Pith. sign in

REVIEW 4 major objections 5 minor 64 references

PocketVina Enables Scalable and Highly Accurate Physically Valid Docking through Multi-Pocket Conditioning

T0 review · 4 major / 5 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 A training-free search-based docking pipeline, PocketVina, achieves the highest rate of physically valid poses on four established benchmarks by docking each ligand into multiple predicted pockets rather than one known site or the whole…

desk verdict The multi-pocket sampling idea is sound and the engineering is solid, but the headline claim rests on an oracle best-RMSD selection that the comparison numbers don't match; ask for top-1-by-Vina-score results before trusting the SOTA claim. read the letter →

arxiv 2506.20043 v1 pith:MJCTZG3N submitted 2025-06-24 q-bio.QM cs.LGq-bio.BM

Ahmet Sarigun , Bora Uyar , Vedran Franke , Altuna Akalin This is my paper
classification q-bio.QMcs.LGq-bio.BM
keywords moleculardockingmulti-pocketconditioningP2RankQuickVina2-GPUphysicallyvalidPoseBustersvirtualscreeningprotein–ligandinteraction
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

PocketVina claims that the best way to get ligand poses that are both geometrically accurate and physically plausible is to combine classical search-based docking with modern pocket prediction: predict several candidate pockets with P2Rank, dock the ligand into each one with an accelerated Vina-family engine, and pool the poses. On four benchmarks, the paper reports the highest share of poses that sit within $2$ Å of the crystal pose and pass all PoseBusters physical-validity checks — $50.96\%$ on the PDBbind2020 time-split set, against $41.41\%$ for the best classical competitor and $36.46\%$ for the best deep-learning one — and the gap widens on proteins unseen by deep-learning training data. The paper argues that a training-free, search-based pipeline therefore generalizes to novel targets better than generative deep-learning models, at a fraction of the training investment. A separate claim is that the same pipeline's built-in affinity score separates active from inactive drug–target pairs on a newly curated 563,000-pair screening benchmark, where the deep-learning comparison protocol's metrics either fail to separate or point in the wrong direction.

What carries the argument

The method's load-bearing parts are three. (1) P2Rank, a random-forest pocket predictor that scores points on the solvent-accessible surface from distance-weighted sums of atomic properties, clusters high-scoring points with a $3$ Å single-linkage cutoff, and returns ranked pocket centers — it supplies the multiple docking targets. (2) QuickVina 2-GPU 2.1, an OpenCL-accelerated AutoDock Vina variant whose grid cache is shared across ligands and whose sampling loop uses random mutations, a first-order consistency check, RILC-BFGS (Reduced Iteration and Low-Complexity BFGS) local optimization with a Lewis–Overton line search, and Metropolis acceptance — it converts each pocket center into docked poses. (3) Multi-pocket conditioning itself, the act of docking into every ranked pocket center and pooling the results so that a wrong or missing top pocket does not sink the prediction. The union is a pipeline with no learnable parameters, and the paper's argument is that this search-based design, not learned pose generation, is what delivers physically valid poses on novel proteins.

What would settle it

Re-run the four benchmarks reporting, per complex, the single pose PocketVina's own binding-affinity score ranks first instead of the best-r.m.s.d. pose, normalized over the full 363, 144, 189, and 428 complexes including those with no predicted pocket; if the top-1 PB-valid rate falls to the level of GNINA (41.41%) or below, the claim that multi-pocket conditioning drives state-of-the-art physical validity would fail. A second check: list the failed complexes by rotatable-bond count and pocket depth — if failures concentrate in flexible ligands or shallow pockets, the reported advantage across flexibility classes would need qualification.

Watch

Extended reading notes

Core claim

The paper's central discovery is that conditioning classical docking on multiple predicted pockets — rather than a single known binding site or the entire protein — resolves the trade-off between speed and generalizability that splits the field into classical and deep-learning approaches. Concretely, PocketVina runs P2Rank to get a ranked list of pocket centers, then runs QuickVina 2-GPU 2.1 separately against each center with a $15$–$30$ Å search box and 20 sampled conformations per pocket, and pools all poses. Because Vina's search is physically constrained and its scoring function is trained on no dataset, the best pose per complex passes PoseBusters checks at a far higher rate than the scored poses reported for DiffDock, DiffDock-L, and SurfDock, and the advantage grows on low-sequence-similarity proteins where trained models degrade. The paper reports the highest physically valid success rates on PDBbind2020 time-split and unseen-protein subsets, DockGen-full, Astex, and PoseBusters, and closes with a screening benchmark in which the QuickVina affinity of the sampled poses separates active from inactive drug–target pairs with high statistical significance.

Load-bearing premise

The headline success rates assume the comparison is fair: PocketVina is credited with its most accurate sampled pose per complex while the other methods are credited with the pose their own scoring function selects, and the complexes where pocket prediction finds nothing (59 of 363, 24 of 144, 18 of 189) are counted as failures without being identified individually.

Editorial extensions

If this is right

  • High-throughput virtual screening can run without re-training: the paper reports docking roughly 560,000 protein–ligand pairs on the TargetDock-AI benchmark in about three days on seven GPUs.
  • Physical validity becomes an achievable headline metric for docking rather than a known failure mode of generative models, with $<2$ Å-and-PB-valid success rates of roughly 50–90% across the four benchmarks.
  • For targets with no known ligands or low sequence similarity to training data, a training-free multi-pocket search is claimed to be the most reliable option — the opposite of the prevailing deep-learning trend.
  • Docking-derived affinity from an untrained scoring function can rank active over inactive drug–target pairs, while post-hoc confidence, clash, and strain metrics from the DL pipeline can separate the groups in the wrong direction.
  • Docking into only the top five predicted pockets captures the strong-binding poses, which gives a simple cost–benefit rule for large-scale screening.

Reading between the lines

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

  • The table notes disclose that PocketVina's numbers credit the best-r.m.s.d. pose among all sampled conformations, whereas the comparison numbers for other methods are poses selected by each method's own scoring function; an extension the authors do not run is to report the pose PocketVina's own affinity score would return to a user and compare that top-1 rate against the same tables.
  • The complexes where P2Rank finds no pocket (59 of 363 on PDBbind2020, 24 of 144 on the unseen subset, 18 of 189 on DockGen) are counted as failures in the normalized rates but not listed; identifying them — by flexibility, pocket depth, or ligand size — would show whether coverage or accuracy is the binding constraint.
  • Because the pipeline has no learned parameters, its per-pose distribution could serve as an untrained baseline for future benchmarking: a generative model that cannot beat a multi-pocket Vina search on physical validity is probably learning dataset geometry rather than binding physics.
  • The top-five-pockets rule suggests a screening protocol the paper does not explicitly benchmark: run only the first five P2Rank pockets and rescore with a more expensive function, which would trade a little of the reported accuracy for a large cut in runtime.
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

4 major / 5 minor

Summary. PocketVina is a search-based docking pipeline that combines P2Rank pocket prediction with QuickVina 2-GPU 2.1 multi-pocket docking, requiring no task-specific training. The paper evaluates it on PDBbind2020 time-split and unseen-protein subsets, DockGen, Astex, and PoseBusters, reporting success rates for RMSD < 1 Å, RMSD < 2 Å, and RMSD < 2 Å combined with PoseBusters physical-validity checks. The central claim is that PocketVina achieves state-of-the-art physically valid docking, especially on novel proteins, and remains competitive on RMSD alone. The authors also introduce TargetDock-AI, a large protein–ligand activity-labeled dataset, and report that PocketVina's Vina-based scores separate active from inactive pairs with high statistical significance and far lower runtime than CompassDock.

Significance. If the central evaluation were measured under a comparable protocol, the contribution would be valuable: a training-free, fully search-based docking pipeline that is fast, memory-efficient, reproducible (code on GitHub, data on Zenodo), and built from unmodified third-party components. The systematic multi-pocket exploration idea is sensible, and the large TargetDock-AI dataset plus runtime comparison are useful engineering results. However, the headline state-of-the-art claim currently rests on an asymmetric evaluation protocol: PocketVina's success rates are computed on the best-RMSD pose among all sampled conformations, while comparators are credited with poses selected by their own scoring functions. Because the paper explicitly disclaims scoring optimization, the quantity a user would actually receive from PocketVina—the top pose by QuickVina score—is never reported. The missing-complex handling and hyperparameter selection also leave the main numerical claim weaker than the abstract suggests.

major comments (4)
  1. [Table 1, Table 2, and Fig. 2 notes] The headline success rates are computed on the best-RMSD pose per complex, not on the pose PocketVina would return by its own scoring function. The table notes state this explicitly: 'PocketVina results reflect the best r.m.s.d. values among all sampled poses, whereas literature-reported results correspond to poses selected by each method's scoring function.' This makes the comparison apples-to-oranges. Sampling 20 conformations per pocket across multiple pockets inflates the chance that at least one sampled pose is near the crystal, and the PoseBusters physical-validity rate is then evaluated on that oracle-selected pose. The paper never reports the natural user-facing metric: top-1-by-QuickVina-score RMSD and PB-valid success. Without that number, the abstract's claim of 'state-of-the-art performance when jointly considering ligand RMSD and physical validity' is not supported. I ask the authors to report top-1 results and, if possible, the fraction of complexes where the best-RMSD pose coincides with the top-scored pose.
  2. [Table 1 and Table 2 notes; Results] PocketVina produced poses for only 304 of 363 PDBbind2020 time-split complexes, 120 of 144 unseen-protein complexes, and 171 of 189 DockGen complexes. While normalizing success rates by the full denominator is defensible, the median RMSD is computed only over the subset that produced poses, and the paper does not report which complexes failed or why. If the failures correlate with difficult targets (e.g., no P2Rank pocket near the ligand, or docking failures), the subset median is biased upward and the reported generalization claim is weakened. Please provide per-complex failure counts, the reasons for failure, and the full RMSD distribution including failed complexes.
  3. [Supplementary S2 and Supplementary Fig. S2] The main-text protocol is under-specified with respect to hyperparameter selection. The methods state that three P2Rank configurations were tested and that box sizes from 15 Å to 30 Å were swept, while the ablation in Supplementary Fig. S2 shows that the 'alphafold' P2Rank model combined with 25 Å or 30 Å boxes yields the best RMSD curves. The main Tables 1 and 2 do not state which box size was used for each benchmark, and the paper does not explain whether the reported numbers correspond to a fixed pre-chosen configuration or to a configuration selected after observing test-set performance. Because box size directly controls search coverage, this is load-bearing for the reported success rates. Please specify the exact configuration used for each main-table result and, ideally, demonstrate that the choice was made on a validation split rather than on the test sets.
  4. [Results, 'Relationship between pocket ranks and binding affinity'] The pocket-rank versus Vina-score analysis extracts scores only from the best-RMSD pose per complex, as acknowledged in the text. Conclusions such as 'top-ranked pockets are more likely to correspond to biophysically relevant binding sites' and 'docking into the top five pockets is typically sufficient' therefore do not follow from the data presented: a best-RMSD-based pose selection is not the procedure a virtual-screening user would apply, and the analysis does not measure enrichment or ranking quality in a screening sense. At minimum, the same analysis should be repeated using top-scored poses, and the screening recommendation should be supported by an actual hit-rate or enrichment comparison over pocket-rank groups.
minor comments (5)
  1. [Introduction] The sentence 'they contain a highly fraction of implausible structures' should read 'a high fraction'.
  2. [Throughout] The dataset name is written inconsistently as 'PDBbind2020' in most places and 'PDBBind2020' in Table 1 and some figure captions; please unify.
  3. [Fig. 3 notes] The footnote markers '1Fig. 3-a Note' and '2Fig. 3-b Note' appear to be leftover formatting artifacts; they should be cleaned up.
  4. [Table 3] The TargetDock-AI discrimination result is summarized only as a checkmark and p-value; reporting an effect size such as AUC or a Mann–Whitney U statistic would make the comparison more informative.
  5. [Methods, Eq. (9)] The Armijo condition as written uses '>' but standard Armijo requires '<'; please verify the inequality direction and the surrounding description, since the current text appears to contain a typographical error in the search-direction update.

Circularity Check

1 steps flagged · score 6.0 of 10

Headline PB-valid rates are oracle best-RMSD selections over the sampled pose bag, not poses selected by PocketVina's scoring; comparators use scoring-selected poses.

  1. fitted input called prediction [Table 1 note (page 7); repeated in Table 2 note and Figure 2 note]
    "PocketVina results reflect the best r.m.s.d. values among all sampled poses, whereas literature-reported results correspond to poses selected by each method’s scoring function."

    The headline success rates are computed by selecting, per complex, the sampled pose with the lowest RMSD to the crystal ligand, and then applying the RMSD<2 Å and PoseBusters filters. Since the selection rule is exactly the evaluation metric, the reported 50.96%, 52.08%, and 39.68% values are oracle upper bounds over the sampled pose set (20 poses per pocket, across multiple P2Rank pockets), not the accuracy of any pose PocketVina's own scoring function would return. This is a fitted/selected quantity presented as the method's prediction: it does not test the scoring or ranking pipeline, is not comparable to literature numbers that report each method's top-ranked pose, and grows with the number of sampled poses.

full rationale

The only load-bearing circularity is the oracle-selection protocol for the headline docking success rates. Table 1, Table 2, and Figure 2 notes state that PocketVina's numbers use the best-RMSD pose among all sampled conformations, while all comparators use poses selected by their own scoring functions. Because the selection criterion is the reported metric itself, the headline RMSD and PB-valid percentages are by construction the best possible over the sampled pose set; they measure sampling coverage, not end-to-end predictive accuracy, and they grow with the number of poses sampled. This is a fitted/selected quantity presented as a prediction, and it carries the state-of-the-art claim. I do not count as circular: the use of third-party P2Rank and QuickVina 2-GPU 2.1, the TargetDock-AI active/inactive discrimination (which uses QuickVina's own scores on QuickVina poses and is compared against an external baseline), the citation of the authors' own CompassDock as a baseline (not load-bearing), or the missing-complex normalization (a reporting choice, though subset-only medians should be interpreted carefully). The paper is otherwise self-contained and transparent about the protocol; no self-citation chain or ansatz-smuggling is present. Score 6 reflects that the single central evaluation metric reduces by construction to an oracle selection, while the underlying sampler retains independent content.

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

No new physical entities or forces are introduced. The load-bearing choices are configuration parameters (P2Rank model, box size, sampling count) and the benchmark protocol assumptions.

free parameters (4)
  • P2Rank model choice (default vs alphafold vs alphafold-conservation) = alphafold model reported as default in main text; default model used for TargetDock-AI
    The paper's ablation (Supplementary Figure S2) shows that the 'alphafold' P2Rank configuration yields lower RMSD than 'default' and 'alphafold-conservation'; this is a hand-chosen configuration that affects the reported headline numbers.
  • Search box size (15-30 Å sweep) = 25 Å and 30 Å reported as best in the ablation; fixed 20 Å for TargetDock-AI
    The box size is a free parameter swept in 1 Å increments, and the ablation shows that larger boxes (25-30 Å) give lower RMSD; the final reported numbers therefore use a configuration chosen after looking at the benchmarks.
  • Number of conformations sampled per pocket (mode 20) = 20
    The number of sampled conformations is a user-set parameter; more sampling generally improves the best-RMSD metrics.
  • Number of pockets docked per protein = all pockets predicted by P2Rank (not stated explicitly as a fixed number)
    Docking into more pockets increases the chance that the best-RMSD pose is found; the paper does not fix a limit in the main protocol, though Figure 4 suggests only the top 5 pockets are needed for screening.
assumptions (4)
  • domain assumption P2Rank's random forest pocket predictions are a reasonable proxy for true ligand binding sites across the four benchmark sets.
    The entire pipeline depends on the pocket centers produced by P2Rank; the paper's own ablation (Supplementary Figure S2) tests three P2Rank configurations and box sizes, but does not report what fraction of true binding sites fall within the predicted pocket boxes.
  • domain assumption PoseBusters validity checks are the correct definition of physical validity.
    The central claim is defined through PoseBusters' PB-valid criteria; the paper adopts this metric from prior work without independent validation.
  • domain assumption Literature-reported results for Glide, GNINA, Vina, SurfDock, DiffDock-L, etc. are directly comparable to PocketVina's numbers despite differing protocols (post-processing, pose selection, protein preprocessing).
    The tables in the paper compare PocketVina's best-RMSD pose to literature numbers from SurfDock [31] and PoseBusters [32]; the footnote says post-processed results were excluded for some methods, but the alignment is not verified by the authors.
  • domain assumption PDBbind2020 time-split test set, PoseBusters, Astex, and DockGen are representative of real-world docking targets.
    Standard benchmarks are used, but all four sets contain crystal holo-structures, which makes the 'novel target' claim weaker than a fully unbound target test.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PocketVina Enables Scalable and Highly Accurate Physically Valid Docking through Multi-Pocket Conditioning." pith.science (2026). https://pith.science/paper/MJCTZG3N

@misc{pith2026250620043,
  author       = {Pith},
  title        = {Pith review of: PocketVina Enables Scalable and Highly Accurate Physically Valid Docking through Multi-Pocket Conditioning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MJCTZG3N}},
  note         = {Machine review of arXiv:2506.20043}
}
read the original abstract

Sampling physically valid ligand-binding poses remains a major challenge in molecular docking, particularly for unseen or structurally diverse targets. We introduce PocketVina, a fast and memory-efficient, search-based docking framework that combines pocket prediction with systematic multi-pocket exploration. We evaluate PocketVina across four established benchmarks--PDBbind2020 (timesplit and unseen), DockGen, Astex, and PoseBusters--and observe consistently strong performance in sampling physically valid docking poses. PocketVina achieves state-of-the-art performance when jointly considering ligand RMSD and physical validity (PB-valid), while remaining competitive with deep learning-based approaches in terms of RMSD alone, particularly on structurally diverse and previously unseen targets. PocketVina also maintains state-of-the-art physically valid docking accuracy across ligands with varying degrees of flexibility. We further introduce TargetDock-AI, a benchmarking dataset we curated, consisting of over 500000 protein-ligand pairs, and a partition of the dataset labeled with PubChem activity annotations. On this large-scale dataset, PocketVina successfully discriminates active from inactive targets, outperforming a deep learning baseline while requiring significantly less GPU memory and runtime. PocketVina offers a robust and scalable docking strategy that requires no task-specific training and runs efficiently on standard GPUs, making it well-suited for high-throughput virtual screening and structure-based drug discovery.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

64 extracted references · 55 canonical work pages

  1. [1]

    Nature medicine23(4), 405–408 (2017)

    Corsello, S.M., Bittker, J.A., Liu, Z., Gould, J., McCarren, P., Hirschman, J.E., Johnston, S.E., Vrcic, A., Wong, B., Khan, M.,et al.: The drug repurposing hub: a next-generation drug library and information resource. Nature medicine23(4), 405–408 (2017)

    work page 2017

  2. [2]

    Nature Protocols 18(11), 3460–3511 (2023)

    Tran-Nguyen, V.-K., Junaid, M., Simeon, S., Ballester, P.J.: A practical guide to machine-learning scoring for structure-based virtual screening. Nature Protocols 18(11), 3460–3511 (2023)

    work page 2023

  3. [3]

    Journal of computational chemistry31(2), 455–461 (2010)

    Trott, O., Olson, A.J.: Autodock vina: improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading. Journal of computational chemistry31(2), 455–461 (2010)

    2010

  4. [4]

    0: New docking methods, expanded force field, and python bindings

    Eberhardt, J., Santos-Martins, D., Tillack, A.F., Forli, S.: Autodock vina 1.2. 0: New docking methods, expanded force field, and python bindings. Journal of chemical information and modeling61(8), 3891–3898 (2021)

    work page 2021

  5. [5]

    Friesner, R.A., Banks, J.L., Murphy, R.B., Halgren, T.A., Klicic, J.J., Mainz, D.T., Repasky, M.P., Knoll, E.H., Shelley, M., Perry, J.K.,et al.: Glide: a new approach for rapid, accurate docking and scoring. 1. method and assessment of docking accuracy. Journal of medicinal chemistry47(7), 1739–1749 (2004)

    work page 2004

  6. [6]

    Proteins: Structure, Function, and Bioinformatics52(4), 609–623 (2003)

    Verdonk, M.L., Cole, J.C., Hartshorn, M.J., Murray, C.W., Taylor, R.D.: Improved protein–ligand docking using gold. Proteins: Structure, Function, and Bioinformatics52(4), 609–623 (2003)

    2003

  7. [7]

    Journal of chemical information and modeling53(8), 1893–1904 (2013)

    Koes, D.R., Baumgartner, M.P., Camacho, C.J.: Lessons learned in empirical scoring with smina from the csar 2011 benchmarking exercise. Journal of chemical information and modeling53(8), 1893–1904 (2013)

    work page 2013

  8. [8]

    Journal of chemical information and modeling57(4), 942–957 (2017)

    Ragoza, M., Hochuli, J., Idrobo, E., Sunseri, J., Koes, D.R.: Protein–ligand scor- ing with convolutional neural networks. Journal of chemical information and modeling57(4), 942–957 (2017)

    work page 2017

Show all 64 references
  1. [9]

    Nature Machine Intelligence3(12), 1033–1039 (2021)

    M´ endez-Lucio, O., Ahmad, M., Rio-Chanona, E.A., Wegner, J.K.: A geometric deep learning approach to predict binding conformations of bioactive molecules. Nature Machine Intelligence3(12), 1033–1039 (2021)

    2021

  2. [10]

    Nature Machine Intelligence6(6), 688–700 (2024)

    Cao, D., Chen, G., Jiang, J., Yu, J., Zhang, R., Chen, M., Zhang, W., Chen, L., Zhong, F., Zhang, Y.,et al.: Generic protein–ligand interaction scoring by integrating physical prior knowledge and data augmentation modelling. Nature Machine Intelligence6(6), 688–700 (2024)

    2024

  3. [11]

    Wiley Interdisciplinary Reviews: Computa- tional Molecular Science11(1), 1478 (2021)

    Li, H., Sze, K.-H., Lu, G., Ballester, P.J.: Machine-learning scoring functions 22 for structure-based virtual screening. Wiley Interdisciplinary Reviews: Computa- tional Molecular Science11(1), 1478 (2021)

    2021

  4. [12]

    Wiley Interdisciplinary Reviews: Computational Molecular Science12(2), 1567 (2022)

    Xiong, G., Shen, C., Yang, Z., Jiang, D., Liu, S., Lu, A., Chen, X., Hou, T., Cao, D.: Featurization strategies for protein–ligand interactions and their applications in scoring function development. Wiley Interdisciplinary Reviews: Computational Molecular Science12(2), 1567 (2022)

    2022

  5. [13]

    nature596(7873), 583–589 (2021)

    Jumper, J., Evans, R., Pritzel, A., Green, T., Figurnov, M., Ronneberger, O., Tunyasuvunakool, K., Bates, R., ˇZ ´ ıdek, A., Potapenko, A.,et al.: Highly accurate protein structure prediction with alphafold. nature596(7873), 583–589 (2021)

    2021

  6. [14]

    Nature616(7958), 673–685 (2023)

    Sadybekov, A.V., Katritch, V.: Computational approaches streamlining drug discovery. Nature616(7958), 673–685 (2023)

    2023

  7. [15]

    Science379(6637), 1123–1130 (2023)

    Lin, Z., Akin, H., Rao, R., Hie, B., Zhu, Z., Lu, W., Smetanin, N., Verkuil, R., Kabeli, O., Shmueli, Y.,et al.: Evolutionary-scale prediction of atomic-level protein structure with a language model. Science379(6637), 1123–1130 (2023)

    2023

  8. [16]

    Nature620(7976), 1089–1100 (2023)

    Watson, J.L., Juergens, D., Bennett, N.R., Trippe, B.L., Yim, J., Eisenach, H.E., Ahern, W., Borst, A.J., Ragotte, R.J., Milles, L.F.,et al.: De novo design of protein structure and function with rfdiffusion. Nature620(7976), 1089–1100 (2023)

    2023

  9. [17]

    arxiv (2022)

    St¨ ark, H., Ganea, O., Pattanaik, L., Barzilay, R., Jaakkola, T.: Equibind: geo- metric deep learning for drug binding structure prediction. arxiv (2022). arXiv preprint arXiv:2202.0514610

    2022 arXiv

  10. [18]

    Advances in neural information processing systems35, 7236–7249 (2022)

    Lu, W., Wu, Q., Zhang, J., Rao, J., Li, C., Zheng, S.: Tankbind: Trigonometry- aware neural networks for drug-protein binding structure prediction. Advances in neural information processing systems35, 7236–7249 (2022)

    2022

  11. [19]

    arXiv preprint arXiv:2210.06069 (2022)

    Zhang, Y., Cai, H., Shi, C., Zhong, B., Tang, J.: E3bind: An end-to-end equivariant network for protein-ligand docking. arXiv preprint arXiv:2210.06069 (2022)

    2022 arXiv

  12. [20]

    Zhou, G., Gao, Z., Ding, Q., Zheng, H., Xu, H., Wei, Z., Zhang, L., Ke, G.: Uni-mol: A universal 3d molecular representation learning framework (2023)

    2023

  13. [21]

    Nature Computational Science3(9), 789–804 (2023)

    Zhang, X., Zhang, O., Shen, C., Qu, W., Chen, S., Cao, H., Kang, Y., Wang, Z., Wang, E., Zhang, J.,et al.: Efficient and accurate large library ligand docking with karmadock. Nature Computational Science3(9), 789–804 (2023)

    2023

  14. [22]

    arXiv preprint arXiv:2210.01776 (2022)

    Corso, G., St¨ ark, H., Jing, B., Barzilay, R., Jaakkola, T.: Diffdock: Diffusion steps, twists, and turns for molecular docking. arXiv preprint arXiv:2210.01776 (2022)

    2022 arXiv

  15. [23]

    ArXiv, 2402 (2024)

    Corso, G., Deng, A., Fry, B., Polizzi, N., Barzilay, R., Jaakkola, T.: Deep confident 23 steps to new pockets: Strategies for docking generalization. ArXiv, 2402 (2024)

    2024

  16. [24]

    Bioinformatics31(13), 2214–2216 (2015)

    Alhossary, A., Handoko, S.D., Mu, Y., Kwoh, C.-K.: Fast, accurate, and reliable molecular docking with quickvina 2. Bioinformatics31(13), 2214–2216 (2015)

    2015

  17. [25]

    Nature566(7743), 224–229 (2019)

    Lyu, J., Wang, S., Balius, T.E., Singh, I., Levit, A., Moroz, Y.S., O’Meara, M.J., Che, T., Algaa, E., Tolmachova, K.,et al.: Ultra-large library docking for discovering new chemotypes. Nature566(7743), 224–229 (2019)

    2019

  18. [26]

    Nature protocols16(10), 4799–4832 (2021)

    Bender, B.J., Gahbauer, S., Luttens, A., Lyu, J., Webb, C.M., Stein, R.M., Fink, E.A., Balius, T.E., Carlsson, J., Irwin, J.J.,et al.: A practical guide to large-scale docking. Nature protocols16(10), 4799–4832 (2021)

    2021

  19. [27]

    Nature Reviews Chemistry6(4), 287–295 (2022)

    Ackloo, S., Al-Awar, R., Amaro, R.E., Arrowsmith, C.H., Azevedo, H., Batey, R.A., Bengio, Y., Betz, U.A., Bologa, C.G., Chodera, J.D.,et al.: Cache (critical assessment of computational hit-finding experiments): A public–private partner- ship benchmarking initiative to enable ...

    2022

  20. [28]

    Nature 610(7932), 582–591 (2022)

    Kaplan, A.L., Confair, D.N., Kim, K., Barros- ´Alvarez, X., Rodriguiz, R.M., Yang, Y., Kweon, O.S., Che, T., McCorvy, J.D., Kamber, D.N.,et al.: Bespoke library docking for 5-ht2a receptor agonists with antidepressant activity. Nature 610(7932), 582–591 (2022)

    2022

  21. [29]

    Accounts of chemical research57(10), 1500–1509 (2024)

    Zhang, X., Shen, C., Zhang, H., Kang, Y., Hsieh, C.-Y., Hou, T.: Advancing ligand docking through deep learning: challenges and prospects in virtual screening. Accounts of chemical research57(10), 1500–1509 (2024)

    2024

  22. [30]

    Yu, Y., Lu, S., Gao, Z., Zheng, H., Ke, G.: Do deep learning models really outperform traditional approaches in molecular docking? arXiv preprint arXiv:2302.07134 (2023)

    2023 arXiv

  23. [31]

    Nature Methods22(2), 310–322 (2025)

    Cao, D., Chen, M., Zhang, R., Wang, Z., Huang, M., Yu, J., Jiang, X., Fan, Z., Zhang, W., Zhou, H.,et al.: Surfdock is a surface-informed diffusion gener- ative model for reliable and accurate protein–ligand complex prediction. Nature Methods22(2), 310–322 (2025)

    2025

  24. [32]

    Chemical Science15(9), 3130–3139 (2024)

    Buttenschoen, M., Morris, G.M., Deane, C.M.: Posebusters: Ai-based docking methods fail to generate physically valid poses or generalise to novel sequences. Chemical Science15(9), 3130–3139 (2024)

    2024

  25. [33]

    Harris, C., Didi, K., Jamasb, A.R., Joshi, C.K., Mathis, S.V., Lio, P., Blundell, T.: Benchmarking generated poses: How rational is structure-based drug design with generative models? arXiv preprint arXiv:2308.07413 (2023)

    2023 arXiv

  26. [34]

    Journal of medicinal chemistry47(12), 2977–2980 (2004)

    Wang, R., Fang, X., Lu, Y., Wang, S.: The pdbbind database: Collection of binding affinities for protein- ligand complexes with known three-dimensional 24 structures. Journal of medicinal chemistry47(12), 2977–2980 (2004)

    2004

  27. [35]

    arXiv preprint arXiv:2406.06841 (2024)

    Sarigun, A., Franke, V., Uyar, B., Akalin, A.: Compassdock: Comprehensive accu- rate assessment approach for deep learning-based molecular docking in inference and fine-tuning. arXiv preprint arXiv:2406.06841 (2024)

    2024 arXiv

  28. [36]

    bioRxiv, 2024–07 (2024)

    Durairaj, J., Adeshina, Y., Cao, Z., Zhang, X., Oleinikovas, V., Duignan, T., McClure, Z., Robin, X., Kovtun, D., Rossi, E., et al.: Plinder: The protein-ligand interactions dataset and evaluation resource. bioRxiv, 2024–07 (2024)

    2024

  29. [37]

    Bioinformatics28(2), 286–287 (2012)

    Ngan, C.-H., Hall, D.R., Zerbe, B., Grove, L.E., Kozakov, D., Vajda, S.: Ftsite: high accuracy detection of ligand binding sites on unbound protein structures. Bioinformatics28(2), 286–287 (2012)

    2012

  30. [38]

    Nucleic acids research41(W1), 314–321 (2013)

    Tan, K.P., Nguyen, T.B., Patel, S., Varadarajan, R., Madhusudhan, M.S.: Depth: a web server to compute depth, cavity sizes, detect potential small-molecule ligand-binding cavities and predict the pka of ionizable residues in proteins. Nucleic acids research41(W1), 314–321 (2013)

    2013

  31. [39]

    BMC bioinformatics10, 1–11 (2009)

    Le Guilloux, V., Schmidtke, P., Tuffery, P.: Fpocket: an open source platform for ligand pocket detection. BMC bioinformatics10, 1–11 (2009)

    2009

  32. [40]

    Bioinformatics25(23), 3185–3186 (2009)

    Ghersi, D., Sanchez, R.: Easymifs and sitehound: a toolkit for the identification of ligand-binding sites in protein structures. Bioinformatics25(23), 3185–3186 (2009)

    2009

  33. [41]

    Bioinformatics27(15), 2083–2088 (2011)

    Zhang, Z., Li, Y., Lin, B., Schroeder, M., Huang, B.: Identification of cavities on protein surface using multiple computational approaches for drug binding site prediction. Bioinformatics27(15), 2083–2088 (2011)

    2011

  34. [42]

    Bioinfor- matics33(19), 3036–3042 (2017)

    Jim´ enez, J., Doerr, S., Mart ´ ınez-Rosell, G., Rose, A.S., De Fabritiis, G.: Deepsite: protein-binding site predictor using 3d-convolutional neural networks. Bioinfor- matics33(19), 3036–3042 (2017)

    2017

  35. [43]

    Journal of cheminformatics10, 1–12 (2018)

    Kriv´ ak, R., Hoksza, D.: P2rank: machine learning based tool for rapid and accurate prediction of ligand binding sites from protein structure. Journal of cheminformatics10, 1–12 (2018)

    2018

  36. [44]

    Nucleic Acids Research, 421 (2025)

    Pol´ ak, L., ˇSkoda, P., Riedlov´ a, K., Kriv´ ak, R., Novotn` y, M., Hoksza, D.: Prankweb 4: a modular web server for protein–ligand binding site prediction and downstream analysis. Nucleic Acids Research, 421 (2025)

    2025

  37. [45]

    Future medicinal chemistry6(16), 1741–1755 (2014) 25

    Chen, Y.-C., Totrov, M., Abagyan, R.: Docking to multiple pockets or ligand fields for screening, activity prediction and scaffold hopping. Future medicinal chemistry6(16), 1741–1755 (2014) 25

    2014

  38. [46]

    IEEE/ACM Transactions on Computational Biology and Bioinformatics (2024)

    Tang, S., Ding, J., Zhu, X., Wang, Z., Zhao, H., Wu, J.: Vina-gpu 2.1: towards fur- ther optimizing docking speed and precision of autodock vina and its derivatives. IEEE/ACM Transactions on Computational Biology and Bioinformatics (2024)

    2024

  39. [47]

    Molecules27(9), 3041 (2022)

    Tang, S., Chen, R., Lin, M., Lin, Q., Zhu, Y., Ding, J., Hu, H., Ling, M., Wu, J.: Accelerating autodock vina with gpus. Molecules27(9), 3041 (2022)

    2022

  40. [48]

    Journal of chemical information and modeling63(7), 1982–1998 (2023)

    Ding, J., Tang, S., Mei, Z., Wang, L., Huang, Q., Hu, H., Ling, M., Wu, J.: Vina-gpu 2.0: further accelerating autodock vina and its derivatives with graphics processing units. Journal of chemical information and modeling63(7), 1982–1998 (2023)

    2023

  41. [49]

    Journal of medicinal chemistry50(4), 726–741 (2007)

    Hartshorn, M.J., Verdonk, M.L., Chessari, G., Brewerton, S.C., Mooij, W.T., Mortenson, P.N., Murray, C.W.: Diverse, high-quality test set for the validation of protein- ligand docking performance. Journal of medicinal chemistry50(4), 726–741 (2007)

    2007

  42. [50]

    Zenodo (2006)

    Landrum, G., et al.: RDKit: Open-source cheminformatics. Zenodo (2006)

    2006

  43. [51]

    Journal of cheminformatics3, 1–14 (2011)

    O’Boyle, N.M., Banck, M., James, C.A., Morley, C., Vandermeersch, T., Hutchi- son, G.R.: Open babel: An open chemical toolbox. Journal of cheminformatics3, 1–14 (2011)

    2011

  44. [52]

    Weininger, D.: Smiles, a chemical language and information system. 1. introduc- tion to methodology and encoding rules. Journal of chemical information and computer sciences28(1), 31–36 (1988)

    1988

  45. [53]

    Su, M., Feng, G., Liu, Z., Li, Y., Wang, R.: Tapping on the black box: how is the scoring power of a machine-learning scoring function dependent on the training set? Journal of chemical information and modeling60(3), 1122–1136 (2020)

    2020

  46. [54]

    arXiv preprint arXiv:2402.11459 (2024)

    Huang, Y., Zhang, O., Wu, L., Tan, C., Lin, H., Gao, Z., Li, S., Li, S., et al.: Re-dock: towards flexible and realistic molecular docking with diffusion bridge. arXiv preprint arXiv:2402.11459 (2024)

    2024 arXiv

  47. [55]

    Physical Chemistry Chemical Physics18(18), 12964–12975 (2016)

    Wang, Z., Sun, H., Yao, X., Li, D., Xu, L., Li, Y., Tian, S., Hou, T.: Compre- hensive evaluation of ten docking programs on a diverse set of protein–ligand complexes: the prediction accuracy of sampling power and scoring power. Physical Chemistry Chemical Physics18(18), 12964...

    2016

  48. [56]

    Nucleic acids research44(D1), 1202–1213 (2016)

    Kim, S., Thiessen, P.A., Bolton, E.E., Chen, J., Fu, G., Gindulyte, A., Han, L., He, J., He, S., Shoemaker, B.A.,et al.: Pubchem substance and compound databases. Nucleic acids research44(D1), 1202–1213 (2016)

    2016

  49. [57]

    Journal of Chemical Information and Modeling62(10), 2499–2509 26 (2022)

    Pan, X., Wang, H., Zhang, Y., Wang, X., Li, C., Ji, C., Zhang, J.Z.: Aa-score: a new scoring function based on amino acid-specific interaction for molecu- lar docking. Journal of Chemical Information and Modeling62(10), 2499–2509 26 (2022)

    2022

  50. [58]

    Harris, C., Didi, K., Jamasb, A.R., Joshi, C.K., Mathis, S.V., Lio, P., Blundell, T.L.: Posecheck: Generative models for 3d structure-based drug design produce unrealistic poses (2023)

    2023

  51. [59]

    Journal of chemical information and modeling61(11), 5362–5376 (2021)

    Ricci-Lopez, J., Aguila, S.A., Gilson, M.K., Brizuela, C.A.: Improving structure- based virtual screening with ensemble docking and machine learning. Journal of chemical information and modeling61(11), 5362–5376 (2021)

    2021

  52. [60]

    Royal Society Open Science11(7), 240487 (2024)

    Higham, J.P., Colquhoun, D.: The affinity–efficacy problem: an essential part of pharmacology education. Royal Society Open Science11(7), 240487 (2024)

    2024

  53. [61]

    Journal of computational chemistry16(3), 273–284 (1995)

    Eisenhaber, F., Lijnzaad, P., Argos, P., Sander, C., Scharf, M.: The double cubic lattice method: Efficient approaches to numerical integration of surface area and volume and to dot surface contouring of molecular assemblies. Journal of computational chemistry16(3), 273–284 (1995)

    1995

  54. [62]

    Journal of chemical information and computer sciences43(2), 493–500 (2003)

    Steinbeck, C., Han, Y., Kuhn, S., Horlacher, O., Luttmann, E., Willighagen, E.: The chemistry development kit (cdk): An open-source java library for chemo- and bioinformatics. Journal of chemical information and computer sciences43(2), 493–500 (2003)

    2003

  55. [63]

    Nucleic acids research50(W1), 454–464 (2022)

    Zheng, W., Wuyun, Q., Zhou, X., Li, Y., Freddolino, P.L., Zhang, Y.: Lomets3: integrating deep learning and profile alignment for advanced protein template recognition and function annotation. Nucleic acids research50(W1), 454–464 (2022)

    2022

  56. [64]

    active” if the compound was experimentally confirmed to bind the target, and “inactive

    DeLano, W.L.,et al.: Pymol: An open-source molecular graphics tool. CCP4 Newsl. Protein Crystallogr40(1), 82–92 (2002) 27 Simple Moderate Complex 0.0 0.2 0.4 0.6 0.8 1.0Success rate a) Method PocketVina SurfDock Glide SP SMINA DiffDock Metric RMSD < 2 Å RMSD < 2 Å & PB-valid 0...

    2002

Pith tools

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