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. 2023, 63, 10, 2960-2974
RETURN TO ISSUEPREVMachine Learning and...Machine Learning and Deep LearningNEXT

A Small Step Toward Generalizability: Training a Machine Learning Scoring Function for Structure-Based Virtual Screening

Cite this: J. Chem. Inf. Model. 2023, 63, 10, 2960–2974
Publication Date (Web):May 11, 2023
https://doi.org/10.1021/acs.jcim.3c00322

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

CC-BY 4.0.
  • Open Access

Article Views

3763

Altmetric

-

Citations

-
LEARN ABOUT THESE METRICS
PDF (2 MB)
Get e-Alerts
Supporting Info (1)»
SUBJECTS:

Abstract

High Resolution Image
Download MS PowerPoint Slide

Over the past few years, many machine learning-based scoring functions for predicting the binding of small molecules to proteins have been developed. Their objective is to approximate the distribution which takes two molecules as input and outputs the energy of their interaction. Only a scoring function that accounts for the interatomic interactions involved in binding can accurately predict binding affinity on unseen molecules. However, many scoring functions make predictions based on data set biases rather than an understanding of the physics of binding. These scoring functions perform well when tested on similar targets to those in the training set but fail to generalize to dissimilar targets. To test what a machine learning-based scoring function has learned, input attribution, a technique for learning which features are important to a model when making a prediction on a particular data point, can be applied. If a model successfully learns something beyond data set biases, attribution should give insight into the important binding interactions that are taking place. We built a machine learning-based scoring function that aimed to avoid the influence of bias via thorough train and test data set filtering and show that it achieves comparable performance on the Comparative Assessment of Scoring Functions, 2016 (CASF-2016) benchmark to other leading methods. We then use the CASF-2016 test set to perform attribution and find that the bonds identified as important by PointVS, unlike those extracted from other scoring functions, have a high correlation with those found by a distance-based interaction profiler. We then show that attribution can be used to extract important binding pharmacophores from a given protein target when supplied with a number of bound structures. We use this information to perform fragment elaboration and see improvements in docking scores compared to using structural information from a traditional, data-based approach. This not only provides definitive proof that the scoring function has learned to identify some important binding interactions but also constitutes the first deep learning-based method for extracting structural information from a target for molecule design.

This publication is licensed under

CC-BY 4.0.
  • cc licence
  • by licence

Introduction

ARTICLE SECTIONS
Jump To
The recent explosion of machine learning (ML) across all scientific disciplines has been accompanied by concerns regarding the generalizability of methods used. Various studies have found data leakage to be present in ML applications, resulting in overly optimistic performances being reported for the tools in question. (1−4) The field of drug discovery is by no means exempt from this: models that learn unintended features from training data sets are extremely sensitive to small data set distribution shifts and cannot make reliable predictions on out-of-distribution data points. Practically, this renders them incapable of assisting in the development of drugs for novel targets. (5−7)
Typically led by human experts, drug development is a time-consuming and expensive process, with recent estimates placing the time required to reach clinical trials at 8.3 years, and the median cost at $985 million. (8) Computational techniques offer a promising alternate route to human-led design. One such computational technique is docking. (9,10) Docking algorithms take as input the coordinates of protein and ligand atoms and predict the possible conformations of the protein–ligand complex. There exists a function describing the energy of the complex, and as conformational space is searched, this function is minimized. In the case of a deterministic algorithm with a single starting conformation, a single output pose is generated along with an estimation of the binding affinity; for a probabilistic search, or any search with multiple starting conformations, the result is a list of binding poses. These are ranked in the order of estimated binding affinity. Affinities are calculated using a scoring function, which uses a combination of atomic interactions to approximate the binding energy.
Over the past decade, machine learning models have shown promise in predicting both the pose and energetics of protein–ligand binding. In particular, deep learning models, which can be made to approximate extremely complex functions, (11) have been used to predict binding affinities (12,13) as well as design de novo molecules satisfying a variety of constraints. (14,15) The binding affinity of a given molecule to a protein target is a function of the positions and identities of its atoms, which can be approximated by a neural network in a machine learning-based scoring function (MLBSF). Given the coordinates of a protein–ligand complex (obtained by docking or otherwise), such a function considers all interactions between the two molecules. In this way, the interaction energy between two molecules can be predicted.
However, as described above, there are growing concerns regarding the generalizability of machine learning models. Indeed, recent research into MLBSFs has highlighted their tendency to learn data set biases, rather than physical interactions. (5−7,16) In these cases, the MLBSFs perform less of an assessment of atomic contributions and something more analogous to a nearest-neighbors search in ligand-pocket space. This is useful if the application is to proteins or scaffolds with a large amount of previous binding data available, but not for new targets or ligands, as the true functional relationship governing binding is not learned. This problem is often exacerbated by a lack of data set filtering. One of the most widely used benchmarks for scoring functions, both classical and machine learning-based, is the Comparative Assessment of Scoring Functions, 2016 (CASF-2016). (17) The usual method of training a MLBSF is to train on the PDBBind (18) general set of crystal structures and binding affinities after removing the structures in CASF-2016. As we found when building our scoring function, this approach suffers from information leakage concerning almost every test structure, resulting in an overestimation of the accuracy of MLBSFs.
There is an active area of research surrounding how machine learning models make predictions, (19−23) a technique known as input attribution. This can be applied to MLBSFs, with the idea being that an ideal MLBSF, one which has learned to distinguish atomic interactions, rather than just data set biases, should be able to identify the most important binding interactions taking place in a given bound structure. This should mean that the function predicts well on unseen targets.
In practice, the type of attribution that can be used with an MLBSF depends on its architecture. The two prevalent deep learning architectures used as scoring functions are convolutional neural networks (CNNs) (24−26) and graph neural networks (GNNs). (27−29) Attribution for CNN-MLBSFs is limited, because spatial relationships are lost as information flows deeper into the network. (30) Masking, a method by which a feature is allocated a score equal to the difference between the model’s prediction made when the feature is removed from the input, and the score when it is present, can still be used, but this is a computationally expensive method.
Another method for performing attribution is attention analysis. Attention mechanisms have occasionally been used in graph-based MLBSFs to encourage models to differentiate the contribution of each interaction to binding affinity. (29,31) These work by allocating scores to the graph’s edges, which can be thought of as indicating how much importance is ascribed to each edge by the model. In contrast to CNNs, GNNs preserve the concept of the atom until the final layers of the network, so edges can be interrogated for information. This is helpful in instances where an attention mechanism has been used to weight the edges, as we gain direct insight into how important the network considers each edge to be. However, none of the current leading models for affinity prediction use attention layers for intermolecular interactions, so attribution cannot be carried out on them in this way.
If input attribution could be carried out on MLBSFs, it could be used to extract binding insights from a protein, for example, for fragment-based drug discovery (FBDD); a strategy that seeks to identify simple, small molecules (fragments) that interact with protein targets and grow them into active leads. In order to meaningfully add atoms to these molecules, structural knowledge about the target needs to be considered. This can be extracted from known binders, but this limits the exploration of chemical space as well as restricting the applicability of the approach to targets with known ligands. Extracting structural information directly from the protein itself is therefore advantageous. Currently, there is only one available generative model for fragment elaboration that relies solely on protein structure. (32) This method uses a data-driven approach (33) to identify “hotspots”: regions in the protein pocket that contribute disproportionately to binding. Currently, there exists no deep learning-based tool to identify such regions.
In order to have full control over the architecture and training set of the MLBSF we perform attribution on, we built PointVS, an E(n)-equivariant graph neural network. It is an MLBSF designed to predict both binding affinity and pose score. We assess its performance on the CASF-2016 benchmark and find it achieves comparable results to other leading scoring functions in spite of the additional filtering of the training data set. We then use attribution to show that the model successfully identifies binding interactions in agreement with a distance-based interaction profiler. (34) To further investigate the power of the attribution method, we use it with fragment screen data to obtain information that can be leveraged to perform fragment elaboration. We see that using this information extraction method results in improved docking scores compared to using a data-driven approach. We make our code as well as our unbiased test and train splits available on GitHub at github.com/oxpig/PointVS.

Methods

ARTICLE SECTIONS
Jump To
An overall schema of the methods used to debias and test PointVS is shown in Figure 1.

Figure 1

Figure 1. An overall schema of the methods used to debias and test PointVS. We first thoroughly filter the testing and training sets (a) before benchmarking the performance of PointVS on the docking power and scoring power tests (b). We then use attribution to gain insights into important binding regions in the protein pocket (c), which we use for fragment elaboration (d).

High Resolution Image
Download MS PowerPoint Slide

Building and Testing PointVS

Model

PointVS is a lightweight E(n)-equivariant graph neural network layer model, consisting of an initial projection to take the number of node features from 12 to 32 and then 48 EGNN layers, followed by a global average pooling of the final node embeddings. This is followed by a sigmoid layer which gives a label y ∈ [0, 1] during the first stage of training on pose prediction. During finetuning on affinity data, this final layer is replaced by a randomly initiated fully connected layer and ReLU activation, which outputs y∈(R+)3 (see Figure 2). It also uses a shallow neural network as an attention mechanism, (35) which learns to score network edges, in this case, representing atomic interactions, by their importance.

Figure 2

Figure 2. Architecture and input format of PointVS. Each of the n atoms in the input to the model (left) is given a one-hot encoded feature vector with a single bit added to indicate whether the atom is from the ligand or the receptor, as well as a position p0∈R3. There are m edges, defined by the edge indices ei ∈ {0, ···, n – 1}2×m, which are the indices of connected atoms, and the corresponding edge attributes ea ∈ {0, 1}3×m. These are one-hot encodings representing ligand–ligand, ligand–protein, and protein–protein edges. There are skip connections between each of the EGNN layers, and the linear and global average final pooling (GAP) layers only act upon the node features f.

High Resolution Image
Download MS PowerPoint Slide
Our EGNN takes four inputs: positions, node embeddings, edge indices, and edge embeddings. If there are n atoms in an input structure, the positions tensor is an n × 3 tensor containing the x, y, and z coordinates of each atom. The node embeddings are an n × 12 tensor of one-hot encoded atom types, with a bit to distinguish between ligand and receptor atoms.
Edges are generated on-the-fly using a variable distance cutoff. We use a cutoff of 10 Å for ligand–protein edges and 2 Å for ligand–ligand and protein–protein edges. In this way, intramolecular edges closely mimic the covalent structure of input molecules with a much more expansive connectivity to describe intermolecular interactions. The edge tensor is an 3 × m matrix of m edges with one-hot encoding for ligand–ligand, ligand–receptor, and receptor–receptor interactions.
The binding pocket is defined as the collection of protein atoms within 6 Å of any ligand atom; the rest of the protein is ignored.

Data Sets

A full list of the data sets used is given in Table 1, and more information about each of them is provided in the Supporting Information (SI).
Table 1. Data Sets, Their Sizes, and the Number of Unique PDB Structures They Containa
data setsizeunique PDB IDstask/split
Redocked780,57219,595C/Train
Redocked\gninaSetPose80633,04717,528C/Train
Redocked\Core80629,96316,900C/Train
General19,15719,157R/Train
General\Core8014,55514,555R/Train
gninaSetPose20,411411C/Test
Core285285R/Test
a

Redocked includes no structures with the same PDB code as any structures in the gninaSetPose or Core sets, and the General set shares no PDB codes with the Core set. In the task/split column, “C” refers to pose classification and “R” refers to affinity regression.

Due to the lack of affinity data for complexes with known structures, we choose to pretrain on a pose classification task in order to learn some of the physics of binding before finetuning on affinity data.
Our training data set for the pose classifier, which separates poses into binders and nonbinders, is Redocked2020 (36) (herein Redocked), with the test set, gninaSetPose, generated according to McNutt et al. (24) so that classification performance can be directly compared with their work. This test is similar to the docking power test from CASF-16, (17) which we also conduct.
For regression on binding affinity values, we train a pose classifier on the Redocked set, finetune on affinity data from the PDBBind General Set, and test on the PDBBind Core Set. This replicates the scoring power test from CASF-16. As the Core Set is a subset of the General Set, we remove Core Set structures from the training data.
As mentioned above, a prominent problem in scoring functions is the memorization of data set biases. The effects of such biases can be reduced (although not entirely removed) by reducing data leakage between the train and test data sets. To this end, we prepared filtered subsets of both the Redocked set and the PDBBind v.2020 General Set for both of these stages of training, as well as random subsets of the original training sets of the same size as the filtered sets.
The filtered data sets were constructed by removing proteins and ligands if they met either of the following criteria:
1.

Tanimoto similarity of the 2048-bit Morgan fingerprint between the ligand and any of the 285 test set ligands greater than 0.8.

2.

Sequence identity between the protein and any of the 285 test set proteins greater than 0.8.

These filters were also applied to the Redocked set with respect to the gninaSetPose test set. All files which could not be parsed by either RDKit (37) or OpenBabel (38) were also discarded. This process reduces the training set sizes from 780,572 to 633,047 and 629,963 (Redocked\gninaSetPose80 and Redocked\Core80) and from 19,157 to 14,555 (General\Core80).
At these similarity thresholds, there may still be some similar bound structures in the train and test set. However, there is no cutoff that would ensure no leakage without removing so much data from the training set that the drop in performance would be attributed to data reduction rather than removal of data leakage: for example, moving only the sequence identity threshold to 30%, a much more realistic cutoff for removing bias, would result in the training set containing only 11,418 bound structures.
To ensure that any changes in the performance of PointVS can be attributed to the removal of bias rather than a reduction in the size of the training data set, we constructed training sets of the same size as the debiased ones. These were randomly sampled from the original training sets such that they could still contain bound structures similar to those in the test set. We refer to these data sets as Redocked\gninaSetPoseR, Redocked\CoreR, and General\CoreR.

Performance Metrics

We carried out the CASF-16 docking power and scoring power tests, as well as the pose ranking test as defined by the authors of gnina. (24)
Docking power refers to a scoring function’s ability to identify a ligand’s native binding pose among decoys. It is quantified using Top-N, which is the percentage of systems with a “good” pose ranked in the top N. A pose is considered good if the RMSD to the crystal pose is less than 2 Å. The gnina pose ranking test also assesses docking power but uses the gninaSetPose rather than the CASF-16 test set.
The scoring power is defined by Su et al. as “the ability of a scoring function to produce binding scores in a linear correlation with experimental binding data”. (17) This is best measured using the Pearson correlation coefficient (PCC) between the score given by the scoring function and the measured binding affinity data. For training our networks, we once again used both filtered and unfiltered versions of the PDBBind General data set to test the influence of data set bias.

Attribution

A number of previous studies have identified the tendency of MLBSFs to make predictions based on data set biases rather than an understanding of the physics of binding: data set clustering and cross-validation have been used to demonstrate that many CNN-based MLBSFs, for instance, are prone to distinguishing binders from nonbinders based not on protein–ligand interactions but on ligand features alone. (5−7,39) Ensuring that predictive power is diminished in the absence of receptor information in the test set causes at least some receptor information to be used. (40) However, only directly observing which parts of the input space are used in making predictions can give affirmative proof that machine learning models are learning to identify physical interactions that separate active from inactive poses.
Interpretability in machine learning for pose prediction and virtual screening is an ongoing problem; the architecture of most CNNs causes the concept of individual atoms to be lost as information flows deeper into the network, making them fundamentally unsuitable to attribution at the atomic level. In contrast, graph-based architectures such as PointVS maintain distinct nodes (atoms) until the final layers. These, having passed through the entire network, are rich in information and are directly related to the relevant atom in the input. As the node features are directly attributable to their input atom and the final pose score is a function of the node features, there is a direct relationship between the atoms and the PointVS score, although there is no incentive during training to localize class contributions on particular atoms. The edges between atoms can also be probed for how much importance is ascribed to atomic interactions, as an intuitive way to describe noncovalent bonds. As set out in Figure 1, we can also use this knowledge of importance assigned by PointVS to edges to identify important protein atoms for binding.
To check whether PointVS has learned to identify important binding interactions and, further, whether these can be extracted with attribution, we compare the results of performing attribution on PointVS to those found when using other leading MLBSFs. We use three approaches for attribution: atom masking, bond masking, and edge attention.

Atom Masking

Atom masking is a process by which the importance of different atoms can be ascertained. It is carried out by calculating the difference between the score given when an atom i is removed from the set of input atoms, X, and the score when it is present. As all architectures of MLBSFs take as input atom information, atom masking can be carried out on any MLBSF.

Bond Masking

Similar to atom masking, bond masking is a process by which the importance of different bonds can be ascertained. It is carried out by calculating the difference between the score given when an edge e is removed from the input graph encoding the protein–ligand complex and the score when it is present. While atom masking can be carried out with CNNs and GNNs alike, bond masking can only be carried out with architectures that take as input not only information about the atoms in the complex but also information about the connections between atoms. Graph representations of molecules, which use nodes and node features to describe atoms, and edges and edge features to describe interactions between them, are well-suited to this type of attribution. CNNs, which take only atom information as input, are not.

Edge Attention

GNN-based MLBSFs can be built to include an attention-like mechanism, where the edges are given a score between 0 and 1 by a shallow multilayer perceptron (MLP), which takes as input the edge embeddings themselves. These weights can be thought of as indicating how much each edge is weighted by the network, so attribution can easily be performed by extracting the edge attention weights.
In order to facilitate attribution, PointVS was built to include an attention-like mechanism. After the edges are assigned attention scores, the edge message passed to the ith node at each layer mi is the sum of connecting edge embeddings mij in the neighborhood of the node N(i), weighted by this attention score eij, as in equation 8 in Satorras et al.: (41)
mi=∑j∈N(i)eijmij
(1)
We carry out attribution with PointVS as well as two other MLBSFs for comparison: a convolution-based method, gnina, (24) and another graph-based model, InteractionGraphNet. (29)

gnina

gnina (24,42) is one of the leading tools for predicting protein–ligand binding and a popular convolution-based tool. It consists of an ensemble of CNNs, which take as input a bound structure and output either a pose score or binding affinity.
In recognition of the difficulty associated with performing attribution with convolution-based scoring functions, the authors published a paper describing various methods that can be used to carry out attribution with gnina. (30) The code implementing these methods, gninavis, is currently unavailable, so we instead implemented a standard atomic masking procedure. This was carried out by generating a PDB file for each atom in the bound structure, from which said atom was removed. These were all scored by gnina, and their respective results compared to the score of the complete PDB.

InteractionGraphNet

We also carried out bond masking on InteractionGraphNet (IGN). (29) This uses two graphs, one intramolecular and one intermolecular, to describe the bound structure. The intramolecular graph uses an attention mechanism similar to that of PointVS to ascribe weights to edges, but as this is only applied to edges within the same molecule, these weights cannot be probed for information about binding interactions.
The authors of the paper highlighted that IGN’s architecture was chosen to try and force the model to learn the key features of protein–ligand interactions rather than data set bias. However, a recent paper that performed clustering and cross-validation on an array of MLBSFs (43) showed that, although IGN was one of the best predictors when assessed on a conventional test set, it saw a substantial decrease in performance when cross-validation was carried out. The scoring power PCC (defined above) decreased from 0.80 on the default test set to 0.47 when a pocket similarity-based clustering (a process that ensured that no bound structures with similar pocket structures were in both the train and test sets) was used. This decrease in performance is a sign of IGN making predictions based on data set biases rather than an understanding of the physics of binding.
We performed bond masking attribution on IGN as follows. First, we supplied a bound structure to the model. The two graph types, intramolecular and intermolecular, are then constructed. Edges in the intermolecular are initialized if they are found to connect atoms that are less than 8 Å apart. We then perform masking on any edge that joins atoms at a distance of less than 4 Å (for consistency with PointVS) by generating a new graph with said edge deleted. These are then passed through the network, and their scores compared to the score of the full graph.

Fragment Elaboration

To elaborate a fragment toward more potent, lead-like compounds, information about important areas of the protein pocket being targeted is key. We set up a series of tests to investigate if the attribution scores from PointVS can identify these important sites.
By performing attribution on a crystal structure of a given protein with a small molecule bound, we obtain binding information in the form of an importance score for every protein atom less than 6 Å away from any ligand atom. If a fragment screen has been carried out on said target, the attribution process can then be repeated for several crystal structures to obtain an average importance score for each of the protein atoms. We use this list of protein atoms and associated scores as hotspots to perform fragment elaboration. To assess them, we say that the quality of a hotspot, how important the point it is placed on actually is for binding, is proportional to the binding affinity of the molecules generated in the elaboration process.
We compare the ability of PointVS hotspots to highlight important binding regions to traditional, data-based fragment hotspot maps. To obtain these, we use the Hotspots API, (33) which implements the algorithm described in Radoux et al. (44) We then process the output of the API as in Hadfield et al. (32) A description of this process is given in the SI.
For each target we perform fragment elaboration tests on, we obtain hotspot maps with PointVS using the crystal structures of their fragment screens available on the fragalysis platform. (45) The structures used are listed in SI Table 1. To obtain traditional hotspot maps, we run the Hotspots API on one structure of the target randomly selected from the fragalysis set of bound structures.
We also use the fragalysis set of crystal ligand structures listed in SI Table 1 to obtain fragments to elaborate. To do this, we enumerate all cuts of acyclic single bonds that are not part of functional groups and add a “dummy” atom at the site of the cut. This atom is where atoms will be added to the fragment.
To perform fragment elaboration using the two sets of hotspot maps, we use STRIFE, a generative model for fragment elaboration. We follow the method outlined for its use in Hadfield et al. (32) We take as input a fragment, a specified exit atom, and information about one hotspot, namely, its coordinates and type. STRIFE then generates molecules matching the pharmacophoric profile specified.
To achieve this, STRIFE has two main stages. The first of these, exploration, aims to generate a set of elaborations that contain an acceptor or donor in close proximity to a hotspot. These elaborations are then docked using GOLD’s constrained docking functionality, (10) and those which successfully place a functional group within a certain distance (2 Å for the API hotspots, which are in the protein pocket, and 3 Å for PointVS hotspots, which are on protein atoms) of the hotspot being targeted are selected as “quasi-actives”. In the next stage, refinement, these quasi-actives are used to derive a fine-grained pharmacophoric profile: specifically, they are used to calculate the number of hydrogen bond donors and acceptors present, as well as the number of aromatic groups, and the distance between the exit atom and these groups. STRIFE then generates elaborations using those pharmacophoric profiles and docks them. These docked elaborations constitute its final output.
For a given hotspot, we perform elaboration on a set of fragments. The size of this test set depends on the number of bound structures of ligands available for the target in question (see SI Table 1). For some fragment–hotspot pairings, elaborations cannot be generated. This can be due to the hotspot in question being too close or too far from the exit point on the fragment or at an angle that is unfavorable to elaborate along. In these cases, the generation stage will be unable to produce any quasi-actives; in other words, it will be unable to generate any molecules that, when docked, have a hydrogen bond acceptor or donor within a certain distance from the hotspot. An example of this is shown in Table 5, which shows the number of fragments successfully elaborated in the pocket of Mpro for every hotspot tested. We see that the hotspot ranked ninth by the Hotspots API, for instance, cannot be reached by elaborating on any of the 109 fragments tested, and hence, no molecules are output.
The number of final elaborations generated is dependent on the number of quasi-actives that the first stage, exploration, generates. Although STRIFE is asked to generate 250 elaborations per fragment, there are often fragment–hotspot pairings where the number of quasi-actives identified is insufficient for the refinement phase to produce 250 distinct molecules. The mean number of elaborations generated for a given pair is then less than 250, as seen in Table 5. Nevertheless, as each hotspot is tested on a number of fragments occupying various positions in the pocket, we still obtain a large number of generated molecules for a given hotspot.
To assess the generated molecules, we dock them using GOLD, (10) the CCDC’s protein–ligand docking software. We use the constrained docking functionality, constraining using the fragment, for which we have a crystal structure, from which each molecule receives a score. We then calculate a ligand efficiency by dividing this score by the number of heavy atoms present in the molecule. This is to compensate for the tendency of docking algorithms to favor larger molecules. From the ligand efficiency, we derive a standardized ligand efficiency. This is obtained by standardizing the ligand efficiencies of the generated molecules and the ground truth molecule for a given fragment–hotspot combination to have zero mean and unit variance (here, ground truth molecule refers to the molecule that was segmented to obtain the test fragment). We then sort the standardized ligand efficiencies of the elaborations from highest to lowest scoring and take the mean of the top α. In this work, we use α = 20. Subtracting the ground truth ligand efficiency value from this value then provides us with ΔSLE20 for every fragment–hotspot pair, so for a given hotspot, we take the mean over the ΔSLE20’s of all the fragments successfully elaborated toward it. We use this metric as it provides an insight into how the elaboration process when using a particular hotspot has impacted the original molecule’s docking score: a positive ΔSLEα means the top elaborations are improvements on the original molecule, and a negative ΔSLEα shows that the elaboration process has decreased the ligand efficiency with respect to the starting molecule.

Results

ARTICLE SECTIONS
Jump To
We tested the ability of our MLBSF, PointVS, to perform pose selection and affinity prediction. Through the use of attribution, we also verified not only that PointVS has successfully learned to recognize binding interactions but also that, when given a number of bound structures of a given target, it could be used to extract information about important binding pharmacophores that could then be used in automated fragment elaboration.

Training and Testing PointVS

Bias in CASF-16

We implemented a filtering step to ensure that the training set and test set structures used by PointVS have no overlap (see Methods). In order to compare PointVS with other machine learning methods which have not included this filtering step, (24,46,47) we also constructed a smaller test set out of the PDBBind Core set by applying the same filters, but excluding test set structures rather than train set structures. However, we found that such a test set would only contain a single structure: 284 of 285 of the Core set proteins have a counterpart with 90% sequence similarity in the General set. Further, 273 of them (95.7%) have a counterpart with an identical sequence, making any fair and unbiased comparison to these methods impossible.

Docking and Scoring Power Tests

We carried out the pose ranking test as defined by the authors of gnina (24) as well as the CASF-16 docking power and scoring power tests. The results of these tests are shown in Figure 3 and Tables 2 and 3.

Figure 3

Figure 3. Top-N vs N pose ranking performance on the gninaSetPose set for Autodock Vina, PointVS, and gnina. Terms in brackets in the legend refer to the training sets; filtering the training set by protein and ligand similarity results in slightly degraded performance. The Top-1 values for PointVS trained on Redocked\gninaSetPose80 and gnina trained on Redocked are the same (68%), with PointVS trained on Redocked achieving 70%.

High Resolution Image
Download MS PowerPoint Slide
Table 2. Top-1 Pose Ranking Performance for Different Models and Training Sets (in Brackets) on the PDBBind Core Set (the CASF-16 Docking Power Test Set)a
 Top-1
modelcrystal pose includedcrystal pose not included
PointVS (Redocked\CoreR)90.584.9
PointVS (Redocked\Core80)91.284.2
PointVS (Redocked)91.385.3
gnina (Redocked)91.283.2
Autodock Vina90.284.6
a

The nonmachine learning scoring function is shown in italic. The performance is shown for both the case where the crystal structure of the ligand is included in the set of poses being ranked and the case where it is not included, in which case the native pose is defined as any pose less than 2 Å RMSD away from the crystal pose. The highest Top-1 for each case is shown in bold. Not included in the table are ΔVinaXGB and ΔVinaRF, for which only the performances on the crystal pose included test are provided by the authors as 92 and 90, respectively.

Table 3. Pearson Correlation Coefficients between Measured and Predicted Affinity for Different Scoring Functions on the PDBBind Core Seta
 scoring functionPCC
Biasedgnina (Redocked)0.753 ± 0.008
 gnina (General)0.816 ± 0.008
 PointVS (General)0.805 ± 0.010
 PointVS (General\CoreR)0.803 ± 0.012
 ΔVinaXGB0.796
 ΔVinaRF0.732
DebiasedPointVS (General\Core80)0.754 ± 0.015
 X-Score0.631
 Autodock Vina0.601
a

Scoring functions are split into biased and debiased methods according to the overlap between their training sets, which are shown in brackets where applicable, and the Core set. Nonmachine learning scoring functions are shown in italic. The highest PCCs obtained with both biased and debiased methods are shown in bold.

Top-N pose ranking performance on the gninaSetPose set is shown in Figure 3. When the training sets are the same, PointVS ranks a good pose at the top in 70% of cases, as opposed to 68% with gnina. When PointVS is trained on the smaller Redocked\gninaSetPose80 set, such that no similar proteins or ligands as the test set are seen during training, this drops to 68%. This suggests that, for pose prediction, filtering the data set to remove overlap between train and test structures only causes a minor decrease in docking power.
The results of the CASF-16 docking power test (Table 2) support this conclusion: the difference in performance between all of the methods is minimal and is not affected significantly by whether or not the training data has been filtered for structural similarity.
On the scoring power test, the three best methods by scoring without data set filtering (gnina, PointVS, and ΔVinaXGB) achieve very similar performance, and PointVS outperforms the best nonmachine-learning methods from the original CASF-16 work, (17) Autodock Vina, significantly (0.805 vs 0.601). However, the performance of PointVS when trained on the filtered data set General\Core80 is a better indicator of its true performance on unseen protein–ligand combinations with no analogues in the training data.
To test whether the drop in the performance of PointVS when similar structures are removed from the training set is related to bias or change in training data size, we trained PointVS on General\CoreR. This has the same training set size as General\Core80 but can still contain similar structures to those in the test set. PointVS trained on the General\CoreR achieves a PCC of 0.803 on the CASF-16 set. This is almost identical with the PCC of PointVS trained on the entire General set (PCC of 0.805) and significantly higher than PointVS trained on the General\Core80 set (PCC of 0.754). This points to the performance of MLBSFs when trained and tested using biased data sets being boosted and contrasts with our findings in pose prediction performance, in which we saw minimal change when filtering was performed.
The three machine learning methods, which use three distinct featurization methods and three separate architectures (PointVS, gnina, and ΔVinaXGB), obtain similar PCCs when trained on the unfiltered General set, suggesting that the limiting factor in predictive performance is what information the training data holds about the test set data, rather than the architecture or featurization.
PointVS does not outperform other leading scoring functions: in fact, there exist several scoring functions which report higher performance for the tests carried out. (48−50) However, given our findings surrounding the data leakage present in the CASF-16 tests, we suggest that performances reported for scoring functions trained on the unfiltered data set are overly optimistic. Without the other leading models being retrained on the filtered data set and their performances reevaluated, there is no way to provide a fair and unbiased comparison to them.
Furthermore, although our data set filtering has gone some way to reducing the bias that PointVS learns, it is unlikely that we have removed all dependence on it. Our similarity thresholds (80% sequence identity for proteins and 0.8 Tanimoto similarity for ligands) will still allow very similar bound structures to be included in the train and test data sets. However, as discussed in the Methods, there is no threshold that can be used to filter the PDBBind General set that would result in all data leakage being removed while maintaining a large enough training set for this to not be the limiting factor in performance. It is likely, then, that PointVS has still learned some unintentional biases. Even the performance of the debiased model (particularly on the scoring power test, which appears to be more influenced by bias) is probably overly optimistic.
PointVS appears to perform at a similar level to leading scoring functions even when trained and tested using a less biased setup. In the next sections of the paper, we take advantage of the fact that we can examine its edge attention scores in order to investigate if it has learned about binding interactions.

Attribution: Identifying Important Binding Sites

PointVS appears to offer a substantial improvement in predictive performance of binding affinity over nonmachine-learning methods. However, MLBSFs are notoriously difficult to interpret. Good performance when trained on debiased training data is circumstantial evidence that the interatomic interactions involved in binding are being learned, but demonstrating that the attribution scores of PointVS relate to important binding interactions is a more direct measure. We next present several results suggesting that PointVS is capable of identifying such interactions.

Human Tankyrase-2 Inhibitors

The Protein–Ligand Interaction Profiler (PLIP) (34) uses simple geometric rules to predict interactions at the interface between a protein and a ligand. We used PLIP to identify the “most important” bonds for three human Tankyrase-2 inhibitors and examined whether the atoms forming these bonds were highlighted when attribution was carried out on gnina, (24) IGN, (29) and PointVS.
The structures of three tankyrase inhibitors are shown in Figure 4. The top row (structures a, b, and c) shows the PLIP analysis of each structure. The structures in the other three rows show the results of attribution with gnina (second row, structures d, e, and f), IGN (third row, structures g, h, and i), and PointVS (bottom row, structures j, k, and l). The structures in the second row are colored by the scores obtained from atom masking with gnina, with atoms contributing negatively to binding being shown in red (negative score), neutral atoms in white, and positively contributing atoms in green (positive score). Attribution with IGN and PointVS, which was carried out with bond masking and edge attention analysis, respectively, resulted in scores corresponding to interactions between pairs of atoms rather than individual atoms. The five highest-scoring edges for structures using both of these methods are shown in the bottom two rows, with dark red lines representing high scores from IGN in structures g, h, and i and darker pink representing higher scores from PointVS in structures j, k, and l.

Figure 4

Figure 4. Three Tankyrase-2 inhibitors: 5C5P (a, d, g, j), 4J21 (b, e, h, k), and 4J22 (c, f, i, l). The top row shows the Protein–Ligand Interaction Profiler (PLIP) analysis of each structure, where dark blue lines are hydrogen bonds, dotted gray are hydrophobic interactions, dotted green are π–π interactions, solid green are halogen bonds, and lilac are water bridges. The second row shows the results of performing atom masking with gnina, with green representing a positive attribution score (identified as making a positive contribution to binding) and red, a negative score (making a negative contribution). The third row shows the results of bond masking with IGN, with lines between atoms showing the top five highest-scoring edges for each structure, with darker red representing higher scores. The bottom row shows the results of edge-attention attribution performed with PointVS, with the lines also showing the top five highest-scoring edges, and darker pink representing higher scores.

High Resolution Image
Download MS PowerPoint Slide
Considering the first column, the two ligand atoms that form hydrogen bonds with the protein as calculated by PLIP (Figure 4a) are scored by gnina as the sixth and eighth most important of 20 (Figure 4d). Of the two hydrogen bonding protein atoms, gnina recognizes one of them as being extremely important (the highest scoring protein atom) and the other as neutral. The second structure, Figure 4e, shows a similar story, with the three important ligand atoms as given by PLIP (Figure 4b) scoring fourth, seventh, and eighth out of 20, but the protein atoms they, respectively, bind to scoring negatively, neutrally, and very slightly positively. In the third structure, the three bound ligand atoms as identified by PLIP (Figure 4c) are scored 11th, 13th, and 14th of 30 ligand atoms, with the corresponding protein atoms scoring negatively, weakly positively, and moderately positively (Figure 4f). Interestingly, the last two structures both contain halogen bonds, and in both cases, gnina ranks the ligand atom as important and the protein atom as a negative contribution. Overall, gnina struggles to identify important individual atoms (in both the ligand and protein) for binding. The atoms being scored individually rather than by pairs as with bond masking or edge attention (see below) makes it more difficult to identify the bonds that are being formed: of the bonds identified by PLIP, only in the first structure was there an example where both the protein and ligand atoms were given a high score by gnina, making it the only case where an important bond was identified.
When we performed atom masking on PointVS, we found that the results were similar to those from gnina, in that the atoms contributing to bonds were scored highly but not necessarily enough to recognize their true importance. However, as described below, we found that edge-based attribution methods provided a clearer insight into important interactions.
The first of the results achieved with edge-based attribution methods is the third row of Figure 4, which shows the edges found to be important when carrying out bond masking with IGN. In the first structure, Figure 4g, the bond highlighted as most important connects a ligand atom that PLIP highlights as forming a hydrogen bond (Figure 4a) with a different protein atom. At first glance, it appears that IGN has incorrectly predicted the formation of a hydrogen bond between these atoms, but it should be noted that the atoms that this edge connects are both in aromatic rings that PLIP identifies as being involved in a π-stacking interaction. IGN also correctly identifies a hydrophobic interaction, which is ranked as the second most important edge. In the second structure (Figure 4h), we again see edges connecting rings involved in π-stacking in Figure 4b, this time ranked first and fifth most important. However, none of the other interactions shown to be taking place in Figure 4b, neither the hydrogen nor the halogen bonds, are recognized. This contrasts the third structure, Figure 4i, which shows that IGN has identified the ligand atom that forms a halogen bond as extremely important, with all five of the highest-scoring edges connecting it to various protein atoms. The edge that connects it to the oxygen atom that PLIP identifies as being the other contributor to this bond (Figure 4c) is ranked third most important. In contrast to the first two structures, little importance is ascribed to the edges connecting the aromatic rings involved in π-stacking.
Altogether, the edges identified as important by performing bond masking with IGN show some agreement with PLIP. This is especially true of π-stacking interactions, of which it identifies two of three, and the least true of hydrogen bonds, of which it identifies none. In all three structures, IGN appears to successfully identify one of the interactions taking place (π-stacking in the first two, Figure 4g,h, and halogen bonding in the third, Figure 4i) but fails to identify any beyond those.
Considering the first example in the final row of Figure 4, we see that the bonds ranked first and second most important by PointVS (Figure 4j) correspond to the hydrogen bonds identified by PLIP (Figure 4a). Edges between the two aromatic rings involved in π-stacking are also given high importance scores, being ranked fourth and fifth most important. We see a similar pattern for the second structure, where the atoms involved in hydrogen bonds as defined by PLIP (Figure 4b) are connected by edges assigned high importance by PointVS (ranked first and second most important, Figure 4k). We also again see high scores assigned to the edges connecting the rings involved in π-stacking. PointVS does, however, fail to recognize the edge joining the atoms involved in the halogen bond shown in green in Figure 4b. In the third example structure, PointVS also identifies the hydrogen bonds identified by PLIP (Figure 4c) as being important, again ranking them as the most and second most important edges (Figure 4l). The other edges assigned high importance correspond to hydrophobic interactions.
The results above suggest that of the attribution methods tested, the two edge-based approaches are more effective at identifying important interactions taking place. However, of the two tested, only edge attention analysis with PointVS was able to reliably identify more than one of the bonds that PLIP identified as forming. This highlights that edge attention is a useful method for attribution but, further, that PointVS has gone some way to learning to identify important binding interactions.

Large Scale Attribution Tests

To further verify that performing attribution with PointVS results in similar bonds being highlighted to those specified by PLIP, we performed attribution on the PDBBind Core set (the CASF-16 test set). For each bound structure, we found the top ten highest-scoring protein atoms (corresponding to the protein atoms connected to the ten highest-scoring edges). We then calculated the Spearman’s rank correlation, ρ, between the scores assigned to the top five and top ten scoring protein atoms and the distance between the protein atom and the corresponding ligand atom. This results in a ρ value of 0.719 for the top five protein atoms and a ρ value of 0.788 for the top ten protein atoms. This correlation implies that protein–ligand atom pairs that are close together (which in turn PLIP will consider more likely to be forming a bond) are commonly identified by PointVS. To ensure that this test is an effective way to learn how much a model bases its predictions on an understanding of binding rather than learned biases, we carried out the same process using the biased PointVS model and obtained a ρ value of 0.534 for the top five protein atoms and a ρ value of 0.583 for the top ten protein atoms. This decrease in performance suggests that this test effectively measures the impact bias has had on the predictions made by a model.
We carried out the same process for IGN with bond masking and gnina with atom masking, but as both of these are significantly more computationally challenging than edge attention analysis, we used only 20 randomly selected structures from the Core set (see SI for list of PDB IDs). We also performed attribution with PointVS on this subset for comparison. The results of calculating both the rank correlation for the top five atoms, ρ5, and the top ten protein atoms, ρ10, are shown in Table 4.
Table 4. Mean Rank Correlation Calculated between the Scores of the Top Five (ρ5) and Top Ten (ρ10) Ranked Protein Atoms by PointVS, gnina, and InteractionGraphNet (IGN) and the Distance between Them and the Nearest Polar Ligand Atoma
 ρ5ρ10
PointVS0.6400.788
gnina0.2340.093
IGN0.209–0.071
a

The means were calculated over a subset of 20 randomly selected bound structures from the PDBBind Core set (see SI for PDB IDs).

On this smaller set, we again see a correlation between PointVS scores and distance. In contrast, we see a much weaker correlation when using gnina or IGN (Table 4). To assess the extent to which this can be attributed to the different attribution methods, we also carried out atom and bond masking with PointVS (see SI) and saw a decrease in performance compared to using edge attention.
These results once again suggest that PointVS is capable of identifying bonds that at minimum make sense geometrically, providing further evidence that the edge attribution scores from PointVS are identifying important binding interactions.

Hotspot Identification Using PointVS

Having shown that PointVS is potentially capable of identifying important binding interactions, we next assessed whether it can be used as a method for extracting structural information from a number of bound structures and, hence, used to guide molecule design. To do this, we extracted hotspot maps from PointVS by performing edge attention attribution on fragment screen data and compared the resulting hotspots to those found using a data-driven approach, the Hotspots API. (33)
Our first test used the noncovalent bound structures available for the SARS-CoV-2 main protease (Mpro). The data set available for this target is far more extensive than most targets will benefit from, as it is composed of not only fragment-like molecules but also follow-up compounds. These follow-up compounds, which have been curated with the specific intention of targeting important protein atoms, consist of closely related compounds with the same basic scaffold. Intuitively, the presence of high-quality compounds should result in PointVS being able to extract higher-quality hotspots, as the bound ligands were designed to interact with protein atoms that have been experimentally validated to be important. We used this data set as a starting point to verify that, given a selection of bound structures, PointVS can be used to identify these important protein atoms in a way that is useful for fragment elaboration.
The data was extracted from the fragalysis platform. (51) A full list detailing the structures used is given in SI Table 1. The hotspots identified by the Hotspots API and attribution with PointVS are shown in Figure 5a,b. The hotspots of the two methods will not fall on exactly the same points, as the PointVS hotspots correspond to protein atoms and the API hotspots, to points in the binding pocket. Still, the two methods clearly highlight very separate regions of the binding pocket as important.

Figure 5

Figure 5. Donor and acceptor hotspot maps in the binding pocket of Mpro, colored purple and orange, respectively, and numbered according to their rank. The hotspots on the left (a) were obtained with the Hotspots API, the hotspots in the center (b) were obtained with PointVS by performing attribution on 152 structures of bound fragments, and the hotspots on the right (c) were obtained by extracting the most common protein atoms identified by PLIP as involved in hydrogen bonds with ligand atoms across the 152 bound structures (in this case identified over five times).

High Resolution Image
Download MS PowerPoint Slide

Impact of Fragment Screen Similarity

As mentioned previously, the Mpro data set contains a significant number of closely related follow-up compounds with the same basic scaffold. To ensure that PointVS is providing a novel insight into important binding regions in the pocket of Mpro and not highlighting important protein atoms by simply counting the number of times they are within close proximity of a ligand donor or acceptor atom, we used PLIP, which does use simple geometric rules to predict interactions, to extract the protein atoms that could be forming hydrogen bonds with ligand atoms for every bound structure. We then defined any protein atom that was highlighted more than five times across the 152 bound structures as a PLIP hotspot. This cutoff of five interactions resulted in only four hotspots being extracted. These top four are shown in Figure 5c and correspond to protein atoms that were highlighted by PLIP 146, 40, 9, and 5 times out of the possible 152.
All but one of the PLIP hotspots are within 2 Å of one of the top five PointVS hotspots. However, the top-ranking PLIP hotspot (which corresponds to a protein atom that forms a hydrogen bond in 146 of the 152 bound structures) is ranked second by PointVS, and the second PLIP hotspot (which PLIP identifies as interacting with ligand atoms in 40 structures) is ranked the highest by PointVS. As the first PLIP hotspot is at most the length of one hydrogen bond away from a ligand acceptor in almost every structure, we would expect a geometry-based method to clearly identify it as the top hotspot, as PLIP does. This suggests that PointVS is identifying important protein atoms with a more nuanced approach than simply counting interactions identified with geometric rules. These results do not mean the PointVS hotspot extraction method is immune to being biased when supplied with fragment screens containing clusters of ligand donor and acceptor atoms; however, it is a promising sign that it is not entirely biased by such clusters and that PointVS is able to identify other areas in the pocket as important. This is a useful finding in light of recent research into fragment screen libraries highlighting that even screens designed to be diverse structurally offer a limited exploration of protein pockets. (52)

Dependency on Fragment Screen Size

Given that PointVS identifies sites that are different from a traditional hotspot method and a geometry-based method, we next tested whether the high-scoring sites identified by PointVS are dependent on the size of the input fragment set.
To do this, we performed the same hotspot extraction process as before, but instead of using all the fragment screen data (SI Table 1) at once, we randomly sampled smaller sets from it. We varied the size of the set sampled between 10 and 80. We then took the top five highest-scoring hotspots generated using PointVS with each randomly sampled set of fragment hits. The results of this are shown in Figure 6, where we took samples of size 10, 30, and 80 from the available bound structures (further images using other data set sizes are given in SI Figure 2). This random sampling process was carried out 40 times. We see that when the hotspot maps are extracted from a smaller number of bound structures, a wider range of protein atoms are identified as being important; PointVS is less able to consistently pick out the same five protein atoms every time, and increasing the number of bound structures given to PointVS results in more consistent hotspot maps being generated. Nevertheless, even when using the smallest data sets containing only 10 bound structures, the five hotspots that are identified when using the full fragment screen are the most commonly included, hence their darker color in Figure 6. This suggests that attribution with PointVS could also be a useful tool for less well-studied targets.

Figure 6

Figure 6. Top five scoring hotspot maps in the binding pocket of Mpro found by performing attribution on 40 sets of randomly sampled sets of 10 (a), 30 (b), and 80 (c) bound structures. Both donor and acceptor hotspots are shown in red. The spheres representing the hotspots are transparent and overlaid, so the more opaque a hotspot appears, the more times it was ranked in the top five.

High Resolution Image
Download MS PowerPoint Slide

Fragment Elaboration

Having demonstrated that the PointVS hotspots are consistent among themselves and that they highlight different areas in the Mpro binding site to the API hotspots, we assessed whether they correctly highlight important areas by performing fragment elaboration. If PointVS is successfully identifying important interactions, molecules generated to satisfy the pharmacophoric constraints imposed by PointVS hotspots should be as good, if not better, than molecules generated to satisfy the API hotspots.
We performed attribution on the full set of Mpro bound crystal structures (see SI) to obtain PointVS hotspot maps and used the Hotspots API to obtain a second set of hotspot maps. We used each of the top ten ranked hotspots from each set of maps to elaborate on 109 fragments using each of the top ten ranked hotspots from each set of maps (see Methods). The number of fragments that were successfully elaborated, referring to cases where STRIFE (32) was successfully able to grow toward the hotspot and the placed donor or acceptor atom remained near the hotspot after docking, using each hotspot is shown in Table 5, alongside the mean number of elaborations made per fragment.
Table 5. Number of Fragments That Were Successfully Elaborated on Using PointVS and API Hotspots for Mpro and the Mean Number of Elaborations That Were Generated for Each Fragmenta
 PointVSHotspots API
rankfragments successfully elaboratedelaborations per fragmentfragments successfully elaboratedelaborations per fragment
15721526189
28218257183
35120259207
4141999158
52122328179
66018323162
7462283220
87821654197
92421500
104721828179
a

STRIFE was provided with 109 fragments to elaborate on for each hotspot and asked to generate 250 elaborations for each one.

To assess the binding affinity of the generated elaborations, we docked them, as well as the fragment they were grown from, and used both docking scores to calculate a standardized ligand efficiency score, ΔSLE. This is the difference in ligand efficiency between the starting fragment and the elaborated compound (see Methods). We then ranked the elaborations by these scores and took the average of the top 20 scores to obtain ΔSLE20. To see the performance of a hotspot over all fragments successfully elaborated with it, we took the mean of the ΔSLE20’s obtained.
The highest scoring elaborations by our ΔSLE20 metric were generated using the top five PointVS hotspots (Table 6). Elaborations made using the hotspots ranked 6–10 by PointVS have worse ΔSLE20’s. In contrast, the Hotspots API successfully identified ten high-scoring hotspots (indeed, on average, these top ten hotspots produced elaborations with higher standardized ligand efficiencies than the PointVS hotspots) but was unable to successfully rank them among themselves. The accurate ranking of hotspots is important; as in real-world fragment-to-lead campaigns, only a small number of them can easily be targeted.
Table 6. Mean Standardized Ligand Efficiency of the Top 20 Highest-Scoring Molecules (ΔSLE20) Generated Using Hotspots Ranked 1–5 and 6–10 by PointVS and the Hotspots API for Mproa
methodPointVSAPI
Hotspots 1–5: ΔSLE200.3040.213
Hotspots 6–10: ΔSLE200.1460.269
a

The highest ΔSLE20 for both ranges of hotspots is shown in bold.

To test whether PointVS hotspots are also reliable for other proteins, we repeated the above process for two other targets: SARS-CoV-2 nonstructural protein 3 (Mac1) (53) and SARS-CoV-2 nonstructural protein 14 (NSP14). (54) These both have fragment screen data sets available on fragalysis (55,56) (see SI Table 1). In both cases, these are considerably smaller than the Mpro data set and do not contain follow-up compounds.
Each of the top ten hotspots produced by PointVS and the API were tested, the results of which are shown in Table 7. Both sets of hotspots for Mac1 and NSP14 were tested on 35 and 30 fragments, respectively. The number of successfully elaborated fragments and mean number of elaborations generated per fragment for these targets are provided in SI Table 2.
Table 7. Mean Standardized Ligand Efficiency of the Top 20 Elaborations (ΔSLE20) Made Using PointVS Hotspots and API Hotspots on Three Different Targets: Mpro, Mac1, and NSP14a
 Mpro: 152Mac1: 58NSP14: 19
ΔSLE20PointVSAPIPointVSAPIPointVSAPI
Hotspots 1–50.3410.2131.7491.5811.7081.240
Hotspots 6–100.1890.2691.7271.7481.1431.193
a

The numbers after the target names refer to the number of bound fragment structures available for each target. The list of fragalysis codes corresponding to the structures used is given in SI Table 1. The highest ΔSLE20 for both ranges of hotspots is shown in bold.

On average, the standardized ligand efficiency of molecules generated with PointVS hotspot maps is greater than those generated with the API hotspots. Within the PointVS scores, we see that the Mpro hotspots are lower scoring than those for the other two targets. This can be attributed to the 109 starting fragments for Mpro being of higher quality, having been extracted from curated follow-up compounds rather than fragments. The improvement that is made on them by elaborating is thus less significant. We also see that, unlike with the Hotspots API, the hotspots ranked 1–5 by PointVS consistently outperform those ranked 6–10, suggesting that the attribution method is good at not only picking out important atoms but also ranking them among themselves.

Conclusion

ARTICLE SECTIONS
Jump To
In this paper, we described the development of PointVS, an EGNN-based method for affinity prediction, and the first ML-based method for extracting important binding information from a target for molecule design.
During the development of PointVS, we identified that a commonly used benchmark for MLBSFs, CASF-16, overestimates their accuracy when trained using the most commonly used training data set. In light of this, PointVS was trained and tested on a filtered data set and, hence, encouraged to learn the rules governing intermolecular binding rather than memorize training data. We showed that, when trained on this filtered data set, PointVS achieves comparable results to other leading scoring functions, providing some evidence that it is successfully identifying important binding interactions. We provide further proof of this by performing attribution and showing that PointVS is able to identify important interactions in line with those found by PLIP, a distance-based tool for profiling protein–ligand interactions.
Finally, we investigated how knowledge of important binding regions could be leveraged in other stages of the drug development pipeline; namely, fragment elaboration. We show that, when provided with a set of bound structures, performing attribution on such structures yielded hotspot maps describing important protein atoms. Using these in unison with a fragment elaboration tool, STRIFE, resulted in improved docking scores for the elaborated molecules compared to when hotspots were obtained with a data-based method. This is further evidence that PointVS is not learning to memorize ligand information, but is instead able to recognize important interactions. Further, this constitutes the first ML-based method for extracting structural information from a protein target in a way that is useful for fragment elaboration. More broadly, our work demonstrates that attribution techniques, when applied to debiased models, can be useful for extracting structural information in a way that can be useful in molecule generation.

Data and Software Availability

ARTICLE SECTIONS
Jump To
PointVS is available to download at https://github.com/oxpig/PointVS, as are our unbiased test and train splits.

Supporting Information

ARTICLE SECTIONS
Jump To

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.jcim.3c00322.

  • A brief description of equivariance and invariance, some more in-depth information about the architecture of PointVS and the data sets used to train and test it, a comparison of the different attribution methods used, a description of the process by which the API hotspots were obtained, an additional assessment of the reliance of PointVS hotspots on fragment screen size, and a case study showing how PointVS could have aided in a real-world fragment-to-lead campaign; also all of the IDs of the structures used, from both fragalysis and the PDB (PDF)

Terms & Conditions

Most electronic Supporting Information files are available without a subscription to ACS Web Editions. Such files may be downloaded by article for research use (if there is a public use license linked to the relevant article, that license may permit other uses). Permission may be obtained from ACS for other uses through requests via the RightsLink permission system: http://pubs.acs.org/page/copyright/permissions.html.

Author Information

ARTICLE SECTIONS
Jump To
  • Corresponding Author
  • Authors
  • Author Contributions

    J.S. and L.V. contributed equally. J.S. and C.M.D. conceptualized and designed the study. All authors contributed to the methodology of the study. J.S. implemented the method. J.S. and L.V. conducted the experiments and analyzed the results. N.B., V.C., P.D., F.D., and C.M.D. supervised the project. All authors contributed to the writing, review, and editing of the manuscript.

  • Notes
    The authors declare no competing financial interest.

Acknowledgments

ARTICLE SECTIONS
Jump To

J.S. was supported by funding from the Biotechnology and Biosciences Research Council (BB/S507611/1) and BenevolentAI, and L.V. was supported by funding from the Engineering and Physical Sciences Research Council (EP/W522211/1) and IBM Research. This work was additionally supported by the Rosetrees Trust (ref M940).

References

ARTICLE SECTIONS
Jump To

This article references 56 other publications.

  1. 1
    Barnett, E.; Onete, D.; Salekin, A.; Faraone, S. V. Genomic Machine Learning Meta-regression: Insights on Associations of Study Features with Reported Model Performance. medRxiv 2022,  DOI: 10.1101/2022.01.10.22268751
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  2. 2
    Shim, M.; Lee, S.-H.; Hwang, H.-J. Inflated prediction accuracy of neuropsychiatric biomarkers caused by data leakage in feature selection. Sci. Rep. 2021, 11, 7980,  DOI: 10.1038/s41598-021-87157-3
    [Crossref], [PubMed], [CAS], Google Scholar
    2
    Inflated prediction accuracy of neuropsychiatric biomarkers caused by data leakage in feature selection
    Shim, Miseon; Lee, Seung-Hwan; Hwang, Han-Jeong
    Scientific Reports (2021), 11 (1), 7980CODEN: SRCEC3; ISSN:2045-2322. (Nature Research)
    In recent years, machine learning techniques have been frequently applied to uncovering neuropsychiatric biomarkers with the aim of accurately diagnosing neuropsychiatric diseases and predicting treatment prognosis. However, many studies did not perform cross validation (CV) when using machine learning techniques, or others performed CV in an incorrect manner, leading to significantly biased results due to overfitting problem. The aim of this study is to investigate the impact of CV on the prediction performance of neuropsychiatric biomarkers, in particular, for feature selection performed with high-dimensional features. To this end, we evaluated prediction performances using both simulation data and actual electroencephalog. (EEG) data. The overall prediction accuracies of the feature selection method performed outside of CV were considerably higher than those of the feature selection method performed within CV for both the simulation and actual EEG data. The differences between the prediction accuracies of the two feature selection approaches can be thought of as the amt. of overfitting due to selection bias. Our results indicate the importance of correctly using CV to avoid biased results of prediction performance of neuropsychiatric biomarkers.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3MXptFOlsLg%253D&md5=a43ca92128b177b8a67e2437fbe5c556
  3. 3
    Roberts, M. Common pitfalls and recommendations for using machine learning to detect and prognosticate for COVID-19 using chest radiographs and CT scans. Nat. Mach. Intell. 2021, 3, 199– 217,  DOI: 10.1038/s42256-021-00307-0
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  4. 4
    Tu, F.; Zhu, J.; Zheng, Q.; Zhou, M. Be Careful of When: An Empirical Study on Time-Related Misuse of Issue Tracking Data. In Proceedings of the 2018 26th ACM Joint Meeting on European Software Engineering Conference and Symposium on the Foundations of Software Engineering , New York, NY, USA, 2018; p 307– 318.
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  5. 5
    Sieg, J.; Flachsenberg, F.; Rarey, M. In Need of Bias Control: Evaluating Chemical Data for Machine Learning in Structure-Based Virtual Screening. J. Chem. Inf. Model. 2019, 59, 947– 961,  DOI: 10.1021/acs.jcim.8b00712
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    5
    In Need of Bias Control: Evaluating Chemical Data for Machine Learning in Structure-Based Virtual Screening
    Sieg, Jochen; Flachsenberg, Florian; Rarey, Matthias
    Journal of Chemical Information and Modeling (2019), 59 (3), 947-961CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    A review. Reports of successful applications of machine learning (ML) methods in structure-based virtual screening (SBVS) are increasing. ML methods such as convolutional neural networks show promising results and often outperform traditional methods such as empirical scoring functions in retrospective validation. However, trained ML models are often treated as black boxes and are not straightforwardly interpretable. In most cases, it is unknown which features in the data are decisive and whether a model's predictions are right for the right reason. Hence, the authors reevaluated three widely used benchmark data sets in the context of ML methods and came to the conclusion that not every benchmark data set is suitable. Moreover, the authors demonstrate on two examples from current literature that bias is learned implicitly and unnoticed from std. benchmarks. On the basis of these results, the authors conclude that there is a need for eligible validation expts. and benchmark data sets suited to ML for more bias-controlled validation in ML-based SBVS. Therefore, the authors provide guidelines for setting up validation expts. and give a perspective on how new data sets could be generated.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXktVOgsrY%253D&md5=6cb48f2acd541c10c8565f202a6eb9e8
  6. 6
    Chen, L.; Cruz, A.; Ramsey, S.; Dickson, C. J.; Duca, J. S.; Hornak, V.; Koes, D. R.; Kurtzman, T. Hidden bias in the DUD-E dataset leads to misleading performance of deep learning in structure-based virtual screening. PLoS One 2019, 14, e0220113  DOI: 10.1371/journal.pone.0220113
    [Crossref], [PubMed], [CAS], Google Scholar
    6
    Hidden bias in the DUD-E dataset leads to misleading performance of deep learning in structure-based virtual screening
    Chen, Lieyang; Cruz, Anthony; Ramsey, Steven; Dickson, Callum J.; Duca, Jose S.; Hornak, Viktor; Koes, David R.; Kurtzman, Tom
    PLoS One (2019), 14 (8), e0220113CODEN: POLNCL; ISSN:1932-6203. (Public Library of Science)
    Recently much effort has been invested in using convolutional neural network (CNN) models trained on 3D structural images of protein-ligand complexes to distinguish binding from non-binding ligands for virtual screening. However, the dearth of reliable protein-ligand x-ray structures and binding affinity data has required the use of constructed datasets for the training and evaluation of CNN mol. recognition models. Here, we outline various sources of bias in one such widely-used dataset, the Directory of Useful Decoys: Enhanced (DUD-E). We have constructed and performed tests to investigate whether CNN models developed using DUD-E are properly learning the underlying physics of mol. recognition, as intended, or are instead learning biases inherent in the dataset itself. We find that superior enrichment efficiency in CNN models can be attributed to the analog and decoy bias hidden in the DUD-E dataset rather than successful generalization of the pattern of protein-ligand interactions. Comparing addnl. deep learning models trained on PDBbind datasets, we found that their enrichment performances using DUD-E are not superior to the performance of the docking program AutoDock Vina. Together, these results suggest that biases that could be present in constructed datasets should be thoroughly evaluated before applying them to machine learning based methodol. development.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXhvFeks7rL&md5=637472aa275cbacf3e926983f3795726
  7. 7
    Wallach, I.; Heifets, A. Most Ligand-Based Classification Benchmarks Reward Memorization Rather than Generalization. J. Chem. Inf. Model. 2018, 58, 916– 932,  DOI: 10.1021/acs.jcim.7b00403
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    7
    Most Ligand-Based Classification Benchmarks Reward Memorization Rather than Generalization
    Wallach, Izhar; Heifets, Abraham
    Journal of Chemical Information and Modeling (2018), 58 (5), 916-932CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    Undetected overfitting can occur when there are significant redundancies between training and validation data. The authors describe AVE, a new measure of training-validation redundancy for ligand-based classification problems, that accounts for the similarity among inactive mols. as well as active ones. The authors investigated seven widely used benchmarks for virtual screening and classification, and the authors show that the amt. of AVE bias strongly correlates with the performance of ligand-based predictive methods irresp. of the predicted property, chem. fingerprint, similarity measure, or previously applied unbiasing techniques. Therefore, it may be the case that the previously reported performance of most ligand-based methods can be explained by overfitting to benchmarks rather than good prospective accuracy.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXotlCgs7s%253D&md5=a59fb9ba04b1e6b37cb83594625081ff
  8. 8
    Wouters, O. J.; McKee, M.; Luyten, J. Estimated Research and Development Investment Needed to Bring a New Medicine to Market, 2009–2018. JAMA, J. Am. Med. Assoc. 2020, 323, 844– 853,  DOI: 10.1001/jama.2020.1166
    [Crossref], [PubMed], [CAS], Google Scholar
    8
    Estimated Research and Development Investment Needed to Bring a New Medicine to Market, 2009-2018
    Wouters Olivier J; McKee Martin; Luyten Jeroen
    JAMA (2020), 323 (9), 844-853 ISSN:.
    IMPORTANCE: The mean cost of developing a new drug has been the subject of debate, with recent estimates ranging from $314 million to $2.8 billion. OBJECTIVE: To estimate the research and development investment required to bring a new therapeutic agent to market, using publicly available data. DESIGN AND SETTING: Data were analyzed on new therapeutic agents approved by the US Food and Drug Administration (FDA) between 2009 and 2018 to estimate the research and development expenditure required to bring a new medicine to market. Data were accessed from the US Securities and Exchange Commission, Drugs@FDA database, and ClinicalTrials.gov, alongside published data on clinical trial success rates. EXPOSURES: Conduct of preclinical and clinical studies of new therapeutic agents. MAIN OUTCOMES AND MEASURES: Median and mean research and development spending on new therapeutic agents approved by the FDA, capitalized at a real cost of capital rate (the required rate of return for an investor) of 10.5% per year, with bootstrapped CIs. All amounts were reported in 2018 US dollars. RESULTS: The FDA approved 355 new drugs and biologics over the study period. Research and development expenditures were available for 63 (18%) products, developed by 47 different companies. After accounting for the costs of failed trials, the median capitalized research and development investment to bring a new drug to market was estimated at $985.3 million (95% CI, $683.6 million-$1228.9 million), and the mean investment was estimated at $1335.9 million (95% CI, $1042.5 million-$1637.5 million) in the base case analysis. Median estimates by therapeutic area (for areas with ≥5 drugs) ranged from $765.9 million (95% CI, $323.0 million-$1473.5 million) for nervous system agents to $2771.6 million (95% CI, $2051.8 million-$5366.2 million) for antineoplastic and immunomodulating agents. Data were mainly accessible for smaller firms, orphan drugs, products in certain therapeutic areas, first-in-class drugs, therapeutic agents that received accelerated approval, and products approved between 2014 and 2018. Results varied in sensitivity analyses using different estimates of clinical trial success rates, preclinical expenditures, and cost of capital. CONCLUSIONS AND RELEVANCE: This study provides an estimate of research and development costs for new therapeutic agents based on publicly available data. Differences from previous studies may reflect the spectrum of products analyzed, the restricted availability of data in the public domain, and differences in underlying assumptions in the cost calculations.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A280%3ADC%252BB387oslyhsA%253D%253D&md5=ff63c2f03e59ef0223585a71fb6d5df0
  9. 9
    Morris, G. M.; Huey, R.; Lindstrom, W.; Sanner, M. F.; Belew, R. K.; Goodsell, D. S.; Olson, A. J. AutoDock4 and AutoDockTools4: Automated docking with selective receptor flexibility. J. Comput. Chem. 2009, 30, 2785– 2791,  DOI: 10.1002/jcc.21256
    [Crossref], [PubMed], [CAS], Google Scholar
    9
    AutoDock and AutoDockTools: Automated docking with selective receptor flexibility
    Morris, Garrett M.; Huey, Ruth; Lindstrom, William; Sanner, Michel F.; Belew, Richard K.; Goodsell, David S.; Olson, Arthur J.
    Journal of Computational Chemistry (2009), 30 (16), 2785-2791CODEN: JCCHDD; ISSN:0192-8651. (John Wiley & Sons, Inc.)
    We describe the testing and release of AutoDock4 and the accompanying graphical user interface AutoDockTools. AutoDock4 incorporates limited flexibility in the receptor. Several tests are reported here, including a redocking expt. with 188 diverse ligand-protein complexes and a cross-docking expt. using flexible sidechains in 87 HIV protease complexes. We also report its utility in anal. of covalently bound ligands, using both a grid-based docking method and a modification of the flexible sidechain technique. © 2009 Wiley Periodicals, Inc. J Comput Chem, 2009.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BD1MXht1GitrnK&md5=679ce22fc50e9291c9aa16e7a1855845
  10. 10
    Verdonk, M. L.; Cole, J. C.; Hartshorn, M. J.; Murray, C. W.; Taylor, R. D. Improved protein–ligand docking using GOLD. Proteins: Struct., Funct., Bioinf. 2003, 52, 609– 623,  DOI: 10.1002/prot.10465
    [Crossref], [PubMed], [CAS], Google Scholar
    10
    Improved protein-ligand docking using GOLD
    Verdonk, Marcel L.; Cole, Jason C.; Hartshorn, Michael J.; Murray, Christopher W.; Taylor, Richard D.
    Proteins: Structure, Function, and Genetics (2003), 52 (4), 609-623CODEN: PSFGEY; ISSN:0887-3585. (Wiley-Liss, Inc.)
    The Chemscore function was implemented as a scoring function for the protein-ligand docking program GOLD, and its performance compared to the original Goldscore function and two consensus docking protocols, "Goldscore-CS" and "Chemscore-GS," in terms of docking accuracy, prediction of binding affinities, and speed. In the "Goldscore-CS" protocol, dockings produced with the Goldscore function are scored and ranked with the Chemscore function; in the "Chemscore-GS" protocol, dockings produced with the Chemscore function are scored and ranked with the Goldscore function. Comparisons were made for a "clean" set of 224 protein-ligand complexes, and for two subsets of this set, one for which the ligands are "drug-like," the other for which they are "fragment-like.". For "drug-like" and "fragment-like" ligands, the docking accuracies obtained with Chemscore and Goldscore functions are similar. For larger ligands, Goldscore gives superior results. Docking with the Chemscore function is up to three times faster than docking with the Goldscore function. Both combined docking protocols give significant improvements in docking accuracy over the use of the Goldscore or Chemscore function alone. "Goldscore-CS" gives success rates of up to 81% (top-ranked GOLD soln. within 2.0 Å of the exptl. binding mode) for the "clean list," but at the cost of long search times. For most virtual screening applications, "Chemscore-GS" seems optimal; search settings that give docking speeds of around 0.25-1.3 min/compd. have success rates of about 78% for "drug-like" compds. and 85% for "fragment-like" compds. In terms of producing binding energy ests., the Goldscore function appears to perform better than the Chemscore function and the two consensus protocols, particularly for faster search settings. Even at docking speeds of around 1-2 min/compd., the Goldscore function predicts binding energies with a std. deviation of ∼10.5 kJ/mol.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BD3sXmvFGrsLg%253D&md5=73a627d99dab6de1ae364c8e8e5f0fea
  11. 11
    Hornik, K.; Stinchcombe, M.; White, H. Multilayer feedforward networks are universal approximators. Neural Netw 1989, 2, 359– 366,  DOI: 10.1016/0893-6080(89)90020-8
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  12. 12
    Shen, C.; Ding, J.; Wang, Z.; Cao, D.; Ding, X.; Hou, T. From machine learning to deep learning: Advances in scoring functions for protein–ligand docking. Wiley Interdiscip. Rev.: Comput. Mol. Sci. 2020, 10, e1429  DOI: 10.1002/wcms.1429
    [Crossref], [CAS], Google Scholar
    12
    From machine learning to deep learning: Advances in scoring functions for protein-ligand docking
    Shen, Chao; Ding, Junjie; Wang, Zhe; Cao, Dongsheng; Ding, Xiaoqin; Hou, Tingjun
    Wiley Interdisciplinary Reviews: Computational Molecular Science (2020), 10 (1), e1429CODEN: WIRCAH; ISSN:1759-0884. (Wiley-Blackwell)
    A review. Mol. docking has been regarded as a routine tool for drug discovery, but its accuracy highly depends on the reliability of scoring functions (SFs). With the rapid development of machine learning (ML) techniques, ML-based SFs have gradually emerged as a promising alternative for protein-ligand binding affinity prediction and virtual screening, and most of them have shown significantly better performance than a wide range of classical SFs. Emergence of more data-hungry deep learning (DL) approaches in recent years further fascinates the exploitation of more accurate SFs. Here, we summarize the progress of traditional ML-based SFs in the last few years and provide insights into recently developed DL-based SFs. We believe that the continuous improvement in ML-based SFs can surely guide the early-stage drug design and accelerate the discovery of new drugs. This article is categorized under: Computer and Information Science > Chemoinformatics.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXitlyjsbbN&md5=0dd878e211fad64fada03e883a5ab47a
  13. 13
    Jones, D.; Kim, H.; Zhang, X.; Zemla, A.; Stevenson, G.; Bennett, W. F. D.; Kirshner, D.; Wong, S. E.; Lightstone, F. C.; Allen, J. E. Improved Protein–Ligand Binding Affinity Prediction with Structure-Based Deep Fusion Inference. J. Chem. Inf. Model. 2021, 61, 1583– 1592,  DOI: 10.1021/acs.jcim.0c01306
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    13
    Improved Protein-Ligand Binding Affinity Prediction with Structure-Based Deep Fusion Inference
    Jones, Derek; Kim, Hyojin; Zhang, Xiaohua; Zemla, Adam; Stevenson, Garrett; Bennett, W. F. Drew; Kirshner, Daniel; Wong, Sergio E.; Lightstone, Felice C.; Allen, Jonathan E.
    Journal of Chemical Information and Modeling (2021), 61 (4), 1583-1592CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    Predicting accurate protein-ligand binding affinities is an important task in drug discovery but remains a challenge even with computationally expensive biophysics-based energy scoring methods and state-of-the-art deep learning approaches. Despite the recent advances in the application of deep convolutional and graph neural network-based approaches, it remains unclear what the relative advantages of each approach are and how they compare with physics-based methodologies that have found more mainstream success in virtual screening pipelines. We present fusion models that combine features and inference from complementary representations to improve binding affinity prediction. This, to our knowledge, is the first comprehensive study that uses a common series of evaluations to directly compare the performance of three-dimensional (3D)-convolutional neural networks (3D-CNNs), spatial graph neural networks (SG-CNNs), and their fusion. We use temporal and structure-based splits to assess performance on novel protein targets. To test the practical applicability of our models, we examine their performance in cases that assume that the crystal structure is not available. In these cases, binding free energies are predicted using docking pose coordinates as the inputs to each model. In addn., we compare these deep learning approaches to predictions based on docking scores and mol. mechanic/generalized Born surface area (MM/GBSA) calcns. Our results show that the fusion models make more accurate predictions than their constituent neural network models as well as docking scoring and MM/GBSA rescoring, with the benefit of greater computational efficiency than the MM/GBSA method. Finally, we provide the code to reproduce our results and the parameter files of the trained models used in this work. The software is available as open source at https://github.com/llnl/fast. Model parameter files are available at ftp://gdo-bioinformatics.ucllnl.org/fast/pdbbind2016_model_checkpoints/.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3MXmvFaqtrc%253D&md5=11f92ba4bae7dabc75b9e55c1911352b
  14. 14
    Chenthamarakshan, V.; Das, P.; Hoffman, S.; Strobelt, H.; Padhi, I.; Lim, K. W.; Hoover, B.; Manica, M.; Born, J.; Laino, T.; Mojsilovic, A. CogMol: Target-Specific and Selective Drug Design for COVID-19 Using Deep Generative Models. In 34th Conference on Neural Information Processing Systems , 2020; pp 4320– 4332.
    Google Scholar
    There is no corresponding record for this reference.
  15. 15
    Hoffman, S. C.; Chenthamarakshan, V.; Wadhawan, K.; Chen, P.-Y.; Das, P. Optimizing molecules using efficient queries from property evaluations. Nat. Mach. Intell. 2022, 4, 21,  DOI: 10.1038/s42256-021-00422-y
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  16. 16
    Boyles, F.; Deane, C. M.; Morris, G. M. Learning from Docked Ligands: Ligand-Based Features Rescue Structure-Based Scoring Functions When Trained on Docked Poses. J. Chem. Inf. Model. 2022, 62, 5329– 5341,  DOI: 10.1021/acs.jcim.1c00096
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    16
    Learning from Docked Ligands: Ligand-Based Features Rescue Structure-Based Scoring Functions When Trained on Docked Poses
    Boyles, Fergus; Deane, Charlotte M.; Morris, Garrett M.
    Journal of Chemical Information and Modeling (2022), 62 (22), 5329-5341CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    Machine learning scoring functions for protein-ligand binding affinity have been found to consistently outperform classical scoring functions when trained and tested on crystal structures of bound protein-ligand complexes. However, it is less clear how these methods perform when applied to docked poses of complexes. We explore how the use of docked rather than crystallog. poses for both training and testing affects the performance of machine learning scoring functions. Using the PDBbind Core Sets as benchmarks, we show that the performance of a structure-based machine learning scoring function trained and tested on docked poses is lower than that of the same scoring function trained and tested on crystallog. poses. We construct a hybrid scoring function by combining both structure-based and ligand-based features, and show that its ability to predict binding affinity using docked poses is comparable to that of purely structure-based scoring functions trained and tested on crystal poses. We also present a new, freely available validation set-the Updated DUD-E Diverse Subset-for binding affinity prediction using data from DUD-E and ChEMBL. Despite strong performance on docked poses of the PDBbind Core Sets, we find that our hybrid scoring function sometimes generalizes poorly to a protein target not represented in the training set, demonstrating the need for improved scoring functions and addnl. validation benchmarks.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3MXhvFWms7rN&md5=2c539500f49ad27ff6682ff11e9360eb
  17. 17
    Su, M.; Yang, Q.; Du, Y.; Feng, G.; Liu, Z.; Li, Y.; Wang, R. Comparative Assessment of Scoring Functions: The CASF-2016 Update. J. Chem. Inf. Model. 2019, 59, 895– 913,  DOI: 10.1021/acs.jcim.8b00545
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    17
    Comparative Assessment of Scoring Functions: The CASF-2016 Update
    Su, Minyi; Yang, Qifan; Du, Yu; Feng, Guoqin; Liu, Zhihai; Li, Yan; Wang, Renxiao
    Journal of Chemical Information and Modeling (2019), 59 (2), 895-913CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    In structure-based drug design, scoring functions are often employed to evaluate protein-ligand interactions. A variety of scoring functions have been developed so far, and thus, some objective benchmarks are desired for assessing their strength and weakness. The comparative assessment of scoring functions (CASF) benchmark developed by us provides an answer to this demand. CASF is designed as a "scoring benchmark", where the scoring process is decoupled from the docking process to depict the performance of scoring function more precisely. Here, we describe the latest update of this benchmark, i.e., CASF-2016. Each scoring function is still evaluated by four metrics, including "scoring power", "ranking power", "docking power", and "screening power". Nevertheless, the evaluation methods have been improved considerably in several aspects. A new test set is compiled, which consists of 285 protein-ligand complexes with high-quality crystal structures and reliable binding consts. A panel of 25 scoring functions are tested on CASF-2016 as a demonstration. Our results reveal that the performance of current scoring functions is more promising in terms of docking power than scoring, ranking, and screening power. Scoring power is somewhat correlated with ranking power, so are docking power and screening power. The results obtained on CASF-2016 may provide valuable guidance for the end users to make smart choices among available scoring functions. Moreover, CASF is created as an open-access benchmark so that other researchers can utilize it to test a wider range of scoring functions. The complete CASF-2016 benchmark will be released on the PDBbind-CN web server (http://www.pdbbind-cn.org/casf.asp/) once this article is published.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXitlWhtLjM&md5=573295cfd2298a2dee933027041c3d9c
  18. 18
    Liu, Z.; Su, M.; Han, L.; Liu, J.; Yang, Q.; Li, Y.; Wang, R. Forging the Basis for Developing Protein–Ligand Interaction Scoring Functions. Acc. Chem. Res. 2017, 50, 302– 309,  DOI: 10.1021/acs.accounts.6b00491
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    18
    Forging the Basis for Developing Protein-Ligand Interaction Scoring Functions
    Liu, Zhihai; Su, Minyi; Han, Li; Liu, Jie; Yang, Qifan; Li, Yan; Wang, Renxiao
    Accounts of Chemical Research (2017), 50 (2), 302-309CODEN: ACHRE4; ISSN:0001-4842. (American Chemical Society)
    A review. In structure-based drug design, scoring functions are widely used for fast evaluation of protein-ligand interactions. They are often applied in combination with mol. docking and de novo design methods. Since the early 1990s, a whole spectrum of protein-ligand interaction scoring functions have been developed. Regardless of their tech. difference, scoring functions all need data sets combining protein-ligand complex structures and binding affinity data for parametrization and validation. However, data sets of this kind used to be rather limited in terms of size and quality. However, std. metrics for evaluating scoring function used to be ambiguous. Scoring functions are often tested in mol. docking or even virtual screening trials, which are not totally appropriate for reflecting the genuine quality of scoring functions. Collectively, these underlying obstacles have impeded the invention of more advanced scoring functions. In this Account, the authors describe their long-lasting efforts on overcoming these obstacles, which involve two related projects. On the first project, the authors have created the PDBbind database. It is the first database that systematically annotates the protein-ligand complexes in the Protein Data Bank (PDB) with exptl. binding data. This database has been updated annually since its first public release in 2004. The latest release (version 2016) provides binding data for 16,179 biomol. complexes in PDB. Data sets provided by PDBbind have been applied to many computational and statistical studies on protein-ligand interaction and various subjects. In particular, it has become a major data resource for scoring function development. On the second project, the authors established the Comparative Assessment of Scoring Functions (CASF) benchmark for scoring function evaluation. The authors' key idea is to decouple the "scoring" process from the "sampling" process, so scoring functions can be tested in a relatively pure context to reflect their quality. In the authors' latest work on this track, i.e., CASF-2013, the performance of a scoring function was quantified in four aspects, including "scoring power", "ranking power", "docking power", and "screening power". All four performance tests were conducted on a test set contg. 195 high-quality protein-ligand complexes selected from PDBbind. A panel of 20 std. scoring functions were tested as demonstration. Importantly, CASF is designed to be an open-access benchmark, with which scoring functions developed by different researchers can be compared on the same ground. Indeed, it has become a popular choice for scoring function validation in recent years. Despite the considerable progress that has been made so far, the performance of today's scoring functions still does not meet people's expectation in many aspects. There is a const. demand for more advanced scoring functions. The authors' efforts have helped to overcome some obstacles underlying scoring function development so that the researchers in this field can move forward faster. The authors will continue to improve the PDBbind database and the CASF benchmark in the future to keep them as useful community resources.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC2sXit12qsLc%253D&md5=48d7287ef58c3ed8d9ff7579014fc486
  19. 19
    Yosinski, J.; Clune, J.; Nguyen, A.; Fuchs, T.; Lipson, H. Understanding neural networks through deep visualization. arXiv 2015, arXiv:1506.06579,  DOI: 10.48550/arXiv.1506.06579
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  20. 20
    Samek, W.; Binder, A.; Montavon, G.; Bach, S.; Müller, K.-R. Evaluating the visualization of what a Deep Neural Network has learned. arXiv 2015, arXiv:1509.06321,  DOI: 10.48550/arXiv.1509.06321
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  21. 21
    Poelking, C.; Chessari, G.; Murray, C. W.; Hall, R. J.; Colwell, L.; Verdonk, M. Meaningful machine learning models and machine-learned pharmacophores from fragment screening campaigns. arXiv 2022, arXiv:2204.06348,  DOI: 10.48550/arXiv.2204.06348
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  22. 22
    Sharma, B.; Chenthamarakshan, V.; Dhurandhar, A.; Pereira, S.; Hendler, J. A.; Dordick, J. S.; Das, P. Accurate Clinical Toxicity Prediction using Multi-task Deep Neural Nets and Contrastive Molecular Explanations. Sci. Rep. 2023, 13, 4908,  DOI: 10.1038/s41598-023-31169-8
    [Crossref], [PubMed], [CAS], Google Scholar
    22
    Accurate clinical toxicity prediction using multi-task deep neural nets and contrastive molecular explanations
    Sharma, Bhanushee; Chenthamarakshan, Vijil; Dhurandhar, Amit; Pereira, Shiranee; Hendler, James A.; Dordick, Jonathan S.; Das, Payel
    Scientific Reports (2023), 13 (1), 4908CODEN: SRCEC3; ISSN:2045-2322. (Nature Portfolio)
    Explainable machine learning for mol. toxicity prediction is a promising approach for efficient drug development and chem. safety. A predictive ML model of toxicity can reduce exptl. cost and time while mitigating ethical concerns by significantly reducing animal and clin. testing. Herein, we use a deep learning framework for simultaneously modeling in vitro, in vivo, and clin. toxicity data. Two different mol. input representations are used; Morgan fingerprints and pre-trained SMILES embeddings. A multi-task deep learning model accurately predicts toxicity for all endpoints, including clin., as indicated by the area under the Receiver Operator Characteristic curve and balanced accuracy. In particular, pre-trained mol. SMILES embeddings as input to the multi-task model improved clin. toxicity predictions compared to existing models in MoleculeNet benchmark. Addnl., our multitask approach is comprehensive in the sense that it is comparable to state-of-the-art approaches for specific endpoints in in vitro, in vivo and clin. platforms. Through both the multi-task model and transfer learning, we were able to indicate the minimal need of in vivo data for clin. toxicity predictions. To provide confidence and explain the model's predictions, we adapt a post-hoc contrastive explanation method that returns pertinent pos. and neg. features, which correspond well to known mutagenic and reactive toxicophores, such as unsubstituted bonded heteroatoms, arom. amines, and Michael receptors. Furthermore, toxicophore recovery by pertinent feature anal. captures more of the in vitro (53%) and in vivo (56%), rather than of the clin. (8%), endpoints, and indeed uncovers a preference in known toxicophore data towards in vitro and in vivo exptl. data. To our knowledge, this is the first contrastive explanation, using both present and absent substructures, for predictions of clin. and in vivo mol. toxicity.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3sXmsV2ks7g%253D&md5=63f1d0798d9ba9b71f3d970b58e752ea
  23. 23
    Hadfield, T. E.; Scantlebury, J.; Deane, C. M. Exploring The Ability Of Machine Learning-Based Virtual Screening Models To Identify The Functional Groups Responsible For Binding. bioRxiv 2023,  DOI: 10.1101/2023.04.29.538820
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  24. 24
    McNutt, A. T.; Francoeur, P.; Aggarwal, R.; Masuda, T.; Meli, R.; Ragoza, M.; Sunseri, J.; Koes, D. R. GNINA 1.0: molecular docking with deep learning. J. Cheminf. 2021, 13, 43,  DOI: 10.1186/s13321-021-00522-2
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  25. 25
    Zheng, L.; Fan, J.; Mu, Y. OnionNet: a Multiple-Layer Intermolecular-Contact-Based Convolutional Neural Network for Protein–Ligand Binding Affinity Prediction. ACS Omega 2019, 4, 15956– 15965,  DOI: 10.1021/acsomega.9b01997
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    24
    OnionNet: a Multiple-Layer Intermolecular-Contact-Based Convolutional Neural Network for Protein-Ligand Binding Affinity Prediction
    Zheng, Liangzhen; Fan, Jingrong; Mu, Yuguang
    ACS Omega (2019), 4 (14), 15956-15965CODEN: ACSODF; ISSN:2470-1343. (American Chemical Society)
    Computational drug discovery provides an efficient tool for helping large-scale lead mol. screening. One of the major tasks of lead discovery is identifying mols. with promising binding affinities toward a target, a protein in general. The accuracies of current scoring functions that are used to predict the binding affinity are not satisfactory enough. Thus, machine learning or deep learning based methods have been developed recently to improve the scoring functions. In this study, a deep convolutional neural network model (called OnionNet) is introduced; its features are based on rotation-free element-pair-specific contacts between ligands and protein atoms, and the contacts are further grouped into different distance ranges to cover both the local and nonlocal interaction information between the ligand and the protein. The prediction power of the model is evaluated and compared with other scoring functions using the comparative assessment of scoring functions (CASF-2013) benchmark and the v2016 core set of the PDBbind database. The robustness of the model is further explored by predicting the binding affinities of the complexes generated from docking simulations instead of exptl. detd. PDB structures.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXhslKhtb3J&md5=0eaf1406acc14b7052d62481de5dec78
  26. 26
    Li, Y.; Rezaei, M. A.; Li, C.; Li, X. DeepAtom: A Framework for Protein-Ligand Binding Affinity Prediction. In 2019 IEEE International Conference on Bioinformatics and Biomedicine (BIBM) ; 2019; pp 303– 310.
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  27. 27
    Feinberg, E. N.; Sur, D.; Wu, Z.; Husic, B. E.; Mai, H.; Li, Y.; Sun, S.; Yang, J.; Ramsundar, B.; Pande, V. S. PotentialNet for molecular property prediction. ACS Cent. Sci. 2018, 4, 1520– 1530,  DOI: 10.1021/acscentsci.8b00507
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    26
    PotentialNet for Molecular Property Prediction
    Feinberg, Evan N.; Sur, Debnil; Wu, Zhenqin; Husic, Brooke E.; Mai, Huanghao; Li, Yang; Sun, Saisai; Yang, Jianyi; Ramsundar, Bharath; Pande, Vijay S.
    ACS Central Science (2018), 4 (11), 1520-1530CODEN: ACSCII; ISSN:2374-7951. (American Chemical Society)
    The arc of drug discovery entails a multiparameter optimization problem spanning vast length scales. The key parameters range from soly. (angstroms) to protein-ligand binding (nanometers) to in vivo toxicity (meters). Through feature learning-instead of feature engineering-deep neural networks promise to outperform both traditional physics-based and knowledge-based machine learning models for predicting mol. properties pertinent to drug discovery. To this end, we present the PotentialNet family of graph convolutions. These models are specifically designed for and achieve state-of-the-art performance for protein-ligand binding affinity. We further validate these deep neural networks by setting new stds. of performance in several ligand-based tasks. In parallel, we introduce a new metric, the Regression Enrichment Factor EF(R)χ, to measure the early enrichment of computational models for chem. data. Finally, we introduce a cross-validation strategy based on structural homol. clustering that can more accurately measure model generalizability, which crucially distinguishes the aims of machine learning for drug discovery from std. machine learning tasks.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXitVGrs7fF&md5=570999e7f212d7c737b24f71665893b7
  28. 28
    Morrone, J. A.; Weber, J. K.; Huynh, T.; Luo, H.; Cornell, W. D. Combining Docking Pose Rank and Structure with Deep Learning Improves Protein–Ligand Binding Mode Prediction over a Baseline Docking Approach. J. Chem. Inf. Model. 2020, 60, 4170– 4179,  DOI: 10.1021/acs.jcim.9b00927
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    27
    Combining Docking Pose Rank and Structure with Deep Learning Improves Protein-Ligand Binding Mode Prediction over a Baseline Docking Approach
    Morrone, Joseph A.; Weber, Jeffrey K.; Huynh, Tien; Luo, Heng; Cornell, Wendy D.
    Journal of Chemical Information and Modeling (2020), 60 (9), 4170-4179CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    The authors present a simple, modular graph-based convolutional neural network that takes structural information from protein-ligand complexes as input to generate models for activity and binding mode prediction. Complex structures are generated by a std. docking procedure and fed into a dual-graph architecture that includes sep. subnetworks for the ligand bonded topol. and the ligand-protein contact map. Recent work has indicated that data set bias drives many past promising results derived from combining deep learning and docking. The dual-graph network allows contributions from ligand identity that give rise to such biases to be distinguished from effects of protein-ligand interactions on classification. The authors show that the neural network is capable of learning from protein structural information when, as in the case of binding mode prediction, an unbiased data set is constructed. The authors next develop a deep learning model for binding mode prediction that uses docking ranking as input in combination with docking structures. This strategy mirrors past consensus models and outperforms a baseline docking program (AutoDock Vina) in a variety of tests, including on cross-docking data sets that mimic real-world docking use cases. Furthermore, the magnitudes of network predictions serve as reliable measures of model confidence.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXjsVOgtbs%253D&md5=ed7f0ba1797757a61d64c817db36cc13
  29. 29
    Jiang, D.; Hsieh, C.-Y.; Wu, Z.; Kang, Y.; Wang, J.; Wang, E.; Liao, B.; Shen, C.; Xu, L.; Wu, J.; Cao, D.; Hou, T. InteractionGraphNet: A Novel and Efficient Deep Graph Representation Learning Framework for Accurate Protein–Ligand Interaction Predictions. J. Med. Chem. 2021, 64, 18209– 18232,  DOI: 10.1021/acs.jmedchem.1c01830
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    28
    InteractionGraphNet: A Novel and Efficient Deep Graph Representation Learning Framework for Accurate Protein-Ligand Interaction Predictions
    Jiang, Dejun; Hsieh, Chang-Yu; Wu, Zhenxing; Kang, Yu; Wang, Jike; Wang, Ercheng; Liao, Ben; Shen, Chao; Xu, Lei; Wu, Jian; Cao, Dongsheng; Hou, Tingjun
    Journal of Medicinal Chemistry (2021), 64 (24), 18209-18232CODEN: JMCMAR; ISSN:0022-2623. (American Chemical Society)
    Accurate quantification of protein-ligand interactions remains a key challenge to structure-based drug design. However, traditional machine learning (ML)-based methods based on handcrafted descriptors, one-dimensional protein sequences, and/or two-dimensional graph representations limit their capability to learn the generalized mol. interactions in 3D space. Here, we proposed a novel deep graph representation learning framework named InteractionGraphNet (IGN) to learn the protein-ligand interactions from the 3D structures of protein-ligand complexes. In IGN, two independent graph convolution modules were stacked to sequentially learn the intramol. and intermol. interactions, and the learned intermol. interactions can be efficiently used for subsequent tasks. Extensive binding affinity prediction, large-scale structure-based virtual screening, and pose prediction expts. demonstrated that IGN achieved better or competitive performance against other state-of-the-art ML-based baselines and docking programs. More importantly, such state-of-the-art performance was proven from the successful learning of the key features in protein-ligand interactions instead of just memorizing certain biased patterns from data.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3MXislSnsrnE&md5=bbee6eb73a8c6b95c389fe11e45c7723
  30. 30
    Hochuli, J.; Helbling, A.; Skaist, T.; Ragoza, M.; Koes, D. R. Visualizing convolutional neural network protein-ligand scoring. J. Mol. Graphics Modell. 2018, 84, 96– 108,  DOI: 10.1016/j.jmgm.2018.06.005
    [Crossref], [PubMed], [CAS], Google Scholar
    29
    Visualizing convolutional neural network protein-ligand scoring
    Hochuli, Joshua; Helbling, Alec; Skaist, Tamar; Ragoza, Matthew; Koes, David Ryan
    Journal of Molecular Graphics & Modelling (2018), 84 (), 96-108CODEN: JMGMFI; ISSN:1093-3263. (Elsevier Ltd.)
    Protein-ligand scoring is an important step in a structure-based drug design pipeline. Selecting a correct binding pose and predicting the binding affinity of a protein-ligand complex enables effective virtual screening. Machine learning techniques can make use of the increasing amts. of structural data that are becoming publicly available. Convolutional neural network (CNN) scoring functions in particular have shown promise in pose selection and affinity prediction for protein-ligand complexes. Neural networks are known for being difficult to interpret. Understanding the decisions of a particular network can help tune parameters and training data to maximize performance. Visualization of neural networks helps decomp. complex scoring functions into pictures that are more easily parsed by humans. Here we present three methods for visualizing how individual protein-ligand complexes are interpreted by 3D convolutional neural networks. We also present a visualization of the convolutional filters and their wts. We describe how the intuition provided by these visualizations aids in network design.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXhtF2ltb3P&md5=db697d859132f02c4a2cef4bfc00ea88
  31. 31
    Lim, J.; Ryu, S.; Park, K.; Choe, Y. J.; Ham, J.; Kim, W. Y. Predicting Drug–Target Interaction Using a Novel Graph Neural Network with 3D Structure-Embedded Graph Representation. J. Chem. Inf. Model. 2019, 59, 3981– 3988,  DOI: 10.1021/acs.jcim.9b00387
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    30
    Predicting Drug-Target Interaction Using a Novel Graph Neural Network with 3D Structure-Embedded Graph Representation
    Lim, Jaechang; Ryu, Seongok; Park, Kyubyong; Choe, Yo Joong; Ham, Jiyeon; Kim, Woo Youn
    Journal of Chemical Information and Modeling (2019), 59 (9), 3981-3988CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    We propose a novel deep learning approach for predicting drug-target interaction using a graph neural network. We introduce a distance-aware graph attention algorithm to differentiate various types of intermol. interactions. Furthermore, we ext. the graph feature of intermol. interactions directly from the 3D structural information on the protein-ligand binding pose. Thus, the model can learn key features for accurate predictions of drug-target interaction rather than just memorize certain patterns of ligand mols. As a result, our model shows better performance than docking and other deep learning methods for both virtual screening (AUROC of 0.968 for the DUD-E test set) and pose prediction (AUROC of 0.935 for the PDBbind test set). In addn., it can reproduce the natural population distribution of active mols. and inactive mols.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXhs1egur3F&md5=670cd989a369d67c734f6762a5a906cf
  32. 32
    Hadfield, T. E.; Imrie, F.; Merritt, A.; Birchall, K.; Deane, C. M. Incorporating Target-Specific Pharmacophoric Information into Deep Generative Models for Fragment Elaboration. J. Chem. Inf. Model. 2022, 62, 2280– 2292,  DOI: 10.1021/acs.jcim.1c01311
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    31
    Incorporating Target-Specific Pharmacophoric Information into Deep Generative Models for Fragment Elaboration
    Hadfield, Thomas E.; Imrie, Fergus; Merritt, Andy; Birchall, Kristian; Deane, Charlotte M.
    Journal of Chemical Information and Modeling (2022), 62 (10), 2280-2292CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    Despite recent interest in deep generative models for scaffold elaboration, their applicability to fragment-to-lead campaigns has so far been limited. This is primarily due to their inability to account for local protein structure or a user's design hypothesis. We propose a novel method for fragment elaboration, STRIFE, that overcomes these issues. STRIFE takes as input fragment hotspot maps (FHMs) extd. from a protein target and processes them to provide meaningful and interpretable structural information to its generative model, which in turn is able to rapidly generate elaborations with complementary pharmacophores to the protein. In a large-scale evaluation, STRIFE outperforms existing, structure-unaware, fragment elaboration methods in proposing highly ligand-efficient elaborations. In addn. to automatically extg. pharmacophoric information from a protein target's FHM, STRIFE optionally allows the user to specify their own design hypotheses.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB38XhtFGru7rE&md5=c639b01f0c509bbd2e20308fa044bdb5
  33. 33
    Curran, P. R.; Radoux, C. J.; Smilova, M. D.; Sykes, R. A.; Higueruelo, A. P.; Bradley, A. R.; Marsden, B. D.; Spring, D. R.; Blundell, T. L.; Leach, A. R.; Pitt, W. R.; Cole, J. C. Hotspots API: A Python Package for the Detection of Small Molecule Binding Hotspots and Application to Structure-Based Drug Design. J. Chem. Inf. Model. 2020, 60, 1911– 1916,  DOI: 10.1021/acs.jcim.9b00996
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    32
    Hotspots API: A Python Package for the Detection of Small Molecule Binding Hotspots and Application to Structure-Based Drug Design
    Curran, Peter R.; Radoux, Chris J.; Smilova, Mihaela D.; Sykes, Richard A.; Higueruelo, Alicia P.; Bradley, Anthony R.; Marsden, Brian D.; Spring, David R.; Blundell, Tom L.; Leach, Andrew R.; Pitt, William R.; Cole, Jason C.
    Journal of Chemical Information and Modeling (2020), 60 (4), 1911-1916CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    Methods that survey protein surfaces for binding hotspots can help to evaluate target tractability and guide exploration of potential ligand binding regions. Fragment Hotspot Maps builds upon interaction data mined from the CSD (Cambridge Structural Database) and exploits the idea of identifying hotspots using small chem. fragments, which is now widely used to design new drug leads. Prior to this publication, Fragment Hotspot Maps was only publicly available through a web application. To increase the accessibility of this algorithm we present the Hotspots API (application programming interface), a toolkit that offers programmatic access to the core Fragment Hotspot Maps algorithm, thereby facilitating the interpretation and application of 3the anal. To demonstrate the package's utility, we present a workflow which automatically derives protein hydrogen-bond constraints for mol. docking with GOLD. The Hotspots API is available from https://github.com/prcurran/hotspots under the MIT license and is dependent upon the com. CSD Python API.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXlsFahsbw%253D&md5=9402dded4d637e5b324e329d9e2c9a08
  34. 34
    Salentin, S.; Schreiber, S.; Haupt, V. J.; Adasme, M. F.; Schroeder, M. PLIP: fully automated protein–ligand interaction profiler. Nucleic Acids Res. 2015, 43, W443– W447,  DOI: 10.1093/nar/gkv315
    [Crossref], [PubMed], [CAS], Google Scholar
    33
    PLIP: fully automated protein-ligand interaction profiler
    Salentin, Sebastian; Schreiber, Sven; Haupt, V. Joachim; Adasme, Melissa F.; Schroeder, Michael
    Nucleic Acids Research (2015), 43 (W1), W443-W447CODEN: NARHAD; ISSN:0305-1048. (Oxford University Press)
    The characterization of interactions in protein-ligand complexes is essential for research in structural bioinformatics, drug discovery and biol. However, comprehensive tools are not freely available to the research community. Here, we present the protein-ligand interaction profiler (PLIP), a novel web service for fully automated detection and visualization of relevant non-covalent protein-ligand contacts in 3D structures, freely available at projects.biotec.tu-dresden.de/plip-web. The input is either a Protein Data Bank structure, a protein or ligand name, or a custom protein-ligand complex (e.g. from docking). In contrast to other tools, the rule-based PLIP algorithm does not require any structure prepn. It returns a list of detected interactions on single atom level, covering seven interaction types (hydrogen bonds, hydrophobic contacts, pi-stacking, pi-cation interactions, salt bridges, water bridges and halogen bonds). PLIP stands out by offering publication-ready images, PyMOL sesions files to generate custom images and parsable result files to facilitate successive data processing. The full python source code is available for download on the website. PLIP's command-line mode allows for high-throughput interaction profiling.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC2sXhtVymtbrI&md5=4878881a7bd715e5585682876e102e65
  35. 35
    Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, L. u.; Polosukhin, I. Attention is All you Need. arXiv 2017, arXiv:1706.03762,  DOI: 10.48550/arXiv.1706.03762
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  36. 36
    Francoeur, P. G.; Masuda, T.; Sunseri, J.; Jia, A.; Iovanisci, R. B.; Snyder, I.; Koes, D. R. Three-Dimensional Convolutional Neural Networks and a Cross-Docked Data Set for Structure-Based Drug Design. J. Chem. Inf. Model. 2020, 60, 4200– 4215,  DOI: 10.1021/acs.jcim.0c00411
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    35
    Three-Dimensional Convolutional Neural Networks and a Cross-Docked Data Set for Structure-Based Drug Design
    Francoeur, Paul G.; Masuda, Tomohide; Sunseri, Jocelyn; Jia, Andrew; Iovanisci, Richard B.; Snyder, Ian; Koes, David R.
    Journal of Chemical Information and Modeling (2020), 60 (9), 4200-4215CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    One of the main challenges in drug discovery is predicting protein-ligand binding affinity. Recently, machine learning approaches have made substantial progress on this task. However, current methods of model evaluation are overly optimistic in measuring generalization to new targets, and there does not exist a std. data set of sufficient size to compare performance between models. We present a new data set for structure-based machine learning, the CrossDocked2020 set, with 22.5 million poses of ligands docked into multiple similar binding pockets across the Protein Data Bank, and perform a comprehensive evaluation of grid-based convolutional neural network (CNN) models on this data set. We also demonstrate how the partitioning of the training data and test data can impact the results of models trained with the PDBbind data set, how performance improves by adding more lower-quality training data, and how training with docked poses imparts pose sensitivity to the predicted affinity of a complex. Our best performing model, an ensemble of five densely connected CNNs, achieves a root mean squared error of 1.42 and Pearson R of 0.612 on the affinity prediction task, an AUC of 0.956 at binding pose classification, and a 68.4% accuracy at pose selection on the CrossDocked2020 set. By providing data splits for clustered cross-validation and the raw data for the CrossDocked2020 set, we establish the first standardized data set for training machine learning models to recognize ligands in noncognate target structures while also greatly expanding the no. of poses available for training. In order to facilitate community adoption of this data set for benchmarking protein-ligand binding affinity prediction, we provide our models, wts., and the CrossDocked2020 set at https://github.com/gnina/models.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXhslWqs73N&md5=5a5465187dbdef41d556baddd1a8e4d9
  37. 37
    Landrum, G. RDKit: Open-source cheminformatics ; 2016; https://www.bibsonomy.org/bibtex/28d01fceeccd6bf2486e47d7c4207b108/salotz.
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  38. 38
    O’Boyle, N. M.; Banck, M.; James, C. A.; Morley, C.; Vandermeersch, T.; Hutchison, G. R. Open Babel: An open chemical toolbox. J. Cheminf. 2011, 3, 33,  DOI: 10.1186/1758-2946-3-33
    [Crossref], [PubMed], [CAS], Google Scholar
    37
    Open Babel: an open chemical toolbox
    O'Boyle, Noel M.; Banck, Michael; James, Craig A.; Morley, Chris; Vandermeersch, Tim; Hutchison, Geoffrey R.
    Journal of Cheminformatics (2011), 3 (), 33CODEN: JCOHB3; ISSN:1758-2946. (Chemistry Central Ltd.)
    Background: A frequent problem in computational modeling is the interconversion of chem. structures between different formats. While std. interchange formats exist (for example, Chem. Markup Language) and de facto stds. have arisen (for example, SMILES format), the need to interconvert formats is a continuing problem due to the multitude of different application areas for chem. data, differences in the data stored by different formats (0D vs. 3D, for example), and competition between software along with a lack of vendor-neutral formats. Results: We discuss, for the first time, Open Babel, an open-source chem. toolbox that speaks the many languages of chem. data. Open Babel version 2.3 interconverts over 110 formats. The need to represent such a wide variety of chem. and mol. data requires a library that implements a wide range of cheminformatics algorithms, from partial charge assignment and aromaticity detection, to bond order perception and canonicalization. We detail the implementation of Open Babel, describe key advances in the 2.3 release, and outline a variety of uses both in terms of software products and scientific research, including applications far beyond simple format interconversion. Conclusions: Open Babel presents a soln. to the proliferation of multiple chem. file formats. In addn., it provides a variety of useful utilities from conformer searching and 2D depiction, to filtering, batch conversion, and substructure and similarity searching. For developers, it can be used as a programming library to handle chem. data in areas such as org. chem., drug design, materials science, and computational chem. It is freely available under an open-source license.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC3MXhsVWjurbF&md5=74e4f19b7f87417f916d57f7abcfb761
  39. 39
    Yang, J.; Shen, C.; Huang, N. Predicting or Pretending: Artificial Intelligence for Protein-Ligand Interactions Lack of Sufficiently Large and Unbiased Datasets. Frontiers in Pharmacology 2020, 11, 1,  DOI: 10.3389/fphar.2020.00069
    [Crossref], [PubMed], Google Scholar
    There is no corresponding record for this reference.
  40. 40
    Scantlebury, J.; Brown, N.; Von Delft, F.; Deane, C. M. Data Set Augmentation Allows Deep Learning-Based Virtual Screening to Better Generalize to Unseen Target Classes and Highlight Important Binding Interactions. J. Chem. Inf. Model. 2020, 60, 3722– 3730,  DOI: 10.1021/acs.jcim.0c00263
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    39
    Data Set Augmentation Allows Deep Learning-Based Virtual Screening to Better Generalize to Unseen Target Classes and Highlight Important Binding Interactions
    Scantlebury, Jack; Brown, Nathan; Von Delft, Frank; Deane, Charlotte M.
    Journal of Chemical Information and Modeling (2020), 60 (8), 3722-3730CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    Current deep learning methods for structure-based virtual screening take the structures of both the protein and the ligand as input but make little or no use of the protein structure when predicting ligand binding. Here, a relatively simple method of data set augmentation forces such deep learning methods to take into account information from the protein. Models trained in this way are more generalizable (make better predictions on protein/ligand complexes from a different distribution to the training data). They also assign more meaningful importance to the protein and ligand atoms involved in binding. Overall, the authors' results show that data set augmentation can help deep learning-based virtual screening to learn phys. interactions rather than data set biases.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXhsVaqt77O&md5=8859d41da975e883cac7c2489befab60
  41. 41
    Satorras, V. G.; Hoogeboom, E.; Welling, M. E(n) Equivariant Graph Neural Networks. arXiv 2021, arXiv:2102.09844,  DOI: 10.48550/arXiv.2102.09844
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  42. 42
    Ragoza, M.; Hochuli, J.; Idrobo, E.; Sunseri, J.; Koes, D. R. Protein-Ligand Scoring with Convolutional Neural Networks. J. Chem. Inf. Model. 2017, 57, 942– 957,  DOI: 10.1021/acs.jcim.6b00740
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    41
    Protein-Ligand Scoring with Convolutional Neural Networks
    Ragoza, Matthew; Hochuli, Joshua; Idrobo, Elisa; Sunseri, Jocelyn; Koes, David Ryan
    Journal of Chemical Information and Modeling (2017), 57 (4), 942-957CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    Computational approaches to drug discovery can reduce the time and cost assocd. with exptl. assays and enable the screening of novel chemotypes. Structure-based drug design methods rely on scoring functions to rank and predict binding affinities and poses. The ever-expanding amt. of protein-ligand binding and structural data enables the use of deep machine learning techniques for protein-ligand scoring. We describe convolutional neural network (CNN) scoring functions that take as input a comprehensive three-dimensional (3D) representation of a protein-ligand interaction. A CNN scoring function automatically learns the key features of protein-ligand interactions that correlate with binding. We train and optimize our CNN scoring functions to discriminate between correct and incorrect binding poses and known binders and nonbinders. We find that our CNN scoring function outperforms the AutoDock Vina scoring function when ranking poses both for pose prediction and virtual screening.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC2sXlsVems7Y%253D&md5=9cae97167da0c93e1d896f85339cca7e
  43. 43
    Zhu, H.; Yang, J.; Huang, N. Assessment of the Generalization Abilities of Machine-Learning Scoring Functions for Structure-Based Virtual Screening. J. Chem. Inf. Model. 2022, 62, 5485– 5502,  DOI: 10.1021/acs.jcim.2c01149
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    42
    Assessment of the Generalization Abilities of Machine-Learning Scoring Functions for Structure-Based Virtual Screening
    Zhu, Hui; Yang, Jincai; Huang, Niu
    Journal of Chemical Information and Modeling (2022), 62 (22), 5485-5502CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    In structure-based virtual screening (SBVS), it is crit. that scoring functions capture protein-ligand at. interactions. By focusing on the local domains of ligand binding pockets, a standardized pocket Pfam-based clustering (Pfam-cluster) approach was developed to assess the cross-target generalization ability of machine-learning scoring functions (MLSFs). Subsequently, 12 typical MLSFs were evaluated using random cross-validation (Random-CV), protein sequence similarity-based cross-validation (Seq-CV), and pocket Pfam-based cross-validation (Pfam-CV) methods. Surprisingly, all of the tested models showed decreased performances from Random-CV to Seq-CV to Pfam-CV expts., not showing satisfactory generalization capacity. Our interpretable anal. suggested that the predictions on novel targets by MLSFs were dependent on buried solvent-accessible surface area (SASA)-related features of complex structures, with greater predicted binding affinities on complexes owning larger protein-ligand interfaces. By combining buried SASA-related features with target-specific patterns that were only shared among structurally similar compds. in the same cluster, the random forest (RF)-Score attained a good performance in the Random-CV test. Based on these findings, we strongly advise assessing the generalization ability of MLSFs with the Pfam-cluster approach and being cautious with the features learned by MLSFs.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB38Xis1Kru7rK&md5=d6197f1f9a4da0df36b420bb9746b446
  44. 44
    Radoux, C. J.; Olsson, T. S.; Pitt, W. R.; Groom, C. R.; Blundell, T. L. Identifying interactions that determine fragment binding at protein hotspots. J. Med. Chem. 2016, 59, 4314– 4325,  DOI: 10.1021/acs.jmedchem.5b01980
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    43
    Identifying Interactions that Determine Fragment Binding at Protein Hotspots
    Radoux, Chris J.; Olsson, Tjelvar S. G.; Pitt, Will R.; Groom, Colin R.; Blundell, Tom L.
    Journal of Medicinal Chemistry (2016), 59 (9), 4314-4325CODEN: JMCMAR; ISSN:0022-2623. (American Chemical Society)
    Locating a ligand-binding site is an important first step in structure-guided drug discovery, but current methods do little to suggest which interactions within a pocket are the most important for binding. Here the authors illustrate a method that samples at. hotspots with simple mol. probes to produce fragment hotspot maps. These maps specifically highlight fragment-binding sites and their corresponding pharmacophores. For ligand-bound structures, they provide an intuitive visual guide within the binding site, directing medicinal chemists where to grow the mol. and alerting them to suboptimal interactions within the original hit. The fragment hotspot map calcn. is validated using exptl. binding positions of 21 fragments and subsequent lead mols. The ligands are found in high scoring areas of the fragment hotspot maps, with fragment atoms having a median percentage rank of 97%. Protein kinase B and pantothenate synthetase are examd. in detail. In each case, the fragment hotspot maps are able to rationalize a Free-Wilson anal. of SAR data from a fragment-based drug design project.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC28XltlOms7Y%253D&md5=fa25eabfe291f9519f5701bcbc6c52c3
  45. 45
    Skyner, R.; von Delft, F. Xchem, fragalysis; https://fragalysis.diamond.ac.uk/ (accessed: 2022-07-30).
    Google Scholar
    There is no corresponding record for this reference.
  46. 46
    Wang, C.; Zhang, Y. Improving scoring-docking-screening powers of protein–ligand scoring functions using random forest. J. Comput. Chem. 2017, 38, 169– 177,  DOI: 10.1002/jcc.24667
    [Crossref], [PubMed], [CAS], Google Scholar
    45
    Improving scoring-docking-screening powers of protein-ligand scoring functions using random forest
    Wang, Cheng; Zhang, Yingkai
    Journal of Computational Chemistry (2017), 38 (3), 169-177CODEN: JCCHDD; ISSN:0192-8651. (John Wiley & Sons, Inc.)
    The development of new protein-ligand scoring functions using machine learning algorithms, such as random forest, has been of significant interest. By efficiently using expanded feature sets and a large set of exptl. data, random forest based scoring functions (RFbScore) can achieve better correlations to exptl. protein-ligand binding data with known crystal structures; however, more extensive tests indicate that such enhancement in scoring power comes with significant under-performance in docking and screening power tests compared to traditional scoring functions. To improve scoring-docking-screening powers of protein-ligand docking functions simultaneously, the authors have introduced a ΔvinaRF parameterization and feature selection framework based on random forest. The authors' developed scoring function ΔvinaRF20, which employs 20 descriptors in addn. to the AutoDock Vina score, can achieve superior performance in all power tests of both CASF-2013 and CASF-2007 benchmarks compared to classical scoring functions. The ΔvinaRF20 scoring function and its code are freely available on the web at: https://www.nyu.edu/projects/yzhang/DeltaVina.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC28XhvVygurzM&md5=1a3ea0b5f159c9c0edf315fed2110ed0
  47. 47
    Lu, J.; Hou, X.; Wang, C.; Zhang, Y. Incorporating Explicit Water Molecules and Ligand Conformation Stability in Machine-Learning Scoring Functions. J. Chem. Inf. Model. 2019, 59, 4540– 4549,  DOI: 10.1021/acs.jcim.9b00645
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    46
    Incorporating Explicit Water Molecules and Ligand Conformation Stability in Machine-Learning Scoring Functions
    Lu, Jianing; Hou, Xuben; Wang, Cheng; Zhang, Yingkai
    Journal of Chemical Information and Modeling (2019), 59 (11), 4540-4549CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
    Structure-based drug design is critically dependent on accuracy of mol. docking scoring functions, and there is a significant interest to advance scoring functions with machine learning approaches. In this work, by judiciously expanding the training set, exploring new features related to explicit mediating water mols. as well as ligand conformation stability, and applying extreme gradient boosting (XGBoost) with Δ-Vina parametrization, we have improved robustness and applicability of machine-learning scoring functions. The new scoring function ΔvinaXGB can not only perform consistently among the top compared to classical scoring functions for the CASF-2016 benchmark but also achieves significantly better prediction accuracy in different types of structures that mimic real docking applications.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXitVSqs73O&md5=57fea8fa6135a109c3bcfd5ddffb1b33
  48. 48
    Moon, S.; Zhung, W.; Yang, S.; Lim, J.; Kim, W. Y. PIGNet: a physics-informed deep learning model toward generalized drug–target interaction predictions. Chem. Sci. 2022, 13, 3661– 3673,  DOI: 10.1039/D1SC06946B
    [Crossref], [PubMed], [CAS], Google Scholar
    47
    PIGNet: a physics-informed deep learning model toward generalized drug-target interaction predictions
    Moon, Seokhyun; Zhung, Wonho; Yang, Soojung; Lim, Jaechang; Kim, Woo Youn
    Chemical Science (2022), 13 (13), 3661-3673CODEN: CSHCCN; ISSN:2041-6520. (Royal Society of Chemistry)
    Recently, deep neural network (DNN)-based drug-target interaction (DTI) models were highlighted for their high accuracy with affordable computational costs. Yet, the models' insufficient generalization remains a challenging problem in the practice of in silico drug discovery. We propose two key strategies to enhance generalization in the DTI model. The first is to predict the atom-atom pairwise interactions via physics-informed equations parameterized with neural networks and provides the total binding affinity of a protein-ligand complex as their sum. We further improved the model generalization by augmenting a broader range of binding poses and ligands to training data. We validated our model, PIGNet, in the comparative assessment of scoring functions (CASF) 2016, demonstrating the outperforming docking and screening powers than previous methods. Our physics-informing strategy also enables the interpretation of predicted affinities by visualizing the contribution of ligand substructures, providing insights for further ligand optimization.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB38XmtlKktb0%253D&md5=c7ca118db214dde8563a0945db0a2331
  49. 49
    Shen, C.; Zhang, X.; Deng, Y.; Gao, J.; Wang, D.; Xu, L.; Pan, P.; Hou, T.; Kang, Y. Boosting Protein–Ligand Binding Pose Prediction and Virtual Screening Based on Residue–Atom Distance Likelihood Potential and Graph Transformer. J. Med. Chem. 2022, 65, 10691– 10706,  DOI: 10.1021/acs.jmedchem.2c00991
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    48
    Boosting Protein-Ligand Binding Pose Prediction and Virtual Screening Based on Residue-Atom Distance Likelihood Potential and Graph Transformer
    Shen, Chao; Zhang, Xujun; Deng, Yafeng; Gao, Junbo; Wang, Dong; Xu, Lei; Pan, Peichen; Hou, Tingjun; Kang, Yu
    Journal of Medicinal Chemistry (2022), 65 (15), 10691-10706CODEN: JMCMAR; ISSN:0022-2623. (American Chemical Society)
    The past few years have witnessed enormous progress toward applying machine learning approaches to the development of protein-ligand scoring functions. However, the robust performance and wide applicability of scoring functions remain a big challenge for increasing the success rate of docking-based virtual screening. Herein, a novel scoring function named RTMScore was developed by introducing a tailored residue-based graph representation strategy and several graph transformer layers for the learning of protein and ligand representations, followed by a mixt. d. network to obtain residue-atom distance likelihood potential. Our approach was resolutely validated on the CASF-2016 benchmark, and the results indicate that RTMScore can outperform almost all of the other state-of-the-art methods in terms of both the docking and screening powers. Further evaluation confirms the robustness of our approach that can not only retain its docking power on cross-docked poses but also achieve improved performance as a rescoring tool in larger-scale virtual screening.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB38XhvF2htrbO&md5=217e5d659ff8c69c9c9559c99344aef0
  50. 50
    Liao, Z.; You, R.; Huang, X.; Yao, X.; Huang, T.; Zhu, S. DeepDock: Enhancing Ligand-protein Interaction Prediction by a Combination of Ligand and Structure Information. In 2019 IEEE International Conference on Bioinformatics and Biomedicine (BIBM) ; 2019; pp 311– 317.
    [Crossref], Google Scholar
    There is no corresponding record for this reference.
  51. 51
    https://fragalysis.diamond.ac.uk/viewer/react/download/tag/5740113a-7603-4af4-9523-7f902186d4a2 (accessed: 2022-10-14).
    Google Scholar
    There is no corresponding record for this reference.
  52. 52
    Carbery, A.; Skyner, R.; von Delft, F.; Deane, C. M. Fragment Libraries Designed to Be Functionally Diverse Recover Protein Binding Information More Efficiently Than Standard Structurally Diverse Libraries. J. Med. Chem. 2022, 65, 11404– 11413,  DOI: 10.1021/acs.jmedchem.2c01004
    [ACS Full Text ACS Full Text], [CAS], Google Scholar
    51
    Fragment Libraries Designed to Be Functionally Diverse Recover Protein Binding Information More Efficiently Than Standard Structurally Diverse Libraries
    Carbery, Anna; Skyner, Rachael; von Delft, Frank; Deane, Charlotte M.
    Journal of Medicinal Chemistry (2022), 65 (16), 11404-11413CODEN: JMCMAR; ISSN:0022-2623. (American Chemical Society)
    Current fragment-based drug design relies on the efficient exploration of chem. space by using structurally diverse libraries of small fragments. However, structurally dissimilar compds. can exploit the same interactions and thus be functionally similar. Using three-dimensional structures of many fragments bound to multiple targets, we examd. if a better strategy for selecting fragments for screening libraries exists. We show that structurally diverse fragments can be described as functionally redundant, often making the same interactions. Ranking fragments by the no. of novel interactions they made, we show that functionally diverse selections of fragments substantially increase the amt. of information recovered for unseen targets compared to the amts. recovered by other methods of selection. Using these results, we design small functionally efficient libraries that can give significantly more information about new protein targets than similarly sized structurally diverse libraries. By covering more functional space, we can generate more diverse sets of drug leads.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB38XitFWktLnO&md5=20c210891befe38eb6ec658cba2a01b8
  53. 53
    Schuller, M. Fragment binding to the Nsp3 macrodomain of SARS-CoV-2 identified through crystallographic screening and computational docking. Sci. Adv. 2021, 7, eabf8711  DOI: 10.1126/sciadv.abf8711
    [Crossref], [PubMed], Google Scholar
    There is no corresponding record for this reference.
  54. 54
    Imprachim, N.; Yosaatmadja, Y.; Newman, J. A. Crystal structures and fragment screening of SARS-CoV-2 NSP14 reveal details of exoribonuclease activation and mRNA capping and provide starting points for antiviral drug development. Nucleic Acids Research 2023, 51, 475,  DOI: 10.1093/nar/gkac1207
    [Crossref], [PubMed], [CAS], Google Scholar
    53
    Crystal structures and fragment screening of SARS-CoV-2 NSP14 reveal details of exoribonuclease activation and mRNA capping and provide starting points for antiviral drug development
    Imprachim Nergis; Yosaatmadja Yuliana; Newman Joseph A
    Nucleic acids research (2023), 51 (1), 475-487 ISSN:.
    NSP14 is a dual function enzyme containing an N-terminal exonuclease domain (ExoN) and C-terminal Guanine-N7-methyltransferase (N7-MTase) domain. Both activities are essential for the viral life cycle and may be targeted for anti-viral therapeutics. NSP14 forms a complex with NSP10, and this interaction enhances the nuclease but not the methyltransferase activity. We have determined the structure of SARS-CoV-2 NSP14 in the absence of NSP10 to 1.7 ÅA resolution. Comparisons with NSP14/NSP10 complexes reveal significant conformational changes that occur within the NSP14 ExoN domain upon binding of NSP10, including helix to coil transitions that facilitate the formation of the ExoN active site and provide an explanation of the stimulation of nuclease activity by NSP10. We have determined the structure of NSP14 in complex with cap analogue 7MeGpppG, and observe conformational changes within a SAM/SAH interacting loop that plays a key role in viral mRNA capping offering new insights into MTase activity. We perform an X-ray fragment screen on NSP14, revealing 72 hits bound to sites of inhibition in the ExoN and MTase domains. These fragments serve as excellent starting point tools for structure guided development of NSP14 inhibitors that may be used to treat COVID-19 and potentially other future viral threats.
    >> More from SciFinder ®
    https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A280%3ADC%252BB28rps1ahtg%253D%253D&md5=b3ad340352d17592df9e8d0fffa58af3
  55. 55
    https://fragalysis.diamond.ac.uk/viewer/react/download/tag/01f41754-b2bb-4817-acc3-a5ebe820316d (accessed: 2022-10-14).
    Google Scholar
    There is no corresponding record for this reference.
  56. 56
    https://fragalysis.diamond.ac.uk/viewer/react/download/tag/3df36d6b-3a5b-400d-97eb-af3b0e0df42d (accessed: 2022-10-14).
    Google Scholar
    There is no corresponding record for this reference.

Cited By

This article has not yet been cited by other publications.

Download PDF

  • Abstract

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 1

    Figure 1. An overall schema of the methods used to debias and test PointVS. We first thoroughly filter the testing and training sets (a) before benchmarking the performance of PointVS on the docking power and scoring power tests (b). We then use attribution to gain insights into important binding regions in the protein pocket (c), which we use for fragment elaboration (d).

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 2

    Figure 2. Architecture and input format of PointVS. Each of the n atoms in the input to the model (left) is given a one-hot encoded feature vector with a single bit added to indicate whether the atom is from the ligand or the receptor, as well as a position p0∈R3. There are m edges, defined by the edge indices ei ∈ {0, ···, n – 1}2×m, which are the indices of connected atoms, and the corresponding edge attributes ea ∈ {0, 1}3×m. These are one-hot encodings representing ligand–ligand, ligand–protein, and protein–protein edges. There are skip connections between each of the EGNN layers, and the linear and global average final pooling (GAP) layers only act upon the node features f.

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 3

    Figure 3. Top-N vs N pose ranking performance on the gninaSetPose set for Autodock Vina, PointVS, and gnina. Terms in brackets in the legend refer to the training sets; filtering the training set by protein and ligand similarity results in slightly degraded performance. The Top-1 values for PointVS trained on Redocked\gninaSetPose80 and gnina trained on Redocked are the same (68%), with PointVS trained on Redocked achieving 70%.

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 4

    Figure 4. Three Tankyrase-2 inhibitors: 5C5P (a, d, g, j), 4J21 (b, e, h, k), and 4J22 (c, f, i, l). The top row shows the Protein–Ligand Interaction Profiler (PLIP) analysis of each structure, where dark blue lines are hydrogen bonds, dotted gray are hydrophobic interactions, dotted green are π–π interactions, solid green are halogen bonds, and lilac are water bridges. The second row shows the results of performing atom masking with gnina, with green representing a positive attribution score (identified as making a positive contribution to binding) and red, a negative score (making a negative contribution). The third row shows the results of bond masking with IGN, with lines between atoms showing the top five highest-scoring edges for each structure, with darker red representing higher scores. The bottom row shows the results of edge-attention attribution performed with PointVS, with the lines also showing the top five highest-scoring edges, and darker pink representing higher scores.

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 5

    Figure 5. Donor and acceptor hotspot maps in the binding pocket of Mpro, colored purple and orange, respectively, and numbered according to their rank. The hotspots on the left (a) were obtained with the Hotspots API, the hotspots in the center (b) were obtained with PointVS by performing attribution on 152 structures of bound fragments, and the hotspots on the right (c) were obtained by extracting the most common protein atoms identified by PLIP as involved in hydrogen bonds with ligand atoms across the 152 bound structures (in this case identified over five times).

    High Resolution Image
    Download MS PowerPoint Slide

    Figure 6

    Figure 6. Top five scoring hotspot maps in the binding pocket of Mpro found by performing attribution on 40 sets of randomly sampled sets of 10 (a), 30 (b), and 80 (c) bound structures. Both donor and acceptor hotspots are shown in red. The spheres representing the hotspots are transparent and overlaid, so the more opaque a hotspot appears, the more times it was ranked in the top five.

    High Resolution Image
    Download MS PowerPoint Slide

  • References

    ARTICLE SECTIONS
    Jump To

    This article references 56 other publications.

    1. 1
      Barnett, E.; Onete, D.; Salekin, A.; Faraone, S. V. Genomic Machine Learning Meta-regression: Insights on Associations of Study Features with Reported Model Performance. medRxiv 2022,  DOI: 10.1101/2022.01.10.22268751
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    2. 2
      Shim, M.; Lee, S.-H.; Hwang, H.-J. Inflated prediction accuracy of neuropsychiatric biomarkers caused by data leakage in feature selection. Sci. Rep. 2021, 11, 7980,  DOI: 10.1038/s41598-021-87157-3
      [Crossref], [PubMed], [CAS], Google Scholar
      2
      Inflated prediction accuracy of neuropsychiatric biomarkers caused by data leakage in feature selection
      Shim, Miseon; Lee, Seung-Hwan; Hwang, Han-Jeong
      Scientific Reports (2021), 11 (1), 7980CODEN: SRCEC3; ISSN:2045-2322. (Nature Research)
      In recent years, machine learning techniques have been frequently applied to uncovering neuropsychiatric biomarkers with the aim of accurately diagnosing neuropsychiatric diseases and predicting treatment prognosis. However, many studies did not perform cross validation (CV) when using machine learning techniques, or others performed CV in an incorrect manner, leading to significantly biased results due to overfitting problem. The aim of this study is to investigate the impact of CV on the prediction performance of neuropsychiatric biomarkers, in particular, for feature selection performed with high-dimensional features. To this end, we evaluated prediction performances using both simulation data and actual electroencephalog. (EEG) data. The overall prediction accuracies of the feature selection method performed outside of CV were considerably higher than those of the feature selection method performed within CV for both the simulation and actual EEG data. The differences between the prediction accuracies of the two feature selection approaches can be thought of as the amt. of overfitting due to selection bias. Our results indicate the importance of correctly using CV to avoid biased results of prediction performance of neuropsychiatric biomarkers.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3MXptFOlsLg%253D&md5=a43ca92128b177b8a67e2437fbe5c556
    3. 3
      Roberts, M. Common pitfalls and recommendations for using machine learning to detect and prognosticate for COVID-19 using chest radiographs and CT scans. Nat. Mach. Intell. 2021, 3, 199– 217,  DOI: 10.1038/s42256-021-00307-0
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    4. 4
      Tu, F.; Zhu, J.; Zheng, Q.; Zhou, M. Be Careful of When: An Empirical Study on Time-Related Misuse of Issue Tracking Data. In Proceedings of the 2018 26th ACM Joint Meeting on European Software Engineering Conference and Symposium on the Foundations of Software Engineering , New York, NY, USA, 2018; p 307– 318.
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    5. 5
      Sieg, J.; Flachsenberg, F.; Rarey, M. In Need of Bias Control: Evaluating Chemical Data for Machine Learning in Structure-Based Virtual Screening. J. Chem. Inf. Model. 2019, 59, 947– 961,  DOI: 10.1021/acs.jcim.8b00712
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      5
      In Need of Bias Control: Evaluating Chemical Data for Machine Learning in Structure-Based Virtual Screening
      Sieg, Jochen; Flachsenberg, Florian; Rarey, Matthias
      Journal of Chemical Information and Modeling (2019), 59 (3), 947-961CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      A review. Reports of successful applications of machine learning (ML) methods in structure-based virtual screening (SBVS) are increasing. ML methods such as convolutional neural networks show promising results and often outperform traditional methods such as empirical scoring functions in retrospective validation. However, trained ML models are often treated as black boxes and are not straightforwardly interpretable. In most cases, it is unknown which features in the data are decisive and whether a model's predictions are right for the right reason. Hence, the authors reevaluated three widely used benchmark data sets in the context of ML methods and came to the conclusion that not every benchmark data set is suitable. Moreover, the authors demonstrate on two examples from current literature that bias is learned implicitly and unnoticed from std. benchmarks. On the basis of these results, the authors conclude that there is a need for eligible validation expts. and benchmark data sets suited to ML for more bias-controlled validation in ML-based SBVS. Therefore, the authors provide guidelines for setting up validation expts. and give a perspective on how new data sets could be generated.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXktVOgsrY%253D&md5=6cb48f2acd541c10c8565f202a6eb9e8
    6. 6
      Chen, L.; Cruz, A.; Ramsey, S.; Dickson, C. J.; Duca, J. S.; Hornak, V.; Koes, D. R.; Kurtzman, T. Hidden bias in the DUD-E dataset leads to misleading performance of deep learning in structure-based virtual screening. PLoS One 2019, 14, e0220113  DOI: 10.1371/journal.pone.0220113
      [Crossref], [PubMed], [CAS], Google Scholar
      6
      Hidden bias in the DUD-E dataset leads to misleading performance of deep learning in structure-based virtual screening
      Chen, Lieyang; Cruz, Anthony; Ramsey, Steven; Dickson, Callum J.; Duca, Jose S.; Hornak, Viktor; Koes, David R.; Kurtzman, Tom
      PLoS One (2019), 14 (8), e0220113CODEN: POLNCL; ISSN:1932-6203. (Public Library of Science)
      Recently much effort has been invested in using convolutional neural network (CNN) models trained on 3D structural images of protein-ligand complexes to distinguish binding from non-binding ligands for virtual screening. However, the dearth of reliable protein-ligand x-ray structures and binding affinity data has required the use of constructed datasets for the training and evaluation of CNN mol. recognition models. Here, we outline various sources of bias in one such widely-used dataset, the Directory of Useful Decoys: Enhanced (DUD-E). We have constructed and performed tests to investigate whether CNN models developed using DUD-E are properly learning the underlying physics of mol. recognition, as intended, or are instead learning biases inherent in the dataset itself. We find that superior enrichment efficiency in CNN models can be attributed to the analog and decoy bias hidden in the DUD-E dataset rather than successful generalization of the pattern of protein-ligand interactions. Comparing addnl. deep learning models trained on PDBbind datasets, we found that their enrichment performances using DUD-E are not superior to the performance of the docking program AutoDock Vina. Together, these results suggest that biases that could be present in constructed datasets should be thoroughly evaluated before applying them to machine learning based methodol. development.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXhvFeks7rL&md5=637472aa275cbacf3e926983f3795726
    7. 7
      Wallach, I.; Heifets, A. Most Ligand-Based Classification Benchmarks Reward Memorization Rather than Generalization. J. Chem. Inf. Model. 2018, 58, 916– 932,  DOI: 10.1021/acs.jcim.7b00403
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      7
      Most Ligand-Based Classification Benchmarks Reward Memorization Rather than Generalization
      Wallach, Izhar; Heifets, Abraham
      Journal of Chemical Information and Modeling (2018), 58 (5), 916-932CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      Undetected overfitting can occur when there are significant redundancies between training and validation data. The authors describe AVE, a new measure of training-validation redundancy for ligand-based classification problems, that accounts for the similarity among inactive mols. as well as active ones. The authors investigated seven widely used benchmarks for virtual screening and classification, and the authors show that the amt. of AVE bias strongly correlates with the performance of ligand-based predictive methods irresp. of the predicted property, chem. fingerprint, similarity measure, or previously applied unbiasing techniques. Therefore, it may be the case that the previously reported performance of most ligand-based methods can be explained by overfitting to benchmarks rather than good prospective accuracy.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXotlCgs7s%253D&md5=a59fb9ba04b1e6b37cb83594625081ff
    8. 8
      Wouters, O. J.; McKee, M.; Luyten, J. Estimated Research and Development Investment Needed to Bring a New Medicine to Market, 2009–2018. JAMA, J. Am. Med. Assoc. 2020, 323, 844– 853,  DOI: 10.1001/jama.2020.1166
      [Crossref], [PubMed], [CAS], Google Scholar
      8
      Estimated Research and Development Investment Needed to Bring a New Medicine to Market, 2009-2018
      Wouters Olivier J; McKee Martin; Luyten Jeroen
      JAMA (2020), 323 (9), 844-853 ISSN:.
      IMPORTANCE: The mean cost of developing a new drug has been the subject of debate, with recent estimates ranging from $314 million to $2.8 billion. OBJECTIVE: To estimate the research and development investment required to bring a new therapeutic agent to market, using publicly available data. DESIGN AND SETTING: Data were analyzed on new therapeutic agents approved by the US Food and Drug Administration (FDA) between 2009 and 2018 to estimate the research and development expenditure required to bring a new medicine to market. Data were accessed from the US Securities and Exchange Commission, Drugs@FDA database, and ClinicalTrials.gov, alongside published data on clinical trial success rates. EXPOSURES: Conduct of preclinical and clinical studies of new therapeutic agents. MAIN OUTCOMES AND MEASURES: Median and mean research and development spending on new therapeutic agents approved by the FDA, capitalized at a real cost of capital rate (the required rate of return for an investor) of 10.5% per year, with bootstrapped CIs. All amounts were reported in 2018 US dollars. RESULTS: The FDA approved 355 new drugs and biologics over the study period. Research and development expenditures were available for 63 (18%) products, developed by 47 different companies. After accounting for the costs of failed trials, the median capitalized research and development investment to bring a new drug to market was estimated at $985.3 million (95% CI, $683.6 million-$1228.9 million), and the mean investment was estimated at $1335.9 million (95% CI, $1042.5 million-$1637.5 million) in the base case analysis. Median estimates by therapeutic area (for areas with ≥5 drugs) ranged from $765.9 million (95% CI, $323.0 million-$1473.5 million) for nervous system agents to $2771.6 million (95% CI, $2051.8 million-$5366.2 million) for antineoplastic and immunomodulating agents. Data were mainly accessible for smaller firms, orphan drugs, products in certain therapeutic areas, first-in-class drugs, therapeutic agents that received accelerated approval, and products approved between 2014 and 2018. Results varied in sensitivity analyses using different estimates of clinical trial success rates, preclinical expenditures, and cost of capital. CONCLUSIONS AND RELEVANCE: This study provides an estimate of research and development costs for new therapeutic agents based on publicly available data. Differences from previous studies may reflect the spectrum of products analyzed, the restricted availability of data in the public domain, and differences in underlying assumptions in the cost calculations.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A280%3ADC%252BB387oslyhsA%253D%253D&md5=ff63c2f03e59ef0223585a71fb6d5df0
    9. 9
      Morris, G. M.; Huey, R.; Lindstrom, W.; Sanner, M. F.; Belew, R. K.; Goodsell, D. S.; Olson, A. J. AutoDock4 and AutoDockTools4: Automated docking with selective receptor flexibility. J. Comput. Chem. 2009, 30, 2785– 2791,  DOI: 10.1002/jcc.21256
      [Crossref], [PubMed], [CAS], Google Scholar
      9
      AutoDock and AutoDockTools: Automated docking with selective receptor flexibility
      Morris, Garrett M.; Huey, Ruth; Lindstrom, William; Sanner, Michel F.; Belew, Richard K.; Goodsell, David S.; Olson, Arthur J.
      Journal of Computational Chemistry (2009), 30 (16), 2785-2791CODEN: JCCHDD; ISSN:0192-8651. (John Wiley & Sons, Inc.)
      We describe the testing and release of AutoDock4 and the accompanying graphical user interface AutoDockTools. AutoDock4 incorporates limited flexibility in the receptor. Several tests are reported here, including a redocking expt. with 188 diverse ligand-protein complexes and a cross-docking expt. using flexible sidechains in 87 HIV protease complexes. We also report its utility in anal. of covalently bound ligands, using both a grid-based docking method and a modification of the flexible sidechain technique. © 2009 Wiley Periodicals, Inc. J Comput Chem, 2009.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BD1MXht1GitrnK&md5=679ce22fc50e9291c9aa16e7a1855845
    10. 10
      Verdonk, M. L.; Cole, J. C.; Hartshorn, M. J.; Murray, C. W.; Taylor, R. D. Improved protein–ligand docking using GOLD. Proteins: Struct., Funct., Bioinf. 2003, 52, 609– 623,  DOI: 10.1002/prot.10465
      [Crossref], [PubMed], [CAS], Google Scholar
      10
      Improved protein-ligand docking using GOLD
      Verdonk, Marcel L.; Cole, Jason C.; Hartshorn, Michael J.; Murray, Christopher W.; Taylor, Richard D.
      Proteins: Structure, Function, and Genetics (2003), 52 (4), 609-623CODEN: PSFGEY; ISSN:0887-3585. (Wiley-Liss, Inc.)
      The Chemscore function was implemented as a scoring function for the protein-ligand docking program GOLD, and its performance compared to the original Goldscore function and two consensus docking protocols, "Goldscore-CS" and "Chemscore-GS," in terms of docking accuracy, prediction of binding affinities, and speed. In the "Goldscore-CS" protocol, dockings produced with the Goldscore function are scored and ranked with the Chemscore function; in the "Chemscore-GS" protocol, dockings produced with the Chemscore function are scored and ranked with the Goldscore function. Comparisons were made for a "clean" set of 224 protein-ligand complexes, and for two subsets of this set, one for which the ligands are "drug-like," the other for which they are "fragment-like.". For "drug-like" and "fragment-like" ligands, the docking accuracies obtained with Chemscore and Goldscore functions are similar. For larger ligands, Goldscore gives superior results. Docking with the Chemscore function is up to three times faster than docking with the Goldscore function. Both combined docking protocols give significant improvements in docking accuracy over the use of the Goldscore or Chemscore function alone. "Goldscore-CS" gives success rates of up to 81% (top-ranked GOLD soln. within 2.0 Å of the exptl. binding mode) for the "clean list," but at the cost of long search times. For most virtual screening applications, "Chemscore-GS" seems optimal; search settings that give docking speeds of around 0.25-1.3 min/compd. have success rates of about 78% for "drug-like" compds. and 85% for "fragment-like" compds. In terms of producing binding energy ests., the Goldscore function appears to perform better than the Chemscore function and the two consensus protocols, particularly for faster search settings. Even at docking speeds of around 1-2 min/compd., the Goldscore function predicts binding energies with a std. deviation of ∼10.5 kJ/mol.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BD3sXmvFGrsLg%253D&md5=73a627d99dab6de1ae364c8e8e5f0fea
    11. 11
      Hornik, K.; Stinchcombe, M.; White, H. Multilayer feedforward networks are universal approximators. Neural Netw 1989, 2, 359– 366,  DOI: 10.1016/0893-6080(89)90020-8
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    12. 12
      Shen, C.; Ding, J.; Wang, Z.; Cao, D.; Ding, X.; Hou, T. From machine learning to deep learning: Advances in scoring functions for protein–ligand docking. Wiley Interdiscip. Rev.: Comput. Mol. Sci. 2020, 10, e1429  DOI: 10.1002/wcms.1429
      [Crossref], [CAS], Google Scholar
      12
      From machine learning to deep learning: Advances in scoring functions for protein-ligand docking
      Shen, Chao; Ding, Junjie; Wang, Zhe; Cao, Dongsheng; Ding, Xiaoqin; Hou, Tingjun
      Wiley Interdisciplinary Reviews: Computational Molecular Science (2020), 10 (1), e1429CODEN: WIRCAH; ISSN:1759-0884. (Wiley-Blackwell)
      A review. Mol. docking has been regarded as a routine tool for drug discovery, but its accuracy highly depends on the reliability of scoring functions (SFs). With the rapid development of machine learning (ML) techniques, ML-based SFs have gradually emerged as a promising alternative for protein-ligand binding affinity prediction and virtual screening, and most of them have shown significantly better performance than a wide range of classical SFs. Emergence of more data-hungry deep learning (DL) approaches in recent years further fascinates the exploitation of more accurate SFs. Here, we summarize the progress of traditional ML-based SFs in the last few years and provide insights into recently developed DL-based SFs. We believe that the continuous improvement in ML-based SFs can surely guide the early-stage drug design and accelerate the discovery of new drugs. This article is categorized under: Computer and Information Science > Chemoinformatics.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXitlyjsbbN&md5=0dd878e211fad64fada03e883a5ab47a
    13. 13
      Jones, D.; Kim, H.; Zhang, X.; Zemla, A.; Stevenson, G.; Bennett, W. F. D.; Kirshner, D.; Wong, S. E.; Lightstone, F. C.; Allen, J. E. Improved Protein–Ligand Binding Affinity Prediction with Structure-Based Deep Fusion Inference. J. Chem. Inf. Model. 2021, 61, 1583– 1592,  DOI: 10.1021/acs.jcim.0c01306
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      13
      Improved Protein-Ligand Binding Affinity Prediction with Structure-Based Deep Fusion Inference
      Jones, Derek; Kim, Hyojin; Zhang, Xiaohua; Zemla, Adam; Stevenson, Garrett; Bennett, W. F. Drew; Kirshner, Daniel; Wong, Sergio E.; Lightstone, Felice C.; Allen, Jonathan E.
      Journal of Chemical Information and Modeling (2021), 61 (4), 1583-1592CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      Predicting accurate protein-ligand binding affinities is an important task in drug discovery but remains a challenge even with computationally expensive biophysics-based energy scoring methods and state-of-the-art deep learning approaches. Despite the recent advances in the application of deep convolutional and graph neural network-based approaches, it remains unclear what the relative advantages of each approach are and how they compare with physics-based methodologies that have found more mainstream success in virtual screening pipelines. We present fusion models that combine features and inference from complementary representations to improve binding affinity prediction. This, to our knowledge, is the first comprehensive study that uses a common series of evaluations to directly compare the performance of three-dimensional (3D)-convolutional neural networks (3D-CNNs), spatial graph neural networks (SG-CNNs), and their fusion. We use temporal and structure-based splits to assess performance on novel protein targets. To test the practical applicability of our models, we examine their performance in cases that assume that the crystal structure is not available. In these cases, binding free energies are predicted using docking pose coordinates as the inputs to each model. In addn., we compare these deep learning approaches to predictions based on docking scores and mol. mechanic/generalized Born surface area (MM/GBSA) calcns. Our results show that the fusion models make more accurate predictions than their constituent neural network models as well as docking scoring and MM/GBSA rescoring, with the benefit of greater computational efficiency than the MM/GBSA method. Finally, we provide the code to reproduce our results and the parameter files of the trained models used in this work. The software is available as open source at https://github.com/llnl/fast. Model parameter files are available at ftp://gdo-bioinformatics.ucllnl.org/fast/pdbbind2016_model_checkpoints/.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3MXmvFaqtrc%253D&md5=11f92ba4bae7dabc75b9e55c1911352b
    14. 14
      Chenthamarakshan, V.; Das, P.; Hoffman, S.; Strobelt, H.; Padhi, I.; Lim, K. W.; Hoover, B.; Manica, M.; Born, J.; Laino, T.; Mojsilovic, A. CogMol: Target-Specific and Selective Drug Design for COVID-19 Using Deep Generative Models. In 34th Conference on Neural Information Processing Systems , 2020; pp 4320– 4332.
      Google Scholar
      There is no corresponding record for this reference.
    15. 15
      Hoffman, S. C.; Chenthamarakshan, V.; Wadhawan, K.; Chen, P.-Y.; Das, P. Optimizing molecules using efficient queries from property evaluations. Nat. Mach. Intell. 2022, 4, 21,  DOI: 10.1038/s42256-021-00422-y
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    16. 16
      Boyles, F.; Deane, C. M.; Morris, G. M. Learning from Docked Ligands: Ligand-Based Features Rescue Structure-Based Scoring Functions When Trained on Docked Poses. J. Chem. Inf. Model. 2022, 62, 5329– 5341,  DOI: 10.1021/acs.jcim.1c00096
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      16
      Learning from Docked Ligands: Ligand-Based Features Rescue Structure-Based Scoring Functions When Trained on Docked Poses
      Boyles, Fergus; Deane, Charlotte M.; Morris, Garrett M.
      Journal of Chemical Information and Modeling (2022), 62 (22), 5329-5341CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      Machine learning scoring functions for protein-ligand binding affinity have been found to consistently outperform classical scoring functions when trained and tested on crystal structures of bound protein-ligand complexes. However, it is less clear how these methods perform when applied to docked poses of complexes. We explore how the use of docked rather than crystallog. poses for both training and testing affects the performance of machine learning scoring functions. Using the PDBbind Core Sets as benchmarks, we show that the performance of a structure-based machine learning scoring function trained and tested on docked poses is lower than that of the same scoring function trained and tested on crystallog. poses. We construct a hybrid scoring function by combining both structure-based and ligand-based features, and show that its ability to predict binding affinity using docked poses is comparable to that of purely structure-based scoring functions trained and tested on crystal poses. We also present a new, freely available validation set-the Updated DUD-E Diverse Subset-for binding affinity prediction using data from DUD-E and ChEMBL. Despite strong performance on docked poses of the PDBbind Core Sets, we find that our hybrid scoring function sometimes generalizes poorly to a protein target not represented in the training set, demonstrating the need for improved scoring functions and addnl. validation benchmarks.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3MXhvFWms7rN&md5=2c539500f49ad27ff6682ff11e9360eb
    17. 17
      Su, M.; Yang, Q.; Du, Y.; Feng, G.; Liu, Z.; Li, Y.; Wang, R. Comparative Assessment of Scoring Functions: The CASF-2016 Update. J. Chem. Inf. Model. 2019, 59, 895– 913,  DOI: 10.1021/acs.jcim.8b00545
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      17
      Comparative Assessment of Scoring Functions: The CASF-2016 Update
      Su, Minyi; Yang, Qifan; Du, Yu; Feng, Guoqin; Liu, Zhihai; Li, Yan; Wang, Renxiao
      Journal of Chemical Information and Modeling (2019), 59 (2), 895-913CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      In structure-based drug design, scoring functions are often employed to evaluate protein-ligand interactions. A variety of scoring functions have been developed so far, and thus, some objective benchmarks are desired for assessing their strength and weakness. The comparative assessment of scoring functions (CASF) benchmark developed by us provides an answer to this demand. CASF is designed as a "scoring benchmark", where the scoring process is decoupled from the docking process to depict the performance of scoring function more precisely. Here, we describe the latest update of this benchmark, i.e., CASF-2016. Each scoring function is still evaluated by four metrics, including "scoring power", "ranking power", "docking power", and "screening power". Nevertheless, the evaluation methods have been improved considerably in several aspects. A new test set is compiled, which consists of 285 protein-ligand complexes with high-quality crystal structures and reliable binding consts. A panel of 25 scoring functions are tested on CASF-2016 as a demonstration. Our results reveal that the performance of current scoring functions is more promising in terms of docking power than scoring, ranking, and screening power. Scoring power is somewhat correlated with ranking power, so are docking power and screening power. The results obtained on CASF-2016 may provide valuable guidance for the end users to make smart choices among available scoring functions. Moreover, CASF is created as an open-access benchmark so that other researchers can utilize it to test a wider range of scoring functions. The complete CASF-2016 benchmark will be released on the PDBbind-CN web server (http://www.pdbbind-cn.org/casf.asp/) once this article is published.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXitlWhtLjM&md5=573295cfd2298a2dee933027041c3d9c
    18. 18
      Liu, Z.; Su, M.; Han, L.; Liu, J.; Yang, Q.; Li, Y.; Wang, R. Forging the Basis for Developing Protein–Ligand Interaction Scoring Functions. Acc. Chem. Res. 2017, 50, 302– 309,  DOI: 10.1021/acs.accounts.6b00491
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      18
      Forging the Basis for Developing Protein-Ligand Interaction Scoring Functions
      Liu, Zhihai; Su, Minyi; Han, Li; Liu, Jie; Yang, Qifan; Li, Yan; Wang, Renxiao
      Accounts of Chemical Research (2017), 50 (2), 302-309CODEN: ACHRE4; ISSN:0001-4842. (American Chemical Society)
      A review. In structure-based drug design, scoring functions are widely used for fast evaluation of protein-ligand interactions. They are often applied in combination with mol. docking and de novo design methods. Since the early 1990s, a whole spectrum of protein-ligand interaction scoring functions have been developed. Regardless of their tech. difference, scoring functions all need data sets combining protein-ligand complex structures and binding affinity data for parametrization and validation. However, data sets of this kind used to be rather limited in terms of size and quality. However, std. metrics for evaluating scoring function used to be ambiguous. Scoring functions are often tested in mol. docking or even virtual screening trials, which are not totally appropriate for reflecting the genuine quality of scoring functions. Collectively, these underlying obstacles have impeded the invention of more advanced scoring functions. In this Account, the authors describe their long-lasting efforts on overcoming these obstacles, which involve two related projects. On the first project, the authors have created the PDBbind database. It is the first database that systematically annotates the protein-ligand complexes in the Protein Data Bank (PDB) with exptl. binding data. This database has been updated annually since its first public release in 2004. The latest release (version 2016) provides binding data for 16,179 biomol. complexes in PDB. Data sets provided by PDBbind have been applied to many computational and statistical studies on protein-ligand interaction and various subjects. In particular, it has become a major data resource for scoring function development. On the second project, the authors established the Comparative Assessment of Scoring Functions (CASF) benchmark for scoring function evaluation. The authors' key idea is to decouple the "scoring" process from the "sampling" process, so scoring functions can be tested in a relatively pure context to reflect their quality. In the authors' latest work on this track, i.e., CASF-2013, the performance of a scoring function was quantified in four aspects, including "scoring power", "ranking power", "docking power", and "screening power". All four performance tests were conducted on a test set contg. 195 high-quality protein-ligand complexes selected from PDBbind. A panel of 20 std. scoring functions were tested as demonstration. Importantly, CASF is designed to be an open-access benchmark, with which scoring functions developed by different researchers can be compared on the same ground. Indeed, it has become a popular choice for scoring function validation in recent years. Despite the considerable progress that has been made so far, the performance of today's scoring functions still does not meet people's expectation in many aspects. There is a const. demand for more advanced scoring functions. The authors' efforts have helped to overcome some obstacles underlying scoring function development so that the researchers in this field can move forward faster. The authors will continue to improve the PDBbind database and the CASF benchmark in the future to keep them as useful community resources.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC2sXit12qsLc%253D&md5=48d7287ef58c3ed8d9ff7579014fc486
    19. 19
      Yosinski, J.; Clune, J.; Nguyen, A.; Fuchs, T.; Lipson, H. Understanding neural networks through deep visualization. arXiv 2015, arXiv:1506.06579,  DOI: 10.48550/arXiv.1506.06579
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    20. 20
      Samek, W.; Binder, A.; Montavon, G.; Bach, S.; Müller, K.-R. Evaluating the visualization of what a Deep Neural Network has learned. arXiv 2015, arXiv:1509.06321,  DOI: 10.48550/arXiv.1509.06321
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    21. 21
      Poelking, C.; Chessari, G.; Murray, C. W.; Hall, R. J.; Colwell, L.; Verdonk, M. Meaningful machine learning models and machine-learned pharmacophores from fragment screening campaigns. arXiv 2022, arXiv:2204.06348,  DOI: 10.48550/arXiv.2204.06348
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    22. 22
      Sharma, B.; Chenthamarakshan, V.; Dhurandhar, A.; Pereira, S.; Hendler, J. A.; Dordick, J. S.; Das, P. Accurate Clinical Toxicity Prediction using Multi-task Deep Neural Nets and Contrastive Molecular Explanations. Sci. Rep. 2023, 13, 4908,  DOI: 10.1038/s41598-023-31169-8
      [Crossref], [PubMed], [CAS], Google Scholar
      22
      Accurate clinical toxicity prediction using multi-task deep neural nets and contrastive molecular explanations
      Sharma, Bhanushee; Chenthamarakshan, Vijil; Dhurandhar, Amit; Pereira, Shiranee; Hendler, James A.; Dordick, Jonathan S.; Das, Payel
      Scientific Reports (2023), 13 (1), 4908CODEN: SRCEC3; ISSN:2045-2322. (Nature Portfolio)
      Explainable machine learning for mol. toxicity prediction is a promising approach for efficient drug development and chem. safety. A predictive ML model of toxicity can reduce exptl. cost and time while mitigating ethical concerns by significantly reducing animal and clin. testing. Herein, we use a deep learning framework for simultaneously modeling in vitro, in vivo, and clin. toxicity data. Two different mol. input representations are used; Morgan fingerprints and pre-trained SMILES embeddings. A multi-task deep learning model accurately predicts toxicity for all endpoints, including clin., as indicated by the area under the Receiver Operator Characteristic curve and balanced accuracy. In particular, pre-trained mol. SMILES embeddings as input to the multi-task model improved clin. toxicity predictions compared to existing models in MoleculeNet benchmark. Addnl., our multitask approach is comprehensive in the sense that it is comparable to state-of-the-art approaches for specific endpoints in in vitro, in vivo and clin. platforms. Through both the multi-task model and transfer learning, we were able to indicate the minimal need of in vivo data for clin. toxicity predictions. To provide confidence and explain the model's predictions, we adapt a post-hoc contrastive explanation method that returns pertinent pos. and neg. features, which correspond well to known mutagenic and reactive toxicophores, such as unsubstituted bonded heteroatoms, arom. amines, and Michael receptors. Furthermore, toxicophore recovery by pertinent feature anal. captures more of the in vitro (53%) and in vivo (56%), rather than of the clin. (8%), endpoints, and indeed uncovers a preference in known toxicophore data towards in vitro and in vivo exptl. data. To our knowledge, this is the first contrastive explanation, using both present and absent substructures, for predictions of clin. and in vivo mol. toxicity.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3sXmsV2ks7g%253D&md5=63f1d0798d9ba9b71f3d970b58e752ea
    23. 23
      Hadfield, T. E.; Scantlebury, J.; Deane, C. M. Exploring The Ability Of Machine Learning-Based Virtual Screening Models To Identify The Functional Groups Responsible For Binding. bioRxiv 2023,  DOI: 10.1101/2023.04.29.538820
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    24. 24
      McNutt, A. T.; Francoeur, P.; Aggarwal, R.; Masuda, T.; Meli, R.; Ragoza, M.; Sunseri, J.; Koes, D. R. GNINA 1.0: molecular docking with deep learning. J. Cheminf. 2021, 13, 43,  DOI: 10.1186/s13321-021-00522-2
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    25. 25
      Zheng, L.; Fan, J.; Mu, Y. OnionNet: a Multiple-Layer Intermolecular-Contact-Based Convolutional Neural Network for Protein–Ligand Binding Affinity Prediction. ACS Omega 2019, 4, 15956– 15965,  DOI: 10.1021/acsomega.9b01997
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      24
      OnionNet: a Multiple-Layer Intermolecular-Contact-Based Convolutional Neural Network for Protein-Ligand Binding Affinity Prediction
      Zheng, Liangzhen; Fan, Jingrong; Mu, Yuguang
      ACS Omega (2019), 4 (14), 15956-15965CODEN: ACSODF; ISSN:2470-1343. (American Chemical Society)
      Computational drug discovery provides an efficient tool for helping large-scale lead mol. screening. One of the major tasks of lead discovery is identifying mols. with promising binding affinities toward a target, a protein in general. The accuracies of current scoring functions that are used to predict the binding affinity are not satisfactory enough. Thus, machine learning or deep learning based methods have been developed recently to improve the scoring functions. In this study, a deep convolutional neural network model (called OnionNet) is introduced; its features are based on rotation-free element-pair-specific contacts between ligands and protein atoms, and the contacts are further grouped into different distance ranges to cover both the local and nonlocal interaction information between the ligand and the protein. The prediction power of the model is evaluated and compared with other scoring functions using the comparative assessment of scoring functions (CASF-2013) benchmark and the v2016 core set of the PDBbind database. The robustness of the model is further explored by predicting the binding affinities of the complexes generated from docking simulations instead of exptl. detd. PDB structures.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXhslKhtb3J&md5=0eaf1406acc14b7052d62481de5dec78
    26. 26
      Li, Y.; Rezaei, M. A.; Li, C.; Li, X. DeepAtom: A Framework for Protein-Ligand Binding Affinity Prediction. In 2019 IEEE International Conference on Bioinformatics and Biomedicine (BIBM) ; 2019; pp 303– 310.
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    27. 27
      Feinberg, E. N.; Sur, D.; Wu, Z.; Husic, B. E.; Mai, H.; Li, Y.; Sun, S.; Yang, J.; Ramsundar, B.; Pande, V. S. PotentialNet for molecular property prediction. ACS Cent. Sci. 2018, 4, 1520– 1530,  DOI: 10.1021/acscentsci.8b00507
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      26
      PotentialNet for Molecular Property Prediction
      Feinberg, Evan N.; Sur, Debnil; Wu, Zhenqin; Husic, Brooke E.; Mai, Huanghao; Li, Yang; Sun, Saisai; Yang, Jianyi; Ramsundar, Bharath; Pande, Vijay S.
      ACS Central Science (2018), 4 (11), 1520-1530CODEN: ACSCII; ISSN:2374-7951. (American Chemical Society)
      The arc of drug discovery entails a multiparameter optimization problem spanning vast length scales. The key parameters range from soly. (angstroms) to protein-ligand binding (nanometers) to in vivo toxicity (meters). Through feature learning-instead of feature engineering-deep neural networks promise to outperform both traditional physics-based and knowledge-based machine learning models for predicting mol. properties pertinent to drug discovery. To this end, we present the PotentialNet family of graph convolutions. These models are specifically designed for and achieve state-of-the-art performance for protein-ligand binding affinity. We further validate these deep neural networks by setting new stds. of performance in several ligand-based tasks. In parallel, we introduce a new metric, the Regression Enrichment Factor EF(R)χ, to measure the early enrichment of computational models for chem. data. Finally, we introduce a cross-validation strategy based on structural homol. clustering that can more accurately measure model generalizability, which crucially distinguishes the aims of machine learning for drug discovery from std. machine learning tasks.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXitVGrs7fF&md5=570999e7f212d7c737b24f71665893b7
    28. 28
      Morrone, J. A.; Weber, J. K.; Huynh, T.; Luo, H.; Cornell, W. D. Combining Docking Pose Rank and Structure with Deep Learning Improves Protein–Ligand Binding Mode Prediction over a Baseline Docking Approach. J. Chem. Inf. Model. 2020, 60, 4170– 4179,  DOI: 10.1021/acs.jcim.9b00927
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      27
      Combining Docking Pose Rank and Structure with Deep Learning Improves Protein-Ligand Binding Mode Prediction over a Baseline Docking Approach
      Morrone, Joseph A.; Weber, Jeffrey K.; Huynh, Tien; Luo, Heng; Cornell, Wendy D.
      Journal of Chemical Information and Modeling (2020), 60 (9), 4170-4179CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      The authors present a simple, modular graph-based convolutional neural network that takes structural information from protein-ligand complexes as input to generate models for activity and binding mode prediction. Complex structures are generated by a std. docking procedure and fed into a dual-graph architecture that includes sep. subnetworks for the ligand bonded topol. and the ligand-protein contact map. Recent work has indicated that data set bias drives many past promising results derived from combining deep learning and docking. The dual-graph network allows contributions from ligand identity that give rise to such biases to be distinguished from effects of protein-ligand interactions on classification. The authors show that the neural network is capable of learning from protein structural information when, as in the case of binding mode prediction, an unbiased data set is constructed. The authors next develop a deep learning model for binding mode prediction that uses docking ranking as input in combination with docking structures. This strategy mirrors past consensus models and outperforms a baseline docking program (AutoDock Vina) in a variety of tests, including on cross-docking data sets that mimic real-world docking use cases. Furthermore, the magnitudes of network predictions serve as reliable measures of model confidence.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXjsVOgtbs%253D&md5=ed7f0ba1797757a61d64c817db36cc13
    29. 29
      Jiang, D.; Hsieh, C.-Y.; Wu, Z.; Kang, Y.; Wang, J.; Wang, E.; Liao, B.; Shen, C.; Xu, L.; Wu, J.; Cao, D.; Hou, T. InteractionGraphNet: A Novel and Efficient Deep Graph Representation Learning Framework for Accurate Protein–Ligand Interaction Predictions. J. Med. Chem. 2021, 64, 18209– 18232,  DOI: 10.1021/acs.jmedchem.1c01830
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      28
      InteractionGraphNet: A Novel and Efficient Deep Graph Representation Learning Framework for Accurate Protein-Ligand Interaction Predictions
      Jiang, Dejun; Hsieh, Chang-Yu; Wu, Zhenxing; Kang, Yu; Wang, Jike; Wang, Ercheng; Liao, Ben; Shen, Chao; Xu, Lei; Wu, Jian; Cao, Dongsheng; Hou, Tingjun
      Journal of Medicinal Chemistry (2021), 64 (24), 18209-18232CODEN: JMCMAR; ISSN:0022-2623. (American Chemical Society)
      Accurate quantification of protein-ligand interactions remains a key challenge to structure-based drug design. However, traditional machine learning (ML)-based methods based on handcrafted descriptors, one-dimensional protein sequences, and/or two-dimensional graph representations limit their capability to learn the generalized mol. interactions in 3D space. Here, we proposed a novel deep graph representation learning framework named InteractionGraphNet (IGN) to learn the protein-ligand interactions from the 3D structures of protein-ligand complexes. In IGN, two independent graph convolution modules were stacked to sequentially learn the intramol. and intermol. interactions, and the learned intermol. interactions can be efficiently used for subsequent tasks. Extensive binding affinity prediction, large-scale structure-based virtual screening, and pose prediction expts. demonstrated that IGN achieved better or competitive performance against other state-of-the-art ML-based baselines and docking programs. More importantly, such state-of-the-art performance was proven from the successful learning of the key features in protein-ligand interactions instead of just memorizing certain biased patterns from data.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3MXislSnsrnE&md5=bbee6eb73a8c6b95c389fe11e45c7723
    30. 30
      Hochuli, J.; Helbling, A.; Skaist, T.; Ragoza, M.; Koes, D. R. Visualizing convolutional neural network protein-ligand scoring. J. Mol. Graphics Modell. 2018, 84, 96– 108,  DOI: 10.1016/j.jmgm.2018.06.005
      [Crossref], [PubMed], [CAS], Google Scholar
      29
      Visualizing convolutional neural network protein-ligand scoring
      Hochuli, Joshua; Helbling, Alec; Skaist, Tamar; Ragoza, Matthew; Koes, David Ryan
      Journal of Molecular Graphics & Modelling (2018), 84 (), 96-108CODEN: JMGMFI; ISSN:1093-3263. (Elsevier Ltd.)
      Protein-ligand scoring is an important step in a structure-based drug design pipeline. Selecting a correct binding pose and predicting the binding affinity of a protein-ligand complex enables effective virtual screening. Machine learning techniques can make use of the increasing amts. of structural data that are becoming publicly available. Convolutional neural network (CNN) scoring functions in particular have shown promise in pose selection and affinity prediction for protein-ligand complexes. Neural networks are known for being difficult to interpret. Understanding the decisions of a particular network can help tune parameters and training data to maximize performance. Visualization of neural networks helps decomp. complex scoring functions into pictures that are more easily parsed by humans. Here we present three methods for visualizing how individual protein-ligand complexes are interpreted by 3D convolutional neural networks. We also present a visualization of the convolutional filters and their wts. We describe how the intuition provided by these visualizations aids in network design.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1cXhtF2ltb3P&md5=db697d859132f02c4a2cef4bfc00ea88
    31. 31
      Lim, J.; Ryu, S.; Park, K.; Choe, Y. J.; Ham, J.; Kim, W. Y. Predicting Drug–Target Interaction Using a Novel Graph Neural Network with 3D Structure-Embedded Graph Representation. J. Chem. Inf. Model. 2019, 59, 3981– 3988,  DOI: 10.1021/acs.jcim.9b00387
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      30
      Predicting Drug-Target Interaction Using a Novel Graph Neural Network with 3D Structure-Embedded Graph Representation
      Lim, Jaechang; Ryu, Seongok; Park, Kyubyong; Choe, Yo Joong; Ham, Jiyeon; Kim, Woo Youn
      Journal of Chemical Information and Modeling (2019), 59 (9), 3981-3988CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      We propose a novel deep learning approach for predicting drug-target interaction using a graph neural network. We introduce a distance-aware graph attention algorithm to differentiate various types of intermol. interactions. Furthermore, we ext. the graph feature of intermol. interactions directly from the 3D structural information on the protein-ligand binding pose. Thus, the model can learn key features for accurate predictions of drug-target interaction rather than just memorize certain patterns of ligand mols. As a result, our model shows better performance than docking and other deep learning methods for both virtual screening (AUROC of 0.968 for the DUD-E test set) and pose prediction (AUROC of 0.935 for the PDBbind test set). In addn., it can reproduce the natural population distribution of active mols. and inactive mols.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXhs1egur3F&md5=670cd989a369d67c734f6762a5a906cf
    32. 32
      Hadfield, T. E.; Imrie, F.; Merritt, A.; Birchall, K.; Deane, C. M. Incorporating Target-Specific Pharmacophoric Information into Deep Generative Models for Fragment Elaboration. J. Chem. Inf. Model. 2022, 62, 2280– 2292,  DOI: 10.1021/acs.jcim.1c01311
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      31
      Incorporating Target-Specific Pharmacophoric Information into Deep Generative Models for Fragment Elaboration
      Hadfield, Thomas E.; Imrie, Fergus; Merritt, Andy; Birchall, Kristian; Deane, Charlotte M.
      Journal of Chemical Information and Modeling (2022), 62 (10), 2280-2292CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      Despite recent interest in deep generative models for scaffold elaboration, their applicability to fragment-to-lead campaigns has so far been limited. This is primarily due to their inability to account for local protein structure or a user's design hypothesis. We propose a novel method for fragment elaboration, STRIFE, that overcomes these issues. STRIFE takes as input fragment hotspot maps (FHMs) extd. from a protein target and processes them to provide meaningful and interpretable structural information to its generative model, which in turn is able to rapidly generate elaborations with complementary pharmacophores to the protein. In a large-scale evaluation, STRIFE outperforms existing, structure-unaware, fragment elaboration methods in proposing highly ligand-efficient elaborations. In addn. to automatically extg. pharmacophoric information from a protein target's FHM, STRIFE optionally allows the user to specify their own design hypotheses.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB38XhtFGru7rE&md5=c639b01f0c509bbd2e20308fa044bdb5
    33. 33
      Curran, P. R.; Radoux, C. J.; Smilova, M. D.; Sykes, R. A.; Higueruelo, A. P.; Bradley, A. R.; Marsden, B. D.; Spring, D. R.; Blundell, T. L.; Leach, A. R.; Pitt, W. R.; Cole, J. C. Hotspots API: A Python Package for the Detection of Small Molecule Binding Hotspots and Application to Structure-Based Drug Design. J. Chem. Inf. Model. 2020, 60, 1911– 1916,  DOI: 10.1021/acs.jcim.9b00996
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      32
      Hotspots API: A Python Package for the Detection of Small Molecule Binding Hotspots and Application to Structure-Based Drug Design
      Curran, Peter R.; Radoux, Chris J.; Smilova, Mihaela D.; Sykes, Richard A.; Higueruelo, Alicia P.; Bradley, Anthony R.; Marsden, Brian D.; Spring, David R.; Blundell, Tom L.; Leach, Andrew R.; Pitt, William R.; Cole, Jason C.
      Journal of Chemical Information and Modeling (2020), 60 (4), 1911-1916CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      Methods that survey protein surfaces for binding hotspots can help to evaluate target tractability and guide exploration of potential ligand binding regions. Fragment Hotspot Maps builds upon interaction data mined from the CSD (Cambridge Structural Database) and exploits the idea of identifying hotspots using small chem. fragments, which is now widely used to design new drug leads. Prior to this publication, Fragment Hotspot Maps was only publicly available through a web application. To increase the accessibility of this algorithm we present the Hotspots API (application programming interface), a toolkit that offers programmatic access to the core Fragment Hotspot Maps algorithm, thereby facilitating the interpretation and application of 3the anal. To demonstrate the package's utility, we present a workflow which automatically derives protein hydrogen-bond constraints for mol. docking with GOLD. The Hotspots API is available from https://github.com/prcurran/hotspots under the MIT license and is dependent upon the com. CSD Python API.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXlsFahsbw%253D&md5=9402dded4d637e5b324e329d9e2c9a08
    34. 34
      Salentin, S.; Schreiber, S.; Haupt, V. J.; Adasme, M. F.; Schroeder, M. PLIP: fully automated protein–ligand interaction profiler. Nucleic Acids Res. 2015, 43, W443– W447,  DOI: 10.1093/nar/gkv315
      [Crossref], [PubMed], [CAS], Google Scholar
      33
      PLIP: fully automated protein-ligand interaction profiler
      Salentin, Sebastian; Schreiber, Sven; Haupt, V. Joachim; Adasme, Melissa F.; Schroeder, Michael
      Nucleic Acids Research (2015), 43 (W1), W443-W447CODEN: NARHAD; ISSN:0305-1048. (Oxford University Press)
      The characterization of interactions in protein-ligand complexes is essential for research in structural bioinformatics, drug discovery and biol. However, comprehensive tools are not freely available to the research community. Here, we present the protein-ligand interaction profiler (PLIP), a novel web service for fully automated detection and visualization of relevant non-covalent protein-ligand contacts in 3D structures, freely available at projects.biotec.tu-dresden.de/plip-web. The input is either a Protein Data Bank structure, a protein or ligand name, or a custom protein-ligand complex (e.g. from docking). In contrast to other tools, the rule-based PLIP algorithm does not require any structure prepn. It returns a list of detected interactions on single atom level, covering seven interaction types (hydrogen bonds, hydrophobic contacts, pi-stacking, pi-cation interactions, salt bridges, water bridges and halogen bonds). PLIP stands out by offering publication-ready images, PyMOL sesions files to generate custom images and parsable result files to facilitate successive data processing. The full python source code is available for download on the website. PLIP's command-line mode allows for high-throughput interaction profiling.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC2sXhtVymtbrI&md5=4878881a7bd715e5585682876e102e65
    35. 35
      Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, L. u.; Polosukhin, I. Attention is All you Need. arXiv 2017, arXiv:1706.03762,  DOI: 10.48550/arXiv.1706.03762
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    36. 36
      Francoeur, P. G.; Masuda, T.; Sunseri, J.; Jia, A.; Iovanisci, R. B.; Snyder, I.; Koes, D. R. Three-Dimensional Convolutional Neural Networks and a Cross-Docked Data Set for Structure-Based Drug Design. J. Chem. Inf. Model. 2020, 60, 4200– 4215,  DOI: 10.1021/acs.jcim.0c00411
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      35
      Three-Dimensional Convolutional Neural Networks and a Cross-Docked Data Set for Structure-Based Drug Design
      Francoeur, Paul G.; Masuda, Tomohide; Sunseri, Jocelyn; Jia, Andrew; Iovanisci, Richard B.; Snyder, Ian; Koes, David R.
      Journal of Chemical Information and Modeling (2020), 60 (9), 4200-4215CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      One of the main challenges in drug discovery is predicting protein-ligand binding affinity. Recently, machine learning approaches have made substantial progress on this task. However, current methods of model evaluation are overly optimistic in measuring generalization to new targets, and there does not exist a std. data set of sufficient size to compare performance between models. We present a new data set for structure-based machine learning, the CrossDocked2020 set, with 22.5 million poses of ligands docked into multiple similar binding pockets across the Protein Data Bank, and perform a comprehensive evaluation of grid-based convolutional neural network (CNN) models on this data set. We also demonstrate how the partitioning of the training data and test data can impact the results of models trained with the PDBbind data set, how performance improves by adding more lower-quality training data, and how training with docked poses imparts pose sensitivity to the predicted affinity of a complex. Our best performing model, an ensemble of five densely connected CNNs, achieves a root mean squared error of 1.42 and Pearson R of 0.612 on the affinity prediction task, an AUC of 0.956 at binding pose classification, and a 68.4% accuracy at pose selection on the CrossDocked2020 set. By providing data splits for clustered cross-validation and the raw data for the CrossDocked2020 set, we establish the first standardized data set for training machine learning models to recognize ligands in noncognate target structures while also greatly expanding the no. of poses available for training. In order to facilitate community adoption of this data set for benchmarking protein-ligand binding affinity prediction, we provide our models, wts., and the CrossDocked2020 set at https://github.com/gnina/models.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXhslWqs73N&md5=5a5465187dbdef41d556baddd1a8e4d9
    37. 37
      Landrum, G. RDKit: Open-source cheminformatics ; 2016; https://www.bibsonomy.org/bibtex/28d01fceeccd6bf2486e47d7c4207b108/salotz.
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    38. 38
      O’Boyle, N. M.; Banck, M.; James, C. A.; Morley, C.; Vandermeersch, T.; Hutchison, G. R. Open Babel: An open chemical toolbox. J. Cheminf. 2011, 3, 33,  DOI: 10.1186/1758-2946-3-33
      [Crossref], [PubMed], [CAS], Google Scholar
      37
      Open Babel: an open chemical toolbox
      O'Boyle, Noel M.; Banck, Michael; James, Craig A.; Morley, Chris; Vandermeersch, Tim; Hutchison, Geoffrey R.
      Journal of Cheminformatics (2011), 3 (), 33CODEN: JCOHB3; ISSN:1758-2946. (Chemistry Central Ltd.)
      Background: A frequent problem in computational modeling is the interconversion of chem. structures between different formats. While std. interchange formats exist (for example, Chem. Markup Language) and de facto stds. have arisen (for example, SMILES format), the need to interconvert formats is a continuing problem due to the multitude of different application areas for chem. data, differences in the data stored by different formats (0D vs. 3D, for example), and competition between software along with a lack of vendor-neutral formats. Results: We discuss, for the first time, Open Babel, an open-source chem. toolbox that speaks the many languages of chem. data. Open Babel version 2.3 interconverts over 110 formats. The need to represent such a wide variety of chem. and mol. data requires a library that implements a wide range of cheminformatics algorithms, from partial charge assignment and aromaticity detection, to bond order perception and canonicalization. We detail the implementation of Open Babel, describe key advances in the 2.3 release, and outline a variety of uses both in terms of software products and scientific research, including applications far beyond simple format interconversion. Conclusions: Open Babel presents a soln. to the proliferation of multiple chem. file formats. In addn., it provides a variety of useful utilities from conformer searching and 2D depiction, to filtering, batch conversion, and substructure and similarity searching. For developers, it can be used as a programming library to handle chem. data in areas such as org. chem., drug design, materials science, and computational chem. It is freely available under an open-source license.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC3MXhsVWjurbF&md5=74e4f19b7f87417f916d57f7abcfb761
    39. 39
      Yang, J.; Shen, C.; Huang, N. Predicting or Pretending: Artificial Intelligence for Protein-Ligand Interactions Lack of Sufficiently Large and Unbiased Datasets. Frontiers in Pharmacology 2020, 11, 1,  DOI: 10.3389/fphar.2020.00069
      [Crossref], [PubMed], Google Scholar
      There is no corresponding record for this reference.
    40. 40
      Scantlebury, J.; Brown, N.; Von Delft, F.; Deane, C. M. Data Set Augmentation Allows Deep Learning-Based Virtual Screening to Better Generalize to Unseen Target Classes and Highlight Important Binding Interactions. J. Chem. Inf. Model. 2020, 60, 3722– 3730,  DOI: 10.1021/acs.jcim.0c00263
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      39
      Data Set Augmentation Allows Deep Learning-Based Virtual Screening to Better Generalize to Unseen Target Classes and Highlight Important Binding Interactions
      Scantlebury, Jack; Brown, Nathan; Von Delft, Frank; Deane, Charlotte M.
      Journal of Chemical Information and Modeling (2020), 60 (8), 3722-3730CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      Current deep learning methods for structure-based virtual screening take the structures of both the protein and the ligand as input but make little or no use of the protein structure when predicting ligand binding. Here, a relatively simple method of data set augmentation forces such deep learning methods to take into account information from the protein. Models trained in this way are more generalizable (make better predictions on protein/ligand complexes from a different distribution to the training data). They also assign more meaningful importance to the protein and ligand atoms involved in binding. Overall, the authors' results show that data set augmentation can help deep learning-based virtual screening to learn phys. interactions rather than data set biases.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB3cXhsVaqt77O&md5=8859d41da975e883cac7c2489befab60
    41. 41
      Satorras, V. G.; Hoogeboom, E.; Welling, M. E(n) Equivariant Graph Neural Networks. arXiv 2021, arXiv:2102.09844,  DOI: 10.48550/arXiv.2102.09844
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    42. 42
      Ragoza, M.; Hochuli, J.; Idrobo, E.; Sunseri, J.; Koes, D. R. Protein-Ligand Scoring with Convolutional Neural Networks. J. Chem. Inf. Model. 2017, 57, 942– 957,  DOI: 10.1021/acs.jcim.6b00740
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      41
      Protein-Ligand Scoring with Convolutional Neural Networks
      Ragoza, Matthew; Hochuli, Joshua; Idrobo, Elisa; Sunseri, Jocelyn; Koes, David Ryan
      Journal of Chemical Information and Modeling (2017), 57 (4), 942-957CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      Computational approaches to drug discovery can reduce the time and cost assocd. with exptl. assays and enable the screening of novel chemotypes. Structure-based drug design methods rely on scoring functions to rank and predict binding affinities and poses. The ever-expanding amt. of protein-ligand binding and structural data enables the use of deep machine learning techniques for protein-ligand scoring. We describe convolutional neural network (CNN) scoring functions that take as input a comprehensive three-dimensional (3D) representation of a protein-ligand interaction. A CNN scoring function automatically learns the key features of protein-ligand interactions that correlate with binding. We train and optimize our CNN scoring functions to discriminate between correct and incorrect binding poses and known binders and nonbinders. We find that our CNN scoring function outperforms the AutoDock Vina scoring function when ranking poses both for pose prediction and virtual screening.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC2sXlsVems7Y%253D&md5=9cae97167da0c93e1d896f85339cca7e
    43. 43
      Zhu, H.; Yang, J.; Huang, N. Assessment of the Generalization Abilities of Machine-Learning Scoring Functions for Structure-Based Virtual Screening. J. Chem. Inf. Model. 2022, 62, 5485– 5502,  DOI: 10.1021/acs.jcim.2c01149
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      42
      Assessment of the Generalization Abilities of Machine-Learning Scoring Functions for Structure-Based Virtual Screening
      Zhu, Hui; Yang, Jincai; Huang, Niu
      Journal of Chemical Information and Modeling (2022), 62 (22), 5485-5502CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      In structure-based virtual screening (SBVS), it is crit. that scoring functions capture protein-ligand at. interactions. By focusing on the local domains of ligand binding pockets, a standardized pocket Pfam-based clustering (Pfam-cluster) approach was developed to assess the cross-target generalization ability of machine-learning scoring functions (MLSFs). Subsequently, 12 typical MLSFs were evaluated using random cross-validation (Random-CV), protein sequence similarity-based cross-validation (Seq-CV), and pocket Pfam-based cross-validation (Pfam-CV) methods. Surprisingly, all of the tested models showed decreased performances from Random-CV to Seq-CV to Pfam-CV expts., not showing satisfactory generalization capacity. Our interpretable anal. suggested that the predictions on novel targets by MLSFs were dependent on buried solvent-accessible surface area (SASA)-related features of complex structures, with greater predicted binding affinities on complexes owning larger protein-ligand interfaces. By combining buried SASA-related features with target-specific patterns that were only shared among structurally similar compds. in the same cluster, the random forest (RF)-Score attained a good performance in the Random-CV test. Based on these findings, we strongly advise assessing the generalization ability of MLSFs with the Pfam-cluster approach and being cautious with the features learned by MLSFs.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB38Xis1Kru7rK&md5=d6197f1f9a4da0df36b420bb9746b446
    44. 44
      Radoux, C. J.; Olsson, T. S.; Pitt, W. R.; Groom, C. R.; Blundell, T. L. Identifying interactions that determine fragment binding at protein hotspots. J. Med. Chem. 2016, 59, 4314– 4325,  DOI: 10.1021/acs.jmedchem.5b01980
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      43
      Identifying Interactions that Determine Fragment Binding at Protein Hotspots
      Radoux, Chris J.; Olsson, Tjelvar S. G.; Pitt, Will R.; Groom, Colin R.; Blundell, Tom L.
      Journal of Medicinal Chemistry (2016), 59 (9), 4314-4325CODEN: JMCMAR; ISSN:0022-2623. (American Chemical Society)
      Locating a ligand-binding site is an important first step in structure-guided drug discovery, but current methods do little to suggest which interactions within a pocket are the most important for binding. Here the authors illustrate a method that samples at. hotspots with simple mol. probes to produce fragment hotspot maps. These maps specifically highlight fragment-binding sites and their corresponding pharmacophores. For ligand-bound structures, they provide an intuitive visual guide within the binding site, directing medicinal chemists where to grow the mol. and alerting them to suboptimal interactions within the original hit. The fragment hotspot map calcn. is validated using exptl. binding positions of 21 fragments and subsequent lead mols. The ligands are found in high scoring areas of the fragment hotspot maps, with fragment atoms having a median percentage rank of 97%. Protein kinase B and pantothenate synthetase are examd. in detail. In each case, the fragment hotspot maps are able to rationalize a Free-Wilson anal. of SAR data from a fragment-based drug design project.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC28XltlOms7Y%253D&md5=fa25eabfe291f9519f5701bcbc6c52c3
    45. 45
      Skyner, R.; von Delft, F. Xchem, fragalysis; https://fragalysis.diamond.ac.uk/ (accessed: 2022-07-30).
      Google Scholar
      There is no corresponding record for this reference.
    46. 46
      Wang, C.; Zhang, Y. Improving scoring-docking-screening powers of protein–ligand scoring functions using random forest. J. Comput. Chem. 2017, 38, 169– 177,  DOI: 10.1002/jcc.24667
      [Crossref], [PubMed], [CAS], Google Scholar
      45
      Improving scoring-docking-screening powers of protein-ligand scoring functions using random forest
      Wang, Cheng; Zhang, Yingkai
      Journal of Computational Chemistry (2017), 38 (3), 169-177CODEN: JCCHDD; ISSN:0192-8651. (John Wiley & Sons, Inc.)
      The development of new protein-ligand scoring functions using machine learning algorithms, such as random forest, has been of significant interest. By efficiently using expanded feature sets and a large set of exptl. data, random forest based scoring functions (RFbScore) can achieve better correlations to exptl. protein-ligand binding data with known crystal structures; however, more extensive tests indicate that such enhancement in scoring power comes with significant under-performance in docking and screening power tests compared to traditional scoring functions. To improve scoring-docking-screening powers of protein-ligand docking functions simultaneously, the authors have introduced a ΔvinaRF parameterization and feature selection framework based on random forest. The authors' developed scoring function ΔvinaRF20, which employs 20 descriptors in addn. to the AutoDock Vina score, can achieve superior performance in all power tests of both CASF-2013 and CASF-2007 benchmarks compared to classical scoring functions. The ΔvinaRF20 scoring function and its code are freely available on the web at: https://www.nyu.edu/projects/yzhang/DeltaVina.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC28XhvVygurzM&md5=1a3ea0b5f159c9c0edf315fed2110ed0
    47. 47
      Lu, J.; Hou, X.; Wang, C.; Zhang, Y. Incorporating Explicit Water Molecules and Ligand Conformation Stability in Machine-Learning Scoring Functions. J. Chem. Inf. Model. 2019, 59, 4540– 4549,  DOI: 10.1021/acs.jcim.9b00645
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      46
      Incorporating Explicit Water Molecules and Ligand Conformation Stability in Machine-Learning Scoring Functions
      Lu, Jianing; Hou, Xuben; Wang, Cheng; Zhang, Yingkai
      Journal of Chemical Information and Modeling (2019), 59 (11), 4540-4549CODEN: JCISD8; ISSN:1549-9596. (American Chemical Society)
      Structure-based drug design is critically dependent on accuracy of mol. docking scoring functions, and there is a significant interest to advance scoring functions with machine learning approaches. In this work, by judiciously expanding the training set, exploring new features related to explicit mediating water mols. as well as ligand conformation stability, and applying extreme gradient boosting (XGBoost) with Δ-Vina parametrization, we have improved robustness and applicability of machine-learning scoring functions. The new scoring function ΔvinaXGB can not only perform consistently among the top compared to classical scoring functions for the CASF-2016 benchmark but also achieves significantly better prediction accuracy in different types of structures that mimic real docking applications.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BC1MXitVSqs73O&md5=57fea8fa6135a109c3bcfd5ddffb1b33
    48. 48
      Moon, S.; Zhung, W.; Yang, S.; Lim, J.; Kim, W. Y. PIGNet: a physics-informed deep learning model toward generalized drug–target interaction predictions. Chem. Sci. 2022, 13, 3661– 3673,  DOI: 10.1039/D1SC06946B
      [Crossref], [PubMed], [CAS], Google Scholar
      47
      PIGNet: a physics-informed deep learning model toward generalized drug-target interaction predictions
      Moon, Seokhyun; Zhung, Wonho; Yang, Soojung; Lim, Jaechang; Kim, Woo Youn
      Chemical Science (2022), 13 (13), 3661-3673CODEN: CSHCCN; ISSN:2041-6520. (Royal Society of Chemistry)
      Recently, deep neural network (DNN)-based drug-target interaction (DTI) models were highlighted for their high accuracy with affordable computational costs. Yet, the models' insufficient generalization remains a challenging problem in the practice of in silico drug discovery. We propose two key strategies to enhance generalization in the DTI model. The first is to predict the atom-atom pairwise interactions via physics-informed equations parameterized with neural networks and provides the total binding affinity of a protein-ligand complex as their sum. We further improved the model generalization by augmenting a broader range of binding poses and ligands to training data. We validated our model, PIGNet, in the comparative assessment of scoring functions (CASF) 2016, demonstrating the outperforming docking and screening powers than previous methods. Our physics-informing strategy also enables the interpretation of predicted affinities by visualizing the contribution of ligand substructures, providing insights for further ligand optimization.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB38XmtlKktb0%253D&md5=c7ca118db214dde8563a0945db0a2331
    49. 49
      Shen, C.; Zhang, X.; Deng, Y.; Gao, J.; Wang, D.; Xu, L.; Pan, P.; Hou, T.; Kang, Y. Boosting Protein–Ligand Binding Pose Prediction and Virtual Screening Based on Residue–Atom Distance Likelihood Potential and Graph Transformer. J. Med. Chem. 2022, 65, 10691– 10706,  DOI: 10.1021/acs.jmedchem.2c00991
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      48
      Boosting Protein-Ligand Binding Pose Prediction and Virtual Screening Based on Residue-Atom Distance Likelihood Potential and Graph Transformer
      Shen, Chao; Zhang, Xujun; Deng, Yafeng; Gao, Junbo; Wang, Dong; Xu, Lei; Pan, Peichen; Hou, Tingjun; Kang, Yu
      Journal of Medicinal Chemistry (2022), 65 (15), 10691-10706CODEN: JMCMAR; ISSN:0022-2623. (American Chemical Society)
      The past few years have witnessed enormous progress toward applying machine learning approaches to the development of protein-ligand scoring functions. However, the robust performance and wide applicability of scoring functions remain a big challenge for increasing the success rate of docking-based virtual screening. Herein, a novel scoring function named RTMScore was developed by introducing a tailored residue-based graph representation strategy and several graph transformer layers for the learning of protein and ligand representations, followed by a mixt. d. network to obtain residue-atom distance likelihood potential. Our approach was resolutely validated on the CASF-2016 benchmark, and the results indicate that RTMScore can outperform almost all of the other state-of-the-art methods in terms of both the docking and screening powers. Further evaluation confirms the robustness of our approach that can not only retain its docking power on cross-docked poses but also achieve improved performance as a rescoring tool in larger-scale virtual screening.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB38XhvF2htrbO&md5=217e5d659ff8c69c9c9559c99344aef0
    50. 50
      Liao, Z.; You, R.; Huang, X.; Yao, X.; Huang, T.; Zhu, S. DeepDock: Enhancing Ligand-protein Interaction Prediction by a Combination of Ligand and Structure Information. In 2019 IEEE International Conference on Bioinformatics and Biomedicine (BIBM) ; 2019; pp 311– 317.
      [Crossref], Google Scholar
      There is no corresponding record for this reference.
    51. 51
      https://fragalysis.diamond.ac.uk/viewer/react/download/tag/5740113a-7603-4af4-9523-7f902186d4a2 (accessed: 2022-10-14).
      Google Scholar
      There is no corresponding record for this reference.
    52. 52
      Carbery, A.; Skyner, R.; von Delft, F.; Deane, C. M. Fragment Libraries Designed to Be Functionally Diverse Recover Protein Binding Information More Efficiently Than Standard Structurally Diverse Libraries. J. Med. Chem. 2022, 65, 11404– 11413,  DOI: 10.1021/acs.jmedchem.2c01004
      [ACS Full Text ACS Full Text], [CAS], Google Scholar
      51
      Fragment Libraries Designed to Be Functionally Diverse Recover Protein Binding Information More Efficiently Than Standard Structurally Diverse Libraries
      Carbery, Anna; Skyner, Rachael; von Delft, Frank; Deane, Charlotte M.
      Journal of Medicinal Chemistry (2022), 65 (16), 11404-11413CODEN: JMCMAR; ISSN:0022-2623. (American Chemical Society)
      Current fragment-based drug design relies on the efficient exploration of chem. space by using structurally diverse libraries of small fragments. However, structurally dissimilar compds. can exploit the same interactions and thus be functionally similar. Using three-dimensional structures of many fragments bound to multiple targets, we examd. if a better strategy for selecting fragments for screening libraries exists. We show that structurally diverse fragments can be described as functionally redundant, often making the same interactions. Ranking fragments by the no. of novel interactions they made, we show that functionally diverse selections of fragments substantially increase the amt. of information recovered for unseen targets compared to the amts. recovered by other methods of selection. Using these results, we design small functionally efficient libraries that can give significantly more information about new protein targets than similarly sized structurally diverse libraries. By covering more functional space, we can generate more diverse sets of drug leads.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A528%3ADC%252BB38XitFWktLnO&md5=20c210891befe38eb6ec658cba2a01b8
    53. 53
      Schuller, M. Fragment binding to the Nsp3 macrodomain of SARS-CoV-2 identified through crystallographic screening and computational docking. Sci. Adv. 2021, 7, eabf8711  DOI: 10.1126/sciadv.abf8711
      [Crossref], [PubMed], Google Scholar
      There is no corresponding record for this reference.
    54. 54
      Imprachim, N.; Yosaatmadja, Y.; Newman, J. A. Crystal structures and fragment screening of SARS-CoV-2 NSP14 reveal details of exoribonuclease activation and mRNA capping and provide starting points for antiviral drug development. Nucleic Acids Research 2023, 51, 475,  DOI: 10.1093/nar/gkac1207
      [Crossref], [PubMed], [CAS], Google Scholar
      53
      Crystal structures and fragment screening of SARS-CoV-2 NSP14 reveal details of exoribonuclease activation and mRNA capping and provide starting points for antiviral drug development
      Imprachim Nergis; Yosaatmadja Yuliana; Newman Joseph A
      Nucleic acids research (2023), 51 (1), 475-487 ISSN:.
      NSP14 is a dual function enzyme containing an N-terminal exonuclease domain (ExoN) and C-terminal Guanine-N7-methyltransferase (N7-MTase) domain. Both activities are essential for the viral life cycle and may be targeted for anti-viral therapeutics. NSP14 forms a complex with NSP10, and this interaction enhances the nuclease but not the methyltransferase activity. We have determined the structure of SARS-CoV-2 NSP14 in the absence of NSP10 to 1.7 ÅA resolution. Comparisons with NSP14/NSP10 complexes reveal significant conformational changes that occur within the NSP14 ExoN domain upon binding of NSP10, including helix to coil transitions that facilitate the formation of the ExoN active site and provide an explanation of the stimulation of nuclease activity by NSP10. We have determined the structure of NSP14 in complex with cap analogue 7MeGpppG, and observe conformational changes within a SAM/SAH interacting loop that plays a key role in viral mRNA capping offering new insights into MTase activity. We perform an X-ray fragment screen on NSP14, revealing 72 hits bound to sites of inhibition in the ExoN and MTase domains. These fragments serve as excellent starting point tools for structure guided development of NSP14 inhibitors that may be used to treat COVID-19 and potentially other future viral threats.
      >> More from SciFinder ®
      https://chemport.cas.org/services/resolver?origin=ACS&resolution=options&coi=1%3ACAS%3A280%3ADC%252BB28rps1ahtg%253D%253D&md5=b3ad340352d17592df9e8d0fffa58af3
    55. 55
      https://fragalysis.diamond.ac.uk/viewer/react/download/tag/01f41754-b2bb-4817-acc3-a5ebe820316d (accessed: 2022-10-14).
      Google Scholar
      There is no corresponding record for this reference.
    56. 56
      https://fragalysis.diamond.ac.uk/viewer/react/download/tag/3df36d6b-3a5b-400d-97eb-af3b0e0df42d (accessed: 2022-10-14).
      Google Scholar
      There is no corresponding record for this reference.

  • Supporting Information

    Supporting Information

    ARTICLE SECTIONS
    Jump To

    The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.jcim.3c00322.

    • A brief description of equivariance and invariance, some more in-depth information about the architecture of PointVS and the data sets used to train and test it, a comparison of the different attribution methods used, a description of the process by which the API hotspots were obtained, an additional assessment of the reliance of PointVS hotspots on fragment screen size, and a case study showing how PointVS could have aided in a real-world fragment-to-lead campaign; also all of the IDs of the structures used, from both fragalysis and the PDB (PDF)


    Terms & Conditions

    Most electronic Supporting Information files are available without a subscription to ACS Web Editions. Such files may be downloaded by article for research use (if there is a public use license linked to the relevant article, that license may permit other uses). Permission may be obtained from ACS for other uses through requests via the RightsLink permission system: http://pubs.acs.org/page/copyright/permissions.html.

PDF [ 2MB]

Pair your accounts.

Export articles to Mendeley

Get article recommendations from ACS based on references in your Mendeley library.

Pair your accounts.

Export articles to Mendeley

Get article recommendations from ACS based on references in your Mendeley library.

You’ve supercharged your research process with ACS and Mendeley!

STEP 1:
Click to create an ACS ID

Please note: If you switch to a different device, you may be asked to login again with only your ACS ID.

Please note: If you switch to a different device, you may be asked to login again with only your ACS ID.

Please note: If you switch to a different device, you may be asked to login again with only your ACS ID.

MENDELEY PAIRING EXPIRED
Your Mendeley pairing has expired. Please reconnect