Pith. sign in

REVIEW 6 major objections 4 minor 48 references

Contrastive Multi-Task Learning with Solvent-Aware Augmentation for Drug Discovery

T0 review · 6 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read SolvCLIP claims that solvent-aware multi-task pre-training of a frozen-encoder interaction module improves protein-ligand binding affinity, docking-pose, and virtual screening in one shared representation.

desk verdict A promising solvent-aware pre-training idea whose central claim is not actually tested, because the solvent conditioning itself is never ablated. read the letter →

arxiv 2508.01799 v2 pith:WTGSW4AP submitted 2025-08-03 q-bio.BM cs.AIcs.LG

Jing Lan , Hexiao Ding , Hongzhao Chen , Yufeng Jiang , Nga-Chun Ng , Gerald W.Y. Cheng , Zongxi Li , Jing Cai
show 2 more authors
Liang-ting Lin Jung Sun Yoo
This is my paper · ORCID
classification q-bio.BMcs.AIcs.LG
keywords solvent-awareaugmentationprotein-ligandinteractioncontrastivepre-trainingbindingaffinitypredictionmoleculardockingvirtualscreeningconformationalensemblesSE(3)-equivarianttransformer
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

SolvCLIP, the framework introduced here, claims that a single pre-trained representation can be made solvent-aware and can then transfer to the three workhorse tasks of structure-based drug discovery: binding affinity regression, docking pose prediction, and virtual screening. To test this, the authors pre-train a frozen-encoder interaction module on 311,123 solvent-augmented protein-ligand complexes, then fine-tune task heads. They report a binding-affinity RMSE of 1.239 on the Atom3D LBA test set, an 82% success rate at <2.0 Å on the Astex docking benchmark, and a 97.1% AUC on DUD-E virtual screening. The paper's claim is that the gain comes from learning ligand conformational flexibility under different solvent conditions, not from task-specific architectures or extra supervised labels.

What carries the argument

The load-bearing object is a solvent-aware interaction module: an $N$-layer $\mathrm{SE}(3)$-equivariant transformer that takes frozen atom-level representations of a pocket and a ligand, plus a ligand conformer ensemble, and refines them into pairwise binding representations. It is trained with three objectives: Masked Molecular Reconstruction (MMR), which masks atoms and reconstructs atom-type and embedding targets under a Gaussian cross-entropy loss; Interatomic Distance Matrix Prediction (IDMP), an L2 regression of pocket-ligand pairwise distances; and Contrastive Learning (CL), which aligns the distance matrices of the same complex in unbound-like and solvent-augmented states while treating swapped-pocket or swapped-ligand complexes as negatives. The solvent views come from a GNN-based implicit-solvent model trained on classical molecular dynamics, producing conformational ensembles across the paper's solvent set (its Table I lists 38 solvents), with an unbound-like primary conformer generated by Gaussian torsion and coordinate perturbation. Together the objectives are meant to capture local geometry, global spatial organization, and solvent invariance in one representation.

What would settle it

Two checks would settle it: (1) compute ligand and pocket similarity between every BioLip pre-training complex and the ligands and pockets of each test set (Atom3D LBA, CASF-2016, PoseBusters, Astex, DUD-E actives) and rerun the benchmarks after removing near-duplicates; (2) re-train the full model with the solvent-conditioned conformers replaced by the same number of random RDKit conformers lacking solvent identity. If the reported gains disappear in either check, the solvent-aware generalization claim is not supported.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that ligand conformational ensembles generated in different solvents, when injected into pre-training as augmented views, teach a protein-ligand interaction module a representation that is simultaneously geometry-aware and solvent-invariant. The authors build this by keeping two pre-trained SE(3)-equivariant encoders (pocket and ligand) frozen and training only an interaction transformer with three self-supervised objectives: masked atom reconstruction, interatomic distance regression, and contrastive alignment of the same complex across solvent states. After fine-tuning, the same representation improves affinity, pose, and screening benchmarks relative to ProNet, Uni-Mol, and DrugCLIP baselines, and the paper's ablation attributes the gains to the joint objectives rather than to any single one. The case study of glycosidase inhibitor azafagomine in PDB entry 2J7H, with a predicted RMSD of 0.157 Å, is offered as evidence of atomic-level fidelity.

Load-bearing premise

The argument presumes that the 311,123 solvent-augmented pre-training complexes contain no near-duplicates of the downstream test complexes, because the paper only filters the fine-tuning training set for similarity to one test set and does not filter the pre-training corpus itself.

Editorial extensions

If this is right

  • A single pre-trained module can be fine-tuned for affinity, pose, and screening, so a drug-discovery pipeline could use one representation instead of separate task-specific models.
  • Early-enrichment gains on DUD-E (RE@0.5% of 109.20 vs 73.97 for DrugCLIP zero-shot) imply solvent-aware pre-training helps rank true actives at the top of a screened library, which is the operationally important regime in virtual screening.
  • The ablation results imply that masked local reconstruction is the largest single contributor, interatomic-distance prediction adds stability, and contrastive learning protects early enrichment; all three are claimed to be synergistic.
  • Because the encoders stay frozen and only the interaction module and task head are updated during fine-tuning, the same pre-trained weights can cheaply be adapted to new tasks or targets.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper does not compare solvent-conditioned conformers against an equal-sized set of random, solvent-agnostic conformers; that comparison would isolate whether solvent identity itself, rather than mere conformational diversity, drives the gains.
  • If the solvent signal transfers, the same augmentation recipe could be applied to other structure-based tasks such as binding-site detection or protein-ligand co-folding, where conformational flexibility also matters.
  • Because the upper bound of the solvent signal is set by the GNN implicit-solvent model, the method's ceiling is tied to that surrogate's accuracy; replacing it with explicit-solvent or quantum-derived ensembles is a natural stress test.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

6 major / 4 minor

Summary. The paper proposes SolvCLIP, a self-supervised pre-training framework for protein–ligand interaction prediction. The method augments ligand representations with conformational ensembles generated under many solvent conditions, and combines three pre-training objectives: masked molecular reconstruction (MMR), interatomic distance matrix prediction (IDMP), and a contrastive loss (CL) that aligns same-complex distance matrices across solvent conditions. The authors evaluate the framework on ligand binding affinity prediction (Atom3D LBA), molecular docking (CASF-2016, PoseBusters, Astex), and virtual screening (DUD-E), reporting improved RMSE and correlation, docking success rates, and AUC/enrichment over strong baselines. An ablation study on DUD-E is provided, together with a docking case study on PDB entry 2J7H.

Significance. If the reported results are reliable, solvent-aware multi-task pre-training would be a useful contribution to structure-based drug design, and the paper would advance the practical use of conformational ensembles in protein–ligand representation learning. I credit the authors for making code available, for using a large pre-training corpus, for evaluating on several standard benchmarks, and for including a DUD-E ablation with mean ± std. However, the current evidence does not establish the central claim: the solvent-aware augmentation itself is never ablated, and the pre-training corpus is not screened for overlap with downstream test sets. The selective reporting of docking cutoffs and the undefined '3.7% gain' further weaken the quantitative conclusions. If the missing controls can be supplied and the leakage risk resolved, the work would be worth revisiting; as presented, the empirical support is insufficient.

major comments (6)
  1. [Section III-A-1; Section IV-A-2] The pre-training corpus (January 2025 BioLip, reduced to 311,123 solvent-augmented complexes) is not screened for similarity or identity with any downstream test set. The fine-tuning protocol for CASF-2016 follows Uni-Mol's similarity filtering, but no analogous filter is applied to the pre-training data against Atom3D LBA, CASF-2016, PoseBusters/Astex, or DUD-E. Because BioLip is a PDB-derived database, many pre-training complexes are likely to be identical or near-duplicates of benchmark complexes; without an overlap analysis, the reported gains could reflect memorization rather than solvent-aware generalization. Please provide a similarity/identity analysis between the pre-training set and each test set, and either rerun pre-training after removing overlapping complexes or report results on the non-overlapping subset.
  2. [Section II-A, Eq. (11); Section IV-B, Table V] The central novelty, solvent-aware augmentation, is never ablated. In Table V, the rows remove MMR, IDMP, or CL, but every configuration still uses the full solvent-conformer ensemble as input to MMR and CL; therefore the full model's gain over the no-CL row could come from the contrastive loss, from greater conformational diversity, or from additional reconstruction targets rather than from solvent conditioning per se. To support the claim that solvent-aware augmentation causes the improvement, include a control trained only on the primary/unbound-like ligand conformation with no solvent conformers, and also control for the number of conformers so that solvent diversity is distinguished from mere multi-conformer diversity.
  3. [Abstract; Section IV-A-1, Table II] The abstract's '3.7% gain in binding affinity prediction' is not defined anywhere in the text. Table II reports RMSE, Pearson, and Spearman, while the results paragraph mentions a '7.7% RMSE reduction' and an '8.7% RMSE reduction' without stating the metric to which 3.7% refers or the baseline. Please specify the exact metric and baseline, show the computation, and provide an uncertainty estimate for this quantity.
  4. [Tables II, III, IV; Section IV-B] The main performance tables do not report error bars, repeated runs, or significance tests. Only Table V (ablation on DUD-E) gives mean ± std; Tables II, III, and IV report single values, so the differences against strong baselines (e.g., Pearson 0.793 vs. 0.765 for ProNet in Table II) cannot be evaluated statistically. Please add standard deviations across multiple seeds or bootstrap confidence intervals, and where direct comparisons are made, report paired significance tests.
  5. [Section IV-A-2, Table III, Fig. 4] The docking presentation is selective. On CASF-2016, Table III shows that the proposed method is below Uni-Mol Docking at the <2.0 Å (76.5% vs. 80.4%) and <3.0 Å (86.3% vs. 87.0%) cutoffs, yet the text claims the highest success rate based only on the <5.0 Å cutoff; the highlighted Astex 82% success rate in Fig. 4 is not accompanied by baseline numbers for the same set. Please report all cutoffs and all methods on PoseBusters and Astex in a common table, including per-method success rates and confidence intervals, and avoid presenting only the cutoff that favors the proposed method.
  6. [Section IV-A-3] The DUD-E evaluation protocol is described as 'strict three-fold cross-validation to prevent data leakage,' but the splitting unit is not stated. If the split is at the level of individual active/decoy compounds rather than targets, the model can see the same target in both training and test folds, which would artificially inflate AUC and enrichment (e.g., AUC 97.1% and RE@0.5% 109.2). Please clarify whether the cross-validation is target-level or compound-level; if it is compound-level, re-run the evaluation with target-level splits or report per-target metrics.
minor comments (4)
  1. [Section II-A; Section III-A-2; Eq. (11)] The text and Eq. (11) refer to 39 solvent conformers and 39 organic solvents, while Section III-A-2 and Table I list 38 distinct solvents; please reconcile this inconsistency.
  2. [Abstract; Section IV-A-2] The phrase 'PoseBusters Astex docking benchmarks' is ambiguous: the abstract reports 82% on Astex, while the text also reports 64% on PoseBusters (428 complexes). Please state clearly that Astex is a separate or subset benchmark and give the numbers and baselines for each.
  3. [Section IV-A-1] The fine-tuning protocol for the LBA task is not described (epochs, learning rate schedule, early stopping, number of seeds). This information is needed to assess whether the reported RMSE and correlations are stable or tuned to the test set.
  4. [Section IV-C, Fig. 4(b)] The case study reports an RMSD of 0.157 Å for PDB entry 2J7H, described as below the coordinate error of the crystal structure; please state how this entry was selected and whether it is representative, since a single favorable example is not evidence of general docking accuracy.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: all reported results are external benchmark measurements; the main weaknesses are an omitted ablation and possible pre-training/test overlap, neither of which makes any output equal to an input by construction.

full rationale

I walked the derivation chain from pre-training objectives (Eqs. 9-12) through fine-tuning to the reported metrics (Tables II-V, Fig. 4). The model is pre-trained on BioLip with RDKit/AMBER-MD conformers and the GNN implicit-solvent generator of Ref. [15]; downstream evaluations use Atom3D LBA, PDBbind/CASF-2016, PoseBusters/Astex, and DUD-E, all external benchmarks. No reported number is the identity of a fitted parameter, and no equation reduces a prediction to its training input. The contrastive loss (Eq. 12) is explicitly an objective that aligns distance matrices, so any resulting solvent invariance is an intended property trained into the representation, not a separately derived law. The paper has no same-author citations that carry a load-bearing uniqueness or ansatz argument. Two weaknesses are evident but they are not circularity: (1) the ablation in Table V removes MMR, IDMP, and CL but never removes the solvent-aware conformer ensemble, so the unique contribution is not directly isolated; (2) the January 2025 BioLip pre-training set is not filtered for similarity to downstream test sets, so memorization cannot be excluded. These concerns affect evidential strength, not logical circularity. I therefore assign a score of 0.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central empirical claim rests on several inherited tools (frozen Uni-Mol/DrugCLIP encoders, an external GNN implicit solvent model) and hand-selected hyperparameters. The most distinctive assumption is that solvent variation can be treated as noise to be removed via contrastive alignment, which is embedded in the loss rather than tested. No new physical entities are introduced.

free parameters (4)
  • Contrastive temperature tau = not reported
    Temperature in the contrastive loss (Eq. 12) controls the scale of similarity scores; no sensitivity analysis is provided, and it directly shapes the learned representation.
  • Number of solvent conformers per ligand = 39
    Each ligand is augmented with 39 conformers across 38 solvents; this number is a design choice not ablated, despite the central role of solvent augmentation in the method.
  • Masking hyperparameters (probability, ratio, noise) = 0.15, 0.8, 0.8
    Masked molecular reconstruction uses these values; they are chosen by hand and are not varied or justified empirically.
  • Pocket radius = 8 Å
    Residues within 8 Å of ligand heavy atoms define the binding pocket; this geometric cutoff affects all input graphs and is a standard but arbitrary choice.
assumptions (4)
  • standard math SE(3)-equivariant transformer maintains equivariance when composing pocket and ligand subgraphs
    The framework relies on the encoders and interaction module being SE(3)-equivariant, as claimed in Section II-A; this is inherited from Uni-Mol/DrugCLIP and not reproved.
  • domain assumption Distance matrices between pocket and ligand atoms capture the information needed for binding affinity and pose
    The IDMP objective and contrastive loss both operate on distance matrices, assuming these are sufficient structural descriptors for the downstream tasks.
  • domain assumption The GNN-based implicit solvent model generates physically accurate solvent-specific conformer ensembles
    Section III-A-2 uses the approach of Katzberger et al. [15]; the accuracy of these conformers is taken as given and not validated in this paper.
  • ad hoc to paper Contrastive positive pairs (same complex, different solvents) should be represented similarly
    Eq. (12) forces distance-matrix representations of different solvent conformers to align, implicitly assuming solvent does not change binding-relevant geometry; this is a modeling choice specific to this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Contrastive Multi-Task Learning with Solvent-Aware Augmentation for Drug Discovery." pith.science (2026). https://pith.science/paper/WTGSW4AP

@misc{pith2026250801799,
  author       = {Pith},
  title        = {Pith review of: Contrastive Multi-Task Learning with Solvent-Aware Augmentation for Drug Discovery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WTGSW4AP}},
  note         = {Machine review of arXiv:2508.01799}
}
read the original abstract

Accurate prediction of protein-ligand interactions is essential for computer-aided drug discovery. However, existing methods often fail to capture solvent-dependent conformational changes and lack the ability to jointly learn multiple related tasks. To address these limitations, we introduce a pre-training method that incorporates ligand conformational ensembles generated under diverse solvent conditions as augmented input. This design enables the model to learn both structural flexibility and environmental context in a unified manner. The training process integrates molecular reconstruction to capture local geometry, interatomic distance prediction to model spatial relationships, and contrastive learning to build solvent-invariant molecular representations. Together, these components lead to significant improvements, including a 3.7% gain in binding affinity prediction, an 82% success rate on the PoseBusters Astex docking benchmarks, and an area under the curve of 97.1% in virtual screening. The framework supports solvent-aware, multi-task modeling and produces consistent results across benchmarks. A case study further demonstrates sub-angstrom docking accuracy with a root-mean-square deviation of 0.157 angstroms, offering atomic-level insight into binding mechanisms and advancing structure-based drug design.

Figures

Figures reproduced from arXiv: 2508.01799 by the authors.

Figure 1
Figure 1. Solvent-Aware Data Augmentation Pipeline. (a) Protein sequences and ligand SMILES are transformed into solvent-aware 3D conformer en￾sembles using RDKit and AMBER-MD; (b) The SE(3)-Transformer encodes pocket-ligand graphs based on interatomic distances and edge types; (c) Contrastive learning is employed to align the pocket-ligand complex with its solvent-augmented counterpart. A. Overview To capture intricate prote… view at source ↗
Figure 2
Figure 2. Overview of the pre-training framework. (a) The complex is decomposed into the apo pocket and the ligand. The ligand is augmented with solvent-aware conformers; (b) Two encoders are employed to generate initial pocket and ligand representations, with their parameters kept frozen throughout both training and inference; (c) A solvent-aware interaction module refines the embeddings for downstream tasks via contrastive … view at source ↗
Figure 3
Figure 3. Contrastive Learning (CL). Align interatomic distance matrix of the same pocket-ligand pair (solvent vs. non-solvent) as positives; negatives swap ligand or pocket within the batch. III. EXPERIMENTS A. Training Data 1) Dataset Construction: We curated 348,970 protein￾ligand complexes from the January 2025 BioLip [27] release through a multi-stage filtration process: retaining only drug￾like ligands bound to protein … view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Performance comparison on PoseBusters and Astex. (a) overall success rates (<2.0 A RMSD) across 428 PoseBusters complexes and 85 Astex ˚ entries for all evaluated methods; “Ours” is highlighted. (b) case-study of structural superposition for PDB entry 2J7H (Thermotoga …

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 48 canonical work pages

  1. [1]

    Drugclip: Contrastive protein–molecule representation learning for virtual screening,

    B. Gao et al. , “Drugclip: Contrastive protein–molecule representation learning for virtual screening,” Advances in Neural Information Pro- cessing Systems, vol. 36, pp. 44 595–44 614, 2023

    work page 2023

  2. [2]

    Normalized protein–ligand distance likelihood score for end-to-end blind docking and virtual screening,

    S. Xia, Y . Gu, and Y . Zhang, “Normalized protein–ligand distance likelihood score for end-to-end blind docking and virtual screening,” Journal of Chemical Information and Modeling, vol. 65, no. 3, pp. 1101– 1114, 2025

    work page 2025

  3. [3]

    Deep Confident Steps to New Pockets: Strategies for Docking Generalization

    G. Corso et al. , “Deep confident steps to new pockets: Strategies for docking generalization,” arXiv, p. arXiv:2402.18396v1, 2024

    work page Pith review arXiv 2024

  4. [4]

    Molecular docking: State-of-the-art scoring functions and search algorithms,

    R. M. de Angelo et al. , “Molecular docking: State-of-the-art scoring functions and search algorithms,” in Computer-Aided and Machine Learning-Driven Drug Design: From Theory to Applications. Springer, 2025, pp. 163–198

    work page 2025

  5. [5]

    Recent advances in ai-driven protein–ligand interaction predictions,

    J. Sim et al. , “Recent advances in ai-driven protein–ligand interaction predictions,” Current Opinion in Structural Biology , vol. 92, p. 103020, 2025

    work page 2025

  6. [6]

    Reliable and accurate solution to the induced fit docking problem for protein–ligand binding,

    E. B. Miller et al. , “Reliable and accurate solution to the induced fit docking problem for protein–ligand binding,” Journal of Chemical Theory and Computation , vol. 17, no. 4, pp. 2630–2639, 2021

    work page 2021

  7. [7]

    Molecular dynamics simulations for the protein–ligand complex structures obtained by computational docking studies using implicit or explicit solvents,

    K. Kato et al., “Molecular dynamics simulations for the protein–ligand complex structures obtained by computational docking studies using implicit or explicit solvents,” Chemical Physics Letters , vol. 781, p. 139022, 2021

    work page 2021

  8. [8]

    Uni-mol: A universal 3d molecular representation learning framework,

    G. Zhou et al. , “Uni-mol: A universal 3d molecular representation learning framework,” 2023

    work page 2023

Show all 48 references
  1. [9]

    Equibind: Geometric deep learning for drug binding structure prediction,

    H. St ¨ark et al. , “Equibind: Geometric deep learning for drug binding structure prediction,” in International Conference on Machine Learning. PMLR, 2022

    2022

  2. [10]

    E3bind: An end-to-end equivariant network for protein– ligand docking,

    Y . Zhang et al., “E3bind: An end-to-end equivariant network for protein– ligand docking,” arXiv preprint arXiv:2210.06069 , 2022

    2022 arXiv

  3. [11]

    Tankbind: Trigonometry-aware neural networks for drug– protein binding structure prediction,

    W. Lu et al., “Tankbind: Trigonometry-aware neural networks for drug– protein binding structure prediction,” Advances in Neural Information Processing Systems, vol. 35, pp. 7236–7249, 2022

    2022

  4. [12]

    Efficient and accurate large library ligand docking with KarmaDock,

    X. Zhang et al., “Efficient and accurate large library ligand docking with KarmaDock,” Nature Computational Science, vol. 3, no. 9, pp. 789–804, 2023

    2023

  5. [13]

    Cosp: Co-supervised pretraining of pocket and ligand,

    Z. Gao, C. Tan, J. Xia, and S. Z. Li, “Cosp: Co-supervised pretraining of pocket and ligand,” in Machine Learning and Knowledge Discovery in Databases: Research Track: European Conference, ECML PKDD 2023, Turin, Italy, September 18–22, 2023, Proceedings, Part I , 2023, p. 405–421

    2023

  6. [14]

    Contrastive learning in protein language space predicts interactions between drugs and protein targets,

    R. Singh et al., “Contrastive learning in protein language space predicts interactions between drugs and protein targets,” Proceedings of the National Academy of Sciences , vol. 120, no. 24, p. e2220778120, 2023

    2023

  7. [15]

    Rapid access to small-molecule conformational ensembles in organic solvents enabled by graph neural network-based implicit solvent model,

    P. Katzberger et al. , “Rapid access to small-molecule conformational ensembles in organic solvents enabled by graph neural network-based implicit solvent model,” Journal of the American Chemical Society , vol. 147, no. 16, pp. 13 264–13 275, 2025

    2025

  8. [16]

    A quantitative analysis of ligand binding at the protein–lipid bilayer interface,

    A. P. Barkdull, M. Holcomb, and S. Forli, “A quantitative analysis of ligand binding at the protein–lipid bilayer interface,” Communications Chemistry, vol. 8, no. 1, p. 89, 2025

    2025

  9. [17]

    Geometry-enhanced molecular represen- tation learning for property prediction,

    X. Fang, L. Liu, J. Lei et al., “Geometry-enhanced molecular represen- tation learning for property prediction,” Nature Machine Intelligence , vol. 4, pp. 127–134, 2022

    2022

  10. [18]

    Force fields for protein simulations,

    J. W. Ponder and D. A. Case, “Force fields for protein simulations,” Protein Simulations, vol. 66, pp. 27–85, 2003

    2003

  11. [19]

    The rosetta all-atom energy function for macromolecular modeling and design,

    R. F. Alford, A. Leaver-Fay, J. R. Jeliazkov, M. J. O’Meara, F. P. DiMaio, H. Park, M. V . Shapovalov, P. D. Renfrew, V . K. Mulligan, K. Kappel, J. W. Labonte, M. S. Pacella, R. Bonneau, P. Bradley, R. L. J. Dunbrack, R. Das, D. Baker, B. Kuhlman, T. Kortemme, and J. J. Gray,...

    2017

  12. [20]

    A machine learning approach to predicting protein–ligand binding affinity with applications to molecular docking,

    P. J. Ballester and J. B. O. Mitchell, “A machine learning approach to predicting protein–ligand binding affinity with applications to molecular docking,” Bioinformatics, vol. 26, no. 9, pp. 1169–1175, 03 2010

    2010

  13. [21]

    Pignet: a physics- informed deep learning model toward generalized drug–target interaction predictions,

    S. Moon, W. Zhung, S. Yang, J. Lim, and W. Y . Kim, “Pignet: a physics- informed deep learning model toward generalized drug–target interaction predictions,” Chem. Sci., vol. 13, pp. 3661–3673, 2022

    2022

  14. [22]

    Protein-ligand binding representation learning from fine-grained interactions,

    S. Feng, M. Li, Y . Jia, W.-Y . Ma, and Y . Lan, “Protein-ligand binding representation learning from fine-grained interactions,” The Twelfth International Conference on Learning Representations , 2024

    2024

  15. [23]

    Mole- BERT: Rethinking pre-training graph neural networks for molecules,

    J. Xia, C. Zhao, B. Hu, Z. Gao, C. Tan, Y . Liu, S. Li, and S. Z. Li, “Mole- BERT: Rethinking pre-training graph neural networks for molecules,” The Eleventh International Conference on Learning Representations , 2023

    2023

  16. [24]

    Q-gem: Quantum chemistry knowledge fusion geometry- enhanced molecular representation for property prediction,

    Z. Yang, L. Wang, T. Huang, Y . Wang, M. Gao, T. Hou, J. Ding, and J. Xiao, “Q-gem: Quantum chemistry knowledge fusion geometry- enhanced molecular representation for property prediction,” Advanced Science, p. e04867, 2025

    2025

  17. [25]

    Token-mol 1.0: tokenized drug design with large language models,

    J. Wang, R. Qin, M. Wang et al., “Token-mol 1.0: tokenized drug design with large language models,” Nature Communications, vol. 16, p. 4416, 2025

    2025

  18. [26]

    A simple frame- work for contrastive learning of visual representations,

    T. Chen, S. Kornblith, M. Norouzi, and G. E. Hinton, “A simple frame- work for contrastive learning of visual representations,” Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13–18 July 2020, Virtual Event , pp. 1597–1607, 2020

    2020

  19. [27]

    Biolip: a semi-manually curated database for biologically relevant ligand–protein interactions,

    J. Yang, A. Roy, and Y . Zhang, “Biolip: a semi-manually curated database for biologically relevant ligand–protein interactions,” Nucleic Acids Research, vol. 41, no. D1, pp. D1096–D1103, 10 2012

    2012

  20. [28]

    ATOM3d: Tasks on molecules in three dimensions,

    R. J. L. Townshend, M. V ¨ogele, P. A. Suriana, A. Derry, A. Powers, Y . Laloudakis, S. Balachandar, B. Jing, B. M. Anderson, S. Eismann, R. Kondor, R. Altman, and R. O. Dror, “ATOM3d: Tasks on molecules in three dimensions,” Thirty-fifth Conference on Neural Information Proce...

    2021

  21. [29]

    The PDBbind database: Methodologies and updates,

    R. Wang, X. Fang, Y . Lu, C.-Y . Yang, and S. Wang, “The PDBbind database: Methodologies and updates,” Journal of Medicinal Chemistry, vol. 48, no. 12, pp. 4111–4119, jun 2005

    2005

  22. [30]

    Deepdta: deep drug–target binding affinity prediction,

    H. ¨Ozt¨urk, A. ¨Ozg¨ur, and E. Ozkirimli, “Deepdta: deep drug–target binding affinity prediction,” Bioinformatics, vol. 34, no. 17, pp. i821– i829, 09 2018

    2018

  23. [31]

    Learning hierarchical protein representations via complete 3d graph networks,

    L. Wang, H. Liu, Y . Liu, J. Kurtin, and S. Ji, “Learning hierarchical protein representations via complete 3d graph networks,” The Eleventh International Conference on Learning Representations , 2023

    2023

  24. [32]

    Autodock4 and autodocktools4: Automated docking with selective receptor flexibility,

    G. M. Morris, R. Huey, W. Lindstrom, M. F. Sanner, R. K. Belew, D. S. Goodsell, and A. J. Olson, “Autodock4 and autodocktools4: Automated docking with selective receptor flexibility,” Journal of Computational Chemistry, vol. 30, no. 16, pp. 2785–2791, 2009

    2009

  25. [33]

    Autodock vina: Improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading,

    O. Trott and A. J. Olson, “Autodock vina: Improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading,” Journal of Computational Chemistry, vol. 31, no. 2, pp. 455–461, 2010

    2010

  26. [34]

    Vinardo: A scoring function based on autodock vina improves scoring, docking, and virtual screening,

    R. Quiroga and M. A. Villarreal, “Vinardo: A scoring function based on autodock vina improves scoring, docking, and virtual screening,” PLOS ONE, vol. 11, no. 5, pp. 1–18, 05 2016

    2016

  27. [35]

    Lessons learned in empirical scoring with smina from the csar 2011 benchmarking exercise,

    D. R. Koes, M. P. Baumgartner, and C. J. Camacho, “Lessons learned in empirical scoring with smina from the csar 2011 benchmarking exercise,” Journal of Chemical Information and Modeling, vol. 53, no. 8, pp. 1893–1904, aug 2013

    2011

  28. [36]

    Posebusters: Ai-based docking methods fail to generate physically valid poses or generalise to novel sequences,

    M. Buttenschoen, G. M. Morris, and C. M. Deane, “Posebusters: Ai-based docking methods fail to generate physically valid poses or generalise to novel sequences,” Chem. Sci. , vol. 15, pp. 3130–3139, 2024

    2024

  29. [37]

    Directory of useful decoys, enhanced (dud-e): Better ligands and decoys for better benchmarking,

    M. M. Mysinger, M. Carchia, J. J. Irwin, and B. K. Shoichet, “Directory of useful decoys, enhanced (dud-e): Better ligands and decoys for better benchmarking,” Journal of Medicinal Chemistry , vol. 55, no. 14, pp. 6582–6594, jul 2012

    2012

  30. [38]

    Attentionsitedti: an interpretable graph- based model for drug-target interaction prediction using nlp sentence- level relation classification,

    M. Yazdani-Jahromi, N. Yousefi, A. Tayebi, E. Kolanthai, C. J. Neal, S. Seal, and O. O. Garibay, “Attentionsitedti: an interpretable graph- based model for drug-target interaction prediction using nlp sentence- level relation classification,” Briefings in Bioinformatics , vol....

    2022

  31. [39]

    Predicting drug–protein interaction using quasi-visual question answering system,

    S. Zheng, Y . Li, S. Chen, J. Xu, and Y . Yang, “Predicting drug–protein interaction using quasi-visual question answering system,” Nature Ma- chine Intelligence, vol. 2, no. 2, pp. 134–140, Feb. 2020

    2020

  32. [40]

    Glycosidase inhibition: An assessment of the binding of 18 putative transition-state mimics,

    T. M. Gloster, P. Meloncelli, R. V . Stick, D. Zechel, A. Vasella, and G. J. Davies, “Glycosidase inhibition: An assessment of the binding of 18 putative transition-state mimics,” Journal of the American Chemical Society, vol. 129, no. 8, pp. 2345–2354, feb 2007. APPENDIX A. E...

    2007

  33. [41]

    Mean Squared Error (MSE): MSE = 1 N NX i=1 (yi − ˆyi)2 where yi is the observed value, ˆyi is the predicted value, and N is the number of samples

  34. [42]

    Root Mean Squared Error (RMSE): RMSE = vuut 1 N NX i=1 (yi − ˆyi)2

  35. [43]

    Pearson Correlation Coefficient ( r): r = PN i=1(yi − ¯y)(ˆyi − ¯ˆy)qPN i=1(yi − ¯y)2 · qPN i=1(ˆyi − ¯ˆy)2 where ¯y and ¯ˆy are the means of observed and predicted values, respectively

  36. [44]

    Spearman’s Rank Correlation Coefficient ( ρ): ρ = 1 − 6 PN i=1 d2 i N (N 2 − 1) where di is the difference between the ranks of yi and ˆyi

  37. [45]

    Area Under the ROC Curve (AUC): For binary classifi- cation, AUC is computed as: AUC = Z 1 0 TPR(F P R−1(x))dx where TPR (True Positive Rate) and FPR (False Positive Rate) are defined as: TPR = TP TP + FN , FPR = FP FP + TN

  38. [46]

    ROC Enrichment at k% (RE @k): The enrichment factor for the top k% of predictions is: RE@k = Number of actives in top k% Total actives × 100 k Reported for k = {0.5, 1, 2, 5}: RE@0.5 = TP@0.5% Nactive × 200, RE@1 = TP@1% Nactive × 100 RE@2 = TP@2% Nactive × 50, RE@5 = TP@5% Na...

  39. [47]

    Concordance Index (CI): CI = 1 Z X δj >δi h (bi − bj) h(x) =    0 x <0 0.5 x = 0 1 x >0

  40. [48]

    r2 m Coefficient of Determination: r2 m = r2 × 1 − q r2 − r2 0 where r denotes the squared correlation coefficients between the observed and predicted values with intercepts and r0 is the coefficient without intercepts

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.