CoChat
Sign Up
Cited by 155Computational Drug Discovery MethodsMachine Learning in Materials ScienceProtein Degradation and Inhibitors

Mathematical deep learning for pose and binding affinity prediction and ranking in D3R Grand Challenges

Open access
Nguyen, Duc Duy, Cang, Zixuan, Wu, Kedi, Wang, Menglun, Cao, Yin, Wei, Guo-Wei·2019·Journal of Computer-Aided Molecular Design
DOI 10.1007/s10822-018-0146-6PMC PMC7163798
Open this paper in CoChat — free account

Abstract

Advanced mathematics, such as multiscale weighted colored subgraph and element specific persistent homology, and machine learning including deep neural networks were integrated to construct mathematical deep learning models for pose and binding affinity prediction and ranking in the last two D3R Grand Challenges in computer-aided drug design and discovery. D3R Grand Challenge 2 focused on the pose prediction, binding affinity ranking and free energy prediction for Farnesoid X receptor ligands. Our models obtained the top place in absolute free energy prediction for free energy set 1 in stage 2. The latest competition, D3R Grand Challenge 3 (GC3), is considered as the most difficult challenge so far. It has five subchallenges involving Cathepsin S and five other kinase targets, namely VEGFR2, JAK2, p38-α, TIE2, and ABL1. There is a total of 26 official competitive tasks for GC3. Our predictions were ranked 1st in 10 out of these 26 tasks.

AI snapshot

Written from the paper’s full text.

TL;DR

The paper reports mathematical deep learning models combining multiscale weighted colored subgraphs and element-specific persistent homology with machine learning for D3R Grand Challenge 2 and 3 pose, affinity ranking, and free energy tasks. In GC3, the authors’ submissions ranked 1st in 10 of 26 official competitive tasks, and in GC2 they achieved the top place in free energy prediction for free energy set 1 in stage 2.

Key findings

  • In D3R GC3, the authors’ predictions were ranked 1st in 10 out of 26 official competitive tasks.
  • For GC3 Subchallenge 1 (Cathepsin S), the best Kendall’s τ improved from 0.23 in stage 1 to 0.54 in stage 2 (receipt ID 6jekk); free energy RMSEc was 0.33 kcal/mol in stage 1 (receipt ID fomca) and 0.29 kcal/mol in stage 2 (receipt ID v4jv4).
  • For GC3 Subchallenge 4 (TIE2), the model ranked first on all evaluated metrics: affinity ranking excluding Kds > 10 μM gave Kendall’s τ = 0.57 and Spearman = 0.76 (receipt ID uuihe); active/inactive classification gave MCC = 0.78; Set 1 absolute free energy gave MCC = 1.0 (receipt ID vwbp8); Set 2 gave RMSEc = 1.02 kcal/mol (receipt ID 5g8ed).
  • In GC2, the submission with receipt ID 5bvwx ranked second in free energy set 1 of stage 1 with RMSEc = 0.68 kcal/mol, and in stage 2 the model obtained first place for free energy set 1 with Kendall’s τ = 0.41 (receipt ID 4rbjk).

Methods

  • The models used multiscale weighted colored subgraph (MWCS) and element-specific persistent homology descriptors, combined with machine learning methods including random forest, gradient boosting trees, convolutional neural networks, and multi-task learning; training used PDBbind data.
  • Ligands were prepared from SMILES using LIGPREP at pH 7.5, with Gasteiger charges added via MGLTools v1.5.6; docking used Autodock Vina, GOLD, GLIDE, induced fit docking, and several protocols named align-close, align-target, close-dock, cross-dock, constraint-IFD, and free-IFD.
  • The paper evaluates submissions in D3R GC2 and GC3. GC2 involved FXR with 36 crystallographic ligands, 102 compounds for affinity ranking, and free energy subsets of 18 and 15 ligands. GC3 had five subchallenges involving Cathepsin S and kinases VEGFR2, JAK2, p38-α, TIE2, and ABL1, with 26 official competitive tasks; reported metrics include Kendall’s τ, Spearman correlation, RMSEc, and MCC.

Limitations

  • The paper states that docking accuracy remains a bottleneck for affinity prediction, that none of the tried docking protocols showed a dominant role in binding affinity accuracy, and that excluding pose prediction error improved Kendall’s τ from 0.21 to 0.54 on Cathepsin S affinity ranking.
  • The paper also states that machine learning-based scoring function performance depends crucially on training data quality and statistical distribution, and on selected features that may not accurately or completely describe protein–ligand binding interactions.
  • Evident from the reported results: the models did not perform well on GC3 Subchallenge 2, with the best result being second place on VEGFR2 active/inactive classification (MCC = 0.48), and affinity ranking in Subchallenge 3 was described as unremarkable; GC2 pose ranking power was also described as not impressive.

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

Integrity signals

Registry

Crossref · 10.1007/s10822-018-0146-6

Cites

70 of 70 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 26 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    In D3R GC3, the paper's submissions were ranked 1st in 10 of 26 official evaluation tasks.

    “Our submissions were ranked 1st in 10 of these 26 tasks as shown in Table 1 for additional information.”
    • Also stated in: § Conclusion · plain
      “In GC3, our submissions were ranked 1st in 10 out of 26 official evaluation tasks.”
    Quote in the paperSubstantial claim
  • Findingown data

    In GC2 stage 2, the paper's models improved the accuracy of energy prediction for compounds in free energy set 1.

    “In stage 2, our models improved the accuracy of the energy prediction of compounds in the aforementioned free energy set.”
    Quote in the paperSubstantial claim
  • Findingown data

    None of the tried docking protocols (align-close, align-target, close-dock, cross-dock, constraint-IFD, and free-IFD) showed a dominant role in binding affinity accuracy.

    “We have tried a variety of docking protocols, namely align-close, align-target, close-dock, cross-dock, constraint-IFD, and free-IFD, for pose selection in GC3. However, none of them showed a dominant role in binding affinity accuracy.”
    Quote in the paperSubstantial claim
  • Findingown data

    In GC3 Subchallenge 4 Set 1 absolute free energy predictions, the paper's predictor receipt ID vwbp8 ranked first with MCC = 1.0.

    “Specifically, on Set 1, our predictor with receipt ID vwbp8 was ranked the first place with MCC = 1.0.”
    Quote in the paperSubstantial claim
  • Findingown data

    Our predictions were ranked 1st in 10 out of the 26 official competitive tasks in D3R Grand Challenge 3.

    “Our predictions were ranked 1st in 10 out of these 26 tasks.”
    • Also stated in: § Introduction · plain
      “we obtained the first place in 10 out of a total of 26 predictive tasks.”
    Quote in the paperSubstantial claim
  • May combine two claims. 9
  • Classified as probably not a substantial claim. 8

  • May combine two claims.

    Findingreasoning

    In GC3 Subchallenge 1, the best Kendall's tau for the scoring task was 0.23 in stage 1 and 0.54 in stage 2 (receipt ID 6jekk).

    “That explains why our predictors performed decently for scoring task in stage 1 with the best Kendall’s τ = 0.23, but they achieved a state-of-the-art result for the same task in stage 2 with the best Kendall’s τ = 0.54 (receipt ID 6jekk).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    In GC2 stage 2, the paper obtained first place in Kendall's tau (τ = 0.41) with receipt ID 4rbjk for free energy set 1, the highest Kendall's tau among all submissions in two stages for that set.

    “In fact, we obtained the first place in term of Kendall’s tau value (τ = 0.41) with receipt ID 4rbjk. That was also the highest Kendall’s tau value among all submissions in two stages for the free energy set 1.”
    • Also stated in: § Conclusion · plain
      “Overall, in GC2, our predictive models achieved the top place in free energy prediction for free energy set 1 in stage 2.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    In GC3 Subchallenge 1, the paper's prediction with receipt ID fomca obtained RMSEc = 0.33 kcal/mol in stage 1, and receipt ID v4jv4 obtained RMSEc = 0.29 kcal/mol in stage 2.

    “Particularly, in stage 1, our prediction with receipt ID fomca obtained RMSEc = 0.33 kcal/mol. In stage 2, we accomplished RMSEc = 0.29 kcal/mol with receipt ID v4jv4.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    In GC3 Subchallenge 2, the paper's best result was second place on active/inactive classification of the VEGFR2 set, with MCC = 0.48 from receipt ID rtv8m.

    “Our best result is the second place on the active/inactive classification of VERGFR2 set. Our best Matthews correlation coefficient (MCC) on such task is reported to be 0.48 from receipt ID rtv8m.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    In GC3 Subchallenge 1, the paper's models attained first places for the free energy set in stage 1 and for scoring and free energy set in stage 2.

    “With this setting, our models attained the first places for the following tasks: free energy set in stage 1, scoring and free energy set in stage 2.”
    Quote in the paperSubstantial claim

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

    Premisecitation

    Statistics of element specific persistent homology barcodes can be used for featurization.

    “The statistics of element specific persistent homology barcodes can also be used for featurization [23].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [23] Cang ZX, Mu L, Wei GW (2018) Representability of algebraic topology for biomolecules in machine learning based scoring and virtual screening. PLoS Comput Biol 14(1):e1005929. https://doi.org/10.1371/journal.pcbi.1005929 source

      Not checked in the free run.

  • May combine two claims.

    Premisenone given

    Both the paper's graph-theory-based models and algebraic-topology-based models achieve top-class performance with ensemble-of-trees methods.

    “Both our graph theory based models [20, 62] and algebraic topology based models [21, 23] achieve top-class performance with the ensemble of trees methods.”
    Quote in the paperSubstantial claim0 of 4 sources checked
    • cites [20] Nguyen DD, Xiao T, Wang ML, Wei GW (2017) Rigidity strengthening: a mechanism for protein–ligand binding. J Chem Inform Model 57:1715–1721 source

      Not checked in the free run.

    • cites [62] Nguyen DD, Wei GW (2018) Multiscale weighted colored algebraic graphs for biomolecules (to be submitted)

      Not checked in the free run.

    • cites [21] Cang ZX, Wei, GW (2018) “Integration of element specific persistent homology and machine learning for protein–ligand binding affinity prediction. Int J Numer Methods Biomed Eng. https://doi.org/10.1002/cnm.2914 source

      Not checked in the free run.

    • cites [23] Cang ZX, Mu L, Wei GW (2018) Representability of algebraic topology for biomolecules in machine learning based scoring and virtual screening. PLoS Comput Biol 14(1):e1005929. https://doi.org/10.1371/journal.pcbi.1005929 source

      Not checked in the free run.

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

    Interpretationreasoning

    A higher-level bagging of models of different kinds can further improve performance.

    “Therefore, a higher level bagging of models of different kinds can further improve the performance.”
    Quote in the paperNot a substantial claim
    • Rests on an assumption: Random forest is better at lowering bias and gradient boosting trees focus more on reducing variance.

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

    Interpretationreasoning

    The topological analysis provides a description of the molecular system that captures key aspects including multiscale geometry, interaction network characterization, and physical/chemical information, ensuring success of downstream machine learning modeling.

    “The topological analysis provides us a description of the molecular system that captures a collection of key aspects of the system including the multiscale description of geometry, the characterization of interaction network in an arbitrary dimension, and the important physical and chemical information, which ensures the success of the downstream machine learning modeling.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: The intrinsic physics of interest of complex biomolecules and interactions lies on low-dimensional manifolds or subspaces embedded in a high-dimensional data space.

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

    Premisecitation

    Element specific persistent homology barcodes can be transformed to image-like representations and used in 1-dimensional or 2-dimensional convolutional neural networks.

    “The barcodes can also be transformed to representations similar to images and used in a 1-dimensional or a 2-dimensional convolutional neural networks [22, 23].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [22] Cang ZX, Wei GW (2017) TopologyNet: topology based deep convolutional and multi-task neural networks for biomolecular property predictions. PLoS Comput Biol 13(7):e1005690. https://doi.org/10.1371/journal.pcbi.1005690 source

      Not checked in the free run.

    • cites [23] Cang ZX, Mu L, Wei GW (2018) Representability of algebraic topology for biomolecules in machine learning based scoring and virtual screening. PLoS Comput Biol 14(1):e1005929. https://doi.org/10.1371/journal.pcbi.1005929 source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Element specific persistent homology computation on all carbon atoms describes the hydrophobic network or structural basis of a molecule, while computation on nitrogen and oxygen atoms characterizes hydrophilic networks.

    “specific persistent homology computation on the collection of all carbon atoms describes the hydrophobic network or the structural basis of the molecule while computation on the nitrogen and oxygen atoms characterizes the hydrophilic networks [21].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [21] Cang ZX, Wei, GW (2018) “Integration of element specific persistent homology and machine learning for protein–ligand binding affinity prediction. Int J Numer Methods Biomed Eng. https://doi.org/10.1002/cnm.2914 source

      Not checked in the free run.

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

    Premisecitation

    The MWCG approach is over 40% more accurate than the Gaussian network model on a set of 364 proteins.

    “The MWCG approach has been shown to be over 40% more accurate than the Gaussian network model on a set of 364 proteins [24].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [24] Bramer D, Wei G-W (2018) Multiscale weighted colored graphs for protein flexibility and rigidity analysis. J Chem Phys 148(5):054103 source

      Not checked in the free run.

    • The authors state this as an assumption: The intrinsic physics of interest of complex biomolecules and interactions lies on low-dimensional manifolds or subspaces embedded in a high-dimensional data space.

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

    Premisecitation

    The Wasserstein metric can measure similarities among barcodes of different proteins and can be used with nearest neighbors and manifold learning.

    “For example, Wasserstein metric can be applied to measure similarities among the barcodes of different proteins, which can be used with methods such as nearest neighbors and manifold learning [23].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [23] Cang ZX, Mu L, Wei GW (2018) Representability of algebraic topology for biomolecules in machine learning based scoring and virtual screening. PLoS Comput Biol 14(1):e1005929. https://doi.org/10.1371/journal.pcbi.1005929 source

      Not checked in the free run.

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

    Premisecitation

    Topological descriptors integrated with deep learning achieved state-of-the-art virtual screening of active ligands and decoys.

    “Topological descriptors were further integrated with a variety of deep learning algorithms to achieve state-of-the-art analysis and prediction of protein folding stability change upon mutation [38], drug toxicity [39], aqueous solubility [40], partition coefficient [40], binding affinity [21, 22], and the virtual screening of active ligands and decoys [23].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [23] Cang ZX, Mu L, Wei GW (2018) Representability of algebraic topology for biomolecules in machine learning based scoring and virtual screening. PLoS Comput Biol 14(1):e1005929. https://doi.org/10.1371/journal.pcbi.1005929 source

      Not checked in the free run.

    • The authors state this as an assumption: The intrinsic physics of interest of complex biomolecules and interactions lies on low-dimensional manifolds or subspaces embedded in a high-dimensional data space.

  • May combine two claims.

    Interpretationown data

    The paper's results confirm the predictive power and practical usage of its mathematical deep learning models in drug design and discovery.

    “These results confirm the predictive power and practical usage of our mathematical deep learning models in drug design and discovery.”
    Quote in the paperSubstantial claim

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

    Premisecitation

    Machine learning-based scoring functions using advanced algorithms such as random forest and deep convolutional neural networks can characterize non-additive contributions of functional groups in protein-ligand binding interactions.

    “By using advanced machine learning algorithms, such as random forest (RF) and deep convolutional neural network, the machine learning-based SFs can characterize the non-additive contributions of functional groups in protein–ligand binding interactions [18].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [18] Baum B, Muley L, Smolinski M, Heine A, Hangauer D, Klebe G (2010) Non-additivity of functional group contributions in protein–ligand binding: a comprehensive study by crystallography and isothermal titration calorimetry. J Mol Biol 397(4):1042–1054 source

      Not checked in the free run.

    • Rests on an assumption: The performance of machine learning-based scoring functions depends crucially on the quality and statistical distribution of training data and on selected features that may or may not accurately and completely describe protein–ligand binding interactions.
16 of 16

  • The authors state this as an assumption.

    The intrinsic physics of interest of complex biomolecules and interactions lies on low-dimensional manifolds or subspaces embedded in a high-dimensional data space.

    “We assume that the intrinsic physics of interest of complex biomolecules and interactions lies on low-dimensional manifolds or subspaces embedded in a high-dimensional data space.”

  • The paper relies on this without support.

    The performance of machine learning-based scoring functions depends crucially on the quality and statistical distribution of training data and on selected features that may or may not accurately and completely describe protein–ligand binding interactions.

    “However, the performance of machine learning-based SFs depends crucially on the training data quality and statistic distribution. Additionally, it also depends on selected features that might or might not accurately and completely describe the protein–ligand binding interactions.”

  • The authors state this as an assumption.

    Grouping weak learners can improve learning capability.

    “Ensemble of trees methods build models consisting of a collection of decision trees with the assumption that grouping the weak learners can improve the learning capability.”

  • The paper relies on this without support.

    Random forest is better at lowering bias and gradient boosting trees focus more on reducing variance.

    “As these two models attempt to reduce error in two different ways, they behave differently in the bias-variance trade-off where the random forest is better at lowering bias and gradient boosting trees focus more on reducing variance.”

  • The paper relies on this without support.

    Ensemble learning methods are robust and overfitting can be reduced by learning partial problems.

    “The ensemble learning methods are also robust and overfitting can be reduced by learning partial problems.”

  • The paper relies on this without support.

    When the feature is complex or has an underlying dimension in feature space, deep learning models can further improve predictor performance.

    “When the feature is complex or there is some underlying dimension in the feature space, deep learning models can further imporve the performance of the predictor.”

  • The paper relies on this without support.

    A spatial dimension associated to the filtration parameter lies in the persistent homology representation of protein–ligand systems.

    “For example, a spatial dimension associated to the filtration parameter lies in the persistent homology representation of protein–ligand systems.”
7 of 7

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?

Cites

Flagged first, then by citations

Research works printed but not linked: 2

  1. Schrödinger LLC (2017) Schrödinger release 2017-4. Schrödinger LLC, New York
  2. Edelsbrunner H (1992) Weighted alpha shapes. Technical report. University of Illinois, Champaign

Cited by

Sorted by citations

Related

Authors
Nguyen, D. D., Cang, Z., Wu, K., Wang, M., Cao, Y., & Wei, G.-W.
Affiliations
Michigan State University (US)
Corresponding author
Wei, Guo-Wei
ORCID
Venue
Journal of Computer-Aided Molecular Design
Publisher
Springer Science and Business Media LLC
Published
2018-08-16
Volume
33
Issue
1
Pages
71-82
ISSN
  • 0920-654X (ISSN-L)
  • 1573-4951
DOI
10.1007/s10822-018-0146-6
PMC
PMC7163798
OpenAlex
W2963883198
Semantic Scholar
2ca793d6b2ce88d315b20441f9aa4189965d429a
Cited by
155
References
70
Citations by year
2026: 8 · 2025: 25 · 2024: 19 · 2023: 18 · 2022: 20 · 2021: 27 · 2020: 21 · 2019: 15 · 2018: 2
FWCI
12.13
Type
journal article
Language
en
Open access
Licence
www.springer.com/tdm (tdm)
Locations
  • Journal of Computer-Aided Molecular Design (publishedVersion)
  • PubMed (publishedVersion)
  • PubMed Central (acceptedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University)
Funders
  • National Science Foundation (1160352, 1302285, 1761320)
  • National Institute of General Medical Sciences (R01GM126189)
Topics
  • Physical Sciences › Computer Science › Computational Theory and Mathematics › Computational Drug Discovery Methods
  • Physical Sciences › Materials Science › Materials Chemistry › Machine Learning in Materials Science
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Protein Degradation and Inhibitors
Keywords
binding affinity ranking, free energy prediction, persistent homology, cathepsin S, kinase inhibitors, VEGFR2, JAK2, Tie2, c-Abl
MeSH terms
Deep Learning, Receptor, Farnesoid X-Activated, Binding Sites, Cathepsins, Ligands, Protein Binding, Protein Conformation, Protein Kinases, Thermodynamics, Drug Design, Computer-Aided Design, Receptors, Cytoplasmic and Nuclear, Crystallography, X-Ray, Databases, Protein, Molecular Docking Simulation

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