CoChat
Sign Up
Cited by 371Computational Drug Discovery MethodsMachine Learning in BioinformaticsProtein Degradation and Inhibitors

Graph Convolutional Neural Networks for Predicting Drug-Target Interactions

Open access
Torng, Wen, Altman, Russ B.·2019·Journal of Chemical Information and Modeling
DOI 10.1021/acs.jcim.9b00628

Published version of a preprint

Open this paper in CoChat — free accountOpen PDF

Abstract

Accurate determination of target-ligand interactions is crucial in the drug discovery process. In this paper, we propose a graph-convolutional (Graph-CNN) framework for predicting protein-ligand interactions. First, we built an unsupervised graph-autoencoder to learn fixed-size representations of protein pockets from a set of representative druggable protein binding sites. Second, we trained two Graph-CNNs to automatically extract features from pocket graphs and 2D ligand graphs, respectively, driven by binding classification labels. We demonstrate that graph-autoencoders can learn fixed-size representations for protein pockets of varying sizes and the Graph-CNN framework can effectively capture protein-ligand binding interactions without relying on target-ligand complexes. Across several metrics, Graph-CNNs achieved better or comparable performance to 3DCNN ligand-scoring, AutoDock Vina, RF-Score, and NNScore on common virtual screening benchmark data sets. Visualization of key pocket residues and ligand atoms contributing to the classification decisions confirms that our networks are able to detect important interface residues and ligand atoms within the pockets and ligands, respectively.

AI snapshot

Written from the paper’s full text.

TL;DR

The paper proposes a two-stage graph-convolutional framework that first learns fixed-size protein-pocket representations with an unsupervised graph autoencoder and then trains parallel Graph-CNNs on pocket graphs and 2D molecular graphs to predict drug-target binding without requiring co-crystal complexes. Across DUD-E and MUV benchmarks, the model achieves better or comparable performance to 3DCNN ligand-scoring, AutoDock Vina, RF-Score, and NNScore.

Key findings

  • On DUD-E, Graph-CNN achieved AUC 0.886 versus 3DCNN 0.868, Vina 0.716, RF-Score 0.622, and NNScore 0.584; at 0.5% FPR, RE was 44.406 versus 42.559, 9.139, 5.628, and 4.166, respectively.
  • On the independent MUV dataset, Graph-CNN achieved average AUC 0.563 versus 3DCNN 0.518, Vina 0.545, RF-Score 0.497, and NNScore 0.430; Graph-CNN attained AUC > 0.6 for 5 of 8 MUV targets, while Vina, 3DCNN, RF-Score, and NNScore had no more than two such targets.
  • A five-fold MUV-trained Graph-CNN model improved average AUC to 0.707, with large gains on MUV targets 832, 846, and 852, suggesting DUD-E training coverage was incomplete for those targets.
  • The unsupervised pocket autoencoder produced latent pocket embeddings that clustered by SCOP family, and visualization of pocket residues and ligand atoms highlighted interactions corresponding to known co-complex contacts.

Methods

  • The framework has two stages: an unsupervised graph autoencoder trained on 965 representative pockets (863 DrugBank-derived high-quality structures representing 262 drugs plus 102 DUD-E pockets) to learn fixed-size pocket fingerprints, followed by a supervised classifier with separate pocket and molecule Graph-CNN modules, an interaction layer, and a softmax binding/non-binding classifier.
  • Training used DUD-E (102 targets, on average 224 actives and over 10,000 decoys per target), plus two constructed variants: DUD-E-Pairwise, with negative pockets assigned via PocketFEATURE at a score threshold of -1.9, and DUD-E-CHEMBL assay negatives, using experimentally validated negatives with IC50 > 50 µM and target confidence 9. External validation used 8 MUV targets after filtering for sequence and binding-site dissimilarity to DUD-E.
  • Evaluation used four-fold cross-validation split by target for DUD-E, AUC and ROC enrichment (RE) at 0.5%, 1%, 2%, and 5% FPR, MUV AUC/RE, t-SNE visualization of pocket embeddings, binding propensity matrices, dummy-pocket ablation, and saliency-map visualization of pocket residue and ligand atom contributions.

Limitations

  • The paper states that DUD-E has limited target number and diversity and provides limited information on ligand-to-target preferences; it also notes that MUV is challenging because it uses experimentally validated negatives, maximally spread actives, and some assays that may not measure direct target-ligand interactions.
  • The paper states that Graph-CNN performance dropped on MUV targets 832, 846, and 852, and that the difference between Graph-CNN-DUD-E and Graph-CNN-MUV on these targets suggests interactions learned from DUD-E are still incomplete; it calls for high-quality, large-scale datasets covering both target-ligand and ligand-target binding propensities.
  • Evident from the text: the full MUV model was trained for one epoch with no rigorous hyperparameter optimization; the pocket representation relies on FEATURE vectors rather than raw atom distributions; and the first pocket and molecule graph convolution layers were fixed during supervised training.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1021/acs.jcim.9b00628

Cites

57 of 57 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 13 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    The model can predict meaningful ligand-target binding profiles.

    “we further examined the ability of our model to predict binding propensity of a given compound to different targets and showed that the model can predict meaningful ligand-target binding profiles.”
    Quote in the paperSubstantial claim
  • Findingown data

    Graph-autoencoders can learn meaningful fixed-size representations for protein pockets of varying sizes.

    “We demonstrate that graph-autoencoders can learn meaningful fixed-size representation for protein pockets of varying sizes”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Protein pockets vary in size and have complex properties.
  • Findingown data

    The Graph-CNN framework can effectively capture protein-ligand binding interactions without relying on target-ligand co-complexes.

    “the Graph-CNN framework can effectively capture protein-ligand binding interactions without relying on target-ligand cocomplexes”
    • Also stated in: § Conclusion · boosted
      “We further demonstrate that the Graph-CNN framework can effectively capture protein-ligand binding interactions without relying on target-ligand co-complexes.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Most protein-ligand pairs of interest do not have experimentally solved co-crystal structures.
  • Findingown data

    The learned pocket embedding in 2D space shows that latent representations reflect biological pocket similarity.

    “We visualized the learned pocket embedding in 2D space and show that the latent representations reflect biological pocket similarity.”
    Quote in the paperNot a substantial claim
  • Findingown data

    Graph-CNNs achieved better or comparable performance to 3DCNN ligand-scoring, AutoDock Vina, RF-Score, and NNScore on the DUD-E and MUV virtual screening benchmark datasets, without requiring co-crystal structures.

    “Across several metrics, Graph-CNNs achieved better or comparable performance to 3DCNN ligand-scoring, AutoDock Vina, RF-Score, and NNScore on common virtual screening benchmark datasets.”
    • Also stated in: § I. Introduction · boosted
      “We then performed head-to-head comparisons of prediction performance between our Graph-CNNs, 3DCNN ligand-scoring 15 , AutoDock Vina 2 , RF-Score 19 , and NNScore 20 , on the Database of Useful Decoys: Enhanced (DUD-E) 21 and Maximum Unbiased Validation (MUV) 22 dataset and showed that our model achieved better or comparable performances on both datasets without requiring co-crystal structures.”
    Quote in the paperSubstantial claim
  • May combine two claims. 18
  • The quote only partly states this claim. 3
  • Classified as probably not a substantial claim. 17

  • May combine two claims.

    Findingown data

    Across several metrics, Graph-CNNs achieved better or comparable performance to 3DCNN ligand-scoring, AutoDock Vina, RF-Score, and NNScore on the DUD-E and MUV benchmark, and showed complementary advantages to ligand-based methods.

    “Across several metrics, Graph-CNNs achieved better or comparable performance to 3DCNN ligand-scoring, AutoDock Vina, RF-Score, and NNScore on the DUD-E and MUV benchmark, and showed complementary advantages to ligand-based methods.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Local high-scored blocks form at the intersections of the corresponding row and column clusters, showing that ligands are predicted to have stronger interactions with targets that are similar to its primary target compared to unrelated targets and vice versa.

    “Local high-scored blocks form at the intersections of the corresponding row and column clusters, showing that ligands are predicted to have stronger interactions with targets that are similar to its primary target compared to unrelated targets and vice versa.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The Graph-CNN-MUV models achieve significant improvements over Graph-CNN-DUD-E on targets 832, 846, and 852, achieving AUC scores comparable to AP.

    “The Graph-CNN-MUV models achieve significant improvements over Graph-CNN-DUDE on targets 832, 846, and 852, achieving AUC scores comparable to AP.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Graph-CNN achieved the best performance among the 5 methods for targets 859, 692, 689, and 466, whereas AP achieved the best performance for targets 548, 832, 852, and 846.

    “Graph-CNN achieved the best performance among the 5 methods for target 859, 692, 689, 466 whereas AP achieved the best performance for target 548, 832, 852, 846.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Graph-CNN-MUV models achieve significant improvements on targets 852, 846, and 832, achieving AUC scores comparable to AP.

    “the Graph-CNN-MUV models achieve significant improvements on the three targets, achieving AUC scores comparable to AP.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Graph-CNN attained AUC greater than 0.6 for 5 out of 8 MUV targets, while Vina, 3DCNN, RF-Score, and NNScore have no more than two targets achieving an AUC greater than 0.6.

    “While Vina, 3DCNN, RF-score, and NNscore have no more than two targets achieving an AUC greater than 0.6, the Graph-CNN models attain AUC greater than 0.6 for 5 out of 8 targets.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The model can generate meaningful ligand-target binding profiles and achieved better generalizability on the MUV dataset compared to other machine-learning and deep learning structure-based methods.

    “We show that our model can generate meaningful ligand-target binding profiles and achieved better generalizability on the MUV dataset compared to other machine-learning and deeplearning structure-based methods.”
    Quote in the paperSubstantial claim

  • The quote only partly states this claim.

    Findingown data

    For SRC (PDB: 3EL8) and compound 1, the identified key residues and atoms from the importance maps highly correspond to the observed interacting residues and ligand atoms in the co-crystalized complex.

    “The identified key residues and atoms highly correspond to the observed interacting residues and ligand atoms in the co-crystalized complex of SRC and pyrazolopyrimidine 5 (Top, PDB: 3EL8).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    All methods, except Vina, performed well on category IV (target 548) where ligands are separated and pocket similarity is relatively high.

    “All methods, except Vina, performed well on category IV (target 548) where ligands are separated and the pocket similarity is relatively high.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisereasoning

    3DCNN methods for protein-ligand interaction prediction require a data-preprocessing step that artificially docks query ligands into their targets, and this cross-docking step can introduce significant noise.

    “However, since most protein-ligand pairs of interests do not have experimentally solved co-crystal structures, to generate protein-ligand co-complex for training and testing, these methods require a data-preprocessing step that artificially docks query ligands into their targets.This cross-docking step can introduce significant noises, stemming from cross-docking ligands into noncognate receptors.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Most protein-ligand pairs of interest do not have experimentally solved co-crystal structures.

  • May combine two claims.

    Premisecitation

    The DUD-E dataset consists of 102 targets across different protein families, with on average 224 actives and over 10,000 decoys per target.

    “The DUD-E dataset consists of 102 targets across different protein families.On average, each target has 224 actives and over 10,000 decoys.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [21] source

      Not checked in the free run.

  • May combine two claims.

    Premisereasoning

    The DUD-E dataset has limited number and diversity of protein targets, so learned features may be biased towards certain subsets of targets.

    “which has limited number and diversity of protein targets.”
    • Also stated in: § I. Introduction · hedged
      “The learned features may therefore be biased towards certain subsets of targets.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisereasoning

    The DUD-E dataset emphasizes assessing binding preferences of a given target to different ligands but does not provide much information regarding preferences of a given ligand to different targets.

    “Because the DUD-E dataset was designed for virtual screening tasks, it emphasizes on assessing binding preferences of a given target to different ligands.However, it does not provide much information regarding preferences of a given ligand to different targets.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based). May combine two claims.

    Interpretationnone given

    Viewing protein pockets as graphs of residue nodes allows pockets of arbitrary size to be represented and imposes relatively weak geometry constraints.

    “Viewing protein pockets as graphs of residue nodes allows pockets of arbitrary size to be represented and imposes relatively weak geometry constraints.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Protein pockets vary in size and have complex properties.

  • The quote only partly states this claim.

    Study design and conductcitation

    The final MUV external validation set comprises eight targets selected to have no more than 80% global sequence identity and no significant binding site structural similarity to any DUD-E target.

    “Following procedures described in Ragoza et al. 15 , we selected MUV targets that have no more than 80% global sequence identity to any DUD-E target.Furthermore, targets with significant binding site structural similarity to any of the DUD-E target binding site were removed.Finally, only unique targets with available crystal structure of the pocket site were included.The above procedure results in”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites Ragoza et al. 15

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationreasoning

    The Graph-CNN framework can be layered on top of 3DCNNs that characterize local pocket residue environments to enable an end-to-end framework that learns biochemical features directly from raw crystal structures.

    “The proposed Graph-CNN framework can be layered on top of 3DCNNs that characterize local pocket residue environments to further enable an end-toend framework that learns biochemical features directly from raw crystal structures.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. The quote only partly states this claim.

    Study design and conductnone given

    The folds in the four-fold cross-validation were split between targets, with all ligands of the same target belonging to the same fold.

    “The folds are split between targets, where all ligands of the same target belong to the same fold.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given

    Each node in the interaction hidden layer represents a favorable or non-favorable interaction type between the target and ligand.

    “Each node in the interaction hidden layer represents a favorable or non-favorable interaction type between the target and ligand.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    Within each training fold, three targets were randomly selected as validation dataset for hyperparameter tuning.

    “Within each training fold, we randomly selected three targets as our validation dataset, and choose our model hyper-parameters based on the validation performances.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    AP is the best performing method reported in a large-scale virtual screening study of ligand-based methods conducted by Riniker et al.

    “AP is the best performing method reported in a large-scale virtual screening study of ligand-based methods conducted by Riniker et al. 7”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites Riniker et al. 7

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Study design and conductown data

    The study used a four-fold cross-validation strategy to train and evaluate Graph-CNNs on the DUD-E dataset.

    “We adopt a four-fold cross-validation strategy to train and evaluate our Graph-CNNs on the DUD-E -Pairwise interaction dataset (section 2.1.2(A)).”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Interpretationnone given

    Multiple graph autoencoder layers can be stacked and trained greedily, allowing features of different complexity to be integrated through a hierarchical manner.

    “Multiple graph autoencoder layers can be stacked and trained greedily, allowing features of different complexity to be integrated through a hierarchical manner.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The pyrimidine ring makes contact with Ile 627.

    “The pyrimidine ring makes contact with Ile 627.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [45] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationreasoning

    Such a design can simultaneously allow detailed characterization at the local amino acid environment level and global flexibility at the pocket level.

    “Such design can simultaneously allow detailed characterization at the local amino acid environment level and global flexibility at the pocket level.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    No two folds had targets with greater than 75% global sequence identity.

    “we ensured that no any two folds have targets with greater than 75% global sequence identity.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    Our ligand interaction map highlights the benzene ring, the nearby oxygen, and the polycyclic ring in compound 2 as the key features, which may play the roles of the pyrimidine ring, the amino-thiazole ring, and the 2-chloro-6-methyl phenyl group, respectively.

    “Our ligand interaction map highlights the benzene ring, the nearby oxygen, and the polycyclic ring in compound 2 as the key features, which may play the roles of the pyrimidine ring, the amino-thiazole ring, and the 2-chloro-6methyl phenyl group, respectively.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductreasoning

    For the MUV cross-validation, the folds were split between targets.

    “Similar to the training of the DUD-E models, the folds are split between targets.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Interpretationreasoning

    The learned pocket representation can enable efficient pocket similarity search, pocket classification, and serve as input for downstream machine learning algorithms.

    “Such representation has the potential to enable efficient pocket similarity search, pocket classification, and can serve as input for downstream machine learning algorithms.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    The 2-chloro-6-methyl phenyl group of dasatinib occupies the hydrophobic pocket and makes van der Waals contacts with Met 674, Ile 683, Ile 697 and Thr 699.

    “The 2-chloro-6-methyl phenyl group of dasatinib occupies the hydrophobic pocket and makes van der Waals contacts with Met 674, Ile 683, Ile 697 and Thr 699.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [45] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    The amino-thiazole ring has van der Waals contacts with Leu 753 and Ala 651, and also forms hydrogen bonds with Met 702.

    “The amino-thiazole ring has van der Waals contacts with Leu 753 and Ala 651, and also forms hydrogen bonds with Met 702.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [45] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductnone given

    The supervised model was trained for a full epoch and no rigorous attempt was made to optimize the hyperparameters.

    “The supervised model was trained for a full epoch and no rigorous attempt was made to optimize the hyperparameters.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The amide group forms hydrogen bond with Thr 699.

    “The amide group forms hydrogen bond with Thr 699.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [45] source

      Not checked in the free run.

32 of 32

  • The paper relies on this without support.

    Protein pockets vary in size and have complex properties.

    “Protein pockets are of varying sizes and with complex properties.”

  • The paper relies on this without support.

    It is challenging to define protein pocket similarity or to embed a protein pocket as a fixed-size vector for downstream machine learning tasks without emphasizing certain aspects of pocket properties while ignoring others.

    “It is challenging to define protein pocket similarity or to embed protein pocket as a fixed-size vector for downstream machine learning tasks without emphasizing certain aspects of pocket properties while ignoring others.”

  • The paper relies on this without support.

    Biochemical interactions start between atoms and can be aggregated over space to form complex interactions.

    “Biochemical interactions start between atoms, and can be similarly aggregated over space to form complex interactions.”

  • The paper relies on this without support.

    Most protein-ligand pairs of interest do not have experimentally solved co-crystal structures.

    “However, since most protein-ligand pairs of interests do not have experimentally solved co-crystal structures”

  • The paper relies on this without support.

    Target-ligand binding interactions are promiscuous.

    “Target-ligand binding interactions are known to be promiscuous.”

  • The paper relies on this without support.

    Ligand-based methods achieve better performances than structure-based methods on the MUV dataset.

    “ligand-based methods are known to achieve better performances than structure-based methods on the MUV dataset”

  • The paper relies on this without support.

    The MUV dataset does not use computational decoys.

    “Unlike DUD-E, the MUV dataset does not use computational decoys.”

  • The paper relies on this without support.

    The actives in the MUV dataset are selected to be maximally spread based on simple descriptors and embedded in decoys.

    “the actives are selected to be maximally spread based on simple descriptors and embedded in decoys.”

  • The paper relies on this without support.

    Both chemical-based and cell-based assays are used to construct the MUV dataset.

    “both chemical-based and cell-based assays are used to construct dataset”

  • The paper relies on this without support.

    Virtual screening datasets often provide information about ligand binding propensity to given targets but have limited information on target preferences of given ligands.

    “Virtual screening dataset often provide information about ligand binding propensity to given targets but have limited information on target preferences of given ligands.”

  • The paper relies on this without support.

    Current benchmark datasets often include limited number of targets, which only cover a subset of potential target space.

    “Moreover, current benchmark datasets often include limited number of targets, which only cover a subset of potential target space.”

  • The paper relies on this without support.

    To learn a generalizable model of protein-ligand binding, the quality of the training dataset is crucial.

    “To learn a generalizable model of protein-ligand binding, the quality of the training dataset is crucial.”
12 of 12

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • Rigorous analyses suggest that GCNNs do not improve predictive performance compared to 'classical QSAR' fingerprints.

    “Rigorous analyses suggest that GCNNs do not improve predictive performance compared to 'classical QSAR' fingerprints (92).”

    From Artificial Intelligence for Drug Discovery: Are We There Yet? (2024) · cites this paper as [92]

    In “External Dataset Validation: MUV” (page unavailable), the paper reports that Graph-CNN had the best performance among five methods on four of eight targets, while the ligand-based AP method led on the other four. Its analysis says Graph-CNNs outperformed ligand-based methods on the harder cases, rather than finding that they do not improve predictive performance over fingerprint-based methods. (LLM, conf=0.88)

    Checked by a reasoning model — may be wrong.

    Experimental — accuracy not yet measured

  • GCNNs have been used to model drug combinations, drug-target interactions, and multiple property predictions.

    “GCNNs have been used to model drug combinations (90), drug-target interactions (91), and multiple property predictions (92).”

    From Artificial Intelligence for Drug Discovery: Are We There Yet? (2024) · cites this paper as [92]

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Conflicting counts for the DUD-E pocket visualization

    Minor

    The methods report 393 pockets from 47 SCOP families for this visualization, while the results report 401 pockets from 48 families.

1 of 1

Cites

Flagged first, then by citations

Research works printed but not linked: 8

  1. Krizhevsky A., 2012, In Adv. Neural Inf. Process. Syst., 2012, 1097
  2. Wallach I., 2015, arXiv, arXiv:1510.0285
  3. Duvenaud D. K., 2015, In Adv. Neural Inf. Process. Syst., 2015, 2224
  4. Fout A., 2017, In Adv. Neural Inf. Process. Syst., 2017, 6530
  5. Yoon S., BMC Bioinformatics, 2007, 8, S10
  6. Glorot X., 2011, In Proceedings of the fourteenth international conference on artificial intelligence and statistics, 2011, 315
  7. Tieleman T., 2012, COURSERA: Neural networks for machine learning, 4, 26
  8. Maaten L. v. d., 2008, Journal of machine learning research, 9, 2579

Cited by

Sorted by citations

Related

Authors
Torng, W., & Altman, R. B.
Affiliations
Stanford University (US)
Corresponding author
  • Torng, Wen
  • Altman, Russ B.
ORCID
Venue
Journal of Chemical Information and Modeling
Publisher
American Chemical Society (ACS)
Published
2019-10-03
Volume
59
Issue
10
Pages
4131-4149
ISSN
  • 1549-9596 (ISSN-L)
  • 1549-960X
DOI
10.1021/acs.jcim.9b00628
OpenAlex
W2978484973
Cited by
371
References
56
Citations by year
2026: 25 · 2025: 54 · 2024: 55 · 2023: 63 · 2022: 59 · 2021: 72 · 2020: 35 · 2019: 6
FWCI
29.27
Type
journal article
Language
en
Open access
Licence
Locations
  • Journal of Chemical Information and Modeling (publishedVersion)
  • bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion, cc-by-nc-nd)
  • PubMed (publishedVersion)
Funders
  • U.S. Department of Health and Human Services (GM102365)
  • National Science Foundation (ACI-1429830)
  • Chan-Zuckerberg Biohub
  • Siebel Scholar Foundation
Topics
  • Physical Sciences › Computer Science › Computational Theory and Mathematics › Computational Drug Discovery Methods
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Machine Learning in Bioinformatics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Protein Degradation and Inhibitors
Keywords
drug-target interaction prediction, protein-ligand binding, graph convolutional networks, graph autoencoder, virtual screening, binding affinity prediction
MeSH terms
Drug Development, Binding Sites, Ligands, Protein Binding, Protein Conformation, Proteins, Neural Networks, Computer, Drug Discovery, Databases, Chemical

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-10.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-10.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review