52 (11), 2864-2875CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    Drug mols. consist of a few tens of atoms connected by covalent bonds. How many such mols. are possible in total and what is their structure. This question is of pressing interest in medicinal chem. to help solve the problems of drug potency, selectivity, and toxicity and reduce attrition rates by pointing to new mol. series. To better define the unknown chem. space, we have enumerated 166.4 billion mols. of up to 17 atoms of C, N, O, S, and halogens forming the chem. universe database GDB-17, covering a size range contg. many drugs and typical for lead compds. GDB-17 contains millions of isomers of known drugs, including analogs with high shape similarity to the parent drug. Compared to known mols. in PubChem, GDB-17 mols. are much richer in nonarom. heterocycles, quaternary centers, and stereoisomers, densely populate the third dimension in shape space, and represent many more scaffold types.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC38XhsFClsL3J&md5=d0bf9a29f3e9ae1e57bb1c953a562ced
  5. 5
    Walters, W. P. Virtual Chemical Libraries: Miniperspective. J. Med. Chem. 2019, 62, 1116– 1124,  DOI: 10.1021/acs.jmedchem.8b01048
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    5
    Virtual Chemical Libraries
    Walters, W. Patrick
    Journal of Medicinal Chemistry (2019), 62 (3), 1116-1124CODEN: JMCMAR; ISSN:0022-2623. (American Chemical Society)
    A review. Advances in computer processing speed and storage capacity have enabled researchers to generate virtual chem. libraries contg. billions of mols. While these nos. appear large, they are only a small fraction of the no. of org. mols. that could potentially be synthesized. This review provides an overview of recent advances in the generation and use of virtual chem. libraries in medicinal chem. The authors also consider the practical implications of these libraries in drug discovery programs and highlight a no. of current and future challenges.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXhsFyqtbnM&md5=ede412935f56e350915f35e79f1b969f
  6. 6
    Hartenfeller, M.; Zettl, H.; Walter, M.; Rupp, M.; Reisen, F.; Proschak, E.; Weggen, S.; Stark, H.; Schneider, G. DOGS: reaction-driven de novo design of bioactive com- pounds. PLoS Comput. Biol. 2012, 8, e1002380  DOI: 10.1371/journal.pcbi.1002380
    [Crossref], [PubMed], [CAS], Google Scholar
    6
    DOGS: reaction-driven de novo design of bioactive compounds
    Hartenfeller, Markus; Zettl, Heiko; Walter, Miriam; Rupp, Matthias; Reisen, Felix; Proschak, Ewgenij; Weggen, Sascha; Stark, Holger; Schneider, Gisbert
    PLoS Computational Biology (2012), 8 (2), e1002380CODEN: PCBLBG; ISSN:1553-7358. (Public Library of Science)
    We present a computational method for the reaction-based de novo design of drug-like mols. The software DOGS (Design of Genuine Structures) features a ligand-based strategy for automated 'in silico' assembly of potentially novel bioactive compds. The quality of the designed compds. is assessed by a graph kernel method measuring their similarity to known bioactive ref. ligands in terms of structural and pharmacophoric features. We implemented a deterministic compd. construction procedure that explicitly considers compd. synthesizability, based on a compilation of 25'144 readily available synthetic building blocks and 58 established reaction principles. This enables the software to suggest a synthesis route for each designed compd. Two prospective case studies are presented together with details on the algorithm and its implementation. De novo designed ligand candidates for the human histamine H4 receptor and γ-secretase were synthesized as suggested by the software. The computational approach proved to be suitable for scaffold-hopping from known ligands to novel chemotypes and for generating bioactive mols. with drug-like properties.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC38Xjt1Sjtr0%253D&md5=cdcf6905d52761614c33ab068b608ac5
  7. 7
    Spiegel, J.; Durrant, J. AutoGrow4: An open-source genetic algorithm for de novo drug design and lead optimization. J. Cheminf. 2020, 12, 25,  DOI: 10.1186/s13321-020-00429-4
    [Crossref], [CAS], Google Scholar
    7
    AutoGrow4: an open-source genetic algorithm for de novo drug design and lead optimization
    Spiegel, Jacob O.; Durrant, Jacob D.
    Journal of Cheminformatics (2020), 12 (1), 25CODEN: JCOHB3; ISSN:1758-2946. (SpringerOpen)
    Abstr.: We here present AutoGrow4, an open-source program for semi-automated computer-aided drug discovery. AutoGrow4 uses a genetic algorithm to evolve predicted ligands on demand and so is not limited to a virtual library of pre-enumerated compds. It is a useful tool for generating entirely novel drug-like mols. and for optimizing preexisting ligands. By leveraging recent computational and cheminformatics advancements, AutoGrow4 is faster, more stable, and more modular than previous versions. It implements new docking-program compatibility, chem. filters, multithreading options, and selection methods to support a wide range of user needs. To illustrate both de novo design and lead optimization, we here apply AutoGrow4 to the catalytic domain of poly(ADP-ribose) polymerase 1 (PARP-1), a well characterized DNA-damage-recognition protein. AutoGrow4 produces drug-like compds. with better predicted binding affinities than FDA-approved PARP-1 inhibitors (pos. controls). The predicted binding modes of the AutoGrow4 compds. mimic those of the known inhibitors, even when AutoGrow4 is seeded with random small mols. AutoGrow4 is available under the terms of the Apache License, Version 2.0. A copy can be downloaded free of charge from <a><a><a> not available: see fulltext].
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXnsVWksrs%253D&md5=10cb3d13c137b5e93fc3f92568edba22
  8. 8
    Jensen, J. H. A graph-based genetic algorithm and generative model/Monte Carlo tree search for the exploration of chemical space. Chem. Sci. 2019, 10, 3567– 3572,  DOI: 10.1039/C8SC05372C
    [Crossref], [PubMed], [CAS], Google Scholar
    8
    A graph-based genetic algorithm and generative model/Monte Carlo tree search for the exploration of chemical space
    Jensen, Jan H.
    Chemical Science (2019), 10 (12), 3567-3572CODEN: CSHCCN; ISSN:2041-6520. (Royal Society of Chemistry)
    This paper presents a comparison of a graph-based genetic algorithm (GB-GA) and machine learning (ML) results for the optimization of log P values with a constraint for synthetic accessibility and shows that the GA is as good as or better than the ML approaches for this particular property. The mols. found by the GB-GA bear little resemblance to the mols. used to construct the initial mating pool, indicating that the GB-GA approach can traverse a relatively large distance in chem. space using relatively few (50) generations. The paper also introduces a new non-ML graph-based generative model (GB-GM) that can be parameterized using very small data sets and combined with a Monte Carlo tree search (MCTS) algorithm. The results are comparable to previously published results (Sci. Technol. Adv. Mater., 2017, 18, 972-976) using a recurrent neural network (RNN) generative model, and the GB-GM-based method is several orders of magnitude faster. The MCTS results seem more dependent on the compn. of the training set than the GA approach for this particular property. Our results suggest that the performance of new ML-based generative models should be compared to that of more traditional, and often simpler, approaches such a GA.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXislClt78%253D&md5=108c19e322025e736330203e0d312237
  9. 9
    Yoshikawa, N.; Terayama, K.; Sumita, M.; Homma, T.; Oono, K.; Tsuda, K. Population-based De Novo Molecule Generation, Using Grammatical Evolution. Chem. Lett. 2018, 47, 1431– 1434,  DOI: 10.1246/cl.180665
    [Crossref], [CAS], Google Scholar
    9
    Population-based de novo molecule generation, using grammatical evolution
    Yoshikawa, Naruki; Terayama, Kei; Sumita, Masato; Homma, Teruki; Oono, Kenta; Tsuda, Koji
    Chemistry Letters (2018), 47 (11), 1431-1434CODEN: CMLTAG; ISSN:0366-7022. (Chemical Society of Japan)
    Automatic mol. design with machine learning and simulations has shown a remarkable ability to generate new and promising drug candidates. We propose a new population-based approach using a grammatical evolution named ChemGE, that can update a large population of mols. concurrently and evaluate with multiple simulators in parallel. In computational expts., ChemGE succeeded in finding hundreds of candidate mols. whose affinity for thymidine kinase is better than that of known binding mols. in a database (DUD-E).
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXitlKnsLvP&md5=ccb65d8a0d2bf9b350b4607ef2e70f67
  10. 10
    Ǵomez-Bombarelli, R.; Wei, J. N.; Duvenaud, D.; Hernández-Lobato, J. M.; Śanchez- Lengeling, B.; Sheberla, D.; Aguilera-Iparraguirre, J.; Hirzel, T. D.; Adams, R. P.; Aspuru-Guzik, A. Automatic Chemical Design Using a Data-Driven Continuous Rep- resentation of Molecules. ACS Cent. Sci. 2018, 4, 268– 276,  DOI: 10.1021/acscentsci.7b00572
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    10
    Automatic Chemical Design Using a Data-Driven Continuous Representation of Molecules
    Gomez-Bombarelli, Rafael; Wei, Jennifer N.; Duvenaud, David; Hernandez-Lobato, Jose Miguel; Sanchez-Lengeling, Benjamin; Sheberla, Dennis; Aguilera-Iparraguirre, Jorge; Hirzel, Timothy D.; Adams, Ryan P.; Aspuru-Guzik, Alan
    ACS Central Science (2018), 4 (2), 268-276CODEN: ACSCII; ISSN:2374-7951. (American Chemical Society)
    We report a method to convert discrete representations of mols. to and from a multidimensional continuous representation. This model allows us to generate new mols. for efficient exploration and optimization through open-ended spaces of chem. compds. A deep neural network was trained on hundreds of thousands of existing chem. structures to construct three coupled functions: an encoder, a decoder, and a predictor. The encoder converts the discrete representation of a mol. into a real-valued continuous vector, and the decoder converts these continuous vectors back to discrete mol. representations. The predictor ests. chem. properties from the latent continuous vector representation of the mol. Continuous representations of mols. allow us to automatically generate novel chem. structures by performing simple operations in the latent space, such as decoding random vectors, perturbing known chem. structures, or interpolating between mols. Continuous representations also allow the use of powerful gradient-based optimization to efficiently guide the search for optimized functional compds. We demonstrate our method in the domain of drug-like mols. and also in a set of mols. with fewer that nine heavy atoms.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXntlWquw%253D%253D&md5=322d9ff569fc9c8831e91d915104d985
  11. 11
    Gawehn, E.; Hiss, J. A.; Schneider, G. Deep learning in drug discovery. Mol. Inf. 2016, 35, 3– 14,  DOI: 10.1002/minf.201501008
    [Crossref], [PubMed], [CAS], Google Scholar
    11
    Deep Learning in Drug Discovery
    Gawehn, Erik; Hiss, Jan A.; Schneider, Gisbert
    Molecular Informatics (2016), 35 (1), 3-14CODEN: MIONBS; ISSN:1868-1743. (Wiley-VCH Verlag GmbH & Co. KGaA)
    Artificial neural networks had their first heyday in mol. informatics and drug discovery approx. two decades ago. Currently, we are witnessing renewed interest in adapting advanced neural network architectures for pharmaceutical research by borrowing from the field of "deep learning". Compared with some of the other life sciences, their application in drug discovery is still limited. Here, we provide an overview of this emerging field of mol. informatics, present the basic concepts of prominent deep learning methods and offer motivation to explore these techniques for their usefulness in computer-assisted drug discovery and design. We specifically emphasize deep neural networks, restricted Boltzmann machine networks and convolutional networks.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC2MXitV2rs7bE&md5=bd2553359c4b824c5272388462772659
  12. 12
    Goodfellow, I.; Bengio, Y.; Courville, A. Deep Learning; MIT Press, 2016.
    Google Scholar
    There is no corresponding record for this reference.
  13. 13
    Chollet, F. Deep learning with Python; Manning Publications Co: Shelter Island, NY, 2018.
    Google Scholar
    There is no corresponding record for this reference.
  14. 14
    Foster, D.; Safari, A. O. M. C. Generative deep learning: teaching machines to paint, write, compose, and play; O’Reilly Media, 2019.
    Google Scholar
    There is no corresponding record for this reference.
  15. 15
    White, D.; Wilson, R. C. Generative models for chemical structures. J. Chem. Inf. Model. 2010, 50, 1257– 1274,  DOI: 10.1021/ci9004089
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    15
    Generative Models for Chemical Structures
    White, David; Wilson, Richard C.
    Journal of Chemical Information and Modeling (2010), 50 (7), 1257-1274CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    We apply recently developed techniques for pattern recognition to construct a generative model for chem. structure. This approach can be viewed as ligand-based de novo design. We construct a statistical model describing the structural variations present in a set of mols. which may be sampled to generate new structurally similar examples. We prevent the possibility of generating chem. invalid mols., according to our implicit hydrogen model, by projecting samples onto the nearest chem. valid mol. By populating the input set with mols. that are active against a target, we show how new mols. may be generated that will likely also be active against the target.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC3cXos1ersr4%253D&md5=f694ddd48c404ebea7792e4a52458ebb
  16. 16
    Sanchez-Lengeling, B.; Aspuru-Guzik, A. Inverse molecular design using machine learning: Generative models for matter engineering. Science 2018, 361, 360– 365,  DOI: 10.1126/science.aat2663
    [Crossref], [PubMed], [CAS], Google Scholar
    16
    Inverse molecular design using machine learning: Generative models for matter engineering
    Sanchez-Lengeling, Benjamin; Aspuru-Guzik, Alan
    Science (Washington, DC, United States) (2018), 361 (6400), 360-365CODEN: SCIEAS; ISSN:0036-8075. (American Association for the Advancement of Science)
    The discovery of new materials can bring enormous societal and technol. progress. In this context, exploring completely the large space of potential materials is computationally intractable. Here, we review methods for achieving inverse design, which aims to discover tailored materials from the starting point of a particular desired functionality. Recent advances from the rapidly growing field of artificial intelligence, mostly from the subfield of machine learning, have resulted in a fertile exchange of ideas, where approaches to inverse mol. design are being proposed and employed at a rapid pace. Among these, deep generative models have been applied to numerous classes of materials: rational design of prospective drugs, synthetic routes to org. compds., and optimization of photovoltaics and redox flow batteries, as well as a variety of other solid-state materials.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXhtlyitr3L&md5=779c4a42ba1e84d99d13ad1b32b9529a
  17. 17
    Yuan, W.; Jiang, D.; Nambiar, D. K.; Liew, L. P.; Hay, M. P.; Bloomstein, J.; Lu, P.; Turner, B.; Le, Q.-T.; Tibshirani, R.; Khatri, P.; Moloney, M. G.; Koong, A. C. Chemical Space Mimicry for Drug Discovery. J. Chem. Inf. Model. 2017, 57, 875– 882,  DOI: 10.1021/acs.jcim.6b00754
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    17
    Chemical Space Mimicry for Drug Discovery
    Yuan, William; Jiang, Dadi; Nambiar, Dhanya K.; Liew, Lydia P.; Hay, Michael P.; Bloomstein, Joshua; Lu, Peter; Turner, Brandon; Le, Quynh-Thu; Tibshirani, Robert; Khatri, Purvesh; Moloney, Mark G.; Koong, Albert C.
    Journal of Chemical Information and Modeling (2017), 57 (4), 875-882CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    The authors describe a new library generation method, Machine-based Identification of Mols. Inside Characterized Space (MIMICS) that generates sets of mols. inspired by a text-based input. MIMICS-generated libraries were found to preserve distributions of properties while simultaneously increasing structural diversity. Newly identified MIMICS-generated compds. were found to be bioactive as inhibitors of specific components of the unfolded protein response (UPR) and the VEGFR2 pathway in cell-based assays, thus confirming that applicability of this methodol. towards drug design applications. Wider application of MIMICS could facilitate the efficient utilization of chem. space.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC2sXjs12gtrs%253D&md5=f237aa1f08c24e416756cc80f018b3d1
  18. 18
    Segler, M. H.; Kogej, T.; Tyrchan, C.; Waller, M. P. Generating focused molecule libraries for drug discovery with recurrent neural networks. ACS Cent. Sci. 2018, 4, 120– 131,  DOI: 10.1021/acscentsci.7b00512
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    18
    Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks
    Segler, Marwin H. S.; Kogej, Thierry; Tyrchan, Christian; Waller, Mark P.
    ACS Central Science (2018), 4 (1), 120-131CODEN: ACSCII; ISSN:2374-7951. (American Chemical Society)
    In de novo drug design, computational strategies are used to generate novel mols. with good affinity to the desired biol. target. In this work, we show that recurrent neural networks can be trained as generative models for mol. structures, similar to statistical language models in natural language processing. We demonstrate that the properties of the generated mols. correlate very well with the properties of the mols. used to train the model. In order to enrich libraries with mols. active toward a given biol. target, we propose to fine-tune the model with small sets of mols., which are known to be active against that target. Against Staphylococcus aureus, the model reproduced 14% of 6051 hold-out test mols. that medicinal chemists designed, whereas against Plasmodium falciparum (Malaria), it reproduced 28% of 1240 test mols. When coupled with a scoring function, our model can perform the complete de novo drug design cycle to generate large sets of novel mols. for drug discovery.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC2sXitVCjsLfP&md5=708f40422c7a911c629525ce5b66088b
  19. 19
    Elton, D. C.; Boukouvalas, Z.; Fuge, M. D.; Chung, P. W. Deep learning for molecular design─a review of the state of the art. Mol. Syst. Des. Eng. 2019, 4, 828– 849,  DOI: 10.1039/C9ME00039A
    [Crossref], [CAS], Google Scholar
    19
    Deep learning for molecular design-a review of the state of the art
    Elton, Daniel C.; Boukouvalas, Zois; Fuge, Mark D.; Chung, Peter W.
    Molecular Systems Design & Engineering (2019), 4 (4), 828-849CODEN: MSDEBG; ISSN:2058-9689. (Royal Society of Chemistry)
    In the space of only a few years, deep generative modeling has revolutionized how we think of artificial creativity, yielding autonomous systems which produce original images, music, and text. Inspired by these successes, researchers are now applying deep generative modeling techniques to the generation and optimization of mols.-in our review we found 45 papers on the subject published in the past two years. These works point to a future where such systems will be used to generate lead mols., greatly reducing resources spent downstream synthesizing and characterizing bad leads in the lab. In this review we survey the increasingly complex landscape of models and representation schemes that have been proposed. The four classes of techniques we describe are recursive neural networks, autoencoders, generative adversarial networks, and reinforcement learning. After first discussing some of the math. fundamentals of each technique, we draw high level connections and comparisons with other techniques and expose the pros and cons of each. Several important high level themes emerge as a result of this work, including the shift away from the SMILES string representation of mols. towards more sophisticated representations such as graph grammars and 3D representations, the importance of reward function design, the need for better stds. for benchmarking and testing, and the benefits of adversarial training and reinforcement learning over max. likelihood based training.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXhtVWktLjN&md5=a0925fdace3a12af31b2f6fc1edef89b
  20. 20
    Schwalbe-Koda, D.; Ǵomez-Bombarelli, R. In Machine Learning Meets Quantum Physics; Schütt, K. T., Chmiela, S., von Lilienfeld, O. A., Tkatchenko, A., Tsuda, K., Müller, K.-R., Eds.; Springer International Publishing: Cham, 2020; pp 445– 467.
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  21. 21
    Zhavoronkov, A.; Vanhaelen, Q.; Oprea, T. I. Will Artificial Intelligence for Drug Discovery Impact Clinical Pharmacology?. Clin. Pharmacol. Ther. (N. Y., NY, U. S.) 2020, 107, 780– 785,  DOI: 10.1002/cpt.1795
    [Crossref], [PubMed], Google Scholar
    There is no corresponding record for this reference.
  22. 22
    Bian, Y.; Xie, X.-Q. Generative chemistry: drug discovery with deep learning gener- ative models. J. Mol. Model. 2021, 27, 71,  DOI: 10.1007/s00894-021-04674-8
    [Crossref], [PubMed], [CAS], Google Scholar
    22
    Generative chemistry: drug discovery with deep learning generative models
    Bian, Yuemin; Xie, Xiang-Qun
    Journal of Molecular Modeling (2021), 27 (3), 71CODEN: JMMOFK; ISSN:0948-5023. (Springer)
    A review. The de novo design of mol. structures using deep learning generative models introduces an encouraging soln. to drug discovery in the face of the continuously increased cost of new drug development. From the generation of original texts, images, and videos, to the scratching of novel mol. structures the creativity of deep learning generative models exhibits the height machine intelligence can achieve. The purpose of this paper is to review the latest advances in generative chem. which relies on generative modeling to expedite the drug discovery process. This review starts with a brief history of artificial intelligence in drug discovery to outline this emerging paradigm. Commonly used chem. databases, mol. representations, and tools in cheminformatics and machine learning are covered as the infrastructure for generative chem. The detailed discussions on utilizing cutting-edge generative architectures, including recurrent neural network, variational autoencoder, adversarial autoencoder, and generative adversarial network for compd. generation are focused. Challenges and future perspectives follow.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3MXhsValsLrP&md5=bf1851d6e96138674bc169295356eee6
  23. 23
    Chen, H.; Engkvist, O.; Wang, Y.; Olivecrona, M.; Blaschke, T. The rise of deep learning in drug discovery. Drug Discovery Today 2018, 23, 1241– 1250,  DOI: 10.1016/j.drudis.2018.01.039
    [Crossref], [PubMed], [CAS], Google Scholar
    23
    The rise of deep learning in drug discovery
    Chen Hongming; Engkvist Ola; Olivecrona Marcus; Blaschke Thomas; Wang Yinhai
    Drug discovery today (2018), 23 (6), 1241-1250 ISSN:.
    Over the past decade, deep learning has achieved remarkable success in various artificial intelligence research areas. Evolved from the previous research on artificial neural networks, this technology has shown superior performance to other machine learning algorithms in areas such as image and voice recognition, natural language processing, among others. The first wave of applications of deep learning in pharmaceutical research has emerged in recent years, and its utility has gone beyond bioactivity predictions and has shown promise in addressing diverse problems in drug discovery. Examples will be discussed covering bioactivity prediction, de novo molecular design, synthesis prediction and biological image analysis.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A280%3ADC%252BC1MvjvFyqtQ%253D%253D&md5=d6cbdd98ede30181802cca1786cd5a95
  24. 24
    Vamathevan, J.; Clark, D.; Czodrowski, P.; Dunham, I.; Ferran, E.; Lee, G.; Li, B.; Madabhushi, A.; Shah, P.; Spitzer, M.; Zhao, S. Applications of machine learning in drug discovery and development. Nat. Rev. Drug Discovery 2019, 18, 463– 477,  DOI: 10.1038/s41573-019-0024-5
    [Crossref], [PubMed], [CAS], Google Scholar
    24
    Applications of machine learning in drug discovery and development
    Vamathevan, Jessica; Clark, Dominic; Czodrowski, Paul; Dunham, Ian; Ferran, Edgardo; Lee, George; Li, Bin; Madabhushi, Anant; Shah, Parantu; Spitzer, Michaela; Zhao, Shanrong
    Nature Reviews Drug Discovery (2019), 18 (6), 463-477CODEN: NRDDAG; ISSN:1474-1776. (Nature Research)
    A review. Drug discovery and development pipelines are long, complex and depend on numerous factors. Machine learning (ML) approaches provide a set of tools that can improve discovery and decision making for well-specified questions with abundant, high-quality data. Opportunities to apply ML occur in all stages of drug discovery. Examples include target validation, identification of prognostic biomarkers and anal. of digital pathol. data in clin. trials. Applications have ranged in context and methodol., with some approaches yielding accurate predictions and insights. The challenges of applying ML lie primarily with the lack of interpretability and repeatability of ML-generated results, which may limit their application. In all areas, systematic and comprehensive high-dimensional data still need to be generated. With ongoing efforts to tackle these issues, as well as increasing awareness of the factors needed to validate ML approaches, the application of ML can promote data-driven decision making and has the potential to speed up the process and reduce failure rates in drug discovery and development.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXosF2rtrY%253D&md5=211782aeea3d8b9f50368f89177a70d2
  25. 25
    Engel, T., Gasteiger, J., Eds. Chemoinformatics: basic concepts and methods; Wiley-VCH: Weinheim, 2018; OCLC: 1012130305.
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  26. 26
    Kim, S.; Chen, J.; Cheng, T.; Gindulyte, A.; He, J.; He, S.; Li, Q.; Shoemaker, B. A.; Thiessen, P. A.; Yu, B.; Zaslavsky, L.; Zhang, J.; Bolton, E. E. PubChem 2019 update: improved access to chemical data. Nucleic Acids Res. 2019, 47, D1102– D1109,  DOI: 10.1093/nar/gky1033
    [Crossref], [PubMed], Google Scholar
    There is no corresponding record for this reference.
  27. 27
    Ash, S.; Cline, M.; Homer, R. W.; Hurst, T.; Smith, G. SYBYL Line Notation (SLN): A Versatile Language for Chemical Structure Representation. J. Chem. Inf. Comput. Sci. 1997, 37, 71– 79,  DOI: 10.1021/ci960109j
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    27
    SYBYL Line Notation (SLN): A Versatile Language for Chemical Structure Representation
    Ash, Sheila; Cline, Malcolm A.; Homer, R. Webster; Hurst, Tad; Smith, Gregory B.
    Journal of Chemical Information and Computer Sciences (1997), 37 (1), 71-79CODEN: JCISD8; ISSN:0095-2338. (American Chemical Society)
    SYBYL Line Notation (SLN) is an ASCII language used to represent chem. structures, including common org. mols., macromols., polymers, and combinatorial libraries. SLN is also used to express substructural (2D) queries and includes a complete facility for Markush representation. This concise language is ideal for database storage of chem. entities as well as for network communication of structures and queries.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADyaK2sXksVCitA%253D%253D&md5=9c53b87aec0a69043773cd771c662152
  28. 28
    Koniver, D. A.; Wiswesser, W. J.; Usdin, E. Wiswesser Line Notation: Simplified Techniques for Converting Chemical Structures to WLN. Science 1972, 176, 1437– 1439,  DOI: 10.1126/science.176.4042.1437
    [Crossref], [PubMed], [CAS], Google Scholar
    28
    Wiswesser Line Notation. Simplified techniques for converting chemical structures to WLN
    Koniver, Deena A.; Wiswesser, William J.; Usdin, Earl
    Science (Washington, DC, United States) (1972), 176 (4042), 1437-9CODEN: SCIEAS; ISSN:0036-8075.
    Techniques were developed for the generation of Wiswesser Line Notations (WLN), which require knowledge neither of rules for manual conversion of structures to line notations nor of computer programing. The desired WLN are obtained simply by drawing the structures of the compds. of interest on a tablet, which is linked to an appropriately programmed computer.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADyaE38Xks1Srt7c%253D&md5=e7b61344874233d0114df299b83e3a9d
  29. 29
    Weininger, D. SMILES, a chemical language and information system. 1. Introduction to methodology and encoding rules. J. Chem. Inf. Model. 1988, 28, 31– 36,  DOI: 10.1021/ci00057a005
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    29
    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.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADyaL1cXnsVeqsA%253D%253D&md5=04592975f9dd3c0ce3c1ad618ba2b17d
  30. 30
    O’Boyle, N. M. Towards a Universal SMILES representation - A standard method to generate canonical SMILES based on the InChI. J. Cheminf. 2012, 4, 22,  DOI: 10.1186/1758-2946-4-22
    [Crossref], [CAS], Google Scholar
    30
    Towards a universal SMILES representation - a standard method to generate canonical SMILES based on the InChI
    O'Boyle, Noel M.
    Journal of Cheminformatics (2012), 4 (), 22CODEN: JCOHB3; ISSN:1758-2946. (Chemistry Central Ltd.)
    Background: There are two line notations of chem. structures that have established themselves in the field: the SMILES string and the InChI string. The InChI aims to provide a unique, or canonical, identifier for chem. structures, while SMILES strings are widely used for storage and interchange of chem. structures, but no std. exists to generate a canonical SMILES string. Results: I describe how to use the InChI canonicalisation to derive a canonical SMILES string in a straightforward way, either incorporating the InChI normalizations (Inchified SMILES) or not (Universal SMILES). This is the first description of a method to generate canonical SMILES that takes stereochem. into account. When tested on the 1.1 m compds. in the ChEMBL database, and a 1 m compd. subset of the PubChem Substance database, no canonicalisation failures were found with Inchified SMILES. Using Universal SMILES, 99.79% of the ChEMBL database was canonicalised successfully and 99.77% of the PubChem subset. Conclusions: The InChI canonicalisation algorithm can successfully be used as the basis for a common std. for canonical SMILES. While challenges remain - such as the development of a std. arom. model for SMILES - the ability to create the same SMILES using different toolkits will mean that for the first time it will be possible to easily compare the chem. models used by different toolkits.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC38XhvVSiur7I&md5=eb59d742c5dec35b5d2c90417acd223d
  31. 31
    Bjerrum, E. J. SMILES enumeration as data augmentation for neural network modeling of molecules. arXiv (Machine Learning) , May 17, 2017, 703.07076, ver. 2.
    Google Scholar
    There is no corresponding record for this reference.
  32. 32
    Bjerrum, E. J.; Sattarov, B. Improving chemical autoencoder latent space and molec- ular de novo generation diversity with heteroencoders. Biomolecules 2018, 8, 131,  DOI: 10.3390/biom8040131
    [Crossref], [CAS], Google Scholar
    32
    Improving chemical autoencoder latent space and molecular de novo generation diversity with heteroencoders
    Bjerrum, Esben Jannik; Sattarov, Boris
    Biomolecules (2018), 8 (4), 131/1-131/17CODEN: BIOMHC; ISSN:2218-273X. (MDPI AG)
    Chem. autoencoders are attractive models as they combine chem. space navigation with possibilities for de novo mol. generation in areas of interest. This enables them to produce focused chem. libraries around a single lead compd. for employment early in a drug discovery project. Here, it is shown that the choice of chem. representation, such as strings from the simplified mol.-input line-entry system (SMILES), has a large influence on the properties of the latent space. It is further explored to what extent translating between different chem. representations influences the latent space similarity to the SMILES strings or circular fingerprints. By employing SMILES enumeration for either the encoder or decoder, it is found that the decoder has the largest influence on the properties of the latent space. Training a sequence to sequence heteroencoder based on recurrent neural networks (RNNs) with long short-term memory cells (LSTM) to predict different enumerated SMILES strings from the same canonical SMILES string gives the largest similarity between latent space distance and mol. similarity measured as circular fingerprints similarity. Using the output from the code layer in quant. structure activity relationship (QSAR) of five mol. datasets shows that heteroencoder derived vectors markedly outperforms autoencoder derived vectors as well as models built using ECFP4 fingerprints, underlining the increased chem. relevance of the latent space.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXit12hurjL&md5=c0e3ae0b5ec0126d70e59045568b1a90
  33. 33
    Arús-Pous, J.; Johansson, S. V.; Prykhodko, O.; Bjerrum, E. J.; Tyrchan, C.; Rey- mond, J.-L.; Chen, H.; Engkvist, O. Randomized SMILES strings improve the quality of molecular generative models. J. Cheminf. 2019, 11, 71,  DOI: 10.1186/s13321-019-0393-0
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  34. 34
    Moret, M.; Friedrich, L.; Grisoni, F.; Merk, D.; Schneider, G. Generative molecular design in low data regimes. Nature Machine Intelligence 2020, 2, 171– 180,  DOI: 10.1038/s42256-020-0160-y
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  35. 35
    van Deursen, R.; Ertl, P.; Tetko, I. V.; Godin, G. GEN: highly efficient SMILES explorer using autodidactic generative examination networks. J. Cheminf. 2020, 12, 22,  DOI: 10.1186/s13321-020-00425-8
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  36. 36
    Prykhodko, O.; Johansson, S. V.; Kotsias, P.-C.; Arús-Pous, J.; Bjerrum, E. J.; En- gkvist, O.; Chen, H. A de novo molecular generation method using latent vector based generative adversarial network. J. Cheminf. 2019, 11, 74,  DOI: 10.1186/s13321-019-0397-9
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  37. 37
    Heller, S. R.; McNaught, A.; Pletnev, I.; Stein, S.; Tchekhovskoi, D. InChI, the IUPAC International Chemical Identifier. J. Cheminf. 2015, 7, 23,  DOI: 10.1186/s13321-015-0068-4
    [Crossref], [CAS], Google Scholar
    37
    InChI, the IUPAC International Chemical Identifier
    Heller Stephen R; Stein Stephen; Tchekhovskoi Dmitrii; McNaught Alan; Pletnev Igor
    Journal of cheminformatics (2015), 7 (), 23 ISSN:1758-2946.
    This paper documents the design, layout and algorithms of the IUPAC International Chemical Identifier, InChI.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A280%3ADC%252BC2MbpslOrtQ%253D%253D&md5=4acc4f470f8cdb9b4f84558fd3302470
  38. 38
    Winter, R.; Montanari, F.; Nóe, F.; Clevert, D.-A. Learning continuous and data- driven molecular descriptors by translating equivalent chemical representations. Chem. Sci. 2019, 10, 1692– 1701,  DOI: 10.1039/C8SC04175J
    [Crossref], [PubMed], [CAS], Google Scholar
    38
    Learning continuous and data-driven molecular descriptors by translating equivalent chemical representations
    Winter, Robin; Montanari, Floriane; Noe, Frank; Clevert, Djork-Arne
    Chemical Science (2019), 10 (6), 1692-1701CODEN: CSHCCN; ISSN:2041-6520. (Royal Society of Chemistry)
    There has been a recent surge of interest in using machine learning across chem. space in order to predict properties of mols. or design mols. and materials with the desired properties. Most of this work relies on defining clever feature representations, in which the chem. graph structure is encoded in a uniform way such that predictions across chem. space can be made. In this work, we propose to exploit the powerful ability of deep neural networks to learn a feature representation from low-level encodings of a huge corpus of chem. structures. Our model borrows ideas from neural machine translation: it translates between two semantically equiv. but syntactically different representations of mol. structures, compressing the meaningful information both representations have in common in a low-dimensional representation vector. Once the model is trained, this representation can be extd. for any new mol. and utilized as a descriptor. In fair benchmarks with respect to various human-engineered mol. fingerprints and graph-convolution models, our method shows competitive performance in modeling quant. structure-activity relationships in all analyzed datasets. Addnl., we show that our descriptor significantly outperforms all baseline mol. fingerprints in two ligand-based virtual screening tasks. Overall, our descriptors show the most consistent performances in all expts. The continuity of the descriptor space and the existence of the decoder that permits deducing a chem. structure from an embedding vector allow for exploration of the space and open up new opportunities for compd. optimization and idea generation.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXit1aqsbnO&md5=57678211dbeac5a8135e2d38c41eeee2
  39. 39
    O’Boyle, N.; Dalke, A. DeepSMILES: An Adaptation of SMILES for Use in Machine- Learning of Chemical Structures; preprint, ChemRxiv , September 19, 2018, ver. 1. DOI: 10.26434/chemrxiv.7097960.v1 .
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  40. 40
    Krenn, M.; Häse, F.; Nigam, A.; Friederich, P.; Aspuru-Guzik, A. Self-referencing em- bedded strings (SELFIES): A 100% robust molecular string representation. Machine Learning: Science and Technology 2020, 1, 045024,  DOI: 10.1088/2632-2153/aba947
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  41. 41
    Faulon, J.-L., Bender, A., Eds. Handbook of chemoinformatics algorithms; Chapman & Hall/CRC mathematical and computational biology series; Chapman & Hall/CRC: Boca Raton, FL, 2010; Chapter 1. OCLC: ocn226357322.
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  42. 42
    Wishart, D. S. DrugBank 5.0: a major update to the DrugBank database for 2018. Nucleic Acids Res. 2018, 46, D1074– D1082,  DOI: 10.1093/nar/gkx1037
    [Crossref], [PubMed], [CAS], Google Scholar
    42
    DrugBank 5.0: a major update to the DrugBank database for 2018
    Wishart, David S.; Feunang, Yannick D.; Guo, An C.; Lo, Elvis J.; Marcu, Ana; Grant, Jason R.; Sajed, Tanvir; Johnson, Daniel; Li, Carin; Sayeeda, Zinat; Assempour, Nazanin; Iynkkaran, Ithayavani; Liu, Yifeng; Maciejewski, Adam; Gale, Nicola; Wilson, Alex; Chin, Lucy; Cummings, Ryan; Le, Diana; Pon, Allison; Knox, Craig; Wilson, Michael
    Nucleic Acids Research (2018), 46 (D1), D1074-D1082CODEN: NARHAD; ISSN:1362-4962. (Oxford University Press)
    DrugBank is a web-enabled database contg. comprehensivemol. information about drugs, their mechanisms, their interactions and their targets. First described in 2006, Drug- Bank has continued to evolve over the past 12 years in response to marked improvements to web stds. and changing needs for drug research and development. This year's update, DrugBank 5.0, represents the most significant upgrade to the database in more than 10 years. In many cases, existing data content has grown by 100% or more over the last update. For instance, the total no. of investigational drugs in the database has grown by almost 300%, the no. of drug-drug interactions has grown by nearly 600% and the no. of SNP-assocd. drug effects has grown more than 3000%. Significant improvements have been made to the quantity, quality and consistency of drug indications, drug binding data as well as drug-drug and drug-food interactions. A great deal of brand new data have also been added to DrugBank 5.0. This includes information on the influence of hundreds of drugs on metabolite levels (pharmacometabolomics), gene expression levels (pharmacotranscriptomics) and protein expression levels (pharmacoproteomics). New data have also been added on the status of hundreds of newdrug clin. trials and existing drug repurposing trials. Many other important improvements in the content, interface and performance of the DrugBank website have been made and these should greatly enhance its ease of use, utility and potential applications in many areas of pharmacol. research, pharmaceutical science and drug education.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXitlGisbvI&md5=986b28c7ea546596a26dd3ba38f05fee
  43. 43
    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– D1107,  DOI: 10.1093/nar/gkr777
    [Crossref], [PubMed], [CAS], Google Scholar
    43
    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.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC3MXhs12htbjN&md5=aedf7793e1ca54b6a4fa272ea3ef7d0e
  44. 44
    Landrum, G. RDKit: open-source cheminformatics software , 2016.
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  45. 45
    Steinbeck, C.; Hoppe, C.; Kuhn, S.; Floris, M.; Guha, R.; Willighagen, E. Recent De- velopments of the Chemistry Development Kit (CDK) - An Open-Source Java Library for Chemo- and Bioinformatics. Curr. Pharm. Des. 2006, 12, 2111– 2120,  DOI: 10.2174/138161206777585274
    [Crossref], [PubMed], [CAS], Google Scholar
    45
    Recent developments of the Chemistry Development Kit (CDK) - an open-source Java library for chemo- and bioinformatics
    Steinbeck, Christoph; Hoppe, Christian; Kuhn, Stefan; Floris, Matteo; Guha, Rajarshi; Willighagen, Egon L.
    Current Pharmaceutical Design (2006), 12 (17), 2111-2120CODEN: CPDEFP; ISSN:1381-6128. (Bentham Science Publishers Ltd.)
    The Chem. Development Kit (CDK) provides methods for common tasks in mol. informatics, including 2D and 3D rendering of chem. structures, I/O routines, SMILES parsing and generation, ring searches, isomorphism checking, structure diagram generation, etc. Implemented in Java, it is used both for server-side computational services, possibly equipped with a web interface, as well as for applications and client-side applets. This article introduces the CDK's new QSAR capabilities and the recently introduced interface to statistical software.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BD28XmslWqsL0%253D&md5=4e7e47ffe75b600ee3f81309a4bbb609
  46. 46
    Sun, J.; Jeliazkova, N.; Chupakhin, V.; Golib-Dzib, J.-F.; Engkvist, O.; Carlsson, L.; Wegner, J.; Ceulemans, H.; Georgiev, I.; Jeliazkov, V.; Kochev, N.; Ashby, T. J.; Chen, H. ExCAPE-DB: an integrated large scale dataset facilitating Big Data analysis in chemogenomics. J. Cheminf. 2017, 9, 17,  DOI: 10.1186/s13321-017-0222-2
    [Crossref], [CAS], Google Scholar
    46
    ExCAPE-DB: an integrated large scale dataset facilitating Big Data analysis in chemogenomics
    Sun, Jiangming; Jeliazkova, Nina; Chupakin, Vladimir; Golib-Dzib, Jose-Felipe; Engkvist, Ola; Carlsson, Lars; Wegner, Joerg; Ceulemans, Hugo; Georgiev, Ivan; Jeliazkov, Vedrin; Kochev, Nikolay; Ashby, Thomas J.; Chen, Hongming
    Journal of Cheminformatics (2017), 9 (), 17/1-17/9CODEN: JCOHB3; ISSN:1758-2946. (Chemistry Central Ltd.)
    Chemogenomics data generally refers to the activity data of chem. compds. on an array of protein targets and represents an important source of information for building in silico target prediction models. The increasing vol. of chemogenomics data offers exciting opportunities to build models based on Big Data. Prepg. a high quality data set is a vital step in realizing this goal and this work aims to compile such a comprehensive chemogenomics dataset. This dataset comprises over 70 million SAR data points from publicly available databases (PubChem and ChEMBL) including structure, target information and activity annotations. Our aspiration is to create a useful chemogenomics resource reflecting industry-scale data not only for building predictive models of in silico polypharmacol. and offtarget effects but also for the validation of chemoinformatics approaches in general.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXlsFGhsbc%253D&md5=1bd39132077c5c91f19bae5ea47c1b27
  47. 47
    Irwin, J. J.; Sterling, T.; Mysinger, M. M.; Bolstad, E. S.; Coleman, R. G. ZINC: A Free Tool to Discover Chemistry for Biology. J. Chem. Inf. Model. 2012, 52, 1757– 1768,  DOI: 10.1021/ci3001277
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    47
    ZINC: A Free Tool to Discover Chemistry for Biology
    Irwin, John J.; Sterling, Teague; Mysinger, Michael M.; Bolstad, Erin S.; Coleman, Ryan G.
    Journal of Chemical Information and Modeling (2012), 52 (7), 1757-1768CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    ZINC is a free public resource for ligand discovery. The database contains over twenty million com. available mols. in biol. relevant representations that may be downloaded in popular ready-to-dock formats and subsets. The Web site also enables searches by structure, biol. activity, phys. property, vendor, catalog no., name, and CAS no. Small custom subsets may be created, edited, shared, docked, downloaded, and conveyed to a vendor for purchase. The database is maintained and curated for a high purchasing success rate and is freely available at zinc.docking.org.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC38XmvFGnsrg%253D&md5=97f2ede64afc6b5e3ea2f279e38e32a0
  48. 48
    Shivanyuk, A.; Ryabukhin, S.; Tolmachev, A.; Bogolyubsky, A.; Mykytenko, D.; Chupryna, A.; Heilman, W.; Kostyuk, A. Enamine real database: Making chemical diversity real. Chem. Today 2007, 25, 58– 59
    [CAS], Google Scholar
    48
    Enamine real database: making chemical diversity real
    Shivanyuk, Alexander N.; Ryabukhin, Sergey V.; Bogolyubsky, Andrey V.; Mykytenko, Dmytro M.; Chupryna, Alexander A.; Heilman, William; Kostyuk, Alexander N.; Tolmachev, Andrey A.
    Chimica Oggi (2007), 25 (6), 58-59CODEN: CHOGDS; ISSN:0392-839X. (Tekno Scienze)
    The Enamine REAL DataBase (RDB) covers rigorously validated chem. space of over 29,000,000 virtual HTS compds., over 10,000,000 of which comply to drug likeness Rule-of-5 stds. The high efficiency of our RDB methodol. is based on 30 optimized reactions, 54 optimized chem. procedures applied to 18,000 proprietary in house and 9000 purchased building blocks and our efficient algorithms for calcg. the synthetic feasibility of all virtual structures. Optimized schemes for RDB prodn. allows the synthesis of 20,000 compds. a month with an av. feasibility rate of 65%.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BD1cXktVSns7w%253D&md5=648ca29e7a3a36591a42d88349789e80
  49. 49
    Huang, R.; Xia, M.; Nguyen, D.-T.; Zhao, T.; Sakamuru, S.; Zhao, J.; Shahane, S. A.; Rossoshek, A.; Simeonov, A. Tox21Challenge to build predictive models of nuclear receptor and stress response pathways as mediated by exposure to environmental chemicals and drugs. Front. Environ. Sci. 2016, 3, 85,  DOI: 10.3389/fenvs.2015.00085
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  50. 50
    Ramakrishnan, R.; Hartmann, M.; Tapavicza, E.; Von Lilienfeld, O. A. Electronic spectra from TDDFT and machine learning in chemical space. J. Chem. Phys. 2015, 143, 084111,  DOI: 10.1063/1.4928757
    [Crossref], [PubMed], [CAS], Google Scholar
    50
    Electronic spectra from TDDFT and machine learning in chemical space
    Ramakrishnan, Raghunathan; Hartmann, Mia; Tapavicza, Enrico; von Lilienfeld, O. Anatole
    Journal of Chemical Physics (2015), 143 (8), 084111/1-084111/8CODEN: JCPSA6; ISSN:0021-9606. (American Institute of Physics)
    Due to its favorable computational efficiency, time-dependent (TD) d. functional theory (DFT) enables the prediction of electronic spectra in a high-throughput manner across chem. space. Its predictions, however, can be quite inaccurate. We resolve this issue with machine learning models trained on deviations of ref. second-order approx. coupled-cluster (CC2) singles and doubles spectra from TDDFT counterparts, or even from DFT gap. We applied this approach to low-lying singlet-singlet vertical electronic spectra of over 20 000 synthetically feasible small org. mols. with up to eight CONF atoms. The prediction errors decay monotonously as a function of training set size. For a training set of 10 000 mols., CC2 excitation energies can be reproduced to within ±0.1 eV for the remaining mols. Anal. of our spectral database via chromophore counting suggests that even higher accuracies can be achieved. Based on the evidence collected, we discuss open challenges assocd. with data-driven modeling of high-lying spectra and transition intensities. (c) 2015 American Institute of Physics.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC2MXhsVSlt73O&md5=b439f44c75bb8f99906d3c920bfe7c6f
  51. 51
    Ramakrishnan, R.; Dral, P. O.; Rupp, M.; Von Lilienfeld, O. A. Quantum chemistry structures and properties of 134 kilo molecules. Sci. Data 2014, 1, 140022,  DOI: 10.1038/sdata.2014.22
    [Crossref], [PubMed], [CAS], Google Scholar
    51
    Quantum chemistry structures and properties of 134 kilo molecules
    Ramakrishnan, Raghunathan; Dral, Pavlo O.; Rupp, Matthias; von Lilienfeld, O. Anatole
    Scientific Data (2014), 1 (), 140022CODEN: SDCABS; ISSN:2052-4463. (Nature Publishing Group)
    Computational de novo design of new drugs and materials requires rigorous and unbiased exploration of chem. compd. space. However, large uncharted territories persist due to its size scaling combinatorially with mol. size. We report computed geometric, energetic, electronic, and thermodn. properties for 134k stable small org. mols. made up of CHONF. These mols. correspond to the subset of all 133,885 species with up to nine heavy atoms (CONF) out of the GDB-17 chem. universe of 166 billion org. mols. We report geometries minimal in energy, corresponding harmonic frequencies, dipole moments, polarizabilities, along with energies, enthalpies, and free energies of atomization. All properties were calcd. at the B3LYP/6-31G(2df,p) level of quantum chem. Furthermore, for the predominant stoichiometry, C7H10O2, there are 6,095 constitutional isomers among the 134k mols. We report energies, enthalpies, and free energies of atomization at the more accurate G4MP2 level of theory for all of them. As such, this data set provides quantum chem. properties for a relevant, consistent, and comprehensive chem. space of small org. mols. This database may serve the benchmarking of existing methods, development of new methods, such as hybrid quantum mechanics/machine learning, and systematic identification of structure-property relationships.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC2MXks1aisLo%253D&md5=feaffe204e7139a5fcd685bc2c6841fc
  52. 52
    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
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    52
    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.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BD2cXjs1Sjs74%253D&md5=86e609172307402d8b0d4589b1270a2f
  53. 53
    Cho, K.; Van Merrïenboer, B.; Bahdanau, D.; Bengio, Y. On the properties of neural machine translation: Encoder-decoder approaches. arXiv (Computation and Language) , October 7, 2014, 1409.1259, ver. 2.
    Google Scholar
    There is no corresponding record for this reference.
  54. 54
    Hochreiter, S.; Schmidhuber, J. Long short-term memory. Neural computation 1997, 9, 1735– 1780,  DOI: 10.1162/neco.1997.9.8.1735
    [Crossref], [PubMed], [CAS], Google Scholar
    54
    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.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A280%3ADyaK1c%252FhvVahsQ%253D%253D&md5=5da426ddc18e5bc1972e520bbcc33bec
  55. 55
    Goodfellow, I. J.; Pouget-Abadie, J.; Mirza, M.; Xu, B.; Warde-Farley, D.; Ozair, S.; Courville, A.; Bengio, Y. Generative adversarial networks. arXiv (Machine Learning) , June 10, 2014, 1406.2661, ver. 1.
    Google Scholar
    There is no corresponding record for this reference.
  56. 56
    Kingma, D. P.; Welling, M. Auto-encoding variational bayes. arXiv (Machine Learning) , May 1, 2014, 1312.6114, ver. 10.
    Google Scholar
    There is no corresponding record for this reference.
  57. 57
    Makhzani, A.; Shlens, J.; Jaitly, N.; Goodfellow, I.; Frey, B. Adversarial autoencoders. arXiv (Machine Learning) , May 25, 2016, 1511.05644, ver. 2..
    Google Scholar
    There is no corresponding record for this reference.
  58. 58
    Olivecrona, M.; Blaschke, T.; Engkvist, O.; Chen, H. Molecular de-novo design through deep reinforcement learning. J. Cheminf. 2017, 9, 48,  DOI: 10.1186/s13321-017-0235-x
    [Crossref], [CAS], Google Scholar
    58
    Molecular de-novo design through deep reinforcement learning
    Olivecrona Marcus; Blaschke Thomas; Engkvist Ola; Chen Hongming
    Journal of cheminformatics (2017), 9 (1), 48 ISSN:1758-2946.
    This work introduces a method to tune a sequence-based generative model for molecular de novo design that through augmented episodic likelihood can learn to generate structures with certain specified desirable properties. We demonstrate how this model can execute a range of tasks such as generating analogues to a query structure and generating compounds predicted to be active against a biological target. As a proof of principle, the model is first trained to generate molecules that do not contain sulphur. As a second example, the model is trained to generate analogues to the drug Celecoxib, a technique that could be used for scaffold hopping or library expansion starting from a single molecule. Finally, when tuning the model towards generating compounds predicted to be active against the dopamine receptor type 2, the model generates structures of which more than 95% are predicted to be active, including experimentally confirmed actives that have not been included in either the generative model nor the activity prediction model. Graphical abstract .
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A280%3ADC%252BC1M7mtVKisg%253D%253D&md5=68f5a1219ef81fbd52a5c8911cacbbfc
  59. 59
    Gupta, A.; Müller, A. T.; Huisman, B. J.; Fuchs, J. A.; Schneider, P.; Schneider, G. Generative Recurrent Networks for De Novo Drug Design. Mol. Inf. 2018, 37, 1700111,  DOI: 10.1002/minf.201700111
    [Crossref], [PubMed], Google Scholar
    There is no corresponding record for this reference.
  60. 60
    Guimaraes, G. L.; Sanchez-Lengeling, B.; Outeiral, C.; Farias, P. L. C.; Aspuru- Guzik, A. Objective-Reinforced Generative Adversarial Networks (ORGAN) for Sequence Generation Models. arXiv (Machine Learning) , February 7, 2018, 1705.10843, ver. 3.
    Google Scholar
    There is no corresponding record for this reference.
  61. 61
    Lim, J.; Ryu, S.; Kim, J. W.; Kim, W. Y. Molecular generative model based on conditional variational autoencoder for de novo molecular design. J. Cheminf. 2018, 10, 31,  DOI: 10.1186/s13321-018-0286-7
    [Crossref], [CAS], Google Scholar
    61
    Molecular generative model based on conditional variational autoencoder for de novo molecular design
    Lim, Jaechang; Ryu, Seongok; Kim, Jin Woo; Kim, Woo Youn
    Journal of Cheminformatics (2018), 10 (), 31/1-31/9CODEN: JCOHB3; ISSN:1758-2946. (Chemistry Central Ltd.)
    We propose a mol. generative model based on the conditional variational autoencoder for de novo mol. design. It is specialized to control multiple mol. properties simultaneously by imposing them on a latent space. As a proof of concept, we demonstrate that it can be used to generate drug-like mols. with five target properties. We were also able to adjust a single property without changing the others and to manipulate it beyond the range of the dataset.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXmtFWhtrw%253D&md5=8a1262077f9d6ffa5ec0385ed1a69f6f
  62. 62
    Merk, D.; Grisoni, F.; Friedrich, L.; Schneider, G. Tuning artificial intelligence on the de novo design of natural-product-inspired retinoid X receptor modulators. Commun. Chem. 2018, 1, 68,  DOI: 10.1038/s42004-018-0068-1
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  63. 63
    Merk, D.; Friedrich, L.; Grisoni, F.; Schneider, G. De Novo Design of Bioactive Small Molecules by Artificial Intelligence. Mol. Inf. 2018, 37, 1700153,  DOI: 10.1002/minf.201700153
    [Crossref], [PubMed], Google Scholar
    There is no corresponding record for this reference.
  64. 64
    Polykovskiy, D.; Zhebrak, A.; Vetrov, D.; Ivanenkov, Y.; Aladinskiy, V.; Mamoshina, P.; Bozdaganyan, M.; Aliper, A.; Zhavoronkov, A.; Kadurin, A. Entangled Conditional Adversarial Autoencoder for de Novo Drug Discovery. Mol. Pharmaceutics 2018, 15, 4398– 4405,  DOI: 10.1021/acs.molpharmaceut.8b00839
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    64
    Entangled Conditional Adversarial Autoencoder for de Novo Drug Discovery
    Polykovskiy, Daniil; Zhebrak, Alexander; Vetrov, Dmitry; Ivanenkov, Yan; Aladinskiy, Vladimir; Mamoshina, Polina; Bozdaganyan, Marine; Aliper, Alexander; Zhavoronkov, Alex; Kadurin, Artur
    Molecular Pharmaceutics (2018), 15 (10), 4398-4405CODEN: MPOHBP; ISSN:1543-8384. (American Chemical Society)
    Modern computational approaches and machine learning techniques accelerate the invention of new drugs. Generative models can discover novel mol. structures within hours, while conventional drug discovery pipelines require months of work. In this article, we propose a new generative architecture, entangled conditional adversarial autoencoder that generates mol. structures based on various properties, such as activity against a specific protein, soly., or ease of synthesis. We apply the proposed model to generate a novel inhibitor of Janus kinase 3, implicated in rheumatoid arthritis, psoriasis and vitiligo. The discovered mol. was tested in vitro and showed good activity and selectivity.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXhs1ensL%252FM&md5=58796965d266a5b78ca53c0f11c73999
  65. 65
    Li, Y.; Vinyals, O.; Dyer, C.; Pascanu, R.; Battaglia, P. Learning Deep Generative Models of Graphs. arXiv (Machine Learning) , March 8, 2018, 1803.03324, ver. 1.
    Google Scholar
    There is no corresponding record for this reference.
  66. 66
    Liu, Q.; Allamanis, M.; Brockschmidt, M.; Gaunt, A. Constrained graph variational autoencoders for molecule design. Adv. Neural Inf. Process. Syst. 2018, 7795– 7804
    Google Scholar
    There is no corresponding record for this reference.
  67. 67
    Mercado, R.; Rastemo, T.; Lindelof, E.; Klambauer, G.; Engkvist, O.; Chen, H.; Bjerrum, E. J. Graph networks for molecular design. Mach. Learn.: Sci. Technol. 2021, 2, 025023,  DOI: 10.1088/2632-2153/abcf91
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  68. 68
    De Cao, N.; Kipf, T. MolGAN: An implicit generative model for small molecular graphs. arXiv (Machine Learning) , May 30, 2018, 1805.11973, ver. 1.
    Google Scholar
    There is no corresponding record for this reference.
  69. 69
    Simonovsky, M.; Komodakis, N. Graphvae: Towards generation of small graphs us- ing variational autoencoders. International Conference on Artificial Neural Networks. 2018, 11139, 412– 422,  DOI: 10.1007/978-3-030-01418-6_41
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  70. 70
    Ma, T.; Chen, J.; Xiao, C. Constrained Generation of Semantically Valid Graphs via Regularizing Variational Autoencoders. Adv. Neural Inf. Process. Syst. 2018, 7113– 7124
    Google Scholar
    There is no corresponding record for this reference.
  71. 71
    Hawkins, P. C. D. Conformation Generation: The State of the Art. J. Chem. Inf. Model. 2017, 57, 1747– 1756,  DOI: 10.1021/acs.jcim.7b00221
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    71
    Conformation Generation: The State of the Art
    Hawkins, Paul C. D.
    Journal of Chemical Information and Modeling (2017), 57 (8), 1747-1756CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    The generation of conformations for small mols. is a problem of continuing interest in cheminformatics and computational drug discovery. This review will present an overview of methods used to sample conformational space, focusing on those methods designed for org. mols. commonly of interest in drug discovery. Different approaches to both the sampling of conformational space and the scoring of conformational stability will be compared and contrasted, with an emphasis on those methods suitable for conformer sampling of large nos. of drug-like mols. Particular attention will be devoted to the appropriate utilization of information from exptl. solid-state structures in validating and evaluating the performance of these tools. The review will conclude with some areas worthy of further investigation.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC2sXhtFagu7bE&md5=776818e66798d1987e015440a28e208e
  72. 72
    Skalic, M.; Jiḿenez, J.; Sabbadin, D.; De Fabritiis, G. Shape-Based Generative Mod- eling for de Novo Drug Design. J. Chem. Inf. Model. 2019, 59, 1205– 1214,  DOI: 10.1021/acs.jcim.8b00706
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    72
    Shape-based generative modeling for de novo drug design
    Skalic, Miha; Jimenez, Jose; Sabbadin, Davide; De Fabritiis, Gianni
    Journal of Chemical Information and Modeling (2019), 59 (3), 1205-1214CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    In this work, we propose a machine learning approach to generate novel mols. starting from a seed compd., its three-dimensional (3D) shape, and its pharmacophoric features. The pipeline draws inspiration from generative models used in image anal. and represents a first example of the de novo design of lead-like mols. guided by shape-based features. A variational autoencoder is used to perturb the 3D representation of a compd., followed by a system of convolutional and recurrent neural networks that generate a sequence of SMILES tokens. The generative design of novel scaffolds and functional groups can cover unexplored regions of chem. space that still possess lead-like properties.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXjtVShu78%253D&md5=8eb58a7d5aa780431d285a63dfed5765
  73. 73
    Gebauer, N.; Gastegger, M.; Schütt, K. T. Symmetry-adapted generation of 3d point sets for the targeted discovery of molecules. NeurIPS . 2019.
    Google Scholar
    There is no corresponding record for this reference.
  74. 74
    Ragoza, M.; Masuda, T.; Koes, D. R. Learning a Continuous Representation of 3D Molecular Structures with Deep Generative Models. arXiv (Quantitative Methods) , November 15, 2020, 2010.08687, ver. 3.
    Google Scholar
    There is no corresponding record for this reference.
  75. 75
    Preuer, K.; Renz, P.; Unterthiner, T.; Hochreiter, S.; Klambauer, G. Fŕechet ChemNet Distance: A Metric for Generative Models for Molecules in Drug Discovery. J. Chem. Inf. Model. 2018, 58, 1736– 1741,  DOI: 10.1021/acs.jcim.8b00234
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    75
    Fr´echet ChemNet Distance: A Metric for Generative Models for Molecules in Drug Discovery
    Preuer, Kristina; Renz, Philipp; Unterthiner, Thomas; Hochreiter, Sepp; Klambauer, Guenter
    Journal of Chemical Information and Modeling (2018), 58 (9), 1736-1741CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    The new wave of successful generative models in machine learning has increased the interest in deep learning driven de novo drug design. However, method comparison is difficult because of various flaws of the currently employed evaluation metrics. The authors propose an evaluation metric for generative models called Fre´chet ChemblNet distance (FCD). The FCD's advantage over previous metrics is that it can detect whether generated mols. are diverse and have similar chem. and biol. properties as real mols.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXhsFejsLrL&md5=dc8e10eb0f85a7f27b48da91dcb21a27
  76. 76
    Arús-Pous, J.; Blaschke, T.; Ulander, S.; Reymond, J.-L.; Chen, H.; Engkvist, O. Exploring the GDB-13 chemical space using deep generative models. J. Cheminf. 2019, 11, 1– 14,  DOI: 10.1186/s13321-019-0341-z
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  77. 77
    Brown, N.; Fiscato, M.; Segler, M. H.; Vaucher, A. C. GuacaMol: benchmarking models for de novo molecular design. J. Chem. Inf. Model. 2019, 59, 1096– 1108,  DOI: 10.1021/acs.jcim.8b00839
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    77
    GuacaMol: Benchmarking Models for de Novo Molecular Design
    Brown, Nathan; Fiscato, Marco; Segler, Marwin H. S.; Vaucher, Alain C.
    Journal of Chemical Information and Modeling (2019), 59 (3), 1096-1108CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    De novo design seeks to generate mols. with required property profiles by virtual design-make-test cycles. With the emergence of deep learning and neural generative models in many application areas, models for mol. design based on neural networks appeared recently and show promising results. However, the new models have not been profiled on consistent tasks, and comparative studies to well-established algorithms have only seldom been performed. To standardize the assessment of both classical and neural models for de novo mol. design, we propose an evaluation framework, GuacaMol, based on a suite of standardized benchmarks. The benchmark tasks encompass measuring the fidelity of the models to reproduce the property distribution of the training sets, the ability to generate novel mols., the exploration and exploitation of chem. space, and a variety of single and multiobjective optimization tasks. The benchmarking open-source Python code and a leaderboard can be found on https://benevolent.ai/guacamol.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXltVWrsbY%253D&md5=d3fb616b81a4b146cf77950a1c92e4d1
  78. 78
    Polykovskiy, D.; Zhebrak, A.; Sanchez-Lengeling, B.; Golovanov, S.; Tatanov, O.; Belyaev, S.; Kurbanov, R.; Artamonov, A.; Aladinskiy, V.; Veselov, M. Molecular sets (MOSES): a benchmarking platform for molecular generation models. Front. Pharmacol. 2020, 11, 11,  DOI: 10.3389/fphar.2020.565644
    [Crossref], [PubMed], Google Scholar
    There is no corresponding record for this reference.
  79. 79
    Renz, P.; Van Rompaey, D.; Wegner, J. K.; Hochreiter, S.; Klambauer, G. On fail- ure modes in molecule generation and optimization. Drug Discovery Today: Technol. 2019, 32–33, 55– 63,  DOI: 10.1016/j.ddtec.2020.09.003
    [Crossref], [PubMed], [CAS], Google Scholar
    79
    On failure modes in molecule generation and optimization
    Renz Philipp; Hochreiter Sepp; Klambauer Gunter; Van Rompaey Dries; Wegner Jorg Kurt
    Drug discovery today. Technologies (2019), 32-33 (), 55-63 ISSN:.
    There has been a wave of generative models for molecules triggered by advances in the field of Deep Learning. These generative models are often used to optimize chemical compounds towards particular properties or a desired biological activity. The evaluation of generative models remains challenging and suggested performance metrics or scoring functions often do not cover all relevant aspects of drug design projects. In this work, we highlight some unintended failure modes in molecular generation and optimization and how these evade detection by current performance metrics.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A280%3ADC%252BB3svkt1yqtg%253D%253D&md5=ca243ca3af904c3fa64f9e0e2c2b2c2c
  80. 80
    Cieplinski, T.; Danel, T.; Podlewska, S.; Jastrzebski, S. We should at least be able to Design Molecules that Dock Well. arXiv (Biomolecules) December 28, 2020, 2006.16955, ver. 3.
    Google Scholar
    There is no corresponding record for this reference.
  81. 81
    Zhang, J.; Mercado, R.; Engkvist, O.; Chen, H. Comparative study of deep generative models on chemical space coverage. ChemRxiv , May 2, 2021, ver. 3.  DOI: 10.26434/chemrxiv.13234289.v3 .
    [Crossref], [PubMed], Google Scholar
    There is no corresponding record for this reference.
  82. 82
    Blaschke, T.; Arús-Pous, J.; Chen, H.; Margreitter, C.; Tyrchan, C.; Engkvist, O.; Papadopoulos, K.; Patronov, A. REINVENT 2.0: An AI Tool for De Novo Drug Design. J. Chem. Inf. Model. 2020, 60, 5918,  DOI: 10.1021/acs.jcim.0c00915
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    82
    REINVENT 2.0: An AI Tool for De Novo Drug Design
    Blaschke, Thomas; Arus-Pous, Josep; Chen, Hongming; Margreitter, Christian; Tyrchan, Christian; Engkvist, Ola; Papadopoulos, Kostas; Patronov, Atanas
    Journal of Chemical Information and Modeling (2020), 60 (12), 5918-5922CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    In the past few years, we have witnessed a renaissance of the field of mol. de novo drug design. The advancements in deep learning and artificial intelligence (AI) have triggered an avalanche of ideas on how to translate such techniques to a variety of domains including the field of drug design. A range of architectures have been devised to find the optimal way of generating chem. compds. by using either graph- or string (SMILES)-based representations. With this application note, we aim to offer the community a prodn.-ready tool for de novo design, called REINVENT. It can be effectively applied on drug discovery projects that are striving to resolve either exploration or exploitation problems while navigating the chem. space. It can facilitate the idea generation process by bringing to the researcher's attention the most promising compds. REINVENT's code is publicly available at https://github.com/MolecularAI/Reinvent.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXitFOnsbbJ&md5=b06445c0516f122adff2e7c82d7ca70c
  83. 83
    Bung, N.; Krishnan, S. R.; Bulusu, G.; Roy, A. De novo design of new chemical entities for SARS-CoV-2 using artificial intelligence. Future Med. Chem. 2021, 13, 575,  DOI: 10.4155/fmc-2020-0262
    [Crossref], [PubMed], [CAS], Google Scholar
    83
    De novo design of new chemical entities for SARS-CoV-2 using artificial intelligence
    Bung, Navneet; Krishnan, Sowmya R.; Bulusu, Gopalakrishnan; Roy, Arijit
    Future Medicinal Chemistry (2021), 13 (6), 575-585CODEN: FMCUA7; ISSN:1756-8919. (Newlands Press Ltd.)
    The novel coronavirus SARS-CoV-2 has severely affected the health and economy of several countries. Multiple studies are in progress to design novel therapeutics against the potential target proteins in SARS-CoV-2, including 3CL protease, an essential protein for virus replication. In this study we employed deep neural network-based generative and predictive models for de novo design of small mols. capable of inhibiting the 3CL protease. The generative model was optimized using transfer learning and reinforcement learning to focus around the chem. space corresponding to the protease inhibitors. Multiple physicochem. property filters and virtual screening score were used for the final screening. We have identified 33 potential compds. as ideal candidates for further synthesis and testing against SARS-CoV-2.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3MXmtlOkurY%253D&md5=4ad8a669ca9c653df25f57c508fdefa0
  84. 84
    Li, Y.; Zhang, L.; Liu, Z. Multi-objective de novo drug design with conditional graph generative model. J. Cheminf. 2018, 10, 33,  DOI: 10.1186/s13321-018-0287-6
    [Crossref], [CAS], Google Scholar
    84
    Multi-objective de novo drug design with conditional graph generative model
    Li, Yibo; Zhang, Liangren; Liu, Zhenming
    Journal of Cheminformatics (2018), 10 (), 33/1-33/24CODEN: JCOHB3; ISSN:1758-2946. (Chemistry Central Ltd.)
    Recently, deep generative models have revealed itself as a promising way of performing de novo mol. design. However, previous research has focused mainly on generating SMILES strings instead of mol. graphs. Although available, current graph generative models are are often too general and computationally expensive. In this work, a new de novo mol. design framework is proposed based on a type of sequential graph generators that do not use atom level recurrent units. Compared with previous graph generative models, the proposed method is much more tuned for mol. generation and has been scaled up to cover significantly larger mols. in the ChEMBL database. It is shown that the graph-based model outperforms SMILES based models in a variety of metrics, esp. in the rate of valid outputs. For the application of drug design tasks, conditional graph generative model is employed. This method offers highe flexibility and is suitable for generation based on multiple objectives. The results have demonstrated that this approach can be effectively applied to solve several drug design problems, including the generation of compds. contg. a given scaffold, compds. with specific drug-likeness and synthetic accessibility requirements, as well as dual inhibitors against JNK3 and GSK-3ss.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXmtFWiu7s%253D&md5=6cef99e5a789668f55ba72e4e65da160
  85. 85
    Blaschke, T.; Engkvist, O.; Bajorath, J.; Chen, H. Memory-assisted reinforcement learning for diverse molecular de novo design. J. Cheminf. 2020, 12, 1– 17,  DOI: 10.1186/s13321-020-00473-0
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  86. 86
    Zhavoronkov, A. Deep learning enables rapid identification of potent DDR1 kinase inhibitors. Nat. Biotechnol. 2019, 37, 1038– 1040,  DOI: 10.1038/s41587-019-0224-x
    [Crossref], [PubMed], [CAS], Google Scholar
    86
    Deep learning enables rapid identification of potent DDR1 kinase inhibitors
    Zhavoronkov, Alex; Ivanenkov, Yan A.; Aliper, Alex; Veselov, Mark S.; Aladinskiy, Vladimir A.; Aladinskaya, Anastasiya V.; Terentiev, Victor A.; Polykovskiy, Daniil A.; Kuznetsov, Maksim D.; Asadulaev, Arip; Volkov, Yury; Zholus, Artem; Shayakhmetov, Rim R.; Zhebrak, Alexander; Minaeva, Lidiya I.; Zagribelnyy, Bogdan A.; Lee, Lennart H.; Soll, Richard; Madge, David; Xing, Li; Guo, Tao; Aspuru-Guzik, Alan
    Nature Biotechnology (2019), 37 (9), 1038-1040CODEN: NABIF9; ISSN:1087-0156. (Nature Research)
    We have developed a deep generative model, generative tensorial reinforcement learning (GENTRL), for de novo small-mol. design. GENTRL optimizes synthetic feasibility, novelty, and biol. activity. We used GENTRL to discover potent inhibitors of discoidin domain receptor 1 (DDR1), a kinase target implicated in fibrosis and other diseases, in 21 days. Four compds. were active in biochem. assays, and two were validated in cell-based assays. One lead candidate was tested and demonstrated favorable pharmacokinetics in mice.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXhs12gurnM&md5=b15262b61b9172ab2bc37e534a70f010
  87. 87
    Popova, M.; Shvets, M.; Oliva, J.; Isayev, O. MolecularRNN: Generating real- istic molecular graphs with optimized properties. arXiv (Machine Learning) , May 31, 2019, 1905.13372, ver. 1.
    Google Scholar
    There is no corresponding record for this reference.
  88. 88
    Sanchez-Lengeling, B.; Outeiral, C.; Guimaraes, G. L.; Aspuru-Guzik, A. Optimizing distributions over molecular space. An objective-reinforced generative adversarial network for inverse-design chemistry (ORGANIC). ChemRxiv , August 18, 2017, ver. 3.  DOI: 10.26434/chemrxiv.5309668.v3 .
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  89. 89
    Putin, E.; Asadulaev, A.; Vanhaelen, Q.; Ivanenkov, Y.; Aladinskaya, A. V.; Aliper, A.; Zhavoronkov, A. Adversarial Threshold Neural Computer for Molecular de Novo De- sign. Mol. Pharmaceutics 2018, 15, 4386– 4397,  DOI: 10.1021/acs.molpharmaceut.7b01137
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    89
    Adversarial threshold neural computer for molecular de novo design
    Putin, Evgeny; Asadulaev, Arip; Vanhaelen, Quentin; Ivanenkov, Yan; Aladinskaya, Anastasia V.; Aliper, Alex; Zhavoronkov, Alex
    Molecular Pharmaceutics (2018), 15 (10), 4386-4397CODEN: MPOHBP; ISSN:1543-8384. (American Chemical Society)
    In this article, we propose the deep neural network Adversarial Threshold Neural Computer (ATNC). The ATNC model is intended for the de novo design of novel small-mol. org. structures. The model is based on generative adversarial network architecture and reinforcement learning. ATNC uses a Differentiable Neural Computer as a generator and has a new specific block, called adversarial threshold (AT). AT acts as a filter between the agent (generator) and the environment (discriminator + objective reward functions). Furthermore, to generate more diverse mols. we introduce a new objective reward function named Internal Diversity Clustering (IDC). In this work, ATNC is tested and compared with the Org. model. Both models were trained on the SMILES string representation of the mols., using four objective functions (internal similarity, Muegge druglikeness filter, presence or absence of sp3-rich fragments, and IDC). The SMILES representations of 15K druglike mols. from the ChemDiv collection were used as a training data set. For the different functions, ATNC outperforms Org. Combined with the IDC, ATNC generates 72% of valid and 77% of unique SMILES strings, while Org. generates only 7% of valid and 86% of unique SMILES strings. For each set of mols. generated by ATNC and Org., we analyzed distributions of four mol. descriptors (no. of atoms, mol. wt., logP, and tpsa) and calcd. five chem. statistical features (internal diversity, no. of unique heterocycles, no. of clusters, no. of singletons, and no. of compds. that have not been passed through medicinal chem. filters). Anal. of key mol. descriptors and chem. statistical features demonstrated that the mols. generated by ATNC elicited better druglikeness properties. We also performed in vitro validation of the mols. generated by ATNC; results indicated that ATNC is an effective method for producing hit compds.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXlt1Kks7c%253D&md5=a83d9ef283d643d2b8e81f050c4e7be2
  90. 90
    Putin, E.; Asadulaev, A.; Ivanenkov, Y.; Aladinskiy, V.; Sanchez-Lengeling, B.; Aspuru-Guzik, A.; Zhavoronkov, A. Reinforced Adversarial Neural Computer for de Novo Molecular Design. J. Chem. Inf. Model. 2018, 58, 1194– 1204,  DOI: 10.1021/acs.jcim.7b00690
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    90
    Reinforced Adversarial Neural Computer for de Novo Molecular Design
    Putin, Evgeny; Asadulaev, Arip; Ivanenkov, Yan; Aladinskiy, Vladimir; Sanchez-Lengeling, Benjamin; Aspuru-Guzik, Alan; Zhavoronkov, Alex
    Journal of Chemical Information and Modeling (2018), 58 (6), 1194-1204CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    In silico modeling is a crucial milestone in modern drug design and development. Although computer-aided approaches in this field are well-studied, the application of deep learning methods in this research area is at the beginning. In this work, we present an original deep neural network (DNN) architecture named RANC (Reinforced Adversarial Neural Computer) for the de novo design of novel small-mol. org. structures based on the generative adversarial network (GAN) paradigm and reinforcement learning (RL). As a generator RANC uses a differentiable neural computer (DNC), a category of neural networks, with increased generation capabilities due to the addn. of an explicit memory bank, which can mitigate common problems found in adversarial settings. The comparative results have shown that RANC trained on the SMILES string representation of the mols. outperforms its first DNN-based counterpart Org. by several metrics relevant to drug discovery: the no. of unique structures, passing medicinal chem. filters (MCFs), Muegge criteria, and high QED scores. RANC is able to generate structures that match the distributions of the key chem. features/descriptors (e.g., MW, logP, TPSA) and lengths of the SMILES strings in the training data set. Therefore, RANC can be reasonably regarded as a promising starting point to develop novel mols. with activity against different biol. targets or pathways. In addn., this approach allows scientists to save time and covers a broad chem. space populated with novel and diverse compds.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXpsVChtrs%253D&md5=b10c44dcadf9fb1afc4e65cc7469730f
  91. 91
    You, J.; Liu, B.; Ying, Z.; Pande, V.; Leskovec, J. Graph convolutional policy network for goal-directed molecular graph generation. Adv. Neural Inf. Process. Syst. 2018, 31, 6410– 6421
    Google Scholar
    There is no corresponding record for this reference.
  92. 92
    Karimi, M.; Hasanzadeh, A.; Shen, Y. Network-principled deep generative models for designing drug combinations as graph sets. Bioinformatics 2020, 36, i445– i454,  DOI: 10.1093/bioinformatics/btaa317
    [Crossref], [PubMed], [CAS], Google Scholar
    92
    Network-principled deep generative models for designing drug combinations as graph sets
    Karimi, Mostafa; Hasanzadeh, Arman; Shen, Yang
    Bioinformatics (2020), 36 (Suppl._1), i445-i454CODEN: BOINFP; ISSN:1367-4811. (Oxford University Press)
    Motivation: Combination therapy has shown to improve therapeutic efficacy while reducing side effects. Importantly, it has become an indispensable strategy to overcome resistance in antibiotics, antimicrobials and anticancer drugs. Facing enormous chem. space and unclear design principles for small-mol. combinations, computational drug-combination design has not seen generative models to meet its potential to accelerate resistance-overcoming drug combination discovery. Results: We have developed the first deep generative model for drug combination design, by jointly embedding graph-structured domain knowledge and iteratively training a reinforcement learning-based chem. graph-set designer. First, we have developed hierarchical variational graph auto-encoders trained end-to-end to jointly embed gene-gene, gene-disease and disease-disease networks. Novel attentional pooling is introduced here for learning disease representations from assocd. genes' representations. Second, targeting diseases in learned representations, we have recast the drug-combination design problem as graph-set generation and developed a deep learning-based model with novel rewards. Specifically, besides chem. validity rewards, we have introduced novel generative adversarial award, being generalized sliced Wasserstein, for chem. diverse mols. with distributions similar to known drugs. We have also designed a network principle-based reward for disease-specific drug combinations. Numerical results indicate that, compared to state-of-the-art graph embedding methods, hierarchical variational graph auto-encoder learns more informative and generalizable disease representations. Results also show that the deep generative models generate drug combinations following the principle across diseases. Case studies on four diseases show that network-principled drug combinations tend to have low toxicity. The generated drug combinations collectively cover the disease module similar to FDA-approved drug combinations and could potentially suggest novel systems pharmacol. strategies. Ourmethod allows for examg. and following network-based principle or hypothesis to efficiently generate disease-specific drug combinations in a vast chem. combinatorial space.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXis1ymt7zE&md5=b737e2db6e7e76d0753debee66246ed3
  93. 93
    Griffiths, R.-R.; Hernández-Lobato, J. M. Constrained Bayesian optimization for auto- matic chemical design using variational autoencoders. Chem. Sci. 2020, 11, 577– 586,  DOI: 10.1039/C9SC04026A
    [Crossref], [PubMed], [CAS], Google Scholar
    93
    Constrained Bayesian optimization for automatic chemical design using variational autoencoders
    Griffiths, Ryan-Rhys; Hernandez-Lobato, Jose Miguel
    Chemical Science (2020), 11 (2), 577-586CODEN: CSHCCN; ISSN:2041-6520. (Royal Society of Chemistry)
    Automatic Chem. Design is a framework for generating novel mols. with optimized properties. The original scheme, featuring Bayesian optimization over the latent space of a variational autoencoder, suffers from the pathol. that it tends to produce invalid mol. structures. First, we demonstrate empirically that this pathol. arises when the Bayesian optimization scheme queries latent space points far away from the data on which the variational autoencoder has been trained. Secondly, by reformulating the search procedure as a constrained Bayesian optimization problem, we show that the effects of this pathol. can be mitigated, yielding marked improvements in the validity of the generated mols. We posit that constrained Bayesian optimization is a good approach for solving this kind of training set mismatch in many generative tasks involving Bayesian optimization over the latent space of a variational autoencoder.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXitFOis7bK&md5=628684b606d9d93ccf21b674438acd6a
  94. 94
    Blaschke, T.; Olivecrona, M.; Engkvist, O.; Bajorath, J.; Chen, H. Application of Generative Autoencoder in De Novo Molecular Design. Mol. Inf. 2018, 37, 1700123,  DOI: 10.1002/minf.201700123
    [Crossref], [PubMed], Google Scholar
    There is no corresponding record for this reference.
  95. 95
    Kusner, M. J.; Paige, B.; Hernández-Lobato, J. M. Grammar variational autoencoder. Proc. 34th Int. Conf. Mach. Learn. 2017, 70, 1945– 1954
    Google Scholar
    There is no corresponding record for this reference.
  96. 96
    Dai, H.; Tian, Y.; Dai, B.; Skiena, S.; Song, L. Syntax-directed variational autoencoder for structured data. arXiv (Machine Learning) , February 24, 2018, 1802.08786, ver 1.
    Google Scholar
    There is no corresponding record for this reference.
  97. 97
    Jin, W.; Barzilay, R.; Jaakkola, T. Junction tree variational autoencoder for molecular graph generation. Proc. 35th Int. Conf. Mach. Learn. 2018, 50, 2323– 2332
    Google Scholar
    There is no corresponding record for this reference.
  98. 98
    Samanta, B.; De, A.; Jana, G.; Chattaraj, P. K.; Ganguly, N.; Rodriguez, M. G. NeVAE: A Deep Generative Model for Molecular Graphs. Proceedings of the AAAI Conference on Artificial Intelligence 2019, 33, 1110– 1117,  DOI: 10.1609/aaai.v33i01.33011110
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  99. 99
    Bresson, X.; Laurent, T. A Two-Step Graph Convolutional Decoder for Molecule Generation. arXiv (Machine Learning) , June 15, 2019, 1906.03412, ver 2.
    Google Scholar
    There is no corresponding record for this reference.
  100. 100
    Maziarka, L.; Pocha, A.; Kaczmarczyk, J.; Rataj, K.; Danel, T.; Warcho-l, M. Mol- CycleGAN: a generative model for molecular optimization. J. Cheminf. 2020, 12, 2,  DOI: 10.1186/s13321-019-0404-1
    [Crossref], [CAS], Google Scholar
    100
    Mol-CycleGAN: a generative model for molecular optimization
    Maziarka, Lukasz; Pocha, Agnieszka; Kaczmarczyk, Jan; Rataj, Krzysztof; Danel, Tomasz; Warchol, Michal
    Journal of Cheminformatics (2020), 12 (1), 2CODEN: JCOHB3; ISSN:1758-2946. (SpringerOpen)
    Designing a mol. with desired properties is one of the biggest challenges in drug development, as it requires optimization of chem. compd. structures with respect to many complex properties. To improve the compd. design process, we introduce Mol-CycleGAN-a CycleGAN-based model that generates optimized compds. with high structural similarity to the original ones. Namely, given a mol. our model generates a structurally similar one with an optimized value of the considered property. We evaluate the performance of the model on selected optimization objectives related to structural properties (presence of halogen groups, no. of arom. rings) and to a physicochem. property (penalized logP). In the task of optimization of penalized logP of drug-like mols. our model significantly outperforms previous results.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXotFarsg%253D%253D&md5=243c1aeef3e517ee774fcfabebf80260
  101. 101
    Sattarov, B.; Baskin, I. I.; Horvath, D.; Marcou, G.; Bjerrum, E. J.; Varnek, A. De Novo Molecular Design by Combining Deep Deep Autoencoder Recurrent Neural Networks with Generative Topographic Mapping. J. Chem. Inf. Model. 2019, 59, 1182– 1196,  DOI: 10.1021/acs.jcim.8b00751
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    101
    De Novo Molecular Design by Combining Deep Autoencoder Recurrent Neural Networks with Generative Topographic Mapping
    Sattarov, Boris; Baskin, Igor I.; Horvath, Dragos; Marcou, Gilles; Bjerrum, Esben Jannik; Varnek, Alexandre
    Journal of Chemical Information and Modeling (2019), 59 (3), 1182-1196CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    Here we show that Generative Topog. Mapping (GTM) can be used to explore the latent space of the SMILES-based autoencoders and generate focused mol. libraries of interest. We have built a sequence-to-sequence neural network with Bidirectional Long Short-Term Memory layers and trained it on the SMILES strings from ChEMBL23. Very high reconstruction rates of the test set mols. were achieved (>98%), which are comparable to the ones reported in related publications. Using GTM, we have visualized the autoencoder latent space on the two-dimensional topog. map. Targeted map zones can be used for generating novel mol. structures by sampling assocd. latent space points and decoding them to SMILES. The sampling method based on a genetic algorithm was introduced to optimize compd. properties "on the fly". The generated focused mol. libraries were shown to contain original and a priori feasible compds. which, pending actual synthesis and testing, showed encouraging behavior in independent structure-based affinity estn. procedures (pharmacophore matching, docking).
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXjtlCisbc%253D&md5=44d2f043cc64e112b2caa7851ed1eb4e
  102. 102
    Winter, R.; Montanari, F.; Steffen, A.; Briem, H.; Nóe, F.; Clevert, D.-A. Efficient multi-objective molecular optimization in a continuous latent space. Chem. Sci. 2019, 10, 8016– 8024,  DOI: 10.1039/C9SC01928F
    [Crossref], [PubMed], [CAS], Google Scholar
    102
    Efficient multi-objective molecular optimization in a continuous latent space
    Winter, Robin; Montanari, Floriane; Steffen, Andreas; Briem, Hans; Noe, Frank; Clevert, Djork-Arne
    Chemical Science (2019), 10 (34), 8016-8024CODEN: CSHCCN; ISSN:2041-6520. (Royal Society of Chemistry)
    One of the main challenges in small mol. drug discovery is finding novel chem. compds. with desirable properties. In this work, we propose a novel method that combines in silico prediction of mol. properties such as biol. activity or pharmacokinetics with an in silico optimization algorithm, namely Particle Swarm Optimization. Our method takes a starting compd. as input and proposes new mols. with more desirable (predicted) properties. It navigates a machine-learned continuous representation of a drug-like chem. space guided by a defined objective function. The objective function combines multiple in silico prediction models, defined desirability ranges and substructure constraints. We demonstrate that our proposed method is able to consistently find more desirable mols. for the studied tasks in relatively short time. We hope that our method can support medicinal chemists in accelerating and improving the lead optimization process.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXhtlaltrjO&md5=c63599dc8a30df8f805ce5457ecc053a
  103. 103
    Chenthamarakshan, V.; Das, P.; Hoffman, C. S.; Strobelt, H.; Padhi, I.; Lim, W. K.; Hoover, B.; Manica, M.; Born, J.; Laino, T.; Mojsilovic, A. CogMol: Target-Specific and Selective Drug Design for COVID-19 Using Deep Generative Models. NeurIPS 2020 2020.
    Google Scholar
    There is no corresponding record for this reference.
  104. 104
    Kotsias, P.-C.; Arús-Pous, J.; Chen, H.; Engkvist, O.; Tyrchan, C.; Bjerrum, E. J. Direct steering of de novo molecular generation with descriptor conditional recurrent neural networks. Nature Machine Intelligence 2020, 2, 254– 265,  DOI: 10.1038/s42256-020-0174-5
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  105. 105
    Shayakhmetov, R.; Kuznetsov, M.; Zhebrak, A.; Kadurin, A.; Nikolenko, S.; Aliper, A.; Polykovskiy, D. Molecular Generation for Desired Transcriptome Changes With Ad- versarial Autoencoders. Front. Pharmacol. 2020, 11, 269,  DOI: 10.3389/fphar.2020.00269
    [Crossref], [PubMed], [CAS], Google Scholar
    105
    Molecular generation for desired transcriptome changes with adversarial autoencoders
    Shayakhmetov, Rim; Kuznetsov, Maksim; Zhebrak, Alexander; Kadurin, Artur; Nikolenko, Sergey; Aliper, Alexander; Polykovskiy, Daniil
    Frontiers in Pharmacology (2020), 11 (), 00269CODEN: FPRHAU; ISSN:1663-9812. (Frontiers Media S.A.)
    Gene expression profiles are useful for assessing the efficacy and side effects of drugs. In this paper, we propose a new generative model that infers drug mols. that could induce a desired change in gene expression. Our model-the Bidirectional Adversarial Autoencoder-explicitly separates cellular processes captured in gene expression changes into two feature sets: those related and unrelated to the drug incubation. The model uses related features to produce a drug hypothesis. We have validated our model on the LINCS L1000 dataset by generating mol. structures in the SMILES format for the desired transcriptional response. In the expts., we have shown that the proposed model can generate novel mol. structures that could induce a given gene expression change or predict a gene expression difference after incubation of a given mol. structure.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXhvFKksbbE&md5=36064731a6cb6395596e418d379f2782
  106. 106
    Ḿendez-Lucio, O.; Baillif, B.; Clevert, D.-A.; Rouquíe, D.; Wichard, J. De novo gener- ation of hit-like molecules from gene expression signatures using artificial intelligence. Nat. Commun. 2020, 11, 1– 10,  DOI: 10.1038/s41467-019-13807-w
    [Crossref], [PubMed], Google Scholar
    There is no corresponding record for this reference.
  107. 107
    Born, J.; Manica, M.; Oskooei, A.; Cadow, J.; Rodŕıguez Mart́ınez, M. PaccMannRL: Designing Anticancer Drugs From Transcriptomic Data via Reinforcement Learning. In Research in Computational Molecular Biology; Springer: Cham, 2020; pp 231– 233.
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  108. 108
    Jin, W.; Yang, K.; Barzilay, R.; Jaakkola, T. Learning Multimodal Graph-to-Graph Translation for Molecular Optimization. arXiv (Machine Learning) , January 28, 2019, 1812.01070, ver. 3.
    Google Scholar
    There is no corresponding record for this reference.
  109. 109
    Masuda, T.; Ragoza, M.; Koes, D. R. Generating 3D Molecular Structures Conditional on a Receptor Binding Site with Deep Generative Models. arXiv (Chemical Physics) , November 23, 2020, 2010.14442, ver. 3.
    Google Scholar
    There is no corresponding record for this reference.
  110. 110
    Kang, S.; Cho, K. Conditional Molecular Design with Deep Generative Models. J. Chem. Inf. Model. 2019, 59, 43– 52,  DOI: 10.1021/acs.jcim.8b00263
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    110
    Conditional Molecular Design with Deep Generative Models
    Kang, Seokho; Cho, Kyunghyun
    Journal of Chemical Information and Modeling (2019), 59 (1), 43-52CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    Although machine learning has been successfully used to propose novel mols. that satisfy desired properties, it is still challenging to explore a large chem. space efficiently. In this paper, we present a conditional mol. design method that facilitates generating new mols. with desired properties. The proposed model, which simultaneously performs both property prediction and mol. generation, is built as a semisupervised variational autoencoder trained on a set of existing mols. with only a partial annotation. We generate new mols. with desired properties by sampling from the generative distribution estd. by the model. We demonstrate the effectiveness of the proposed model by evaluating it on drug-like mols. The model improves the performance of property prediction by exploiting unlabeled mols. and efficiently generates novel mols. fulfilling various target conditions.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXhtlantb3N&md5=d2c3a3ff1f2189698828775e89c2a885
  111. 111
    Lim, J.; Hwang, S.-Y.; Moon, S.; Kim, S.; Kim, W. Y. Scaffold-based molecular design with a graph generative model. Chem. Sci. 2020, 11, 1153– 1164,  DOI: 10.1039/C9SC04503A
    [Crossref], [CAS], Google Scholar
    111
    Scaffold-based molecular design with a graph generative model
    Lim, Jaechang; Hwang, Sang-Yeon; Moon, Seokhyun; Kim, Seungsu; Kim, Woo Youn
    Chemical Science (2020), 11 (4), 1153-1164CODEN: CSHCCN; ISSN:2041-6520. (Royal Society of Chemistry)
    Searching for new mols. in areas like drug discovery often starts from the core structures of known mols. Such a method has called for a strategy of designing deriv. compds. retaining a particular scaffold as a substructure. On this account, our present work proposes a graph generative model that targets its use in scaffold-based mol. design. Our model accepts a mol. scaffold as input and extends it by sequentially adding atoms and bonds. The generated mols. are then guaranteed to contain the scaffold with certainty, and their properties can be controlled by conditioning the generation process on desired properties. The learned rule of extending mols. can well generalize to arbitrary kinds of scaffolds, including those unseen during learning. In the conditional generation of mols., our model can simultaneously control multiple chem. properties despite the search space constrained by fixing the substructure. As a demonstration, we applied our model to designing inhibitors of the epidermal growth factor receptor and show that our model can employ a simple semi-supervised extension to broaden its applicability to situations where only a small amt. of data is available.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXit1Ortr%252FO&md5=be36a65abcce15f18b4c1e529bffd905
  112. 112
    Varnek, A., Ed. Tutorials in chemoinformatics; John Wiley & Sons, Inc: Hoboken, NJ, 2017.
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  113. 113
    Engel, T., Gasteiger, J., Eds. Applied chemoinformatics: achievements and future opportunities; Wiley-VCH: Weinheim, 2018; OCLC: 1034693178.
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  114. 114
    Kadurin, A.; Aliper, A.; Kazennov, A.; Mamoshina, P.; Vanhaelen, Q.; Khrabrov, K.; Zhavoronkov, A. The cornucopia of meaningful leads: Applying deep adversarial au- toencoders for new molecule development in oncology. Oncotarget 2017, 8, 10883– 10890,  DOI: 10.18632/oncotarget.14073
    [Crossref], [PubMed], [CAS], Google Scholar
    114
    The cornucopia of meaningful leads: Applying deep adversarial autoencoders for new molecule development in oncology
    Kadurin Artur; Khrabrov Kuzma; Kadurin Artur; Aliper Alexander; Kazennov Andrey; Mamoshina Polina; Vanhaelen Quentin; Zhavoronkov Alex; Kadurin Artur; Kadurin Artur; Kazennov Andrey; Zhavoronkov Alex; Mamoshina Polina; Zhavoronkov Alex
    Oncotarget (2017), 8 (7), 10883-10890 ISSN:.
    Recent advances in deep learning and specifically in generative adversarial networks have demonstrated surprising results in generating new images and videos upon request even using natural language as input. In this paper we present the first application of generative adversarial autoencoders (AAE) for generating novel molecular fingerprints with a defined set of parameters. We developed a 7-layer AAE architecture with the latent middle layer serving as a discriminator. As an input and output the AAE uses a vector of binary fingerprints and concentration of the molecule. In the latent layer we also introduced a neuron responsible for growth inhibition percentage, which when negative indicates the reduction in the number of tumor cells after the treatment. To train the AAE we used the NCI-60 cell line assay data for 6252 compounds profiled on MCF-7 cell line. The output of the AAE was used to screen 72 million compounds in PubChem and select candidate molecules with potential anti-cancer properties. This approach is a proof of concept of an artificially-intelligent drug discovery engine, where AAEs are used to generate new molecular fingerprints with the desired molecular properties.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A280%3ADC%252BC1c%252FpvFKruw%253D%253D&md5=677ef0264494eb8a7ef8c6584c1202ab
  115. 115
    Alpaydin, E. Introduction to machine learning, 2nd ed.; Adaptive computation and machine learning; MIT Press: Cambridge, Mass, 2010; OCLC: ocn317698631.
    Google Scholar
    There is no corresponding record for this reference.
  116. 116
    Raschka, S. Python machine learning: unlock deeper insights into machine learning with this vital guide to cutting-edge predictive analytics; Community experience distilled; Packt Publishing Open Source: Birmingham, UK; Mumbai, 2016.
    Google Scholar
    There is no corresponding record for this reference.
  117. 117
    Frazier, P. I. A Tutorial on Bayesian Optimization. arXiv (Machine Learning) , July 8, 2018, 1807.02811, ver. 1.
    Google Scholar
    There is no corresponding record for this reference.
  118. 118
    Shahriari, B.; Swersky, K.; Wang, Z.; Adams, R. P.; De Freitas, N. Taking the human out of the loop: A review of Bayesian optimization. Proc. IEEE 2016, 104, 148– 175,  DOI: 10.1109/JPROC.2015.2494218
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  119. 119
    Das, P.; Sercu, T.; Wadhawan, K.; Padhi, I.; Gehrmann, S.; Cipcigan, F.; Chen- thamarakshan, V.; Strobelt, H.; Santos, C. D.; Chen, P.-Y.; Yang, Y. Y.; Tan, J.; Hedrick, J.; Crain, J.; Mojsilovic, A. Accelerating antimicrobial discovery with controllable deep generative models and molecular dynamics. arXiv (Machine Learning) , February 26, 2020, 2005.11248, ver. 2.
    Google Scholar
    There is no corresponding record for this reference.
  120. 120
    Kingma, D. P.; Mohamed, S.; Rezende, D. J.; Welling, M. Semi-supervised learning with deep generative models. Adv. Neural Inf. Process. Syst. 2014, 3581– 3589
    Google Scholar
    There is no corresponding record for this reference.
  121. 121
    Gao, W.; Coley, C. W. The synthesizability of molecules proposed by generative mod- els. J. Chem. Inf. Model. 2020, 60, 5714– 5723,  DOI: 10.1021/acs.jcim.0c00174
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    121
    The Synthesizability of Molecules Proposed by Generative Models
    Gao, Wenhao; Coley, Connor W.
    Journal of Chemical Information and Modeling (2020), 60 (12), 5714-5723CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    The discovery of functional mols. is an expensive and time-consuming process, exemplified by the rising costs of small mol. therapeutic discovery. One class of techniques of growing interest for early stage drug discovery is de novo mol. generation and optimization, catalyzed by the development of new deep learning approaches. These techniques can suggest novel mol. structures intended to maximize a multiobjective function, e.g., suitability as a therapeutic against a particular target, without relying on brute-force exploration of a chem. space. However, the utility of these approaches is stymied by ignorance of synthesizability. To highlight the severity of this issue, we use a data-driven computer-aided synthesis planning program to quantify how often mols. proposed by state-of-the-art generative models cannot be readily synthesized. Our anal. demonstrates that there are several tasks for which these models generate unrealistic mol. structures despite performing well on popular quant. benchmarks. Synthetic complexity heuristics can successfully bias generation toward synthetically tractable chem. space, although doing so necessarily detracts from the primary objective. This anal. suggests that to improve the utility of these models in real discovery workflows, new algorithm development is warranted.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXmsFKrurw%253D&md5=5f8ff43137b0489f41b5c58518bec270
  122. 122
    Horwood, J.; Noutahi, E. Molecular Design in Synthetically Accessible Chemical Space via Deep Reinforcement Learning. ACS Omega 2020, 5, 32984– 32994,  DOI: 10.1021/acsomega.0c04153
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    122
    Molecular Design in Synthetically Accessible Chemical Space via Deep Reinforcement Learning
    Horwood, Julien; Noutahi, Emmanuel
    ACS Omega (2020), 5 (51), 32984-32994CODEN: ACSODF; ISSN:2470-1343. (American Chemical Society)
    The fundamental goal of generative drug design is to propose optimized mols. that meet predefined activity, selectivity, and pharmacokinetic criteria. Despite recent progress, we argue that existing generative methods are limited in their ability to favorably shift the distributions of mol. properties during optimization. We instead propose a novel Reinforcement Learning framework for mol. design in which an agent learns to directly optimize through a space of synthetically accessible drug-like mols. This becomes possible by defining transitions in our Markov decision process as chem. reactions and allows us to leverage synthetic routes as an inductive bias. We validate our method by demonstrating that it outperforms existing state-of-the-art approaches in the optimization of pharmacol. relevant objectives, while results on multi-objective optimization tasks suggest increased scalability to realistic pharmaceutical design problems.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXisFOmtL3P&md5=1037d12603e7a52332c11e66e8885838
  123. 123
    Gottipati, S. K.; Sattarov, B.; Niu, S.; Pathak, Y.; Wei, H.; Liu, S.; Blackburn, S.; Thomas, K.; Coley, C.; Tang, J. Learning to navigate the synthetically accessible chemical space using reinforcement learning. Int. Conf. Mach. Learn. 2020, 3668– 3679
    Google Scholar
    There is no corresponding record for this reference.
  124. 124
    Bradshaw, J.; Paige, B.; Kusner, M. J.; Segler, M.; Hernández-Lobato, J. M. Barking up the right tree: an approach to search over molecule synthesis DAGs. Adv. Neural Inf. Process. Syst. 2020, 6852– 6866
    Google Scholar
    There is no corresponding record for this reference.
  125. 125
    Imrie, F.; Bradley, A. R.; van der Schaar, M.; Deane, C. M. Deep generative models for 3d linker design. J. Chem. Inf. Model. 2020, 60, 1983– 1995,  DOI: 10.1021/acs.jcim.9b01120
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    125
    Deep Generative Models for 3D Linker Design
    Imrie, Fergus; Bradley, Anthony R.; van der Schaar, Mihaela; Deane, Charlotte M.
    Journal of Chemical Information and Modeling (2020), 60 (4), 1983-1995CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    Rational compd. design remains a challenging problem for both computational methods and medicinal chemists. Computational generative methods have begun to show promising results for the design problem. However, they have not yet used the power of three-dimensional (3D) structural information. We have developed a novel graph-based deep generative model that combines state-of-the-art machine learning techniques with structural knowledge. Our method ("DeLinker") takes two fragments or partial structures and designs a mol. incorporating both. The generation process is protein-context-dependent, utilizing the relative distance and orientation between the partial structures. This 3D information is vital to successful compd. design, and we demonstrate its impact on the generation process and the limitations of omitting such information. In a large-scale evaluation, DeLinker designed 60% more mols. with high 3D similarity to the original mol. than a database baseline. When considering the more relevant problem of longer linkers with at least five atoms, the outperformance increased to 200%. We demonstrate the effectiveness and applicability of this approach on a diverse range of design problems: fragment linking, scaffold hopping, and proteolysis targeting chimera (PROTAC) design. As far as we are aware, this is the first mol. generative model to incorporate 3D structural information directly in the design process. The code is available at https://github.com/oxpig/DeLinker.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXltlenurs%253D&md5=430c929761c5020d1013538e028ba1bc
  126. 126
    Yang, Y.; Zheng, S.; Su, S.; Zhao, C.; Xu, J.; Chen, H. SyntaLinker: automatic fragment linking with deep conditional transformer neural networks. Chem. Sci. 2020, 11, 8312– 8322,  DOI: 10.1039/D0SC03126G
    [Crossref], [PubMed], [CAS], Google Scholar
    126
    SyntaLinker: automatic fragment linking with deep conditional transformer neural networks
    Yang, Yuyao; Zheng, Shuangjia; Su, Shimin; Zhao, Chao; Xu, Jun; Chen, Hongming
    Chemical Science (2020), 11 (31), 8312-8322CODEN: CSHCCN; ISSN:2041-6520. (Royal Society of Chemistry)
    Linking fragments to generate a focused compd. library for a specific drug target is one of the challenges in fragment-based drug design (FBDD). Hereby, we propose a new program named SyntaLinker, which is based on a syntactic pattern recognition approach using deep conditional transformer neural networks. This state-of-the-art transformer can link mol. fragments automatically by learning from the knowledge of structures in medicinal chem. databases (e.g.ChEMBL database). Conventionally, linking mol. fragments was viewed as connecting substructures that were predefined by empirical rules. In SyntaLinker, however, the rules of linking fragments can be learned implicitly from known chem. structures by recognizing syntactic patterns embedded in SMILES notations. With deep conditional transformer neural networks, SyntaLinker can generate mol. structures based on a given pair of fragments and addnl. restrictions. Case studies have demonstrated the advantages and usefulness of SyntaLinker in FBDD.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXhsVagsrvK&md5=403cc7a09f96ecbe27afa3ce6bee51a2
  127. 127
    Tan, X. Automated design and optimization of multitarget schizophrenia drug candidates by deep learning. Eur. J. Med. Chem. 2020, 204, 112572,  DOI: 10.1016/j.ejmech.2020.112572
    [Crossref], [PubMed], [CAS], Google Scholar
    127
    Automated design and optimization of multitarget schizophrenia drug candidates by deep learning
    Tan, Xiaoqin; Jiang, Xiangrui; He, Yang; Zhong, Feisheng; Li, Xutong; Xiong, Zhaoping; Li, Zhaojun; Liu, Xiaohong; Cui, Chen; Zhao, Qingjie; Xie, Yuanchao; Yang, Feipu; Wu, Chunhui; Shen, Jingshan; Zheng, Mingyue; Wang, Zhen; Jiang, Hualiang
    European Journal of Medicinal Chemistry (2020), 204 (), 112572CODEN: EJMCA5; ISSN:0223-5234. (Elsevier Masson SAS)
    Complex neuropsychiatric diseases such as schizophrenia require drugs that can target multiple G protein-coupled receptors (GPCRs) to modulate complex neuropsychiatric functions. Here, we report an automated system comprising a deep recurrent neural network (RNN) and a multitask deep neural network (MTDNN) to design and optimize multitarget antipsychotic drugs. The system has successfully generated novel mol. structures with desired multiple target activities, among which high-ranking compd. 3 was synthesized, and demonstrated potent activities against dopamine D2, serotonin 5-HT1A and 5-HT2A receptors. Hit expansion based on the MTDNN was performed, 6 analogs of compd. 3 were evaluated exptl., among which compd. 8 not only exhibited specific polypharmacol. profiles but also showed antipsychotic effect in animal models with low potential for sedation and catalepsy, highlighting their suitability for further preclin. studies. The approach can be an efficient tool for designing lead compds. with multitarget profiles to achieve the desired efficacy in the treatment of complex neuropsychiatric diseases.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXhsVCksbnI&md5=dd01837fd3bdbb488e8781cea7d5103b
  128. 128
    Yang, Y.; Zhang, R.; Li, Z.; Mei, L.; Wan, S.; Ding, H.; Chen, Z.; Xing, J.; Feng, H.; Han, J.; Jiang, H.; Zheng, M.; Luo, C.; Zhou, B. Discovery of Highly Potent, Selec- tive, and Orally Efficacious p300/CBP Histone Acetyltransferases Inhibitors. J. Med. Chem. 2020, 63, 1337– 1360,  DOI: 10.1021/acs.jmedchem.9b01721
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    128
    Discovery of Highly Potent, Selective, and Orally Efficacious p300/CBP Histone Acetyltransferases Inhibitors
    Yang, Yaxi; Zhang, Rukang; Li, Zhaojun; Mei, Lianghe; Wan, Shili; Ding, Hong; Chen, Zhifeng; Xing, Jing; Feng, Huijin; Han, Jie; Jiang, Hualiang; Zheng, Mingyue; Luo, Cheng; Zhou, Bing
    Journal of Medicinal Chemistry (2020), 63 (3), 1337-1360CODEN: JMCMAR; ISSN:0022-2623. (American Chemical Society)
    P300 and CREB-binding protein (CBP) are ubiquitously expressed pleiotropic lysine acetyltransferases and play a key role as transcriptional co-activators that are essential for a multitude of cellular processes. Despite great importance, there is a lack of highly selective, potent, druglike p300/CBP inhibitors. Through the artificial-intelligence-assisted drug discovery pipeline and further optimization, we reported the discovery of novel, highly selective, potent small-mol. inhibitors of p300/CBP histone acetyltransferases (HAT) with desired druglike properties, exemplified by B026. Our data demonstrated that B026, with half maximal inhibitory concn. (IC50) values of 1.8 nM to p300 and 9.5 nM to CBP enzyme inhibitory activity, is the most potent, selective p300/CBP HAT inhibitor. Moreover, B026 achieves significant and dose-dependent tumor growth inhibition in an animal model of human cancer, suggesting that B026 is a highly promising p300/CBP HAT inhibitor and warrants extensive preclin. investigation as a potential clin. development candidate.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXksVKnuw%253D%253D&md5=dd4b2562fd8451d2ce1cff0c3fe027fa
  129. 129
    Grisoni, F.; Huisman, B.; Button, A.; Moret, M.; Atz, K.; Merk, D.; Schneider, G. Combining generative artificial intelligence and on-chip synthesis for de novo drug design. ChemRxiv , December 30, 2020, ver. 1.  DOI: 10.26434/chemrxiv.13498587.v1 .
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  130. 130
    Shaker, N.; Abou-Zleikha, M.; AlAmri, M.; Mehellou, Y. A Generative Deep Learning Approach for the Discovery of SARS CoV2 Protease Inhibitors. ChemRxiv , April 23, 2020, ver. 1.  DOI: 10.26434/chemrxiv.12170337.v1 .
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  131. 131
    Born, J.; Manica, M.; Cadow, J.; Markert, G.; Mill, N. A.; Filipavicius, M.; Mart́ınez, M. R. PaccMannRL on SARS-CoV-2: Designing antiviral candidates with conditional generative models. arXiv (Quantitative Methods) , July 6, 2020, 2005.13285, ver. 3.
    Google Scholar
    There is no corresponding record for this reference.
  132. 132
    Hachmann, J.; Olivares-Amaya, R.; Atahan-Evrenk, S.; Amador-Bedolla, C.; Śanchez- Carrera, R. S.; Gold-Parker, A.; Vogt, L.; Brockway, A. M.; Aspuru-Guzik, A. The Harvard clean energy project: large-scale computational screening and design of or- ganic photovoltaics on the world community grid. J. Phys. Chem. Lett. 2011, 2, 2241– 2251,  DOI: 10.1021/jz200866s
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    132
    The Harvard Clean Energy Project: Large-Scale Computational Screening and Design of Organic Photovoltaics on the World Community Grid
    Hachmann, Johannes; Olivares-Amaya, Roberto; Atahan-Evrenk, Sule; Amador-Bedolla, Carlos; Sanchez-Carrera, Roel S.; Gold-Parker, Aryeh; Vogt, Leslie; Brockway, Anna M.; Aspuru-Guzik, Alan
    Journal of Physical Chemistry Letters (2011), 2 (17), 2241-2251CODEN: JPCLCD; ISSN:1948-7185. (American Chemical Society)
    This Perspective introduces the Harvard Clean Energy Project (CEP), a theory-driven search for the next generation of org. solar cell materials. We give a broad overview of its setup and infrastructure, present first results, and outline upcoming developments. CEP has established an automated, high-throughput, in silico framework to study potential candidate structures for org. photovoltaics. The current project phase is concerned with the characterization of millions of mol. motifs using first-principles quantum chem. The scale of this study requires a correspondingly large computational resource, which is provided by distributed volunteer computing on IBM's World Community Grid. The results are compiled and analyzed in a ref. database and will be made available for public use. In addn. to finding specific candidates with certain properties, it is the goal of CEP to illuminate and understand the structure-property relations in the domain of org. electronics. Such insights can open the door to a rational and systematic design of future high-performance materials. The computational work in CEP is tightly embedded in a collaboration with experimentalists, who provide valuable input and feedback to the project.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC3MXhtVKht7rK&md5=68f7f2d3a8a5b5fe6cb2c3b677f444bf
  133. 133
    Jørgensen, P. B.; Mesta, M.; Shil, S.; Lastra, J. M. G.; Wedel, K.; Thygesen, K. S.; Schmidt, M. N. Machine learning-based screening of complex molecules for polymer solar cells. J. Chem. Phys. 2018, 148, 241735,  DOI: 10.1063/1.5023563
    [Crossref], [PubMed], [CAS], Google Scholar
    133
    Machine learning-based screening of complex molecules for polymer solar cells
    Jorgensen Peter Bjorn; Schmidt Mikkel N; Mesta Murat; Garcia Lastra Juan Maria; Shil Suranjan; Jacobsen Karsten Wedel; Thygesen Kristian Sommer
    The Journal of chemical physics (2018), 148 (24), 241735 ISSN:.
    Polymer solar cells admit numerous potential advantages including low energy payback time and scalable high-speed manufacturing, but the power conversion efficiency is currently lower than for their inorganic counterparts. In a Phenyl-C_61-Butyric-Acid-Methyl-Ester (PCBM)-based blended polymer solar cell, the optical gap of the polymer and the energetic alignment of the lowest unoccupied molecular orbital (LUMO) of the polymer and the PCBM are crucial for the device efficiency. Searching for new and better materials for polymer solar cells is a computationally costly affair using density functional theory (DFT) calculations. In this work, we propose a screening procedure using a simple string representation for a promising class of donor-acceptor polymers in conjunction with a grammar variational autoencoder. The model is trained on a dataset of 3989 monomers obtained from DFT calculations and is able to predict LUMO and the lowest optical transition energy for unseen molecules with mean absolute errors of 43 and 74 meV, respectively, without knowledge of the atomic positions. We demonstrate the merit of the model for generating new molecules with the desired LUMO and optical gap energies which increases the chance of finding suitable polymers by more than a factor of five in comparison to the randomised search used in gathering the training set.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A280%3ADC%252BB3c%252FjsFOhuw%253D%253D&md5=50686fea495291fa2f75ac7b3ed3a95e
  134. 134
    Yuan, Q.; Santana-Bonilla, A.; Zwijnenburg, M. A.; Jelfs, K. E. Molecular generation targeting desired electronic properties via deep generative models. Nanoscale 2020, 12, 6744– 6758,  DOI: 10.1039/C9NR10687A
    [Crossref], [PubMed], [CAS], Google Scholar
    134
    Molecular generation targeting desired electronic properties via deep generative models
    Yuan, Qi; Santana-Bonilla, Alejandro; Zwijnenburg, Martijn A.; Jelfs, Kim E.
    Nanoscale (2020), 12 (12), 6744-6758CODEN: NANOHL; ISSN:2040-3372. (Royal Society of Chemistry)
    As we seek to discover new functional materials, we need ways to explore the vast chem. space of precursor building blocks, not only generating large nos. of possible building blocks to investigate, but trying to find non-obvious options, that we might not suggest by chem. experience alone. Artificial intelligence techniques provide a possible avenue to generate large nos. of org. building blocks for functional materials, and can even do so from very small initial libraries of known building blocks. Specifically, we demonstrate the application of deep recurrent neural networks for the exploration of the chem. space of building blocks for a test case of donor-acceptor oligomers with specific electronic properties. The recurrent neural network learned how to produce novel donor-acceptor oligomers by trading off between selected at. substitutions, such as halogenation or methylation, and mol. features such as the oligomer's size. The electronic and structural properties of the generated oligomers can be tuned by sampling from different subsets of the training database, which enabled us to enrich the library of donor-acceptors towards desired properties. We generated approx. 1700 new donor-acceptor oligomers with a recurrent neural network tuned to target oligomers with a HOMO-LUMO gap <2 eV and a dipole moment <2 Debye, which could have potential application in org. photovoltaics.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXksFKktro%253D&md5=3efbe2d03b744fd333c0423f441e60d7
  135. 135
    Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, L.; Polosukhin, I. Attention is all you need. arXiv (Computation and Language) , December 6, 2017, 1706.03762, ver. 5.
    Google Scholar
    There is no corresponding record for this reference.
  136. 136
    Devlin, J.; Chang, M.-W.; Lee, K.; Toutanova, K. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv (Computation and Language) , May 24, 2019, 1810.04805, ver. 2.
    Google Scholar
    There is no corresponding record for this reference.
  137. 137
    Brown, T. B.; Mann, B.; Ryder, N.; Subbiah, M.; Kaplan, J.; Dhariwal, P.; Nee- lakantan, A.; Shyam, P.; Sastry, G.; Askell, A., Language models are few-shot learners. arXiv (Computation and Language) , July 22, 2020, 2005.14165, ver. 4.
    Google Scholar
    There is no corresponding record for this reference.
  138. 138
    Grechishnikova, D. Transformer neural network for protein-specific de novo drug gen- eration as a machine translation problem. Sci. Rep. 2021, 11, 1– 13,  DOI: 10.1038/s41598-020-79682-4
    [Crossref], [PubMed], Google Scholar
    There is no corresponding record for this reference.
  139. 139
    Zheng, S.; Lei, Z.; Ai, H.; Chen, H.; Deng, D.; Yang, Y. Deep Scaffold Hopping with Multi-modal Transformer Neural Networks. ChemRxiv , September 28, 2020, ver. 1. DOI: 10.26434/chemrxiv.13011767.v1 .
    [Crossref], Google Scholar
    There is no corresponding record for this reference.

Cited By

This article is cited by 32 publications.

  1. Daiki Erikawa, Nobuaki Yasuo, Takamasa Suzuki, Shogo Nakamura, Masakazu Sekijima. Gargoyles: An Open Source Graph-Based Molecular Optimization Method Based on Deep Reinforcement Learning. ACS Omega 2023, Article ASAP.
  2. Giuseppe Lamanna, Pietro Delre, Gilles Marcou, Michele Saviano, Alexandre Varnek, Dragos Horvath, Giuseppe Felice Mangiatordi. GENERA: A Combined Genetic/Deep-Learning Algorithm for Multiobjective Target-Oriented De Novo Design. Journal of Chemical Information and Modeling 2023, 63 (16) , 5107-5119. https://doi.org/10.1021/acs.jcim.3c00963
  3. Xu Qian, Xiaowen Dai, Lin Luo, Mingde Lin, Yuan Xu, Yang Zhao, Dingfang Huang, Haodi Qiu, Li Liang, Haichun Liu, Yingbo Liu, Lingxi Gu, Tao Lu, Yadong Chen, Yanmin Zhang. An Interpretable Multitask Framework BiLAT Enables Accurate Prediction of Cyclin-Dependent Protein Kinase Inhibitors. Journal of Chemical Information and Modeling 2023, 63 (11) , 3350-3368. https://doi.org/10.1021/acs.jcim.3c00473
  4. Tobiasz Ciepliński, Tomasz Danel, Sabina Podlewska, Stanisław Jastrzȩbski. Generative Models Should at Least Be Able to Design Molecules That Dock Well: A New Benchmark. Journal of Chemical Information and Modeling 2023, 63 (11) , 3238-3247. https://doi.org/10.1021/acs.jcim.2c01355
  5. Yiming Wang, Kathleen J. Stebe, Cesar de la Fuente-Nunez, Ravi Radhakrishnan. Computational Design of Peptides for Biomaterials Applications. ACS Applied Bio Materials 2023, Article ASAP.
  6. William Bort, Daniyar Mazitov, Dragos Horvath, Fanny Bonachera, Arkadii Lin, Gilles Marcou, Igor Baskin, Timur Madzhidov, Alexandre Varnek. Inverse QSAR: Reversing Descriptor-Driven Prediction Pipeline Using Attention-Based Conditional Variational Autoencoder. Journal of Chemical Information and Modeling 2022, 62 (22) , 5471-5484. https://doi.org/10.1021/acs.jcim.2c01086
  7. Hanna Türk, Elisabetta Landini, Christian Kunkel, Johannes T. Margraf, Karsten Reuter. Assessing Deep Generative Models in Chemical Composition Space. Chemistry of Materials 2022, 34 (21) , 9455-9467. https://doi.org/10.1021/acs.chemmater.2c01860
  8. Chuan Li, Chenghui Wang, Ming Sun, Yan Zeng, Yuan Yuan, Qiaolin Gou, Guangchuan Wang, Yanzhi Guo, Xuemei Pu. Correlated RNN Framework to Quickly Generate Molecules with Desired Properties for Energetic Materials in the Low Data Regime. Journal of Chemical Information and Modeling 2022, 62 (20) , 4873-4887. https://doi.org/10.1021/acs.jcim.2c00997
  9. Bin Xi, Kin Fai Tse, Tsz Fung Kok, Ho Ming Chan, Man Kit Chan, Ho Yin Chan, Kwan Yue Clinton Wong, Shing Hei Robin Yuen, Junyi Zhu. Machine-Learning-Assisted Acceleration on High-Symmetry Materials Search: Space Group Predictions from Band Structures. The Journal of Physical Chemistry C 2022, 126 (29) , 12264-12273. https://doi.org/10.1021/acs.jpcc.2c03156
  10. Weixin Xie, Fanhao Wang, Yibo Li, Luhua Lai, Jianfeng Pei. Advances and Challenges in De Novo Drug Design Using Three-Dimensional Deep Generative Models. Journal of Chemical Information and Modeling 2022, 62 (10) , 2269-2279. https://doi.org/10.1021/acs.jcim.2c00042
  11. Teresa Maria Creanza, Giuseppe Lamanna, Pietro Delre, Marialessandra Contino, Nicola Corriero, Michele Saviano, Giuseppe Felice Mangiatordi, Nicola Ancona. DeLA-Drug: A Deep Learning Algorithm for Automated Design of Druglike Analogues. Journal of Chemical Information and Modeling 2022, 62 (6) , 1411-1424. https://doi.org/10.1021/acs.jcim.2c00205
  12. Ana L. Chávez-Hernández, José L. Medina-Franco. Natural products subsets: Generation and characterization. Artificial Intelligence in the Life Sciences 2023, 3 , 100066. https://doi.org/10.1016/j.ailsci.2023.100066
  13. Jonghwan Choi, Sangmin Seo, Sanghyun Park. COMA: efficient structure-constrained molecular generation using contractive and margin losses. Journal of Cheminformatics 2023, 15 (1) https://doi.org/10.1186/s13321-023-00679-y
  14. Linde Schoenmaker, Olivier J. M. Béquignon, Willem Jespers, Gerard J. P. van Westen. UnCorrupt SMILES: a novel approach to de novo design. Journal of Cheminformatics 2023, 15 (1) https://doi.org/10.1186/s13321-023-00696-x
  15. Morgan Thomas, Andreas Bender, Chris de Graaf. Integrating structure-based approaches in generative molecular design. Current Opinion in Structural Biology 2023, 79 , 102559. https://doi.org/10.1016/j.sbi.2023.102559
  16. Felix Potlitz, Andreas Link, Lukas Schulig. Advances in the discovery of new chemotypes through ultra-large library docking. Expert Opinion on Drug Discovery 2023, 18 (3) , 303-313. https://doi.org/10.1080/17460441.2023.2171984
  17. Alex Sebastião Constâncio, Denise Fukumi Tsunoda, Helena de Fátima Nunes Silva, Jocelaine Martins da Silveira, Deborah Ribeiro Carvalho, . Deception detection with machine learning: A systematic review and statistical analysis. PLOS ONE 2023, 18 (2) , e0281323. https://doi.org/10.1371/journal.pone.0281323
  18. Tomasz Danel, Jan Łęski, Sabina Podlewska, Igor T. Podolak. Docking-based generative approaches in the search for new drug candidates. Drug Discovery Today 2023, 28 (2) , 103439. https://doi.org/10.1016/j.drudis.2022.103439
  19. Mher Matevosyan, Vardan Harutyunyan, Narek Abelyan, Hamlet Khachatryan, Irina Tirosyan, Yeva Gabrielyan, Valter Sahakyan, Smbat Gevorgyan, Vahram Arakelov, Grigor Arakelov, Hovakim Zakaryan. Design of new chemical entities targeting both native and H275Y mutant influenza a virus by deep reinforcement learning. Journal of Biomolecular Structure and Dynamics 2022, 7819 , 1-15. https://doi.org/10.1080/07391102.2022.2158936
  20. Yueshan Li, Liting Zhang, Yifei Wang, Jun Zou, Ruicheng Yang, Xinling Luo, Chengyong Wu, Wei Yang, Chenyu Tian, Haixing Xu, Falu Wang, Xin Yang, Linli Li, Shengyong Yang. Generative deep learning enables the discovery of a potent and selective RIPK1 inhibitor. Nature Communications 2022, 13 (1) https://doi.org/10.1038/s41467-022-34692-w
  21. Parinaz Naseri, George Goussetis, Nelson J. G. Fonseca, Sean V. Hum. Synthesis of multi-band reflective polarizing metasurfaces using a generative adversarial network. Scientific Reports 2022, 12 (1) https://doi.org/10.1038/s41598-022-20851-y
  22. Mateusz K. Bieniek, Ben Cree, Rachael Pirie, Joshua T. Horton, Natalie J. Tatum, Daniel J. Cole. An open-source molecular builder and free energy preparation workflow. Communications Chemistry 2022, 5 (1) https://doi.org/10.1038/s42004-022-00754-9
  23. Lucian Chan, Rajendra Kumar, Marcel Verdonk, Carl Poelking. A multilevel generative framework with hierarchical self-contrasting for bias control and transparency in structure-based ligand design. Nature Machine Intelligence 2022, 4 (12) , 1130-1142. https://doi.org/10.1038/s42256-022-00564-7
  24. Chong Lu, Shien Liu, Weihua Shi, Jun Yu, Zhou Zhou, Xiaoxiao Zhang, Xiaoli Lu, Faji Cai, Ning Xia, Yikai Wang. Systemic evolutionary chemical space exploration for drug discovery. Journal of Cheminformatics 2022, 14 (1) https://doi.org/10.1186/s13321-022-00598-4
  25. Anthony Hughes, David Winkler, James Carr, P. Lee, Y. Yang, Majid Laleh, Mike Tan. Corrosion Inhibition, Inhibitor Environments, and the Role of Machine Learning. Corrosion and Materials Degradation 2022, 3 (4) , 672-693. https://doi.org/10.3390/cmd3040037
  26. Matthias Unterhuber, Karl-Patrik Kresoja, Philipp Lurz, Holger Thiele. Artificial intelligence in proteomics: new frontiers from risk prediction to treatment?. European Heart Journal 2022, 43 (43) , 4525-4527. https://doi.org/10.1093/eurheartj/ehac391
  27. Yiming Ma, Yue Niu, Huaiyu Yang, Jiayu Dai, Jiawei Lin, Huiqi Wang, Songgu Wu, Qiuxiang Yin, Ling Zhou, Junbo Gong. Prediction and design of cyclodextrin inclusion complexes formation via machine learning-based strategies. Chemical Engineering Science 2022, 261 , 117946. https://doi.org/10.1016/j.ces.2022.117946
  28. Keerthi Krishnan, Ryan Kassab, Steve Agajanian, Gennady Verkhivker. Interpretable Machine Learning Models for Molecular Design of Tyrosine Kinase Inhibitors Using Variational Autoencoders and Perturbation-Based Approach of Chemical Space Exploration. International Journal of Molecular Sciences 2022, 23 (19) , 11262. https://doi.org/10.3390/ijms231911262
  29. Kailasam N. Vennila, Kuppanagounder P. Elango. Multimodal generative neural networks and molecular dynamics based identification of PDK1 PIF-pocket modulators. Molecular Systems Design & Engineering 2022, 7 (9) , 1085-1092. https://doi.org/10.1039/D2ME00051B
  30. Wenfei Fan, Ruochun Jin, Ping Lu, Chao Tian, Ruiqi Xu. Towards event prediction in temporal graphs. Proceedings of the VLDB Endowment 2022, 15 (9) , 1861-1874. https://doi.org/10.14778/3538598.3538608
  31. Jia-Shun Cao, Run-Ze Xu, Jing-Yang Luo, Qian Feng, Fang Fang. Rapid quantification of intracellular polyhydroxyalkanoates via fluorescence techniques: A critical review. Bioresource Technology 2022, 350 , 126906. https://doi.org/10.1016/j.biortech.2022.126906
  32. Shuheng Huang, Hu Mei, Laichun Lu, Minyao Qiu, Xiaoqi Liang, Lei Xu, Zuyin Kuang, Yu Heng, Xianchao Pan. De Novo Molecular Design of Caspase-6 Inhibitors by a GRU-Based Recurrent Neural Network Combined with a Transfer Learning Approach. Pharmaceuticals 2021, 14 (12) , 1249. https://doi.org/10.3390/ph14121249
Download PDF

  • Abstract

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 1

    Figure 1. Acetaminophen (center) under various molecular representations. Top-left: Sequence based representations. Prior to being fed to the models, these sequences are also usually one-hot encoded. Top-right: Graph-based representations. While connection matrices are a suitable input for standard architectures, graphs can also be directly handled using graph neural networks. Bottom: Three dimensional representations, images from PubChem. (26) Graphs may be enhanced by including 3D information as node attributes, such as internal distances and angles, or based on a coordinate system such as Cartesian space. Molecular surfaces can be voxelized into a 3D grid for easier processing.

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 2

    Figure 2. Top-left: Three layer Recurrent Neural Network (RNN) both rolled and unrolled. In each layer, the output of a step, besides flowing to the next layer, also flows to the next step of the layer itself. These recurrent connections are depicted in the unfolded view of the network as vertical arrows. Top-right: Variational Autoencoder (VAE) where the input is encoded to the parameters of a statistical distribution, namely, the means (μ) and standard deviation (σ). In practice, these correspond to two vectors which, on the sampling step, are interpreted as a set of means and standard deviations. Bottom-left: Generative Adversarial Network (GAN) composed by a generator and a discriminator. Training seeks not a minimum but a useful equilibrium between the generator and the discriminator. Bottom-right: Adversarial Autoencoder (AAE) where the attached discriminator must discern between encoded points and samples drawn from a prior statistical distribution.

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 3

    Figure 3. Three layer RNN, unfolded over four time-steps. In autoregressive sequence generation, the process is started with a special start token, here “G”. The model then predicts the next token, which is sampled and used as input for the next step. Generation ends when a stop token is predicted.

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 4

    Figure 4. Left: In sequential graph generation, a graph is built by evaluating a current partial graph, adding a node/edge and repeating until the network outputs a stop signal. Right: In the one-shot generation of graphs, probabilities over the full adjacency matrix and node/edge attribute tensors are produced. The graph is then obtained by taking a sample or the argmax of these outputs.

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 5

    Figure 5. Left: General procedure for the generation of 3D shapes as proposed by Skalic et al. (72) The convolutional decoder of a VAE is used to produce a 3D molecular shape which is converted to SMILES by a captioning network. Right: General process for generating molecules as 3D point sets, proposed by Gebauer et al. (73) It is conceptually similar to the sequential graph generation, operating on point sets with an internal coordinate system.

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 6

    Figure 6. In transfer learning, a general model is first trained on a large data set and then fine-tuned toward generating the desired properties with a smaller, focused, data set.

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 7

    Figure 7. Top: The model is first pretrained through maximum likelihood estimation, learning the structure of the output space along with general chemical rules. Then, using RL, the model is optimized for specific properties such as binding affinity or solubility. While similar in concept to transfer learning, the use of RL allows one to bias the model toward a wider range of objectives. Bottom: Directed generation with RL and GAN. This method leverages adversarial training to produce feasible molecules and RL to bias the generation toward desired properties.

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 8

    Figure 8. Here, the latent space of an AE is used as a reversible and continuous molecular representation allowing for the application of various optimization algorithms.

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 9

    Figure 9. Top: In conditioned generation, the desired properties are introduced as explicit inputs to the model. These properties are precomputed for each compound of the training set and used during training to induce a correlation between the two. This correlation is then leveraged during the generation process to target specific property values. Bottom: In the semisupervised case of conditioned generation, only part of the training set has the desired properties available. To overcome this, a predictor network is trained on the labeled instances and used to predict the properties of unlabeled ones.

    High Resolution Image
    Download MS PowerPoint Slide

  • References

    ARTICLE SECTIONS
    Jump To

    This article references 139 other publications.

    1. 1
      Polishchuk, P. G.; Madzhidov, T. I.; Varnek, A. Estimation of the size of drug-like chemical space based on GDB-17 data. J. Comput.-Aided Mol. Des. 2013, 27, 675,  DOI: 10.1007/s10822-013-9672-4
      [Crossref], [PubMed], [CAS], Google Scholar
      1
      Estimation of the size of drug-like chemical space based on GDB-17 data
      Polishchuk, P. G.; Madzhidov, T. I.; Varnek, A.
      Journal of Computer-Aided Molecular Design (2013), 27 (8), 675-679CODEN: JCADEQ; ISSN:0920-654X. (Springer)
      The goal of this paper is to est. the no. of realistic drug-like mols. which could ever be synthesized. Unlike previous studies based on exhaustive enumeration of mol. graphs or on combinatorial enumeration preselected fragments, we used results of constrained graphs enumeration by Reymond to establish a correlation between the no. of generated structures (M) and the no. of heavy atoms (N): logM = 0.584 × N × logN + 0.356. The no. of atoms limiting drug-like chem. space of mols. which follow Lipinsky's rules (N = 36) has been obtained from the anal. of the PubChem database. This results in M ≈ 1033 which is in between the nos. estd. by Ertl (1023) and by Bohacek (1060).
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC3sXhtlWjtbjM&md5=80a95002031f04319351c7e342bc1a55
    2. 2
      Schneider, G. Automating drug discovery. Nat. Rev. Drug Discovery 2018, 17, 97– 113,  DOI: 10.1038/nrd.2017.232
      [Crossref], [PubMed], [CAS], Google Scholar
      2
      Automating drug discovery
      Schneider, Gisbert
      Nature Reviews Drug Discovery (2018), 17 (2), 97-113CODEN: NRDDAG; ISSN:1474-1776. (Nature Research)
      A review. Small mol. drug discovery can be viewed as a challenging multidimensional problem in which various characteristics of compds. including efficacy, pharmacokinetics and safety need to be optimized in parallel to provide drug candidates. Recent advances in areas such as microfluidics-assisted chem. synthesis and biol. testing, as well as artificial intelligence systems that improve a design hypothesis through feedback anal., are now providing a basis for the introduction of greater automation into aspects of this process. This could potentially accelerate time frames for compd. discovery and optimization and enable more effective searches of chem. space. However, such approaches also raise considerable conceptual, tech. and organizational challenges, as well as scepticism about the current hype around them. This article aims to identify the approaches and technologies that could be implemented robustly by medicinal chemists in the near future and to critically analyze the opportunities and challenges for their more widespread application.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC2sXhvFOntbbK&md5=4fc1e441a0a31d7421642eb91fbac8c7
    3. 3
      DiMasi, J. A.; Grabowski, H. G.; Hansen, R. W. Innovation in the pharmaceutical industry: New estimates of R&D costs. Journal of Health Economics 2016, 47, 20– 33,  DOI: 10.1016/j.jhealeco.2016.01.012
      [Crossref], [PubMed], [CAS], Google Scholar
      3
      Innovation in the pharmaceutical industry: New estimates of R&D costs
      DiMasi Joseph A; Grabowski Henry G; Hansen Ronald W
      Journal of health economics (2016), 47 (), 20-33 ISSN:.
      The research and development costs of 106 randomly selected new drugs were obtained from a survey of 10 pharmaceutical firms. These data were used to estimate the average pre-tax cost of new drug and biologics development. The costs of compounds abandoned during testing were linked to the costs of compounds that obtained marketing approval. The estimated average out-of-pocket cost per approved new compound is $1395 million (2013 dollars). Capitalizing out-of-pocket costs to the point of marketing approval at a real discount rate of 10.5% yields a total pre-approval cost estimate of $2558 million (2013 dollars). When compared to the results of the previous study in this series, total capitalized costs were shown to have increased at an annual rate of 8.5% above general price inflation. Adding an estimate of post-approval R&D costs increases the cost estimate to $2870 million (2013 dollars).
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A280%3ADC%252BC28jls1egsw%253D%253D&md5=a78649853406ddbfe54abb0052ca964b
    4. 4
      Ruddigkeit, L.; van Deursen, R.; Blum, L. C.; Reymond, J.-L. Enumeration of 166 Billion Organic Small Molecules in the Chemical Universe Database GDB-17. J. Chem. Inf. Model. 2012, 52, 2864– 2875,  DOI: 10.1021/ci300415d
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      4
      Enumeration of 166 Billion Organic Small Molecules in the Chemical Universe Database GDB-17
      Ruddigkeit, Lars; van Deursen, Ruud; Blum, Lorenz C.; Reymond, Jean-Louis
      Journal of Chemical Information and Modeling (2012), 52 (11), 2864-2875CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      Drug mols. consist of a few tens of atoms connected by covalent bonds. How many such mols. are possible in total and what is their structure. This question is of pressing interest in medicinal chem. to help solve the problems of drug potency, selectivity, and toxicity and reduce attrition rates by pointing to new mol. series. To better define the unknown chem. space, we have enumerated 166.4 billion mols. of up to 17 atoms of C, N, O, S, and halogens forming the chem. universe database GDB-17, covering a size range contg. many drugs and typical for lead compds. GDB-17 contains millions of isomers of known drugs, including analogs with high shape similarity to the parent drug. Compared to known mols. in PubChem, GDB-17 mols. are much richer in nonarom. heterocycles, quaternary centers, and stereoisomers, densely populate the third dimension in shape space, and represent many more scaffold types.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC38XhsFClsL3J&md5=d0bf9a29f3e9ae1e57bb1c953a562ced
    5. 5
      Walters, W. P. Virtual Chemical Libraries: Miniperspective. J. Med. Chem. 2019, 62, 1116– 1124,  DOI: 10.1021/acs.jmedchem.8b01048
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      5
      Virtual Chemical Libraries
      Walters, W. Patrick
      Journal of Medicinal Chemistry (2019), 62 (3), 1116-1124CODEN: JMCMAR; ISSN:0022-2623. (American Chemical Society)
      A review. Advances in computer processing speed and storage capacity have enabled researchers to generate virtual chem. libraries contg. billions of mols. While these nos. appear large, they are only a small fraction of the no. of org. mols. that could potentially be synthesized. This review provides an overview of recent advances in the generation and use of virtual chem. libraries in medicinal chem. The authors also consider the practical implications of these libraries in drug discovery programs and highlight a no. of current and future challenges.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXhsFyqtbnM&md5=ede412935f56e350915f35e79f1b969f
    6. 6
      Hartenfeller, M.; Zettl, H.; Walter, M.; Rupp, M.; Reisen, F.; Proschak, E.; Weggen, S.; Stark, H.; Schneider, G. DOGS: reaction-driven de novo design of bioactive com- pounds. PLoS Comput. Biol. 2012, 8, e1002380  DOI: 10.1371/journal.pcbi.1002380
      [Crossref], [PubMed], [CAS], Google Scholar
      6
      DOGS: reaction-driven de novo design of bioactive compounds
      Hartenfeller, Markus; Zettl, Heiko; Walter, Miriam; Rupp, Matthias; Reisen, Felix; Proschak, Ewgenij; Weggen, Sascha; Stark, Holger; Schneider, Gisbert
      PLoS Computational Biology (2012), 8 (2), e1002380CODEN: PCBLBG; ISSN:1553-7358. (Public Library of Science)
      We present a computational method for the reaction-based de novo design of drug-like mols. The software DOGS (Design of Genuine Structures) features a ligand-based strategy for automated 'in silico' assembly of potentially novel bioactive compds. The quality of the designed compds. is assessed by a graph kernel method measuring their similarity to known bioactive ref. ligands in terms of structural and pharmacophoric features. We implemented a deterministic compd. construction procedure that explicitly considers compd. synthesizability, based on a compilation of 25'144 readily available synthetic building blocks and 58 established reaction principles. This enables the software to suggest a synthesis route for each designed compd. Two prospective case studies are presented together with details on the algorithm and its implementation. De novo designed ligand candidates for the human histamine H4 receptor and γ-secretase were synthesized as suggested by the software. The computational approach proved to be suitable for scaffold-hopping from known ligands to novel chemotypes and for generating bioactive mols. with drug-like properties.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC38Xjt1Sjtr0%253D&md5=cdcf6905d52761614c33ab068b608ac5
    7. 7
      Spiegel, J.; Durrant, J. AutoGrow4: An open-source genetic algorithm for de novo drug design and lead optimization. J. Cheminf. 2020, 12, 25,  DOI: 10.1186/s13321-020-00429-4
      [Crossref], [CAS], Google Scholar
      7
      AutoGrow4: an open-source genetic algorithm for de novo drug design and lead optimization
      Spiegel, Jacob O.; Durrant, Jacob D.
      Journal of Cheminformatics (2020), 12 (1), 25CODEN: JCOHB3; ISSN:1758-2946. (SpringerOpen)
      Abstr.: We here present AutoGrow4, an open-source program for semi-automated computer-aided drug discovery. AutoGrow4 uses a genetic algorithm to evolve predicted ligands on demand and so is not limited to a virtual library of pre-enumerated compds. It is a useful tool for generating entirely novel drug-like mols. and for optimizing preexisting ligands. By leveraging recent computational and cheminformatics advancements, AutoGrow4 is faster, more stable, and more modular than previous versions. It implements new docking-program compatibility, chem. filters, multithreading options, and selection methods to support a wide range of user needs. To illustrate both de novo design and lead optimization, we here apply AutoGrow4 to the catalytic domain of poly(ADP-ribose) polymerase 1 (PARP-1), a well characterized DNA-damage-recognition protein. AutoGrow4 produces drug-like compds. with better predicted binding affinities than FDA-approved PARP-1 inhibitors (pos. controls). The predicted binding modes of the AutoGrow4 compds. mimic those of the known inhibitors, even when AutoGrow4 is seeded with random small mols. AutoGrow4 is available under the terms of the Apache License, Version 2.0. A copy can be downloaded free of charge from <a><a><a> not available: see fulltext].
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXnsVWksrs%253D&md5=10cb3d13c137b5e93fc3f92568edba22
    8. 8
      Jensen, J. H. A graph-based genetic algorithm and generative model/Monte Carlo tree search for the exploration of chemical space. Chem. Sci. 2019, 10, 3567– 3572,  DOI: 10.1039/C8SC05372C
      [Crossref], [PubMed], [CAS], Google Scholar
      8
      A graph-based genetic algorithm and generative model/Monte Carlo tree search for the exploration of chemical space
      Jensen, Jan H.
      Chemical Science (2019), 10 (12), 3567-3572CODEN: CSHCCN; ISSN:2041-6520. (Royal Society of Chemistry)
      This paper presents a comparison of a graph-based genetic algorithm (GB-GA) and machine learning (ML) results for the optimization of log P values with a constraint for synthetic accessibility and shows that the GA is as good as or better than the ML approaches for this particular property. The mols. found by the GB-GA bear little resemblance to the mols. used to construct the initial mating pool, indicating that the GB-GA approach can traverse a relatively large distance in chem. space using relatively few (50) generations. The paper also introduces a new non-ML graph-based generative model (GB-GM) that can be parameterized using very small data sets and combined with a Monte Carlo tree search (MCTS) algorithm. The results are comparable to previously published results (Sci. Technol. Adv. Mater., 2017, 18, 972-976) using a recurrent neural network (RNN) generative model, and the GB-GM-based method is several orders of magnitude faster. The MCTS results seem more dependent on the compn. of the training set than the GA approach for this particular property. Our results suggest that the performance of new ML-based generative models should be compared to that of more traditional, and often simpler, approaches such a GA.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXislClt78%253D&md5=108c19e322025e736330203e0d312237
    9. 9
      Yoshikawa, N.; Terayama, K.; Sumita, M.; Homma, T.; Oono, K.; Tsuda, K. Population-based De Novo Molecule Generation, Using Grammatical Evolution. Chem. Lett. 2018, 47, 1431– 1434,  DOI: 10.1246/cl.180665
      [Crossref], [CAS], Google Scholar
      9
      Population-based de novo molecule generation, using grammatical evolution
      Yoshikawa, Naruki; Terayama, Kei; Sumita, Masato; Homma, Teruki; Oono, Kenta; Tsuda, Koji
      Chemistry Letters (2018), 47 (11), 1431-1434CODEN: CMLTAG; ISSN:0366-7022. (Chemical Society of Japan)
      Automatic mol. design with machine learning and simulations has shown a remarkable ability to generate new and promising drug candidates. We propose a new population-based approach using a grammatical evolution named ChemGE, that can update a large population of mols. concurrently and evaluate with multiple simulators in parallel. In computational expts., ChemGE succeeded in finding hundreds of candidate mols. whose affinity for thymidine kinase is better than that of known binding mols. in a database (DUD-E).
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXitlKnsLvP&md5=ccb65d8a0d2bf9b350b4607ef2e70f67
    10. 10
      Ǵomez-Bombarelli, R.; Wei, J. N.; Duvenaud, D.; Hernández-Lobato, J. M.; Śanchez- Lengeling, B.; Sheberla, D.; Aguilera-Iparraguirre, J.; Hirzel, T. D.; Adams, R. P.; Aspuru-Guzik, A. Automatic Chemical Design Using a Data-Driven Continuous Rep- resentation of Molecules. ACS Cent. Sci. 2018, 4, 268– 276,  DOI: 10.1021/acscentsci.7b00572
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      10
      Automatic Chemical Design Using a Data-Driven Continuous Representation of Molecules
      Gomez-Bombarelli, Rafael; Wei, Jennifer N.; Duvenaud, David; Hernandez-Lobato, Jose Miguel; Sanchez-Lengeling, Benjamin; Sheberla, Dennis; Aguilera-Iparraguirre, Jorge; Hirzel, Timothy D.; Adams, Ryan P.; Aspuru-Guzik, Alan
      ACS Central Science (2018), 4 (2), 268-276CODEN: ACSCII; ISSN:2374-7951. (American Chemical Society)
      We report a method to convert discrete representations of mols. to and from a multidimensional continuous representation. This model allows us to generate new mols. for efficient exploration and optimization through open-ended spaces of chem. compds. A deep neural network was trained on hundreds of thousands of existing chem. structures to construct three coupled functions: an encoder, a decoder, and a predictor. The encoder converts the discrete representation of a mol. into a real-valued continuous vector, and the decoder converts these continuous vectors back to discrete mol. representations. The predictor ests. chem. properties from the latent continuous vector representation of the mol. Continuous representations of mols. allow us to automatically generate novel chem. structures by performing simple operations in the latent space, such as decoding random vectors, perturbing known chem. structures, or interpolating between mols. Continuous representations also allow the use of powerful gradient-based optimization to efficiently guide the search for optimized functional compds. We demonstrate our method in the domain of drug-like mols. and also in a set of mols. with fewer that nine heavy atoms.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXntlWquw%253D%253D&md5=322d9ff569fc9c8831e91d915104d985
    11. 11
      Gawehn, E.; Hiss, J. A.; Schneider, G. Deep learning in drug discovery. Mol. Inf. 2016, 35, 3– 14,  DOI: 10.1002/minf.201501008
      [Crossref], [PubMed], [CAS], Google Scholar
      11
      Deep Learning in Drug Discovery
      Gawehn, Erik; Hiss, Jan A.; Schneider, Gisbert
      Molecular Informatics (2016), 35 (1), 3-14CODEN: MIONBS; ISSN:1868-1743. (Wiley-VCH Verlag GmbH & Co. KGaA)
      Artificial neural networks had their first heyday in mol. informatics and drug discovery approx. two decades ago. Currently, we are witnessing renewed interest in adapting advanced neural network architectures for pharmaceutical research by borrowing from the field of "deep learning". Compared with some of the other life sciences, their application in drug discovery is still limited. Here, we provide an overview of this emerging field of mol. informatics, present the basic concepts of prominent deep learning methods and offer motivation to explore these techniques for their usefulness in computer-assisted drug discovery and design. We specifically emphasize deep neural networks, restricted Boltzmann machine networks and convolutional networks.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC2MXitV2rs7bE&md5=bd2553359c4b824c5272388462772659
    12. 12
      Goodfellow, I.; Bengio, Y.; Courville, A. Deep Learning; MIT Press, 2016.
      Google Scholar
      There is no corresponding record for this reference.
    13. 13
      Chollet, F. Deep learning with Python; Manning Publications Co: Shelter Island, NY, 2018.
      Google Scholar
      There is no corresponding record for this reference.
    14. 14
      Foster, D.; Safari, A. O. M. C. Generative deep learning: teaching machines to paint, write, compose, and play; O’Reilly Media, 2019.
      Google Scholar
      There is no corresponding record for this reference.
    15. 15
      White, D.; Wilson, R. C. Generative models for chemical structures. J. Chem. Inf. Model. 2010, 50, 1257– 1274,  DOI: 10.1021/ci9004089
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      15
      Generative Models for Chemical Structures
      White, David; Wilson, Richard C.
      Journal of Chemical Information and Modeling (2010), 50 (7), 1257-1274CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      We apply recently developed techniques for pattern recognition to construct a generative model for chem. structure. This approach can be viewed as ligand-based de novo design. We construct a statistical model describing the structural variations present in a set of mols. which may be sampled to generate new structurally similar examples. We prevent the possibility of generating chem. invalid mols., according to our implicit hydrogen model, by projecting samples onto the nearest chem. valid mol. By populating the input set with mols. that are active against a target, we show how new mols. may be generated that will likely also be active against the target.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC3cXos1ersr4%253D&md5=f694ddd48c404ebea7792e4a52458ebb
    16. 16
      Sanchez-Lengeling, B.; Aspuru-Guzik, A. Inverse molecular design using machine learning: Generative models for matter engineering. Science 2018, 361, 360– 365,  DOI: 10.1126/science.aat2663
      [Crossref], [PubMed], [CAS], Google Scholar
      16
      Inverse molecular design using machine learning: Generative models for matter engineering
      Sanchez-Lengeling, Benjamin; Aspuru-Guzik, Alan
      Science (Washington, DC, United States) (2018), 361 (6400), 360-365CODEN: SCIEAS; ISSN:0036-8075. (American Association for the Advancement of Science)
      The discovery of new materials can bring enormous societal and technol. progress. In this context, exploring completely the large space of potential materials is computationally intractable. Here, we review methods for achieving inverse design, which aims to discover tailored materials from the starting point of a particular desired functionality. Recent advances from the rapidly growing field of artificial intelligence, mostly from the subfield of machine learning, have resulted in a fertile exchange of ideas, where approaches to inverse mol. design are being proposed and employed at a rapid pace. Among these, deep generative models have been applied to numerous classes of materials: rational design of prospective drugs, synthetic routes to org. compds., and optimization of photovoltaics and redox flow batteries, as well as a variety of other solid-state materials.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXhtlyitr3L&md5=779c4a42ba1e84d99d13ad1b32b9529a
    17. 17
      Yuan, W.; Jiang, D.; Nambiar, D. K.; Liew, L. P.; Hay, M. P.; Bloomstein, J.; Lu, P.; Turner, B.; Le, Q.-T.; Tibshirani, R.; Khatri, P.; Moloney, M. G.; Koong, A. C. Chemical Space Mimicry for Drug Discovery. J. Chem. Inf. Model. 2017, 57, 875– 882,  DOI: 10.1021/acs.jcim.6b00754
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      17
      Chemical Space Mimicry for Drug Discovery
      Yuan, William; Jiang, Dadi; Nambiar, Dhanya K.; Liew, Lydia P.; Hay, Michael P.; Bloomstein, Joshua; Lu, Peter; Turner, Brandon; Le, Quynh-Thu; Tibshirani, Robert; Khatri, Purvesh; Moloney, Mark G.; Koong, Albert C.
      Journal of Chemical Information and Modeling (2017), 57 (4), 875-882CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      The authors describe a new library generation method, Machine-based Identification of Mols. Inside Characterized Space (MIMICS) that generates sets of mols. inspired by a text-based input. MIMICS-generated libraries were found to preserve distributions of properties while simultaneously increasing structural diversity. Newly identified MIMICS-generated compds. were found to be bioactive as inhibitors of specific components of the unfolded protein response (UPR) and the VEGFR2 pathway in cell-based assays, thus confirming that applicability of this methodol. towards drug design applications. Wider application of MIMICS could facilitate the efficient utilization of chem. space.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC2sXjs12gtrs%253D&md5=f237aa1f08c24e416756cc80f018b3d1
    18. 18
      Segler, M. H.; Kogej, T.; Tyrchan, C.; Waller, M. P. Generating focused molecule libraries for drug discovery with recurrent neural networks. ACS Cent. Sci. 2018, 4, 120– 131,  DOI: 10.1021/acscentsci.7b00512
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      18
      Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks
      Segler, Marwin H. S.; Kogej, Thierry; Tyrchan, Christian; Waller, Mark P.
      ACS Central Science (2018), 4 (1), 120-131CODEN: ACSCII; ISSN:2374-7951. (American Chemical Society)
      In de novo drug design, computational strategies are used to generate novel mols. with good affinity to the desired biol. target. In this work, we show that recurrent neural networks can be trained as generative models for mol. structures, similar to statistical language models in natural language processing. We demonstrate that the properties of the generated mols. correlate very well with the properties of the mols. used to train the model. In order to enrich libraries with mols. active toward a given biol. target, we propose to fine-tune the model with small sets of mols., which are known to be active against that target. Against Staphylococcus aureus, the model reproduced 14% of 6051 hold-out test mols. that medicinal chemists designed, whereas against Plasmodium falciparum (Malaria), it reproduced 28% of 1240 test mols. When coupled with a scoring function, our model can perform the complete de novo drug design cycle to generate large sets of novel mols. for drug discovery.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC2sXitVCjsLfP&md5=708f40422c7a911c629525ce5b66088b
    19. 19
      Elton, D. C.; Boukouvalas, Z.; Fuge, M. D.; Chung, P. W. Deep learning for molecular design─a review of the state of the art. Mol. Syst. Des. Eng. 2019, 4, 828– 849,  DOI: 10.1039/C9ME00039A
      [Crossref], [CAS], Google Scholar
      19
      Deep learning for molecular design-a review of the state of the art
      Elton, Daniel C.; Boukouvalas, Zois; Fuge, Mark D.; Chung, Peter W.
      Molecular Systems Design & Engineering (2019), 4 (4), 828-849CODEN: MSDEBG; ISSN:2058-9689. (Royal Society of Chemistry)
      In the space of only a few years, deep generative modeling has revolutionized how we think of artificial creativity, yielding autonomous systems which produce original images, music, and text. Inspired by these successes, researchers are now applying deep generative modeling techniques to the generation and optimization of mols.-in our review we found 45 papers on the subject published in the past two years. These works point to a future where such systems will be used to generate lead mols., greatly reducing resources spent downstream synthesizing and characterizing bad leads in the lab. In this review we survey the increasingly complex landscape of models and representation schemes that have been proposed. The four classes of techniques we describe are recursive neural networks, autoencoders, generative adversarial networks, and reinforcement learning. After first discussing some of the math. fundamentals of each technique, we draw high level connections and comparisons with other techniques and expose the pros and cons of each. Several important high level themes emerge as a result of this work, including the shift away from the SMILES string representation of mols. towards more sophisticated representations such as graph grammars and 3D representations, the importance of reward function design, the need for better stds. for benchmarking and testing, and the benefits of adversarial training and reinforcement learning over max. likelihood based training.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXhtVWktLjN&md5=a0925fdace3a12af31b2f6fc1edef89b
    20. 20
      Schwalbe-Koda, D.; Ǵomez-Bombarelli, R. In Machine Learning Meets Quantum Physics; Schütt, K. T., Chmiela, S., von Lilienfeld, O. A., Tkatchenko, A., Tsuda, K., Müller, K.-R., Eds.; Springer International Publishing: Cham, 2020; pp 445– 467.
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    21. 21
      Zhavoronkov, A.; Vanhaelen, Q.; Oprea, T. I. Will Artificial Intelligence for Drug Discovery Impact Clinical Pharmacology?. Clin. Pharmacol. Ther. (N. Y., NY, U. S.) 2020, 107, 780– 785,  DOI: 10.1002/cpt.1795
      [Crossref], [PubMed], Google Scholar
      There is no corresponding record for this reference.
    22. 22
      Bian, Y.; Xie, X.-Q. Generative chemistry: drug discovery with deep learning gener- ative models. J. Mol. Model. 2021, 27, 71,  DOI: 10.1007/s00894-021-04674-8
      [Crossref], [PubMed], [CAS], Google Scholar
      22
      Generative chemistry: drug discovery with deep learning generative models
      Bian, Yuemin; Xie, Xiang-Qun
      Journal of Molecular Modeling (2021), 27 (3), 71CODEN: JMMOFK; ISSN:0948-5023. (Springer)
      A review. The de novo design of mol. structures using deep learning generative models introduces an encouraging soln. to drug discovery in the face of the continuously increased cost of new drug development. From the generation of original texts, images, and videos, to the scratching of novel mol. structures the creativity of deep learning generative models exhibits the height machine intelligence can achieve. The purpose of this paper is to review the latest advances in generative chem. which relies on generative modeling to expedite the drug discovery process. This review starts with a brief history of artificial intelligence in drug discovery to outline this emerging paradigm. Commonly used chem. databases, mol. representations, and tools in cheminformatics and machine learning are covered as the infrastructure for generative chem. The detailed discussions on utilizing cutting-edge generative architectures, including recurrent neural network, variational autoencoder, adversarial autoencoder, and generative adversarial network for compd. generation are focused. Challenges and future perspectives follow.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3MXhsValsLrP&md5=bf1851d6e96138674bc169295356eee6
    23. 23
      Chen, H.; Engkvist, O.; Wang, Y.; Olivecrona, M.; Blaschke, T. The rise of deep learning in drug discovery. Drug Discovery Today 2018, 23, 1241– 1250,  DOI: 10.1016/j.drudis.2018.01.039
      [Crossref], [PubMed], [CAS], Google Scholar
      23
      The rise of deep learning in drug discovery
      Chen Hongming; Engkvist Ola; Olivecrona Marcus; Blaschke Thomas; Wang Yinhai
      Drug discovery today (2018), 23 (6), 1241-1250 ISSN:.
      Over the past decade, deep learning has achieved remarkable success in various artificial intelligence research areas. Evolved from the previous research on artificial neural networks, this technology has shown superior performance to other machine learning algorithms in areas such as image and voice recognition, natural language processing, among others. The first wave of applications of deep learning in pharmaceutical research has emerged in recent years, and its utility has gone beyond bioactivity predictions and has shown promise in addressing diverse problems in drug discovery. Examples will be discussed covering bioactivity prediction, de novo molecular design, synthesis prediction and biological image analysis.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A280%3ADC%252BC1MvjvFyqtQ%253D%253D&md5=d6cbdd98ede30181802cca1786cd5a95
    24. 24
      Vamathevan, J.; Clark, D.; Czodrowski, P.; Dunham, I.; Ferran, E.; Lee, G.; Li, B.; Madabhushi, A.; Shah, P.; Spitzer, M.; Zhao, S. Applications of machine learning in drug discovery and development. Nat. Rev. Drug Discovery 2019, 18, 463– 477,  DOI: 10.1038/s41573-019-0024-5
      [Crossref], [PubMed], [CAS], Google Scholar
      24
      Applications of machine learning in drug discovery and development
      Vamathevan, Jessica; Clark, Dominic; Czodrowski, Paul; Dunham, Ian; Ferran, Edgardo; Lee, George; Li, Bin; Madabhushi, Anant; Shah, Parantu; Spitzer, Michaela; Zhao, Shanrong
      Nature Reviews Drug Discovery (2019), 18 (6), 463-477CODEN: NRDDAG; ISSN:1474-1776. (Nature Research)
      A review. Drug discovery and development pipelines are long, complex and depend on numerous factors. Machine learning (ML) approaches provide a set of tools that can improve discovery and decision making for well-specified questions with abundant, high-quality data. Opportunities to apply ML occur in all stages of drug discovery. Examples include target validation, identification of prognostic biomarkers and anal. of digital pathol. data in clin. trials. Applications have ranged in context and methodol., with some approaches yielding accurate predictions and insights. The challenges of applying ML lie primarily with the lack of interpretability and repeatability of ML-generated results, which may limit their application. In all areas, systematic and comprehensive high-dimensional data still need to be generated. With ongoing efforts to tackle these issues, as well as increasing awareness of the factors needed to validate ML approaches, the application of ML can promote data-driven decision making and has the potential to speed up the process and reduce failure rates in drug discovery and development.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXosF2rtrY%253D&md5=211782aeea3d8b9f50368f89177a70d2
    25. 25
      Engel, T., Gasteiger, J., Eds. Chemoinformatics: basic concepts and methods; Wiley-VCH: Weinheim, 2018; OCLC: 1012130305.
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    26. 26
      Kim, S.; Chen, J.; Cheng, T.; Gindulyte, A.; He, J.; He, S.; Li, Q.; Shoemaker, B. A.; Thiessen, P. A.; Yu, B.; Zaslavsky, L.; Zhang, J.; Bolton, E. E. PubChem 2019 update: improved access to chemical data. Nucleic Acids Res. 2019, 47, D1102– D1109,  DOI: 10.1093/nar/gky1033
      [Crossref], [PubMed], Google Scholar
      There is no corresponding record for this reference.
    27. 27
      Ash, S.; Cline, M.; Homer, R. W.; Hurst, T.; Smith, G. SYBYL Line Notation (SLN): A Versatile Language for Chemical Structure Representation. J. Chem. Inf. Comput. Sci. 1997, 37, 71– 79,  DOI: 10.1021/ci960109j
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      27
      SYBYL Line Notation (SLN): A Versatile Language for Chemical Structure Representation
      Ash, Sheila; Cline, Malcolm A.; Homer, R. Webster; Hurst, Tad; Smith, Gregory B.
      Journal of Chemical Information and Computer Sciences (1997), 37 (1), 71-79CODEN: JCISD8; ISSN:0095-2338. (American Chemical Society)
      SYBYL Line Notation (SLN) is an ASCII language used to represent chem. structures, including common org. mols., macromols., polymers, and combinatorial libraries. SLN is also used to express substructural (2D) queries and includes a complete facility for Markush representation. This concise language is ideal for database storage of chem. entities as well as for network communication of structures and queries.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADyaK2sXksVCitA%253D%253D&md5=9c53b87aec0a69043773cd771c662152
    28. 28
      Koniver, D. A.; Wiswesser, W. J.; Usdin, E. Wiswesser Line Notation: Simplified Techniques for Converting Chemical Structures to WLN. Science 1972, 176, 1437– 1439,  DOI: 10.1126/science.176.4042.1437
      [Crossref], [PubMed], [CAS], Google Scholar
      28
      Wiswesser Line Notation. Simplified techniques for converting chemical structures to WLN
      Koniver, Deena A.; Wiswesser, William J.; Usdin, Earl
      Science (Washington, DC, United States) (1972), 176 (4042), 1437-9CODEN: SCIEAS; ISSN:0036-8075.
      Techniques were developed for the generation of Wiswesser Line Notations (WLN), which require knowledge neither of rules for manual conversion of structures to line notations nor of computer programing. The desired WLN are obtained simply by drawing the structures of the compds. of interest on a tablet, which is linked to an appropriately programmed computer.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADyaE38Xks1Srt7c%253D&md5=e7b61344874233d0114df299b83e3a9d
    29. 29
      Weininger, D. SMILES, a chemical language and information system. 1. Introduction to methodology and encoding rules. J. Chem. Inf. Model. 1988, 28, 31– 36,  DOI: 10.1021/ci00057a005
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      29
      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.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADyaL1cXnsVeqsA%253D%253D&md5=04592975f9dd3c0ce3c1ad618ba2b17d
    30. 30
      O’Boyle, N. M. Towards a Universal SMILES representation - A standard method to generate canonical SMILES based on the InChI. J. Cheminf. 2012, 4, 22,  DOI: 10.1186/1758-2946-4-22
      [Crossref], [CAS], Google Scholar
      30
      Towards a universal SMILES representation - a standard method to generate canonical SMILES based on the InChI
      O'Boyle, Noel M.
      Journal of Cheminformatics (2012), 4 (), 22CODEN: JCOHB3; ISSN:1758-2946. (Chemistry Central Ltd.)
      Background: There are two line notations of chem. structures that have established themselves in the field: the SMILES string and the InChI string. The InChI aims to provide a unique, or canonical, identifier for chem. structures, while SMILES strings are widely used for storage and interchange of chem. structures, but no std. exists to generate a canonical SMILES string. Results: I describe how to use the InChI canonicalisation to derive a canonical SMILES string in a straightforward way, either incorporating the InChI normalizations (Inchified SMILES) or not (Universal SMILES). This is the first description of a method to generate canonical SMILES that takes stereochem. into account. When tested on the 1.1 m compds. in the ChEMBL database, and a 1 m compd. subset of the PubChem Substance database, no canonicalisation failures were found with Inchified SMILES. Using Universal SMILES, 99.79% of the ChEMBL database was canonicalised successfully and 99.77% of the PubChem subset. Conclusions: The InChI canonicalisation algorithm can successfully be used as the basis for a common std. for canonical SMILES. While challenges remain - such as the development of a std. arom. model for SMILES - the ability to create the same SMILES using different toolkits will mean that for the first time it will be possible to easily compare the chem. models used by different toolkits.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC38XhvVSiur7I&md5=eb59d742c5dec35b5d2c90417acd223d
    31. 31
      Bjerrum, E. J. SMILES enumeration as data augmentation for neural network modeling of molecules. arXiv (Machine Learning) , May 17, 2017, 703.07076, ver. 2.
      Google Scholar
      There is no corresponding record for this reference.
    32. 32
      Bjerrum, E. J.; Sattarov, B. Improving chemical autoencoder latent space and molec- ular de novo generation diversity with heteroencoders. Biomolecules 2018, 8, 131,  DOI: 10.3390/biom8040131
      [Crossref], [CAS], Google Scholar
      32
      Improving chemical autoencoder latent space and molecular de novo generation diversity with heteroencoders
      Bjerrum, Esben Jannik; Sattarov, Boris
      Biomolecules (2018), 8 (4), 131/1-131/17CODEN: BIOMHC; ISSN:2218-273X. (MDPI AG)
      Chem. autoencoders are attractive models as they combine chem. space navigation with possibilities for de novo mol. generation in areas of interest. This enables them to produce focused chem. libraries around a single lead compd. for employment early in a drug discovery project. Here, it is shown that the choice of chem. representation, such as strings from the simplified mol.-input line-entry system (SMILES), has a large influence on the properties of the latent space. It is further explored to what extent translating between different chem. representations influences the latent space similarity to the SMILES strings or circular fingerprints. By employing SMILES enumeration for either the encoder or decoder, it is found that the decoder has the largest influence on the properties of the latent space. Training a sequence to sequence heteroencoder based on recurrent neural networks (RNNs) with long short-term memory cells (LSTM) to predict different enumerated SMILES strings from the same canonical SMILES string gives the largest similarity between latent space distance and mol. similarity measured as circular fingerprints similarity. Using the output from the code layer in quant. structure activity relationship (QSAR) of five mol. datasets shows that heteroencoder derived vectors markedly outperforms autoencoder derived vectors as well as models built using ECFP4 fingerprints, underlining the increased chem. relevance of the latent space.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXit12hurjL&md5=c0e3ae0b5ec0126d70e59045568b1a90
    33. 33
      Arús-Pous, J.; Johansson, S. V.; Prykhodko, O.; Bjerrum, E. J.; Tyrchan, C.; Rey- mond, J.-L.; Chen, H.; Engkvist, O. Randomized SMILES strings improve the quality of molecular generative models. J. Cheminf. 2019, 11, 71,  DOI: 10.1186/s13321-019-0393-0
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    34. 34
      Moret, M.; Friedrich, L.; Grisoni, F.; Merk, D.; Schneider, G. Generative molecular design in low data regimes. Nature Machine Intelligence 2020, 2, 171– 180,  DOI: 10.1038/s42256-020-0160-y
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    35. 35
      van Deursen, R.; Ertl, P.; Tetko, I. V.; Godin, G. GEN: highly efficient SMILES explorer using autodidactic generative examination networks. J. Cheminf. 2020, 12, 22,  DOI: 10.1186/s13321-020-00425-8
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    36. 36
      Prykhodko, O.; Johansson, S. V.; Kotsias, P.-C.; Arús-Pous, J.; Bjerrum, E. J.; En- gkvist, O.; Chen, H. A de novo molecular generation method using latent vector based generative adversarial network. J. Cheminf. 2019, 11, 74,  DOI: 10.1186/s13321-019-0397-9
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    37. 37
      Heller, S. R.; McNaught, A.; Pletnev, I.; Stein, S.; Tchekhovskoi, D. InChI, the IUPAC International Chemical Identifier. J. Cheminf. 2015, 7, 23,  DOI: 10.1186/s13321-015-0068-4
      [Crossref], [CAS], Google Scholar
      37
      InChI, the IUPAC International Chemical Identifier
      Heller Stephen R; Stein Stephen; Tchekhovskoi Dmitrii; McNaught Alan; Pletnev Igor
      Journal of cheminformatics (2015), 7 (), 23 ISSN:1758-2946.
      This paper documents the design, layout and algorithms of the IUPAC International Chemical Identifier, InChI.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A280%3ADC%252BC2MbpslOrtQ%253D%253D&md5=4acc4f470f8cdb9b4f84558fd3302470
    38. 38
      Winter, R.; Montanari, F.; Nóe, F.; Clevert, D.-A. Learning continuous and data- driven molecular descriptors by translating equivalent chemical representations. Chem. Sci. 2019, 10, 1692– 1701,  DOI: 10.1039/C8SC04175J
      [Crossref], [PubMed], [CAS], Google Scholar
      38
      Learning continuous and data-driven molecular descriptors by translating equivalent chemical representations
      Winter, Robin; Montanari, Floriane; Noe, Frank; Clevert, Djork-Arne
      Chemical Science (2019), 10 (6), 1692-1701CODEN: CSHCCN; ISSN:2041-6520. (Royal Society of Chemistry)
      There has been a recent surge of interest in using machine learning across chem. space in order to predict properties of mols. or design mols. and materials with the desired properties. Most of this work relies on defining clever feature representations, in which the chem. graph structure is encoded in a uniform way such that predictions across chem. space can be made. In this work, we propose to exploit the powerful ability of deep neural networks to learn a feature representation from low-level encodings of a huge corpus of chem. structures. Our model borrows ideas from neural machine translation: it translates between two semantically equiv. but syntactically different representations of mol. structures, compressing the meaningful information both representations have in common in a low-dimensional representation vector. Once the model is trained, this representation can be extd. for any new mol. and utilized as a descriptor. In fair benchmarks with respect to various human-engineered mol. fingerprints and graph-convolution models, our method shows competitive performance in modeling quant. structure-activity relationships in all analyzed datasets. Addnl., we show that our descriptor significantly outperforms all baseline mol. fingerprints in two ligand-based virtual screening tasks. Overall, our descriptors show the most consistent performances in all expts. The continuity of the descriptor space and the existence of the decoder that permits deducing a chem. structure from an embedding vector allow for exploration of the space and open up new opportunities for compd. optimization and idea generation.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXit1aqsbnO&md5=57678211dbeac5a8135e2d38c41eeee2
    39. 39
      O’Boyle, N.; Dalke, A. DeepSMILES: An Adaptation of SMILES for Use in Machine- Learning of Chemical Structures; preprint, ChemRxiv , September 19, 2018, ver. 1. DOI: 10.26434/chemrxiv.7097960.v1 .
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    40. 40
      Krenn, M.; Häse, F.; Nigam, A.; Friederich, P.; Aspuru-Guzik, A. Self-referencing em- bedded strings (SELFIES): A 100% robust molecular string representation. Machine Learning: Science and Technology 2020, 1, 045024,  DOI: 10.1088/2632-2153/aba947
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    41. 41
      Faulon, J.-L., Bender, A., Eds. Handbook of chemoinformatics algorithms; Chapman & Hall/CRC mathematical and computational biology series; Chapman & Hall/CRC: Boca Raton, FL, 2010; Chapter 1. OCLC: ocn226357322.
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    42. 42
      Wishart, D. S. DrugBank 5.0: a major update to the DrugBank database for 2018. Nucleic Acids Res. 2018, 46, D1074– D1082,  DOI: 10.1093/nar/gkx1037
      [Crossref], [PubMed], [CAS], Google Scholar
      42
      DrugBank 5.0: a major update to the DrugBank database for 2018
      Wishart, David S.; Feunang, Yannick D.; Guo, An C.; Lo, Elvis J.; Marcu, Ana; Grant, Jason R.; Sajed, Tanvir; Johnson, Daniel; Li, Carin; Sayeeda, Zinat; Assempour, Nazanin; Iynkkaran, Ithayavani; Liu, Yifeng; Maciejewski, Adam; Gale, Nicola; Wilson, Alex; Chin, Lucy; Cummings, Ryan; Le, Diana; Pon, Allison; Knox, Craig; Wilson, Michael
      Nucleic Acids Research (2018), 46 (D1), D1074-D1082CODEN: NARHAD; ISSN:1362-4962. (Oxford University Press)
      DrugBank is a web-enabled database contg. comprehensivemol. information about drugs, their mechanisms, their interactions and their targets. First described in 2006, Drug- Bank has continued to evolve over the past 12 years in response to marked improvements to web stds. and changing needs for drug research and development. This year's update, DrugBank 5.0, represents the most significant upgrade to the database in more than 10 years. In many cases, existing data content has grown by 100% or more over the last update. For instance, the total no. of investigational drugs in the database has grown by almost 300%, the no. of drug-drug interactions has grown by nearly 600% and the no. of SNP-assocd. drug effects has grown more than 3000%. Significant improvements have been made to the quantity, quality and consistency of drug indications, drug binding data as well as drug-drug and drug-food interactions. A great deal of brand new data have also been added to DrugBank 5.0. This includes information on the influence of hundreds of drugs on metabolite levels (pharmacometabolomics), gene expression levels (pharmacotranscriptomics) and protein expression levels (pharmacoproteomics). New data have also been added on the status of hundreds of newdrug clin. trials and existing drug repurposing trials. Many other important improvements in the content, interface and performance of the DrugBank website have been made and these should greatly enhance its ease of use, utility and potential applications in many areas of pharmacol. research, pharmaceutical science and drug education.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXitlGisbvI&md5=986b28c7ea546596a26dd3ba38f05fee
    43. 43
      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– D1107,  DOI: 10.1093/nar/gkr777
      [Crossref], [PubMed], [CAS], Google Scholar
      43
      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.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC3MXhs12htbjN&md5=aedf7793e1ca54b6a4fa272ea3ef7d0e
    44. 44
      Landrum, G. RDKit: open-source cheminformatics software , 2016.
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    45. 45
      Steinbeck, C.; Hoppe, C.; Kuhn, S.; Floris, M.; Guha, R.; Willighagen, E. Recent De- velopments of the Chemistry Development Kit (CDK) - An Open-Source Java Library for Chemo- and Bioinformatics. Curr. Pharm. Des. 2006, 12, 2111– 2120,  DOI: 10.2174/138161206777585274
      [Crossref], [PubMed], [CAS], Google Scholar
      45
      Recent developments of the Chemistry Development Kit (CDK) - an open-source Java library for chemo- and bioinformatics
      Steinbeck, Christoph; Hoppe, Christian; Kuhn, Stefan; Floris, Matteo; Guha, Rajarshi; Willighagen, Egon L.
      Current Pharmaceutical Design (2006), 12 (17), 2111-2120CODEN: CPDEFP; ISSN:1381-6128. (Bentham Science Publishers Ltd.)
      The Chem. Development Kit (CDK) provides methods for common tasks in mol. informatics, including 2D and 3D rendering of chem. structures, I/O routines, SMILES parsing and generation, ring searches, isomorphism checking, structure diagram generation, etc. Implemented in Java, it is used both for server-side computational services, possibly equipped with a web interface, as well as for applications and client-side applets. This article introduces the CDK's new QSAR capabilities and the recently introduced interface to statistical software.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BD28XmslWqsL0%253D&md5=4e7e47ffe75b600ee3f81309a4bbb609
    46. 46
      Sun, J.; Jeliazkova, N.; Chupakhin, V.; Golib-Dzib, J.-F.; Engkvist, O.; Carlsson, L.; Wegner, J.; Ceulemans, H.; Georgiev, I.; Jeliazkov, V.; Kochev, N.; Ashby, T. J.; Chen, H. ExCAPE-DB: an integrated large scale dataset facilitating Big Data analysis in chemogenomics. J. Cheminf. 2017, 9, 17,  DOI: 10.1186/s13321-017-0222-2
      [Crossref], [CAS], Google Scholar
      46
      ExCAPE-DB: an integrated large scale dataset facilitating Big Data analysis in chemogenomics
      Sun, Jiangming; Jeliazkova, Nina; Chupakin, Vladimir; Golib-Dzib, Jose-Felipe; Engkvist, Ola; Carlsson, Lars; Wegner, Joerg; Ceulemans, Hugo; Georgiev, Ivan; Jeliazkov, Vedrin; Kochev, Nikolay; Ashby, Thomas J.; Chen, Hongming
      Journal of Cheminformatics (2017), 9 (), 17/1-17/9CODEN: JCOHB3; ISSN:1758-2946. (Chemistry Central Ltd.)
      Chemogenomics data generally refers to the activity data of chem. compds. on an array of protein targets and represents an important source of information for building in silico target prediction models. The increasing vol. of chemogenomics data offers exciting opportunities to build models based on Big Data. Prepg. a high quality data set is a vital step in realizing this goal and this work aims to compile such a comprehensive chemogenomics dataset. This dataset comprises over 70 million SAR data points from publicly available databases (PubChem and ChEMBL) including structure, target information and activity annotations. Our aspiration is to create a useful chemogenomics resource reflecting industry-scale data not only for building predictive models of in silico polypharmacol. and offtarget effects but also for the validation of chemoinformatics approaches in general.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXlsFGhsbc%253D&md5=1bd39132077c5c91f19bae5ea47c1b27
    47. 47
      Irwin, J. J.; Sterling, T.; Mysinger, M. M.; Bolstad, E. S.; Coleman, R. G. ZINC: A Free Tool to Discover Chemistry for Biology. J. Chem. Inf. Model. 2012, 52, 1757– 1768,  DOI: 10.1021/ci3001277
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      47
      ZINC: A Free Tool to Discover Chemistry for Biology
      Irwin, John J.; Sterling, Teague; Mysinger, Michael M.; Bolstad, Erin S.; Coleman, Ryan G.
      Journal of Chemical Information and Modeling (2012), 52 (7), 1757-1768CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      ZINC is a free public resource for ligand discovery. The database contains over twenty million com. available mols. in biol. relevant representations that may be downloaded in popular ready-to-dock formats and subsets. The Web site also enables searches by structure, biol. activity, phys. property, vendor, catalog no., name, and CAS no. Small custom subsets may be created, edited, shared, docked, downloaded, and conveyed to a vendor for purchase. The database is maintained and curated for a high purchasing success rate and is freely available at zinc.docking.org.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC38XmvFGnsrg%253D&md5=97f2ede64afc6b5e3ea2f279e38e32a0
    48. 48
      Shivanyuk, A.; Ryabukhin, S.; Tolmachev, A.; Bogolyubsky, A.; Mykytenko, D.; Chupryna, A.; Heilman, W.; Kostyuk, A. Enamine real database: Making chemical diversity real. Chem. Today 2007, 25, 58– 59
      [CAS], Google Scholar
      48
      Enamine real database: making chemical diversity real
      Shivanyuk, Alexander N.; Ryabukhin, Sergey V.; Bogolyubsky, Andrey V.; Mykytenko, Dmytro M.; Chupryna, Alexander A.; Heilman, William; Kostyuk, Alexander N.; Tolmachev, Andrey A.
      Chimica Oggi (2007), 25 (6), 58-59CODEN: CHOGDS; ISSN:0392-839X. (Tekno Scienze)
      The Enamine REAL DataBase (RDB) covers rigorously validated chem. space of over 29,000,000 virtual HTS compds., over 10,000,000 of which comply to drug likeness Rule-of-5 stds. The high efficiency of our RDB methodol. is based on 30 optimized reactions, 54 optimized chem. procedures applied to 18,000 proprietary in house and 9000 purchased building blocks and our efficient algorithms for calcg. the synthetic feasibility of all virtual structures. Optimized schemes for RDB prodn. allows the synthesis of 20,000 compds. a month with an av. feasibility rate of 65%.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BD1cXktVSns7w%253D&md5=648ca29e7a3a36591a42d88349789e80
    49. 49
      Huang, R.; Xia, M.; Nguyen, D.-T.; Zhao, T.; Sakamuru, S.; Zhao, J.; Shahane, S. A.; Rossoshek, A.; Simeonov, A. Tox21Challenge to build predictive models of nuclear receptor and stress response pathways as mediated by exposure to environmental chemicals and drugs. Front. Environ. Sci. 2016, 3, 85,  DOI: 10.3389/fenvs.2015.00085
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    50. 50
      Ramakrishnan, R.; Hartmann, M.; Tapavicza, E.; Von Lilienfeld, O. A. Electronic spectra from TDDFT and machine learning in chemical space. J. Chem. Phys. 2015, 143, 084111,  DOI: 10.1063/1.4928757
      [Crossref], [PubMed], [CAS], Google Scholar
      50
      Electronic spectra from TDDFT and machine learning in chemical space
      Ramakrishnan, Raghunathan; Hartmann, Mia; Tapavicza, Enrico; von Lilienfeld, O. Anatole
      Journal of Chemical Physics (2015), 143 (8), 084111/1-084111/8CODEN: JCPSA6; ISSN:0021-9606. (American Institute of Physics)
      Due to its favorable computational efficiency, time-dependent (TD) d. functional theory (DFT) enables the prediction of electronic spectra in a high-throughput manner across chem. space. Its predictions, however, can be quite inaccurate. We resolve this issue with machine learning models trained on deviations of ref. second-order approx. coupled-cluster (CC2) singles and doubles spectra from TDDFT counterparts, or even from DFT gap. We applied this approach to low-lying singlet-singlet vertical electronic spectra of over 20 000 synthetically feasible small org. mols. with up to eight CONF atoms. The prediction errors decay monotonously as a function of training set size. For a training set of 10 000 mols., CC2 excitation energies can be reproduced to within ±0.1 eV for the remaining mols. Anal. of our spectral database via chromophore counting suggests that even higher accuracies can be achieved. Based on the evidence collected, we discuss open challenges assocd. with data-driven modeling of high-lying spectra and transition intensities. (c) 2015 American Institute of Physics.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC2MXhsVSlt73O&md5=b439f44c75bb8f99906d3c920bfe7c6f
    51. 51
      Ramakrishnan, R.; Dral, P. O.; Rupp, M.; Von Lilienfeld, O. A. Quantum chemistry structures and properties of 134 kilo molecules. Sci. Data 2014, 1, 140022,  DOI: 10.1038/sdata.2014.22
      [Crossref], [PubMed], [CAS], Google Scholar
      51
      Quantum chemistry structures and properties of 134 kilo molecules
      Ramakrishnan, Raghunathan; Dral, Pavlo O.; Rupp, Matthias; von Lilienfeld, O. Anatole
      Scientific Data (2014), 1 (), 140022CODEN: SDCABS; ISSN:2052-4463. (Nature Publishing Group)
      Computational de novo design of new drugs and materials requires rigorous and unbiased exploration of chem. compd. space. However, large uncharted territories persist due to its size scaling combinatorially with mol. size. We report computed geometric, energetic, electronic, and thermodn. properties for 134k stable small org. mols. made up of CHONF. These mols. correspond to the subset of all 133,885 species with up to nine heavy atoms (CONF) out of the GDB-17 chem. universe of 166 billion org. mols. We report geometries minimal in energy, corresponding harmonic frequencies, dipole moments, polarizabilities, along with energies, enthalpies, and free energies of atomization. All properties were calcd. at the B3LYP/6-31G(2df,p) level of quantum chem. Furthermore, for the predominant stoichiometry, C7H10O2, there are 6,095 constitutional isomers among the 134k mols. We report energies, enthalpies, and free energies of atomization at the more accurate G4MP2 level of theory for all of them. As such, this data set provides quantum chem. properties for a relevant, consistent, and comprehensive chem. space of small org. mols. This database may serve the benchmarking of existing methods, development of new methods, such as hybrid quantum mechanics/machine learning, and systematic identification of structure-property relationships.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC2MXks1aisLo%253D&md5=feaffe204e7139a5fcd685bc2c6841fc
    52. 52
      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
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      52
      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.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BD2cXjs1Sjs74%253D&md5=86e609172307402d8b0d4589b1270a2f
    53. 53
      Cho, K.; Van Merrïenboer, B.; Bahdanau, D.; Bengio, Y. On the properties of neural machine translation: Encoder-decoder approaches. arXiv (Computation and Language) , October 7, 2014, 1409.1259, ver. 2.
      Google Scholar
      There is no corresponding record for this reference.
    54. 54
      Hochreiter, S.; Schmidhuber, J. Long short-term memory. Neural computation 1997, 9, 1735– 1780,  DOI: 10.1162/neco.1997.9.8.1735
      [Crossref], [PubMed], [CAS], Google Scholar
      54
      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.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A280%3ADyaK1c%252FhvVahsQ%253D%253D&md5=5da426ddc18e5bc1972e520bbcc33bec
    55. 55
      Goodfellow, I. J.; Pouget-Abadie, J.; Mirza, M.; Xu, B.; Warde-Farley, D.; Ozair, S.; Courville, A.; Bengio, Y. Generative adversarial networks. arXiv (Machine Learning) , June 10, 2014, 1406.2661, ver. 1.
      Google Scholar
      There is no corresponding record for this reference.
    56. 56
      Kingma, D. P.; Welling, M. Auto-encoding variational bayes. arXiv (Machine Learning) , May 1, 2014, 1312.6114, ver. 10.
      Google Scholar
      There is no corresponding record for this reference.
    57. 57
      Makhzani, A.; Shlens, J.; Jaitly, N.; Goodfellow, I.; Frey, B. Adversarial autoencoders. arXiv (Machine Learning) , May 25, 2016, 1511.05644, ver. 2..
      Google Scholar
      There is no corresponding record for this reference.
    58. 58
      Olivecrona, M.; Blaschke, T.; Engkvist, O.; Chen, H. Molecular de-novo design through deep reinforcement learning. J. Cheminf. 2017, 9, 48,  DOI: 10.1186/s13321-017-0235-x
      [Crossref], [CAS], Google Scholar
      58
      Molecular de-novo design through deep reinforcement learning
      Olivecrona Marcus; Blaschke Thomas; Engkvist Ola; Chen Hongming
      Journal of cheminformatics (2017), 9 (1), 48 ISSN:1758-2946.
      This work introduces a method to tune a sequence-based generative model for molecular de novo design that through augmented episodic likelihood can learn to generate structures with certain specified desirable properties. We demonstrate how this model can execute a range of tasks such as generating analogues to a query structure and generating compounds predicted to be active against a biological target. As a proof of principle, the model is first trained to generate molecules that do not contain sulphur. As a second example, the model is trained to generate analogues to the drug Celecoxib, a technique that could be used for scaffold hopping or library expansion starting from a single molecule. Finally, when tuning the model towards generating compounds predicted to be active against the dopamine receptor type 2, the model generates structures of which more than 95% are predicted to be active, including experimentally confirmed actives that have not been included in either the generative model nor the activity prediction model. Graphical abstract .
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A280%3ADC%252BC1M7mtVKisg%253D%253D&md5=68f5a1219ef81fbd52a5c8911cacbbfc
    59. 59
      Gupta, A.; Müller, A. T.; Huisman, B. J.; Fuchs, J. A.; Schneider, P.; Schneider, G. Generative Recurrent Networks for De Novo Drug Design. Mol. Inf. 2018, 37, 1700111,  DOI: 10.1002/minf.201700111
      [Crossref], [PubMed], Google Scholar
      There is no corresponding record for this reference.
    60. 60
      Guimaraes, G. L.; Sanchez-Lengeling, B.; Outeiral, C.; Farias, P. L. C.; Aspuru- Guzik, A. Objective-Reinforced Generative Adversarial Networks (ORGAN) for Sequence Generation Models. arXiv (Machine Learning) , February 7, 2018, 1705.10843, ver. 3.
      Google Scholar
      There is no corresponding record for this reference.
    61. 61
      Lim, J.; Ryu, S.; Kim, J. W.; Kim, W. Y. Molecular generative model based on conditional variational autoencoder for de novo molecular design. J. Cheminf. 2018, 10, 31,  DOI: 10.1186/s13321-018-0286-7
      [Crossref], [CAS], Google Scholar
      61
      Molecular generative model based on conditional variational autoencoder for de novo molecular design
      Lim, Jaechang; Ryu, Seongok; Kim, Jin Woo; Kim, Woo Youn
      Journal of Cheminformatics (2018), 10 (), 31/1-31/9CODEN: JCOHB3; ISSN:1758-2946. (Chemistry Central Ltd.)
      We propose a mol. generative model based on the conditional variational autoencoder for de novo mol. design. It is specialized to control multiple mol. properties simultaneously by imposing them on a latent space. As a proof of concept, we demonstrate that it can be used to generate drug-like mols. with five target properties. We were also able to adjust a single property without changing the others and to manipulate it beyond the range of the dataset.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXmtFWhtrw%253D&md5=8a1262077f9d6ffa5ec0385ed1a69f6f
    62. 62
      Merk, D.; Grisoni, F.; Friedrich, L.; Schneider, G. Tuning artificial intelligence on the de novo design of natural-product-inspired retinoid X receptor modulators. Commun. Chem. 2018, 1, 68,  DOI: 10.1038/s42004-018-0068-1
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    63. 63
      Merk, D.; Friedrich, L.; Grisoni, F.; Schneider, G. De Novo Design of Bioactive Small Molecules by Artificial Intelligence. Mol. Inf. 2018, 37, 1700153,  DOI: 10.1002/minf.201700153
      [Crossref], [PubMed], Google Scholar
      There is no corresponding record for this reference.
    64. 64
      Polykovskiy, D.; Zhebrak, A.; Vetrov, D.; Ivanenkov, Y.; Aladinskiy, V.; Mamoshina, P.; Bozdaganyan, M.; Aliper, A.; Zhavoronkov, A.; Kadurin, A. Entangled Conditional Adversarial Autoencoder for de Novo Drug Discovery. Mol. Pharmaceutics 2018, 15, 4398– 4405,  DOI: 10.1021/acs.molpharmaceut.8b00839
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      64
      Entangled Conditional Adversarial Autoencoder for de Novo Drug Discovery
      Polykovskiy, Daniil; Zhebrak, Alexander; Vetrov, Dmitry; Ivanenkov, Yan; Aladinskiy, Vladimir; Mamoshina, Polina; Bozdaganyan, Marine; Aliper, Alexander; Zhavoronkov, Alex; Kadurin, Artur
      Molecular Pharmaceutics (2018), 15 (10), 4398-4405CODEN: MPOHBP; ISSN:1543-8384. (American Chemical Society)
      Modern computational approaches and machine learning techniques accelerate the invention of new drugs. Generative models can discover novel mol. structures within hours, while conventional drug discovery pipelines require months of work. In this article, we propose a new generative architecture, entangled conditional adversarial autoencoder that generates mol. structures based on various properties, such as activity against a specific protein, soly., or ease of synthesis. We apply the proposed model to generate a novel inhibitor of Janus kinase 3, implicated in rheumatoid arthritis, psoriasis and vitiligo. The discovered mol. was tested in vitro and showed good activity and selectivity.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXhs1ensL%252FM&md5=58796965d266a5b78ca53c0f11c73999
    65. 65
      Li, Y.; Vinyals, O.; Dyer, C.; Pascanu, R.; Battaglia, P. Learning Deep Generative Models of Graphs. arXiv (Machine Learning) , March 8, 2018, 1803.03324, ver. 1.
      Google Scholar
      There is no corresponding record for this reference.
    66. 66
      Liu, Q.; Allamanis, M.; Brockschmidt, M.; Gaunt, A. Constrained graph variational autoencoders for molecule design. Adv. Neural Inf. Process. Syst. 2018, 7795– 7804
      Google Scholar
      There is no corresponding record for this reference.
    67. 67
      Mercado, R.; Rastemo, T.; Lindelof, E.; Klambauer, G.; Engkvist, O.; Chen, H.; Bjerrum, E. J. Graph networks for molecular design. Mach. Learn.: Sci. Technol. 2021, 2, 025023,  DOI: 10.1088/2632-2153/abcf91
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    68. 68
      De Cao, N.; Kipf, T. MolGAN: An implicit generative model for small molecular graphs. arXiv (Machine Learning) , May 30, 2018, 1805.11973, ver. 1.
      Google Scholar
      There is no corresponding record for this reference.
    69. 69
      Simonovsky, M.; Komodakis, N. Graphvae: Towards generation of small graphs us- ing variational autoencoders. International Conference on Artificial Neural Networks. 2018, 11139, 412– 422,  DOI: 10.1007/978-3-030-01418-6_41
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    70. 70
      Ma, T.; Chen, J.; Xiao, C. Constrained Generation of Semantically Valid Graphs via Regularizing Variational Autoencoders. Adv. Neural Inf. Process. Syst. 2018, 7113– 7124
      Google Scholar
      There is no corresponding record for this reference.
    71. 71
      Hawkins, P. C. D. Conformation Generation: The State of the Art. J. Chem. Inf. Model. 2017, 57, 1747– 1756,  DOI: 10.1021/acs.jcim.7b00221
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      71
      Conformation Generation: The State of the Art
      Hawkins, Paul C. D.
      Journal of Chemical Information and Modeling (2017), 57 (8), 1747-1756CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      The generation of conformations for small mols. is a problem of continuing interest in cheminformatics and computational drug discovery. This review will present an overview of methods used to sample conformational space, focusing on those methods designed for org. mols. commonly of interest in drug discovery. Different approaches to both the sampling of conformational space and the scoring of conformational stability will be compared and contrasted, with an emphasis on those methods suitable for conformer sampling of large nos. of drug-like mols. Particular attention will be devoted to the appropriate utilization of information from exptl. solid-state structures in validating and evaluating the performance of these tools. The review will conclude with some areas worthy of further investigation.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC2sXhtFagu7bE&md5=776818e66798d1987e015440a28e208e
    72. 72
      Skalic, M.; Jiḿenez, J.; Sabbadin, D.; De Fabritiis, G. Shape-Based Generative Mod- eling for de Novo Drug Design. J. Chem. Inf. Model. 2019, 59, 1205– 1214,  DOI: 10.1021/acs.jcim.8b00706
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      72
      Shape-based generative modeling for de novo drug design
      Skalic, Miha; Jimenez, Jose; Sabbadin, Davide; De Fabritiis, Gianni
      Journal of Chemical Information and Modeling (2019), 59 (3), 1205-1214CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      In this work, we propose a machine learning approach to generate novel mols. starting from a seed compd., its three-dimensional (3D) shape, and its pharmacophoric features. The pipeline draws inspiration from generative models used in image anal. and represents a first example of the de novo design of lead-like mols. guided by shape-based features. A variational autoencoder is used to perturb the 3D representation of a compd., followed by a system of convolutional and recurrent neural networks that generate a sequence of SMILES tokens. The generative design of novel scaffolds and functional groups can cover unexplored regions of chem. space that still possess lead-like properties.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXjtVShu78%253D&md5=8eb58a7d5aa780431d285a63dfed5765
    73. 73
      Gebauer, N.; Gastegger, M.; Schütt, K. T. Symmetry-adapted generation of 3d point sets for the targeted discovery of molecules. NeurIPS . 2019.
      Google Scholar
      There is no corresponding record for this reference.
    74. 74
      Ragoza, M.; Masuda, T.; Koes, D. R. Learning a Continuous Representation of 3D Molecular Structures with Deep Generative Models. arXiv (Quantitative Methods) , November 15, 2020, 2010.08687, ver. 3.
      Google Scholar
      There is no corresponding record for this reference.
    75. 75
      Preuer, K.; Renz, P.; Unterthiner, T.; Hochreiter, S.; Klambauer, G. Fŕechet ChemNet Distance: A Metric for Generative Models for Molecules in Drug Discovery. J. Chem. Inf. Model. 2018, 58, 1736– 1741,  DOI: 10.1021/acs.jcim.8b00234
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      75
      Fr´echet ChemNet Distance: A Metric for Generative Models for Molecules in Drug Discovery
      Preuer, Kristina; Renz, Philipp; Unterthiner, Thomas; Hochreiter, Sepp; Klambauer, Guenter
      Journal of Chemical Information and Modeling (2018), 58 (9), 1736-1741CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      The new wave of successful generative models in machine learning has increased the interest in deep learning driven de novo drug design. However, method comparison is difficult because of various flaws of the currently employed evaluation metrics. The authors propose an evaluation metric for generative models called Fre´chet ChemblNet distance (FCD). The FCD's advantage over previous metrics is that it can detect whether generated mols. are diverse and have similar chem. and biol. properties as real mols.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXhsFejsLrL&md5=dc8e10eb0f85a7f27b48da91dcb21a27
    76. 76
      Arús-Pous, J.; Blaschke, T.; Ulander, S.; Reymond, J.-L.; Chen, H.; Engkvist, O. Exploring the GDB-13 chemical space using deep generative models. J. Cheminf. 2019, 11, 1– 14,  DOI: 10.1186/s13321-019-0341-z
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    77. 77
      Brown, N.; Fiscato, M.; Segler, M. H.; Vaucher, A. C. GuacaMol: benchmarking models for de novo molecular design. J. Chem. Inf. Model. 2019, 59, 1096– 1108,  DOI: 10.1021/acs.jcim.8b00839
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      77
      GuacaMol: Benchmarking Models for de Novo Molecular Design
      Brown, Nathan; Fiscato, Marco; Segler, Marwin H. S.; Vaucher, Alain C.
      Journal of Chemical Information and Modeling (2019), 59 (3), 1096-1108CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      De novo design seeks to generate mols. with required property profiles by virtual design-make-test cycles. With the emergence of deep learning and neural generative models in many application areas, models for mol. design based on neural networks appeared recently and show promising results. However, the new models have not been profiled on consistent tasks, and comparative studies to well-established algorithms have only seldom been performed. To standardize the assessment of both classical and neural models for de novo mol. design, we propose an evaluation framework, GuacaMol, based on a suite of standardized benchmarks. The benchmark tasks encompass measuring the fidelity of the models to reproduce the property distribution of the training sets, the ability to generate novel mols., the exploration and exploitation of chem. space, and a variety of single and multiobjective optimization tasks. The benchmarking open-source Python code and a leaderboard can be found on https://benevolent.ai/guacamol.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXltVWrsbY%253D&md5=d3fb616b81a4b146cf77950a1c92e4d1
    78. 78
      Polykovskiy, D.; Zhebrak, A.; Sanchez-Lengeling, B.; Golovanov, S.; Tatanov, O.; Belyaev, S.; Kurbanov, R.; Artamonov, A.; Aladinskiy, V.; Veselov, M. Molecular sets (MOSES): a benchmarking platform for molecular generation models. Front. Pharmacol. 2020, 11, 11,  DOI: 10.3389/fphar.2020.565644
      [Crossref], [PubMed], Google Scholar
      There is no corresponding record for this reference.
    79. 79
      Renz, P.; Van Rompaey, D.; Wegner, J. K.; Hochreiter, S.; Klambauer, G. On fail- ure modes in molecule generation and optimization. Drug Discovery Today: Technol. 2019, 32–33, 55– 63,  DOI: 10.1016/j.ddtec.2020.09.003
      [Crossref], [PubMed], [CAS], Google Scholar
      79
      On failure modes in molecule generation and optimization
      Renz Philipp; Hochreiter Sepp; Klambauer Gunter; Van Rompaey Dries; Wegner Jorg Kurt
      Drug discovery today. Technologies (2019), 32-33 (), 55-63 ISSN:.
      There has been a wave of generative models for molecules triggered by advances in the field of Deep Learning. These generative models are often used to optimize chemical compounds towards particular properties or a desired biological activity. The evaluation of generative models remains challenging and suggested performance metrics or scoring functions often do not cover all relevant aspects of drug design projects. In this work, we highlight some unintended failure modes in molecular generation and optimization and how these evade detection by current performance metrics.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A280%3ADC%252BB3svkt1yqtg%253D%253D&md5=ca243ca3af904c3fa64f9e0e2c2b2c2c
    80. 80
      Cieplinski, T.; Danel, T.; Podlewska, S.; Jastrzebski, S. We should at least be able to Design Molecules that Dock Well. arXiv (Biomolecules) December 28, 2020, 2006.16955, ver. 3.
      Google Scholar
      There is no corresponding record for this reference.
    81. 81
      Zhang, J.; Mercado, R.; Engkvist, O.; Chen, H. Comparative study of deep generative models on chemical space coverage. ChemRxiv , May 2, 2021, ver. 3.  DOI: 10.26434/chemrxiv.13234289.v3 .
      [Crossref], [PubMed], Google Scholar
      There is no corresponding record for this reference.
    82. 82
      Blaschke, T.; Arús-Pous, J.; Chen, H.; Margreitter, C.; Tyrchan, C.; Engkvist, O.; Papadopoulos, K.; Patronov, A. REINVENT 2.0: An AI Tool for De Novo Drug Design. J. Chem. Inf. Model. 2020, 60, 5918,  DOI: 10.1021/acs.jcim.0c00915
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      82
      REINVENT 2.0: An AI Tool for De Novo Drug Design
      Blaschke, Thomas; Arus-Pous, Josep; Chen, Hongming; Margreitter, Christian; Tyrchan, Christian; Engkvist, Ola; Papadopoulos, Kostas; Patronov, Atanas
      Journal of Chemical Information and Modeling (2020), 60 (12), 5918-5922CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      In the past few years, we have witnessed a renaissance of the field of mol. de novo drug design. The advancements in deep learning and artificial intelligence (AI) have triggered an avalanche of ideas on how to translate such techniques to a variety of domains including the field of drug design. A range of architectures have been devised to find the optimal way of generating chem. compds. by using either graph- or string (SMILES)-based representations. With this application note, we aim to offer the community a prodn.-ready tool for de novo design, called REINVENT. It can be effectively applied on drug discovery projects that are striving to resolve either exploration or exploitation problems while navigating the chem. space. It can facilitate the idea generation process by bringing to the researcher's attention the most promising compds. REINVENT's code is publicly available at https://github.com/MolecularAI/Reinvent.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXitFOnsbbJ&md5=b06445c0516f122adff2e7c82d7ca70c
    83. 83
      Bung, N.; Krishnan, S. R.; Bulusu, G.; Roy, A. De novo design of new chemical entities for SARS-CoV-2 using artificial intelligence. Future Med. Chem. 2021, 13, 575,  DOI: 10.4155/fmc-2020-0262
      [Crossref], [PubMed], [CAS], Google Scholar
      83
      De novo design of new chemical entities for SARS-CoV-2 using artificial intelligence
      Bung, Navneet; Krishnan, Sowmya R.; Bulusu, Gopalakrishnan; Roy, Arijit
      Future Medicinal Chemistry (2021), 13 (6), 575-585CODEN: FMCUA7; ISSN:1756-8919. (Newlands Press Ltd.)
      The novel coronavirus SARS-CoV-2 has severely affected the health and economy of several countries. Multiple studies are in progress to design novel therapeutics against the potential target proteins in SARS-CoV-2, including 3CL protease, an essential protein for virus replication. In this study we employed deep neural network-based generative and predictive models for de novo design of small mols. capable of inhibiting the 3CL protease. The generative model was optimized using transfer learning and reinforcement learning to focus around the chem. space corresponding to the protease inhibitors. Multiple physicochem. property filters and virtual screening score were used for the final screening. We have identified 33 potential compds. as ideal candidates for further synthesis and testing against SARS-CoV-2.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3MXmtlOkurY%253D&md5=4ad8a669ca9c653df25f57c508fdefa0
    84. 84
      Li, Y.; Zhang, L.; Liu, Z. Multi-objective de novo drug design with conditional graph generative model. J. Cheminf. 2018, 10, 33,  DOI: 10.1186/s13321-018-0287-6
      [Crossref], [CAS], Google Scholar
      84
      Multi-objective de novo drug design with conditional graph generative model
      Li, Yibo; Zhang, Liangren; Liu, Zhenming
      Journal of Cheminformatics (2018), 10 (), 33/1-33/24CODEN: JCOHB3; ISSN:1758-2946. (Chemistry Central Ltd.)
      Recently, deep generative models have revealed itself as a promising way of performing de novo mol. design. However, previous research has focused mainly on generating SMILES strings instead of mol. graphs. Although available, current graph generative models are are often too general and computationally expensive. In this work, a new de novo mol. design framework is proposed based on a type of sequential graph generators that do not use atom level recurrent units. Compared with previous graph generative models, the proposed method is much more tuned for mol. generation and has been scaled up to cover significantly larger mols. in the ChEMBL database. It is shown that the graph-based model outperforms SMILES based models in a variety of metrics, esp. in the rate of valid outputs. For the application of drug design tasks, conditional graph generative model is employed. This method offers highe flexibility and is suitable for generation based on multiple objectives. The results have demonstrated that this approach can be effectively applied to solve several drug design problems, including the generation of compds. contg. a given scaffold, compds. with specific drug-likeness and synthetic accessibility requirements, as well as dual inhibitors against JNK3 and GSK-3ss.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXmtFWiu7s%253D&md5=6cef99e5a789668f55ba72e4e65da160
    85. 85
      Blaschke, T.; Engkvist, O.; Bajorath, J.; Chen, H. Memory-assisted reinforcement learning for diverse molecular de novo design. J. Cheminf. 2020, 12, 1– 17,  DOI: 10.1186/s13321-020-00473-0
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    86. 86
      Zhavoronkov, A. Deep learning enables rapid identification of potent DDR1 kinase inhibitors. Nat. Biotechnol. 2019, 37, 1038– 1040,  DOI: 10.1038/s41587-019-0224-x
      [Crossref], [PubMed], [CAS], Google Scholar
      86
      Deep learning enables rapid identification of potent DDR1 kinase inhibitors
      Zhavoronkov, Alex; Ivanenkov, Yan A.; Aliper, Alex; Veselov, Mark S.; Aladinskiy, Vladimir A.; Aladinskaya, Anastasiya V.; Terentiev, Victor A.; Polykovskiy, Daniil A.; Kuznetsov, Maksim D.; Asadulaev, Arip; Volkov, Yury; Zholus, Artem; Shayakhmetov, Rim R.; Zhebrak, Alexander; Minaeva, Lidiya I.; Zagribelnyy, Bogdan A.; Lee, Lennart H.; Soll, Richard; Madge, David; Xing, Li; Guo, Tao; Aspuru-Guzik, Alan
      Nature Biotechnology (2019), 37 (9), 1038-1040CODEN: NABIF9; ISSN:1087-0156. (Nature Research)
      We have developed a deep generative model, generative tensorial reinforcement learning (GENTRL), for de novo small-mol. design. GENTRL optimizes synthetic feasibility, novelty, and biol. activity. We used GENTRL to discover potent inhibitors of discoidin domain receptor 1 (DDR1), a kinase target implicated in fibrosis and other diseases, in 21 days. Four compds. were active in biochem. assays, and two were validated in cell-based assays. One lead candidate was tested and demonstrated favorable pharmacokinetics in mice.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXhs12gurnM&md5=b15262b61b9172ab2bc37e534a70f010
    87. 87
      Popova, M.; Shvets, M.; Oliva, J.; Isayev, O. MolecularRNN: Generating real- istic molecular graphs with optimized properties. arXiv (Machine Learning) , May 31, 2019, 1905.13372, ver. 1.
      Google Scholar
      There is no corresponding record for this reference.
    88. 88
      Sanchez-Lengeling, B.; Outeiral, C.; Guimaraes, G. L.; Aspuru-Guzik, A. Optimizing distributions over molecular space. An objective-reinforced generative adversarial network for inverse-design chemistry (ORGANIC). ChemRxiv , August 18, 2017, ver. 3.  DOI: 10.26434/chemrxiv.5309668.v3 .
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    89. 89
      Putin, E.; Asadulaev, A.; Vanhaelen, Q.; Ivanenkov, Y.; Aladinskaya, A. V.; Aliper, A.; Zhavoronkov, A. Adversarial Threshold Neural Computer for Molecular de Novo De- sign. Mol. Pharmaceutics 2018, 15, 4386– 4397,  DOI: 10.1021/acs.molpharmaceut.7b01137
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      89
      Adversarial threshold neural computer for molecular de novo design
      Putin, Evgeny; Asadulaev, Arip; Vanhaelen, Quentin; Ivanenkov, Yan; Aladinskaya, Anastasia V.; Aliper, Alex; Zhavoronkov, Alex
      Molecular Pharmaceutics (2018), 15 (10), 4386-4397CODEN: MPOHBP; ISSN:1543-8384. (American Chemical Society)
      In this article, we propose the deep neural network Adversarial Threshold Neural Computer (ATNC). The ATNC model is intended for the de novo design of novel small-mol. org. structures. The model is based on generative adversarial network architecture and reinforcement learning. ATNC uses a Differentiable Neural Computer as a generator and has a new specific block, called adversarial threshold (AT). AT acts as a filter between the agent (generator) and the environment (discriminator + objective reward functions). Furthermore, to generate more diverse mols. we introduce a new objective reward function named Internal Diversity Clustering (IDC). In this work, ATNC is tested and compared with the Org. model. Both models were trained on the SMILES string representation of the mols., using four objective functions (internal similarity, Muegge druglikeness filter, presence or absence of sp3-rich fragments, and IDC). The SMILES representations of 15K druglike mols. from the ChemDiv collection were used as a training data set. For the different functions, ATNC outperforms Org. Combined with the IDC, ATNC generates 72% of valid and 77% of unique SMILES strings, while Org. generates only 7% of valid and 86% of unique SMILES strings. For each set of mols. generated by ATNC and Org., we analyzed distributions of four mol. descriptors (no. of atoms, mol. wt., logP, and tpsa) and calcd. five chem. statistical features (internal diversity, no. of unique heterocycles, no. of clusters, no. of singletons, and no. of compds. that have not been passed through medicinal chem. filters). Anal. of key mol. descriptors and chem. statistical features demonstrated that the mols. generated by ATNC elicited better druglikeness properties. We also performed in vitro validation of the mols. generated by ATNC; results indicated that ATNC is an effective method for producing hit compds.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXlt1Kks7c%253D&md5=a83d9ef283d643d2b8e81f050c4e7be2
    90. 90
      Putin, E.; Asadulaev, A.; Ivanenkov, Y.; Aladinskiy, V.; Sanchez-Lengeling, B.; Aspuru-Guzik, A.; Zhavoronkov, A. Reinforced Adversarial Neural Computer for de Novo Molecular Design. J. Chem. Inf. Model. 2018, 58, 1194– 1204,  DOI: 10.1021/acs.jcim.7b00690
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      90
      Reinforced Adversarial Neural Computer for de Novo Molecular Design
      Putin, Evgeny; Asadulaev, Arip; Ivanenkov, Yan; Aladinskiy, Vladimir; Sanchez-Lengeling, Benjamin; Aspuru-Guzik, Alan; Zhavoronkov, Alex
      Journal of Chemical Information and Modeling (2018), 58 (6), 1194-1204CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      In silico modeling is a crucial milestone in modern drug design and development. Although computer-aided approaches in this field are well-studied, the application of deep learning methods in this research area is at the beginning. In this work, we present an original deep neural network (DNN) architecture named RANC (Reinforced Adversarial Neural Computer) for the de novo design of novel small-mol. org. structures based on the generative adversarial network (GAN) paradigm and reinforcement learning (RL). As a generator RANC uses a differentiable neural computer (DNC), a category of neural networks, with increased generation capabilities due to the addn. of an explicit memory bank, which can mitigate common problems found in adversarial settings. The comparative results have shown that RANC trained on the SMILES string representation of the mols. outperforms its first DNN-based counterpart Org. by several metrics relevant to drug discovery: the no. of unique structures, passing medicinal chem. filters (MCFs), Muegge criteria, and high QED scores. RANC is able to generate structures that match the distributions of the key chem. features/descriptors (e.g., MW, logP, TPSA) and lengths of the SMILES strings in the training data set. Therefore, RANC can be reasonably regarded as a promising starting point to develop novel mols. with activity against different biol. targets or pathways. In addn., this approach allows scientists to save time and covers a broad chem. space populated with novel and diverse compds.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXpsVChtrs%253D&md5=b10c44dcadf9fb1afc4e65cc7469730f
    91. 91
      You, J.; Liu, B.; Ying, Z.; Pande, V.; Leskovec, J. Graph convolutional policy network for goal-directed molecular graph generation. Adv. Neural Inf. Process. Syst. 2018, 31, 6410– 6421
      Google Scholar
      There is no corresponding record for this reference.
    92. 92
      Karimi, M.; Hasanzadeh, A.; Shen, Y. Network-principled deep generative models for designing drug combinations as graph sets. Bioinformatics 2020, 36, i445– i454,  DOI: 10.1093/bioinformatics/btaa317
      [Crossref], [PubMed], [CAS], Google Scholar
      92
      Network-principled deep generative models for designing drug combinations as graph sets
      Karimi, Mostafa; Hasanzadeh, Arman; Shen, Yang
      Bioinformatics (2020), 36 (Suppl._1), i445-i454CODEN: BOINFP; ISSN:1367-4811. (Oxford University Press)
      Motivation: Combination therapy has shown to improve therapeutic efficacy while reducing side effects. Importantly, it has become an indispensable strategy to overcome resistance in antibiotics, antimicrobials and anticancer drugs. Facing enormous chem. space and unclear design principles for small-mol. combinations, computational drug-combination design has not seen generative models to meet its potential to accelerate resistance-overcoming drug combination discovery. Results: We have developed the first deep generative model for drug combination design, by jointly embedding graph-structured domain knowledge and iteratively training a reinforcement learning-based chem. graph-set designer. First, we have developed hierarchical variational graph auto-encoders trained end-to-end to jointly embed gene-gene, gene-disease and disease-disease networks. Novel attentional pooling is introduced here for learning disease representations from assocd. genes' representations. Second, targeting diseases in learned representations, we have recast the drug-combination design problem as graph-set generation and developed a deep learning-based model with novel rewards. Specifically, besides chem. validity rewards, we have introduced novel generative adversarial award, being generalized sliced Wasserstein, for chem. diverse mols. with distributions similar to known drugs. We have also designed a network principle-based reward for disease-specific drug combinations. Numerical results indicate that, compared to state-of-the-art graph embedding methods, hierarchical variational graph auto-encoder learns more informative and generalizable disease representations. Results also show that the deep generative models generate drug combinations following the principle across diseases. Case studies on four diseases show that network-principled drug combinations tend to have low toxicity. The generated drug combinations collectively cover the disease module similar to FDA-approved drug combinations and could potentially suggest novel systems pharmacol. strategies. Ourmethod allows for examg. and following network-based principle or hypothesis to efficiently generate disease-specific drug combinations in a vast chem. combinatorial space.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXis1ymt7zE&md5=b737e2db6e7e76d0753debee66246ed3
    93. 93
      Griffiths, R.-R.; Hernández-Lobato, J. M. Constrained Bayesian optimization for auto- matic chemical design using variational autoencoders. Chem. Sci. 2020, 11, 577– 586,  DOI: 10.1039/C9SC04026A
      [Crossref], [PubMed], [CAS], Google Scholar
      93
      Constrained Bayesian optimization for automatic chemical design using variational autoencoders
      Griffiths, Ryan-Rhys; Hernandez-Lobato, Jose Miguel
      Chemical Science (2020), 11 (2), 577-586CODEN: CSHCCN; ISSN:2041-6520. (Royal Society of Chemistry)
      Automatic Chem. Design is a framework for generating novel mols. with optimized properties. The original scheme, featuring Bayesian optimization over the latent space of a variational autoencoder, suffers from the pathol. that it tends to produce invalid mol. structures. First, we demonstrate empirically that this pathol. arises when the Bayesian optimization scheme queries latent space points far away from the data on which the variational autoencoder has been trained. Secondly, by reformulating the search procedure as a constrained Bayesian optimization problem, we show that the effects of this pathol. can be mitigated, yielding marked improvements in the validity of the generated mols. We posit that constrained Bayesian optimization is a good approach for solving this kind of training set mismatch in many generative tasks involving Bayesian optimization over the latent space of a variational autoencoder.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXitFOis7bK&md5=628684b606d9d93ccf21b674438acd6a
    94. 94
      Blaschke, T.; Olivecrona, M.; Engkvist, O.; Bajorath, J.; Chen, H. Application of Generative Autoencoder in De Novo Molecular Design. Mol. Inf. 2018, 37, 1700123,  DOI: 10.1002/minf.201700123
      [Crossref], [PubMed], Google Scholar
      There is no corresponding record for this reference.
    95. 95
      Kusner, M. J.; Paige, B.; Hernández-Lobato, J. M. Grammar variational autoencoder. Proc. 34th Int. Conf. Mach. Learn. 2017, 70, 1945– 1954
      Google Scholar
      There is no corresponding record for this reference.
    96. 96
      Dai, H.; Tian, Y.; Dai, B.; Skiena, S.; Song, L. Syntax-directed variational autoencoder for structured data. arXiv (Machine Learning) , February 24, 2018, 1802.08786, ver 1.
      Google Scholar
      There is no corresponding record for this reference.
    97. 97
      Jin, W.; Barzilay, R.; Jaakkola, T. Junction tree variational autoencoder for molecular graph generation. Proc. 35th Int. Conf. Mach. Learn. 2018, 50, 2323– 2332
      Google Scholar
      There is no corresponding record for this reference.
    98. 98
      Samanta, B.; De, A.; Jana, G.; Chattaraj, P. K.; Ganguly, N.; Rodriguez, M. G. NeVAE: A Deep Generative Model for Molecular Graphs. Proceedings of the AAAI Conference on Artificial Intelligence 2019, 33, 1110– 1117,  DOI: 10.1609/aaai.v33i01.33011110
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    99. 99
      Bresson, X.; Laurent, T. A Two-Step Graph Convolutional Decoder for Molecule Generation. arXiv (Machine Learning) , June 15, 2019, 1906.03412, ver 2.
      Google Scholar
      There is no corresponding record for this reference.
    100. 100
      Maziarka, L.; Pocha, A.; Kaczmarczyk, J.; Rataj, K.; Danel, T.; Warcho-l, M. Mol- CycleGAN: a generative model for molecular optimization. J. Cheminf. 2020, 12, 2,  DOI: 10.1186/s13321-019-0404-1
      [Crossref], [CAS], Google Scholar
      100
      Mol-CycleGAN: a generative model for molecular optimization
      Maziarka, Lukasz; Pocha, Agnieszka; Kaczmarczyk, Jan; Rataj, Krzysztof; Danel, Tomasz; Warchol, Michal
      Journal of Cheminformatics (2020), 12 (1), 2CODEN: JCOHB3; ISSN:1758-2946. (SpringerOpen)
      Designing a mol. with desired properties is one of the biggest challenges in drug development, as it requires optimization of chem. compd. structures with respect to many complex properties. To improve the compd. design process, we introduce Mol-CycleGAN-a CycleGAN-based model that generates optimized compds. with high structural similarity to the original ones. Namely, given a mol. our model generates a structurally similar one with an optimized value of the considered property. We evaluate the performance of the model on selected optimization objectives related to structural properties (presence of halogen groups, no. of arom. rings) and to a physicochem. property (penalized logP). In the task of optimization of penalized logP of drug-like mols. our model significantly outperforms previous results.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXotFarsg%253D%253D&md5=243c1aeef3e517ee774fcfabebf80260
    101. 101
      Sattarov, B.; Baskin, I. I.; Horvath, D.; Marcou, G.; Bjerrum, E. J.; Varnek, A. De Novo Molecular Design by Combining Deep Deep Autoencoder Recurrent Neural Networks with Generative Topographic Mapping. J. Chem. Inf. Model. 2019, 59, 1182– 1196,  DOI: 10.1021/acs.jcim.8b00751
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      101
      De Novo Molecular Design by Combining Deep Autoencoder Recurrent Neural Networks with Generative Topographic Mapping
      Sattarov, Boris; Baskin, Igor I.; Horvath, Dragos; Marcou, Gilles; Bjerrum, Esben Jannik; Varnek, Alexandre
      Journal of Chemical Information and Modeling (2019), 59 (3), 1182-1196CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      Here we show that Generative Topog. Mapping (GTM) can be used to explore the latent space of the SMILES-based autoencoders and generate focused mol. libraries of interest. We have built a sequence-to-sequence neural network with Bidirectional Long Short-Term Memory layers and trained it on the SMILES strings from ChEMBL23. Very high reconstruction rates of the test set mols. were achieved (>98%), which are comparable to the ones reported in related publications. Using GTM, we have visualized the autoencoder latent space on the two-dimensional topog. map. Targeted map zones can be used for generating novel mol. structures by sampling assocd. latent space points and decoding them to SMILES. The sampling method based on a genetic algorithm was introduced to optimize compd. properties "on the fly". The generated focused mol. libraries were shown to contain original and a priori feasible compds. which, pending actual synthesis and testing, showed encouraging behavior in independent structure-based affinity estn. procedures (pharmacophore matching, docking).
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXjtlCisbc%253D&md5=44d2f043cc64e112b2caa7851ed1eb4e
    102. 102
      Winter, R.; Montanari, F.; Steffen, A.; Briem, H.; Nóe, F.; Clevert, D.-A. Efficient multi-objective molecular optimization in a continuous latent space. Chem. Sci. 2019, 10, 8016– 8024,  DOI: 10.1039/C9SC01928F
      [Crossref], [PubMed], [CAS], Google Scholar
      102
      Efficient multi-objective molecular optimization in a continuous latent space
      Winter, Robin; Montanari, Floriane; Steffen, Andreas; Briem, Hans; Noe, Frank; Clevert, Djork-Arne
      Chemical Science (2019), 10 (34), 8016-8024CODEN: CSHCCN; ISSN:2041-6520. (Royal Society of Chemistry)
      One of the main challenges in small mol. drug discovery is finding novel chem. compds. with desirable properties. In this work, we propose a novel method that combines in silico prediction of mol. properties such as biol. activity or pharmacokinetics with an in silico optimization algorithm, namely Particle Swarm Optimization. Our method takes a starting compd. as input and proposes new mols. with more desirable (predicted) properties. It navigates a machine-learned continuous representation of a drug-like chem. space guided by a defined objective function. The objective function combines multiple in silico prediction models, defined desirability ranges and substructure constraints. We demonstrate that our proposed method is able to consistently find more desirable mols. for the studied tasks in relatively short time. We hope that our method can support medicinal chemists in accelerating and improving the lead optimization process.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXhtlaltrjO&md5=c63599dc8a30df8f805ce5457ecc053a
    103. 103
      Chenthamarakshan, V.; Das, P.; Hoffman, C. S.; Strobelt, H.; Padhi, I.; Lim, W. K.; Hoover, B.; Manica, M.; Born, J.; Laino, T.; Mojsilovic, A. CogMol: Target-Specific and Selective Drug Design for COVID-19 Using Deep Generative Models. NeurIPS 2020 2020.
      Google Scholar
      There is no corresponding record for this reference.
    104. 104
      Kotsias, P.-C.; Arús-Pous, J.; Chen, H.; Engkvist, O.; Tyrchan, C.; Bjerrum, E. J. Direct steering of de novo molecular generation with descriptor conditional recurrent neural networks. Nature Machine Intelligence 2020, 2, 254– 265,  DOI: 10.1038/s42256-020-0174-5
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    105. 105
      Shayakhmetov, R.; Kuznetsov, M.; Zhebrak, A.; Kadurin, A.; Nikolenko, S.; Aliper, A.; Polykovskiy, D. Molecular Generation for Desired Transcriptome Changes With Ad- versarial Autoencoders. Front. Pharmacol. 2020, 11, 269,  DOI: 10.3389/fphar.2020.00269
      [Crossref], [PubMed], [CAS], Google Scholar
      105
      Molecular generation for desired transcriptome changes with adversarial autoencoders
      Shayakhmetov, Rim; Kuznetsov, Maksim; Zhebrak, Alexander; Kadurin, Artur; Nikolenko, Sergey; Aliper, Alexander; Polykovskiy, Daniil
      Frontiers in Pharmacology (2020), 11 (), 00269CODEN: FPRHAU; ISSN:1663-9812. (Frontiers Media S.A.)
      Gene expression profiles are useful for assessing the efficacy and side effects of drugs. In this paper, we propose a new generative model that infers drug mols. that could induce a desired change in gene expression. Our model-the Bidirectional Adversarial Autoencoder-explicitly separates cellular processes captured in gene expression changes into two feature sets: those related and unrelated to the drug incubation. The model uses related features to produce a drug hypothesis. We have validated our model on the LINCS L1000 dataset by generating mol. structures in the SMILES format for the desired transcriptional response. In the expts., we have shown that the proposed model can generate novel mol. structures that could induce a given gene expression change or predict a gene expression difference after incubation of a given mol. structure.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXhvFKksbbE&md5=36064731a6cb6395596e418d379f2782
    106. 106
      Ḿendez-Lucio, O.; Baillif, B.; Clevert, D.-A.; Rouquíe, D.; Wichard, J. De novo gener- ation of hit-like molecules from gene expression signatures using artificial intelligence. Nat. Commun. 2020, 11, 1– 10,  DOI: 10.1038/s41467-019-13807-w
      [Crossref], [PubMed], Google Scholar
      There is no corresponding record for this reference.
    107. 107
      Born, J.; Manica, M.; Oskooei, A.; Cadow, J.; Rodŕıguez Mart́ınez, M. PaccMannRL: Designing Anticancer Drugs From Transcriptomic Data via Reinforcement Learning. In Research in Computational Molecular Biology; Springer: Cham, 2020; pp 231– 233.
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    108. 108
      Jin, W.; Yang, K.; Barzilay, R.; Jaakkola, T. Learning Multimodal Graph-to-Graph Translation for Molecular Optimization. arXiv (Machine Learning) , January 28, 2019, 1812.01070, ver. 3.
      Google Scholar
      There is no corresponding record for this reference.
    109. 109
      Masuda, T.; Ragoza, M.; Koes, D. R. Generating 3D Molecular Structures Conditional on a Receptor Binding Site with Deep Generative Models. arXiv (Chemical Physics) , November 23, 2020, 2010.14442, ver. 3.
      Google Scholar
      There is no corresponding record for this reference.
    110. 110
      Kang, S.; Cho, K. Conditional Molecular Design with Deep Generative Models. J. Chem. Inf. Model. 2019, 59, 43– 52,  DOI: 10.1021/acs.jcim.8b00263
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      110
      Conditional Molecular Design with Deep Generative Models
      Kang, Seokho; Cho, Kyunghyun
      Journal of Chemical Information and Modeling (2019), 59 (1), 43-52CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      Although machine learning has been successfully used to propose novel mols. that satisfy desired properties, it is still challenging to explore a large chem. space efficiently. In this paper, we present a conditional mol. design method that facilitates generating new mols. with desired properties. The proposed model, which simultaneously performs both property prediction and mol. generation, is built as a semisupervised variational autoencoder trained on a set of existing mols. with only a partial annotation. We generate new mols. with desired properties by sampling from the generative distribution estd. by the model. We demonstrate the effectiveness of the proposed model by evaluating it on drug-like mols. The model improves the performance of property prediction by exploiting unlabeled mols. and efficiently generates novel mols. fulfilling various target conditions.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXhtlantb3N&md5=d2c3a3ff1f2189698828775e89c2a885
    111. 111
      Lim, J.; Hwang, S.-Y.; Moon, S.; Kim, S.; Kim, W. Y. Scaffold-based molecular design with a graph generative model. Chem. Sci. 2020, 11, 1153– 1164,  DOI: 10.1039/C9SC04503A
      [Crossref], [CAS], Google Scholar
      111
      Scaffold-based molecular design with a graph generative model
      Lim, Jaechang; Hwang, Sang-Yeon; Moon, Seokhyun; Kim, Seungsu; Kim, Woo Youn
      Chemical Science (2020), 11 (4), 1153-1164CODEN: CSHCCN; ISSN:2041-6520. (Royal Society of Chemistry)
      Searching for new mols. in areas like drug discovery often starts from the core structures of known mols. Such a method has called for a strategy of designing deriv. compds. retaining a particular scaffold as a substructure. On this account, our present work proposes a graph generative model that targets its use in scaffold-based mol. design. Our model accepts a mol. scaffold as input and extends it by sequentially adding atoms and bonds. The generated mols. are then guaranteed to contain the scaffold with certainty, and their properties can be controlled by conditioning the generation process on desired properties. The learned rule of extending mols. can well generalize to arbitrary kinds of scaffolds, including those unseen during learning. In the conditional generation of mols., our model can simultaneously control multiple chem. properties despite the search space constrained by fixing the substructure. As a demonstration, we applied our model to designing inhibitors of the epidermal growth factor receptor and show that our model can employ a simple semi-supervised extension to broaden its applicability to situations where only a small amt. of data is available.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXit1Ortr%252FO&md5=be36a65abcce15f18b4c1e529bffd905
    112. 112
      Varnek, A., Ed. Tutorials in chemoinformatics; John Wiley & Sons, Inc: Hoboken, NJ, 2017.
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    113. 113
      Engel, T., Gasteiger, J., Eds. Applied chemoinformatics: achievements and future opportunities; Wiley-VCH: Weinheim, 2018; OCLC: 1034693178.
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    114. 114
      Kadurin, A.; Aliper, A.; Kazennov, A.; Mamoshina, P.; Vanhaelen, Q.; Khrabrov, K.; Zhavoronkov, A. The cornucopia of meaningful leads: Applying deep adversarial au- toencoders for new molecule development in oncology. Oncotarget 2017, 8, 10883– 10890,  DOI: 10.18632/oncotarget.14073
      [Crossref], [PubMed], [CAS], Google Scholar
      114
      The cornucopia of meaningful leads: Applying deep adversarial autoencoders for new molecule development in oncology
      Kadurin Artur; Khrabrov Kuzma; Kadurin Artur; Aliper Alexander; Kazennov Andrey; Mamoshina Polina; Vanhaelen Quentin; Zhavoronkov Alex; Kadurin Artur; Kadurin Artur; Kazennov Andrey; Zhavoronkov Alex; Mamoshina Polina; Zhavoronkov Alex
      Oncotarget (2017), 8 (7), 10883-10890 ISSN:.
      Recent advances in deep learning and specifically in generative adversarial networks have demonstrated surprising results in generating new images and videos upon request even using natural language as input. In this paper we present the first application of generative adversarial autoencoders (AAE) for generating novel molecular fingerprints with a defined set of parameters. We developed a 7-layer AAE architecture with the latent middle layer serving as a discriminator. As an input and output the AAE uses a vector of binary fingerprints and concentration of the molecule. In the latent layer we also introduced a neuron responsible for growth inhibition percentage, which when negative indicates the reduction in the number of tumor cells after the treatment. To train the AAE we used the NCI-60 cell line assay data for 6252 compounds profiled on MCF-7 cell line. The output of the AAE was used to screen 72 million compounds in PubChem and select candidate molecules with potential anti-cancer properties. This approach is a proof of concept of an artificially-intelligent drug discovery engine, where AAEs are used to generate new molecular fingerprints with the desired molecular properties.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A280%3ADC%252BC1c%252FpvFKruw%253D%253D&md5=677ef0264494eb8a7ef8c6584c1202ab
    115. 115
      Alpaydin, E. Introduction to machine learning, 2nd ed.; Adaptive computation and machine learning; MIT Press: Cambridge, Mass, 2010; OCLC: ocn317698631.
      Google Scholar
      There is no corresponding record for this reference.
    116. 116
      Raschka, S. Python machine learning: unlock deeper insights into machine learning with this vital guide to cutting-edge predictive analytics; Community experience distilled; Packt Publishing Open Source: Birmingham, UK; Mumbai, 2016.
      Google Scholar
      There is no corresponding record for this reference.
    117. 117
      Frazier, P. I. A Tutorial on Bayesian Optimization. arXiv (Machine Learning) , July 8, 2018, 1807.02811, ver. 1.
      Google Scholar
      There is no corresponding record for this reference.
    118. 118
      Shahriari, B.; Swersky, K.; Wang, Z.; Adams, R. P.; De Freitas, N. Taking the human out of the loop: A review of Bayesian optimization. Proc. IEEE 2016, 104, 148– 175,  DOI: 10.1109/JPROC.2015.2494218
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    119. 119
      Das, P.; Sercu, T.; Wadhawan, K.; Padhi, I.; Gehrmann, S.; Cipcigan, F.; Chen- thamarakshan, V.; Strobelt, H.; Santos, C. D.; Chen, P.-Y.; Yang, Y. Y.; Tan, J.; Hedrick, J.; Crain, J.; Mojsilovic, A. Accelerating antimicrobial discovery with controllable deep generative models and molecular dynamics. arXiv (Machine Learning) , February 26, 2020, 2005.11248, ver. 2.
      Google Scholar
      There is no corresponding record for this reference.
    120. 120
      Kingma, D. P.; Mohamed, S.; Rezende, D. J.; Welling, M. Semi-supervised learning with deep generative models. Adv. Neural Inf. Process. Syst. 2014, 3581– 3589
      Google Scholar
      There is no corresponding record for this reference.
    121. 121
      Gao, W.; Coley, C. W. The synthesizability of molecules proposed by generative mod- els. J. Chem. Inf. Model. 2020, 60, 5714– 5723,  DOI: 10.1021/acs.jcim.0c00174
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      121
      The Synthesizability of Molecules Proposed by Generative Models
      Gao, Wenhao; Coley, Connor W.
      Journal of Chemical Information and Modeling (2020), 60 (12), 5714-5723CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      The discovery of functional mols. is an expensive and time-consuming process, exemplified by the rising costs of small mol. therapeutic discovery. One class of techniques of growing interest for early stage drug discovery is de novo mol. generation and optimization, catalyzed by the development of new deep learning approaches. These techniques can suggest novel mol. structures intended to maximize a multiobjective function, e.g., suitability as a therapeutic against a particular target, without relying on brute-force exploration of a chem. space. However, the utility of these approaches is stymied by ignorance of synthesizability. To highlight the severity of this issue, we use a data-driven computer-aided synthesis planning program to quantify how often mols. proposed by state-of-the-art generative models cannot be readily synthesized. Our anal. demonstrates that there are several tasks for which these models generate unrealistic mol. structures despite performing well on popular quant. benchmarks. Synthetic complexity heuristics can successfully bias generation toward synthetically tractable chem. space, although doing so necessarily detracts from the primary objective. This anal. suggests that to improve the utility of these models in real discovery workflows, new algorithm development is warranted.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXmsFKrurw%253D&md5=5f8ff43137b0489f41b5c58518bec270
    122. 122
      Horwood, J.; Noutahi, E. Molecular Design in Synthetically Accessible Chemical Space via Deep Reinforcement Learning. ACS Omega 2020, 5, 32984– 32994,  DOI: 10.1021/acsomega.0c04153
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      122
      Molecular Design in Synthetically Accessible Chemical Space via Deep Reinforcement Learning
      Horwood, Julien; Noutahi, Emmanuel
      ACS Omega (2020), 5 (51), 32984-32994CODEN: ACSODF; ISSN:2470-1343. (American Chemical Society)
      The fundamental goal of generative drug design is to propose optimized mols. that meet predefined activity, selectivity, and pharmacokinetic criteria. Despite recent progress, we argue that existing generative methods are limited in their ability to favorably shift the distributions of mol. properties during optimization. We instead propose a novel Reinforcement Learning framework for mol. design in which an agent learns to directly optimize through a space of synthetically accessible drug-like mols. This becomes possible by defining transitions in our Markov decision process as chem. reactions and allows us to leverage synthetic routes as an inductive bias. We validate our method by demonstrating that it outperforms existing state-of-the-art approaches in the optimization of pharmacol. relevant objectives, while results on multi-objective optimization tasks suggest increased scalability to realistic pharmaceutical design problems.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXisFOmtL3P&md5=1037d12603e7a52332c11e66e8885838
    123. 123
      Gottipati, S. K.; Sattarov, B.; Niu, S.; Pathak, Y.; Wei, H.; Liu, S.; Blackburn, S.; Thomas, K.; Coley, C.; Tang, J. Learning to navigate the synthetically accessible chemical space using reinforcement learning. Int. Conf. Mach. Learn. 2020, 3668– 3679
      Google Scholar
      There is no corresponding record for this reference.
    124. 124
      Bradshaw, J.; Paige, B.; Kusner, M. J.; Segler, M.; Hernández-Lobato, J. M. Barking up the right tree: an approach to search over molecule synthesis DAGs. Adv. Neural Inf. Process. Syst. 2020, 6852– 6866
      Google Scholar
      There is no corresponding record for this reference.
    125. 125
      Imrie, F.; Bradley, A. R.; van der Schaar, M.; Deane, C. M. Deep generative models for 3d linker design. J. Chem. Inf. Model. 2020, 60, 1983– 1995,  DOI: 10.1021/acs.jcim.9b01120
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      125
      Deep Generative Models for 3D Linker Design
      Imrie, Fergus; Bradley, Anthony R.; van der Schaar, Mihaela; Deane, Charlotte M.
      Journal of Chemical Information and Modeling (2020), 60 (4), 1983-1995CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      Rational compd. design remains a challenging problem for both computational methods and medicinal chemists. Computational generative methods have begun to show promising results for the design problem. However, they have not yet used the power of three-dimensional (3D) structural information. We have developed a novel graph-based deep generative model that combines state-of-the-art machine learning techniques with structural knowledge. Our method ("DeLinker") takes two fragments or partial structures and designs a mol. incorporating both. The generation process is protein-context-dependent, utilizing the relative distance and orientation between the partial structures. This 3D information is vital to successful compd. design, and we demonstrate its impact on the generation process and the limitations of omitting such information. In a large-scale evaluation, DeLinker designed 60% more mols. with high 3D similarity to the original mol. than a database baseline. When considering the more relevant problem of longer linkers with at least five atoms, the outperformance increased to 200%. We demonstrate the effectiveness and applicability of this approach on a diverse range of design problems: fragment linking, scaffold hopping, and proteolysis targeting chimera (PROTAC) design. As far as we are aware, this is the first mol. generative model to incorporate 3D structural information directly in the design process. The code is available at https://github.com/oxpig/DeLinker.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXltlenurs%253D&md5=430c929761c5020d1013538e028ba1bc
    126. 126
      Yang, Y.; Zheng, S.; Su, S.; Zhao, C.; Xu, J.; Chen, H. SyntaLinker: automatic fragment linking with deep conditional transformer neural networks. Chem. Sci. 2020, 11, 8312– 8322,  DOI: 10.1039/D0SC03126G
      [Crossref], [PubMed], [CAS], Google Scholar
      126
      SyntaLinker: automatic fragment linking with deep conditional transformer neural networks
      Yang, Yuyao; Zheng, Shuangjia; Su, Shimin; Zhao, Chao; Xu, Jun; Chen, Hongming
      Chemical Science (2020), 11 (31), 8312-8322CODEN: CSHCCN; ISSN:2041-6520. (Royal Society of Chemistry)
      Linking fragments to generate a focused compd. library for a specific drug target is one of the challenges in fragment-based drug design (FBDD). Hereby, we propose a new program named SyntaLinker, which is based on a syntactic pattern recognition approach using deep conditional transformer neural networks. This state-of-the-art transformer can link mol. fragments automatically by learning from the knowledge of structures in medicinal chem. databases (e.g.ChEMBL database). Conventionally, linking mol. fragments was viewed as connecting substructures that were predefined by empirical rules. In SyntaLinker, however, the rules of linking fragments can be learned implicitly from known chem. structures by recognizing syntactic patterns embedded in SMILES notations. With deep conditional transformer neural networks, SyntaLinker can generate mol. structures based on a given pair of fragments and addnl. restrictions. Case studies have demonstrated the advantages and usefulness of SyntaLinker in FBDD.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXhsVagsrvK&md5=403cc7a09f96ecbe27afa3ce6bee51a2
    127. 127
      Tan, X. Automated design and optimization of multitarget schizophrenia drug candidates by deep learning. Eur. J. Med. Chem. 2020, 204, 112572,  DOI: 10.1016/j.ejmech.2020.112572
      [Crossref], [PubMed], [CAS], Google Scholar
      127
      Automated design and optimization of multitarget schizophrenia drug candidates by deep learning
      Tan, Xiaoqin; Jiang, Xiangrui; He, Yang; Zhong, Feisheng; Li, Xutong; Xiong, Zhaoping; Li, Zhaojun; Liu, Xiaohong; Cui, Chen; Zhao, Qingjie; Xie, Yuanchao; Yang, Feipu; Wu, Chunhui; Shen, Jingshan; Zheng, Mingyue; Wang, Zhen; Jiang, Hualiang
      European Journal of Medicinal Chemistry (2020), 204 (), 112572CODEN: EJMCA5; ISSN:0223-5234. (Elsevier Masson SAS)
      Complex neuropsychiatric diseases such as schizophrenia require drugs that can target multiple G protein-coupled receptors (GPCRs) to modulate complex neuropsychiatric functions. Here, we report an automated system comprising a deep recurrent neural network (RNN) and a multitask deep neural network (MTDNN) to design and optimize multitarget antipsychotic drugs. The system has successfully generated novel mol. structures with desired multiple target activities, among which high-ranking compd. 3 was synthesized, and demonstrated potent activities against dopamine D2, serotonin 5-HT1A and 5-HT2A receptors. Hit expansion based on the MTDNN was performed, 6 analogs of compd. 3 were evaluated exptl., among which compd. 8 not only exhibited specific polypharmacol. profiles but also showed antipsychotic effect in animal models with low potential for sedation and catalepsy, highlighting their suitability for further preclin. studies. The approach can be an efficient tool for designing lead compds. with multitarget profiles to achieve the desired efficacy in the treatment of complex neuropsychiatric diseases.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXhsVCksbnI&md5=dd01837fd3bdbb488e8781cea7d5103b
    128. 128
      Yang, Y.; Zhang, R.; Li, Z.; Mei, L.; Wan, S.; Ding, H.; Chen, Z.; Xing, J.; Feng, H.; Han, J.; Jiang, H.; Zheng, M.; Luo, C.; Zhou, B. Discovery of Highly Potent, Selec- tive, and Orally Efficacious p300/CBP Histone Acetyltransferases Inhibitors. J. Med. Chem. 2020, 63, 1337– 1360,  DOI: 10.1021/acs.jmedchem.9b01721
      [ACS Full Text ACS Full Text], [CAS], Google Scholar