Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

Generative molecule evolution using 3D pharmacophore for efficient Structure-Based Drug Design

T0 review · 5 major / 6 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 MEVO generates binders comparable to known actives on four of five targets, including a KRAS G12D inhibitor predicted at -22.31 kcal/mol, using latent diffusion plus pocket-aware evolution.

desk verdict MEVO's architecture is genuinely new, but the FEP validation is too thin to support the headline affinity claims. read the letter →

arxiv 2507.20130 v1 pith:YQME7E7M submitted 2025-07-27 cs.LG q-bio.BM

Yi He , Ailun Wang , Zhi Wang , Yu Liu , Xingyuan Xu , Wen Yan This is my paper · ORCID
classification cs.LGq-bio.BM
keywords structure-baseddrugdesign3DpharmacophorelatentdiffusionmodelVQ-VAEmoleculeevolutionfreeenergyperturbationKRASG12Dpocket-awaregeneration
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

The paper proposes MEVO, a generative framework for structure-based drug design that avoids dependence on scarce protein-ligand complex data. It trains a VQ-VAE (a vector-quantised variational autoencoder) and a discrete latent diffusion model on billion-scale ligand-only chemical databases, then adapts generation to a specific binding pocket through an evolutionary loop scored by the product of a force-field energy change and the fraction of user-specified interactions satisfied. The central claim is that this coupling produces high-affinity binders: absolute free-energy perturbation predicts that MEVO-generated molecules match or beat known binders on four of five benchmark targets, and a KRAS G12D inhibitor designed through the framework reaches -22.31 kcal/mol, on par with the highly active MRTX-1133. If the claim holds, pocket-aware hit discovery becomes data-efficient, and each new target can be handled without retraining or large complex datasets.

What carries the argument

The load-bearing mechanism is the evolutionary loop that connects the data-rich pharmacophore condition to the data-scarce pocket condition. A pharmacophore is the spatial arrangement of chemical features a molecule needs to bind; after each diffusion-generation round, generated molecules are relaxed with molecular mechanics, scored by $S = -\Delta U \times \rho$ (using the amber-ff14sb force field for the protein and the ByteFF force field for the ligand), and top molecules are grouped by which protein residues and interaction types they contact. Seeds sampled from these groups have their pharmacophores extracted and fed back as conditions for the next generation round, so the model explores and merges interaction patterns without human intervention and without retraining. The scoring is cheap enough to run like a docking screen and orders of magnitude faster than absolute free-energy perturbation, which is used only for final validation.

What would settle it

Compute the Spearman correlation between evolutionary scores $S$ and absolute binding free energies, from FEP or experiment, across a held-out set of generated molecules, or prospectively synthesize and assay the top-ranked molecules; if the correlation is near zero or the top-ranked molecules fail to bind at measurable concentrations, the selection mechanism that the affinity claim rests on is falsified.

Watch

Extended reading notes

Core claim

MEVO's central claim is that the bottleneck for generative structure-based drug design is not generative modeling but training data, and that the bottleneck can be bypassed by separating universal chemistry from target-specific adaptation. The framework achieves this by encoding 3D molecules into discrete latent tokens with a near-lossless VQ-VAE trained on roughly ten billion ligand-only compounds, generating candidate molecules with a pharmacophore- and pocket-conditioned discrete diffusion model, and then refining candidates by evolution: each round scores molecules with $S = -\Delta U \times \rho$, where $\Delta U$ is the binding-induced change in force-field potential energy and $\rho$ is the fraction of desired protein-ligand interactions satisfied, and the top-ranked, interaction-diverse molecules become pharmacophore seeds for the next round. Using this loop with no additional training per target, MEVO reports FEP-predicted binding affinities for generated molecules that are comparable to or better than known binders on four of five targets, and reports a KRAS G12D lead with predicted absolute binding free energy -22.31 kcal/mol, interpreted as comparable to MRTX-1133.

Load-bearing premise

The load-bearing premise is that the scoring rule used to choose candidate molecules, the product of a force-field energy change and the fraction of desired interactions satisfied, ranks molecules by true binding strength; the paper gives no enrichment or correlation statistics showing that high scores correspond to the strong affinities later computed by free-energy perturbation.

Editorial extensions

If this is right

  • A user with only a protein structure and a residue list for the pocket can generate and evolve high-affinity candidates without retraining the model for that target.
  • Because the evolutionary score is pluggable, replacing $S$ with a more accurate fast scoring function should directly raise the affinity of evolved molecules.
  • The KRAS G12D case demonstrates that the same framework can be used as a fragment grower, a linker designer, and a lead optimizer within one design campaign.
  • On four of the five benchmark targets, the top-ranked generated molecules include at least one binder whose FEP-predicted affinity matches or exceeds the known reference binders, despite no scaffold information being given.
  • The same training-free evolution pattern could be applied to optimize ligand properties other than affinity by swapping the scoring function.

Reading between the lines

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

  • A testable extension the paper does not run: ablating the $-\Delta U$ energy term and evolving on $\rho$ alone would reveal how much of the enrichment comes from interaction geometry rather than physics; if hit rates stay constant, the force-field term is not doing the ranking work.
  • The recipe generalizes beyond small-molecule ligands: any domain with a vast unlabeled corpus and a cheap scoring function, such as RNA binders, peptide binders, or protein-protein interface modulators, could use the same separation of prior learning from target adaptation.
  • The single -22.31 kcal/mol prediction for ligand 4 is one computed number; converting it into a drug-discovery claim will require synthesis and an experimental assay, and a disappointing assay would test the ranking function rather than the generative framework as a whole.
  • The training-free evolution loop can be bolted onto other conditional generative models, not just the diffusion model used here; any generator that can consume pharmacophore conditions could be evolved with the same scoring function, which is a testable extension the paper does not demonstrate.
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

5 major / 6 minor

Summary. The paper introduces MEVO, a structure-based drug design framework that combines a VQ-VAE latent representation trained on billion-scale ligand databases (Enamine REAL, ZINC20), a discrete latent diffusion model conditioned on pharmacophore and pocket features, and an evolutionary optimization loop that scores molecules by S = -ΔU × ρ using amber-ff14sb and ByteFF force fields. The authors validate the approach by comparing generated molecules to FDA-approved drugs on property distributions, by reporting ABFEP-predicted binding affinities for five protein targets (claiming comparable or superior affinity for 4 of 5), and by a KRAS G12D case study in which a designed ligand (ligand 4) is assigned an ABFEP value of -22.31 kcal/mol and claimed to be comparable to the known inhibitor MRTX-1133. The central claims are that MEVO generates high-affinity binders efficiently and that its training-free evolutionary strategy improves affinity without additional model training.

Significance. If the affinity claims withstand scrutiny, MEVO would be a practically useful contribution: it separates universal chemical knowledge (learned from abundant ligand-only data) from target-specific adaptation (via a training-free evolutionary loop), which is an attractive way to address data scarcity in SBDD. The paper also provides positive evidence of scaffold diversity, PLIF-based interaction preservation, and a transferable scoring idea. However, the validation is entirely computational and the load-bearing affinity evidence is reported without protocol details, uncertainties, or the same-protocol reference value for MRTX-1133, so the significance cannot currently be assessed beyond the plausibility of the framework.

major comments (5)
  1. [§2, KRAS G12D case study] The flagship affinity claim for ligand 4 rests on a single ABFEP prediction of -22.31 kcal/mol with no reported uncertainty, no lambda schedule, simulation length, restraints, replica count, or convergence metrics, and no FEP+ ABFE value for MRTX-1133 computed under the identical protocol. The text states this is 'comparable activity with the highly active MRTX-1133', but -22.31 kcal/mol corresponds to a Kd around 10^-4 pM, roughly 2000-fold tighter than the 0.2 pM value cited for MRTX-1133. Without the same-protocol reference value and error bars, the comparison is not quantitatively interpretable and the claim is not verifiable.
  2. [§2, Hit discovery and Figure 3A] The claim that MEVO-generated molecules are comparable or superior to known binders for 4 out of 5 targets is supported only by a figure with aggregate ABFEP predictions; the numerical values, per-target error bars, and ABFEP protocol are not reported anywhere in the text or supplementary material. The ABFEP values for the reference binders are also computed by the authors but not disclosed, so the reader cannot assess the magnitude of the claimed improvement. Please provide a table listing all ABFEP values, uncertainties, and convergence criteria for both generated and reference ligands.
  3. [Methods, Molecule Evolution, Eqs. (9)-(10)] The evolutionary score S = -ΔU × ρ is used to rank candidates and select the top-10 for FEP evaluation, but the paper provides no correlation or enrichment statistic linking S to the subsequently computed ABFEP values. The statement in §2 that 'this simple interaction score can significantly outperform docking score' is unsupported by any comparison shown in the manuscript. Without such evidence, the top-10 selection may reflect winner's-curse bias rather than genuine affinity enrichment, and the 4-out-of-5 target result cannot be attributed to the evolutionary strategy.
  4. [Methods, Auto-Encoder; §2 Overview] The architecture description is internally inconsistent. The overview and Figure 1B describe a VQ-VAE with discrete latent tokens, but the Methods 'Auto-Encoder' section defines a continuous latent vector in Eq. (1) with no codebook, vector quantization, or discrete bottleneck, and the latent diffusion section (Eqs. (4)-(7)) assumes discrete categorical states without explaining how the continuous z from Eq. (1) is converted to discrete tokens. Please clarify whether the latent space is continuous or discrete and specify the quantization procedure, codebook size, and how the discrete diffusion transition matrices are applied.
  5. [§2, KRAS case study and Figure 4] The KRAS demonstration seeds the generation with a fragment extracted from MRTX-1133 (Figure 4C), then presents ligand 4 as comparable to MRTX-1133. Because the starting point is a fragment of the reference inhibitor, this demonstration does not establish de novo discovery relative to MRTX-1133, and the novelty claim is partially circular. Please report the Tanimoto similarity of ligand 4 to MRTX-1133 and clarify whether the intended claim is scaffold hopping, lead optimization, or fully de novo design.
minor comments (6)
  1. [§1 and Conclusion] Typos: 'efficientde novo' should be 'efficient de novo'; 'develope' should be 'developed'.
  2. [Figure 3A] The figure does not allow the reader to read numerical ABFEP values; a supplementary table with all values is needed.
  3. [§2, Evolutionary strategy paragraph] The claim that the interaction score 'can significantly outperform docking score' is presented without a comparison, citation, or dataset; either provide supporting experiments or remove the claim.
  4. [Methods, Conditional Generation] The statement that pocket-conditioned generation is weaker than pharmacophore-conditioned generation due to data scarcity is not quantified; please provide a comparison or remove the qualitative claim.
  5. [Methods, Latent Diffusion, Eq. (6)] The notation Qt is defined as a cumulative product after Eq. (5), but Eq. (6) uses z0Qt and z0Qt-1; please define the cumulative matrix operation explicitly to avoid ambiguity.
  6. [General] No statement on code or data availability is provided; for reproducibility, the authors should indicate whether the model, training data, and FEP input files will be released.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the affinity claims are validated by independent FEP calculations, and the KRAS seed fragment does not by construction produce the reported binding free energy.

full rationale

The paper's central affinity claims are not circular. For the five-target hit-discovery benchmark, the authors state: 'For each protein target, MEVO only takes the protein structure and protein residue indices defining the binding pocket as input, without using any fragments or scaffolds from the reference ligands.' Candidate selection uses the score S = -ΔU × ρ, but the reported binding affinities come from separate ABFEP/FEP+ calculations, and no parameter of S is fitted to those FEP outcomes. The KRAS demonstration does seed generation with a fragment of MRTX-1133, but the claimed affinity of ligand 4 is the result of an independent absolute binding free energy calculation, not a logical consequence of using that fragment; the seed ensures structural relatedness, not the numerical affinity value. The use of ByteFF is a self-citation with overlapping authors, but it is an externally published force field used as a computational dependency, not an unverified uniqueness theorem or a premise that by itself forces the reported results. The unsupported assertion that the interaction score 'can significantly outperform docking score' and the inconsistency between the claimed VQ-VAE and the continuous Auto-Encoder in Eq. (1) are correctness or verifiability concerns, not circularity. On the evidence quoted, no equation, fitted parameter, or self-citation reduces the predicted affinities to their inputs by construction.

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

The paper does not introduce new physical entities, forces, or conserved quantities. The MEVO framework is a software method, not an invented entity in the sense of this ledger. The free parameters are design choices for the resampling, interaction definition, and evolution loop, none of which are fitted to the FEP validation results.

free parameters (4)
  • HAC resampling ratio = 2:3:3:2 for <20, 20-30, 30-40, 40-50 heavy atoms
    Hand-chosen to rebalance the training set towards larger molecules; affects the generated molecular weight distribution.
  • Critical interaction frequency threshold = >30% occurrence across reference binders
    Hand-set in Methods, Interaction Preservation, to define which interactions are 'critical' for PLIF recovery. Changes the recovery metric.
  • Seed molecule count per evolution round = 0-2 seeds sampled from top-ranked groups
    Hand-set in Methods, Molecule Evolution; controls exploration vs exploitation and affects convergence speed.
  • Top-K selection for FEP validation = top-10 ranked molecules per target
    Hand-set in Results; determines the subset of generated molecules that receive FEP evaluation.
assumptions (4)
  • domain assumption The physics-based score S = -ΔU × ρ is correlated with true binding affinity
    Molecule Evolution section: the entire evolutionary optimization uses S to rank molecules, and this correlation is asserted but not demonstrated.
  • domain assumption Pharmacophore features computed by RDKit capture the chemical features relevant for binding
    Conditional Generation and Dataset preparation: pharmacophore conditions are extracted with RDKit and used as generation constraints.
  • domain assumption HelixDock provides a sufficiently unbiased training distribution for pocket-conditioned generation
    Dataset preparation: HelixDock is the only protein-ligand source for pocket conditioning; docking-generated poses may not reflect true binding modes.
  • standard math D3PM discrete diffusion is a valid generative model for latent molecular tokens
    Latent Diffusion section: the paper relies on the published D3PM framework [3].

how reviews work

0 comments
Cite this review

Pith. "Pith review of Generative molecule evolution using 3D pharmacophore for efficient Structure-Based Drug Design." pith.science (2026). https://pith.science/paper/YQME7E7M

@misc{pith2026250720130,
  author       = {Pith},
  title        = {Pith review of: Generative molecule evolution using 3D pharmacophore for efficient Structure-Based Drug Design},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YQME7E7M}},
  note         = {Machine review of arXiv:2507.20130}
}
abstract

Recent advances in generative models, particularly diffusion and auto-regressive models, have revolutionized fields like computer vision and natural language processing. However, their application to structure-based drug design (SBDD) remains limited due to critical data constraints. To address the limitation of training data for models targeting SBDD tasks, we propose an evolutionary framework named MEVO, which bridges the gap between billion-scale small molecule dataset and the scarce protein-ligand complex dataset, and effectively increase the abundance of training data for generative SBDD models. MEVO is composed of three key components: a high-fidelity VQ-VAE for molecule representation in latent space, a diffusion model for pharmacophore-guided molecule generation, and a pocket-aware evolutionary strategy for molecule optimization with physics-based scoring function. This framework efficiently generate high-affinity binders for various protein targets, validated with predicted binding affinities using free energy perturbation (FEP) methods. In addition, we showcase the capability of MEVO in designing potent inhibitors to KRAS$^{\textrm{G12D}}$, a challenging target in cancer therapeutics, with similar affinity to the known highly active inhibitor evaluated by FEP calculations. With high versatility and generalizability, MEVO offers an effective and data-efficient model for various tasks in structure-based ligand design.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Valid Property-Enhanced Contrastive Learning for Targeted Optimization & Resampling for Novel Drug Design

    cs.LG 2025-08 conditional novelty 5.0 of 10

    VECTOR+ combines contrastive learning and Gaussian mixture sampling to generate novel, synthetically plausible inhibitors from low-data datasets, with improved docking scores over known compounds.

Reference graph

Works this paper leans on

70 extracted references · 10 canonical work pages · cited by 1 Pith paper

  1. [1]

    Kras: biology, inhibition, and mechanisms of inhibitor resistance.Current Oncology, 31(4):2024–2046, 2024

    Leonard J Ash, Ottavia Busia-Bourdain, Daniel Okpattah, Avrosina Kamel, Ariel Liberchuk, and Andrew L Wolfe. Kras: biology, inhibition, and mechanisms of inhibitor resistance.Current Oncology, 31(4):2024–2046, 2024

    2024

  2. [2]

    Nippa, Michael Iff, Jann Ledergerber, Carl C

    Kenneth Atz, Leandro Cotos, Clemens Isert, Maria Håkansson, Dorota Focht, Mattis Hilleke, David F. Nippa, Michael Iff, Jann Ledergerber, Carl C. G. Schiebroek, Valentina Romeo, Jan A. Hiss, Daniel Merk, Petra Schneider, Bernd Kuhn, Uwe Grether, and Gisbert Schneider. Prospective de novo drug design with deep interactome learning. Nature Communications, 15...

    work page doi:10.1038/s41467-024-47613-w 2024

  3. [3]

    Structured denoising diffusion models in discrete state-spaces.Advances in neural information processing systems, 34:17981–17993, 2021

    Jacob Austin, Daniel D Johnson, Jonathan Ho, Daniel Tarlow, and Rianne Van Den Berg. Structured denoising diffusion models in discrete state-spaces.Advances in neural information processing systems, 34:17981–17993, 2021

    2021

  4. [4]

    The properties of known drugs

    Guy W Bemis and Mark A Murcko. The properties of known drugs. 1. molecular frameworks.Journal of medicinal chemistry, 39(15):2887–2893, 1996

    1996

  5. [5]

    High-throughput screening: today’s biochemical and cell-based approaches.Drug discovery today, 25(10):1807–1821, 2020

    Vincent Blay, Bhairavi Tolani, Sunita P Ho, and Michelle R Arkin. High-throughput screening: today’s biochemical and cell-based approaches.Drug discovery today, 25(10):1807–1821, 2020

    2020

  6. [6]

    ACEGEN: Reinforcement Learning of Generative Chemical Agents for Drug Discovery.Journal of Chemical Information and Modeling, 64(15):5900–5911, 2024

    Albert Bou, Morgan Thomas, Sebastian Dittert, Carles Navarro, Maciej Majewski, Ye Wang, Shivam Patel, Gary Tresadern, Mazen Ahmad, Vincent Moens, Woody Sherman, Simone Sciabola, and Gianni De Fabritiis. ACEGEN: Reinforcement Learning of Generative Chemical Agents for Drug Discovery.Journal of Chemical Information and Modeling, 64(15):5900–5911, 2024. ISSN...

    doi:10.1021/acs.jcim.4c00895 2024

  7. [7]

    Prolif: a library to encode molecular interactions as fingerprints.Journal of cheminformatics, 13(1):72, 2021

    Cédric Bouysset and Sébastien Fiorucci. Prolif: a library to encode molecular interactions as fingerprints.Journal of cheminformatics, 13(1):72, 2021

    2021

  8. [8]

    Graphinvent: A generative model for molecular graphs.ICLR Workshop, 2021

    Michael M Bronstein et al. Graphinvent: A generative model for molecular graphs.ICLR Workshop, 2021

    2021

Show all 70 references
  1. [9]

    Burger, Xiaohu Hu, Ilya Balabin, Morné Muller, Megan Stanley, Fourie Joubert, and Thomas M

    Pieter B. Burger, Xiaohu Hu, Ilya Balabin, Morné Muller, Megan Stanley, Fourie Joubert, and Thomas M. Kaiser. FEP Augmentation as a Means to Solve Data Paucity Problems for Machine Learning in Chemical Biology.Journal of Chemical Information and Modeling, 64(9):3812–3825, 2024...

    2024 doi

  2. [10]

    Generic protein–ligand interaction scoring by integrating physical prior knowledge and data augmentation modelling.Nature Machine Intelligence, 6(6):688–700, 2024

    Duanhua Cao, Geng Chen, Jiaxin Jiang, Jie Yu, Runze Zhang, Mingan Chen, Wei Zhang, Lifan Chen, Feisheng Zhong, Yingying Zhang, et al. Generic protein–ligand interaction scoring by integrating physical prior knowledge and data augmentation modelling.Nature Machine Intelligence,...

    2024

  3. [11]

    Shicheng Chen, Odin Zhang, Chenran Jiang, Huifeng Zhao, Xujun Zhang, Mengting Chen, Yun Liu, Qun Su, Zhenxing Wu, Xinyue Wang, Wanglin Qu, Yuanyi Ye, Xin Chai, Ning Wang, Tianyue Wang, Yuan An, Guanlin Wu, Qianqian Yang, Jiean Chen, Wei Xie, Haitao Lin, Dan Li, Chang-Yu Hsieh,...

    2025

  4. [12]

    Enhancing hit discovery in virtual screening through absolute protein–ligand binding free-energy calculations.Journal of Chemical Information and Modeling, 63(10): 3171–3185, 2023

    Wei Chen, Di Cui, Steven V Jerome, Mayako Michino, Eelke B Lenselink, David J Huggins, Alexandre Beautrait, Jeremie Vendome, Robert Abel, Richard A Friesner, et al. Enhancing hit discovery in virtual screening through absolute protein–ligand binding free-energy calculations.Jo...

    2023

  5. [13]

    Structure-based design and synthesis of potent and selective kras g12d inhibitors

    Hengmiao Cheng, Puhui Li, Ping Chen, Adriana Irimia, Jae Hyun Bae, Alexei Brooun, Patrick Fagan, Richard Lam, Bingzhen Lin, Jingchuan Zhang, et al. Structure-based design and synthesis of potent and selective kras g12d inhibitors. ACS Medicinal Chemistry Letters, 14(10):1351–1...

    2023

  6. [14]

    PIDiff: Physics informed diffusion model for protein pocket-specific 3D molecular generation.Computers in Biology and Medicine, 180: 108865, 2024

    Seungyeon Choi, Sangmin Seo, Byung Ju Kim, Chihyun Park, and Sanghyun Park. PIDiff: Physics informed diffusion model for protein pocket-specific 3D molecular generation.Computers in Biology and Medicine, 180: 108865, 2024. ISSN 0010-4825. doi: 10.1016/j.compbiomed.2024.108865....

    2024

  7. [15]

    Autodock vina 1.2

    Jerome Eberhardt, Diogo Santos-Martins, Andreas F Tillack, and Stefano Forli. Autodock vina 1.2. 0: New docking methods, expanded force field, and python bindings.Journal of chemical information and modeling, 61 (8):3891–3898, 2021. 15

    2021

  8. [16]

    Assessing interaction recovery of predicted protein-ligand poses.Journal of Cheminformatics, 17(1):1–13, 2025

    David Errington, Constantin Schneider, Cédric Bouysset, and Frédéric A Dreyer. Assessing interaction recovery of predicted protein-ligand poses.Journal of Cheminformatics, 17(1):1–13, 2025

    2025

  9. [17]

    Francoeur, Tomohide Masuda, Jocelyn Sunseri, Andrew Jia, Richard B

    Paul G. Francoeur, Tomohide Masuda, Jocelyn Sunseri, Andrew Jia, Richard B. Iovanisci, Ian Snyder, and David R. Koes. Three-Dimensional Convolutional Neural Networks and a Cross-Docked Data Set for Structure-Based Drug Design. Journal of Chemical Information and Modeling, 60(9...

    2020 doi

  10. [18]

    Glide: a new approach for rapid, accurate docking and scoring

    Richard A Friesner, Jay L Banks, Robert B Murphy, Thomas A Halgren, Jasna J Klicic, Daniel T Mainz, Matthew P Repasky, Eric H Knoll, Mee Shelley, Jason K Perry, et al. Glide: a new approach for rapid, accurate docking and scoring. 1. method and assessment of docking accuracy.J...

    2004

  11. [19]

    Extra precision glide: Docking and scoring incorporating a model of hydrophobic enclosure for protein- ligand complexes.Journal of medicinal chemistry, 49(21):6177–6196, 2006

    Richard A Friesner, Robert B Murphy, Matthew P Repasky, Leah L Frye, Jeremy R Greenwood, Thomas A Halgren, Paul C Sanschagrin, and Daniel T Mainz. Extra precision glide: Docking and scoring incorporating a model of hydrophobic enclosure for protein- ligand complexes.Journal of...

    2006

  12. [20]

    Targetdiff: Diffusion models for target-aware molecule generation.NeurIPS, 2022

    Wenhao Gao et al. Targetdiff: Diffusion models for target-aware molecule generation.NeurIPS, 2022

    2022

  13. [21]

    DecompDiff: Diffusion Models with Decomposed Priors for Structure-Based Drug Design

    Jiaqi Guan, Xiangxin Zhou, Yuwei Yang, Yu Bao, Jian Peng, Jianzhu Ma, Qiang Liu, Liang Wang, and Quanquan Gu. DecompDiff: Diffusion Models with Decomposed Priors for Structure-Based Drug Design. In Proceedings of the 40th International Conference on Machine Learning, pages 118...

    2023

  14. [22]

    Empirical scoring functions for structure-based virtual screening: applications, critical aspects, and challenges.Frontiersin pharmacology, 9:1089, 2018

    Isabella A Guedes, Felipe SS Pereira, and Laurent E Dardenne. Empirical scoring functions for structure-based virtual screening: applications, critical aspects, and challenges.Frontiersin pharmacology, 9:1089, 2018

    2018

  15. [23]

    Anti-tumor efficacy of a potent and selective non-covalent krasg12d inhibitor

    Jill Hallin, Vickie Bowcut, Andrew Calinisan, David M Briere, Lauren Hargis, Lars D Engstrom, Jade Laguer, James Medwid, Darin Vanderpool, Ella Lifset, et al. Anti-tumor efficacy of a potent and selective non-covalent krasg12d inhibitor. Nature medicine, 28(10):2171–2182, 2022

    2022

  16. [24]

    Kras mutation: from undruggable to druggable in cancer

    Lamei Huang, Zhixing Guo, Fang Wang, and Liwu Fu. Kras mutation: from undruggable to druggable in cancer. Signal transduction and targeted therapy, 6(1):386, 2021

    2021

  17. [25]

    MDM: Molecular Diffusion Model for 3D Molecule Generation.Proceedings of the AAAI Conference on Artificial Intelligence, 37(4):5105–5112, 2023

    Lei Huang, Hengtong Zhang, Tingyang Xu, and Ka-Chun Wong. MDM: Molecular Diffusion Model for 3D Molecule Generation.Proceedings of the AAAI Conference on Artificial Intelligence, 37(4):5105–5112, 2023. ISSN 2374-3468. doi: 10.1609/aaai.v37i4.25639. URL https://ojs.aaai.org/ind...

    2023 doi

  18. [26]

    Hadfield, Anthony R

    Fergus Imrie, Thomas E. Hadfield, Anthony R. Bradley, and Charlotte M. Deane. Deep generative design with 3D pharmacophoric constraints. Chemical Science, 12(43):14577–14589, 2021. ISSN 2041-6539. doi: 10.1039/D1SC02436A. URL https://pubs.rsc.org/en/content/articlelanding/2021...

    2021 doi

  19. [27]

    Zinc20—a free ultralarge-scale chemical database for ligand discovery.Journal of chemical information and modeling, 60(12):6065–6073, 2020

    John J Irwin, Khanh G Tang, Jennifer Young, Chinzorig Dandarchuluun, Benjamin R Wong, Munkhzul Khurel- baatar, Yurii S Moroz, John Mayfield, and Roger A Sayle. Zinc20—a free ultralarge-scale chemical database for ligand discovery.Journal of chemical information and modeling, 6...

    2020

  20. [28]

    Helixdock: A large-scale docking dataset.Scientific Data, 2023

    Dejun Jiang et al. Helixdock: A large-scale docking dataset.Scientific Data, 2023

    2023

  21. [29]

    Drugging an undruggable pocket on kras

    Dirk Kessler, Michael Gmachl, Andreas Mantoulidis, Laetitia J Martin, Andreas Zoephel, Moriz Mayer, Andreas Gollner, David Covini, Silke Fischer, Thomas Gerstberger, et al. Drugging an undruggable pocket on kras. Proceedings of the National Academy of Sciences, 116(32):15823–1...

    2019

  22. [30]

    Drugging all ras isoforms with one pocket.Futuremedicinal chemistry, 12(21):1911–1923, 2020

    Dirk Kessler, Andreas Bergner, Jark Böttcher, Gerhard Fischer, Sandra Döbel, Melanie Hinkel, Barbara Müllauer, Alexander Weiss-Puxbaum, and Darryl B McConnell. Drugging all ras isoforms with one pocket.Futuremedicinal chemistry, 12(21):1911–1923, 2020

    1911

  23. [31]

    Pan-kras inhibitor disables oncogenic signalling and tumour growth.Nature, 619(7968):160–166, 2023

    Dongsung Kim, Lorenz Herdeis, Dorothea Rudolph, Yulei Zhao, Jark Böttcher, Alberto Vides, Carlos I Ayala- Santos, Yasin Pourfarjam, Antonio Cuevas-Navarro, Jenny Y Xue, et al. Pan-kras inhibitor disables oncogenic signalling and tumour growth.Nature, 619(7968):160–166, 2023

    2023

  24. [32]

    G-spherenet: Geometric deep learning for affinity prediction.Nature Communications, 2022

    Shengchao Liu et al. G-spherenet: Geometric deep learning for affinity prediction.Nature Communications, 2022

    2022

  25. [33]

    Pdbbind-cn: A protein-ligand binding affinity database with a national ranking

    Zhihai Liu, Yan Li, Li Han, et al. Pdbbind-cn: A protein-ligand binding affinity database with a national ranking. Nucleic Acids Research, 43(D1):D375–D381, 2014. doi: 10.1093/nar/gku1076. 16

    2014 doi

  26. [34]

    Loeffler, Jiazhen He, Alessandro Tibo, Jon Paul Janet, Alexey Voronov, Lewis H

    Hannes H. Loeffler, Jiazhen He, Alessandro Tibo, Jon Paul Janet, Alexey Voronov, Lewis H. Mervin, and Ola Engkvist. Reinvent 4: Modern AI–driven generative molecule design.Journal of Cheminformatics, 16(1):20, 2024. ISSN 1758-2946. doi: 10.1186/s13321-024-00812-5. URL https://...

    2024 doi

  27. [35]

    Enamine real database: World’s largest collection of building blocks and screening compounds,

    Enamine Ltd. Enamine real database: World’s largest collection of building blocks and screening compounds,

  28. [36]

    3D Molecular Generation via Virtual Dynamics

    Shuqi Lu, Lin Yao, Xi Chen, Hang Zheng, Di He, and Guolin Ke. 3D Molecular Generation via Virtual Dynamics. Transactions on Machine Learning Research, 2024. ISSN 2835-8856. URLhttps://openreview.net/forum?id= QvipGVdE6L

    2024

  29. [37]

    Radchenko, Yurii S

    Andreas Luttens, Israel Cabeza de Vaca, Leonard Sparring, José Brea, Antón Leandro Martínez, Nour Aldin Kahlous, Dmytro S. Radchenko, Yurii S. Moroz, María Isabel Loza, Ulf Norinder, and Jens Carlsson. Rapid traversal of vast chemical space using machine learning-guided dockin...

    2025

  30. [38]

    Ultra-large library docking for discovering new chemotypes

    Jiankun Lyu, Sheng Wang, Trent E Balius, et al. Ultra-large library docking for discovering new chemotypes. Nature, 566:224–229, 2020. doi: 10.1038/s41586-019-0917-9

    2020 doi

  31. [39]

    Impact of high-throughput screening in biomedical research

    Ricardo Macarron, Martyn N Banks, Dejan Bojanic, David J Burns, Dragan A Cirovic, Tina Garyantes, Darren VS Green, Robert P Hertzberg, William P Janzen, Jeff W Paslay, et al. Impact of high-throughput screening in biomedical research. Nature reviews Drug discovery, 10(3):188–195, 2011

    2011

  32. [40]

    Relevance of the trillion-sized chemical space “explore” as a source for drug discovery.ACS Medicinal Chemistry Letters, 14(4):466–472, 2023

    Alexander Neumann, Lester Marrison, and Raphael Klein. Relevance of the trillion-sized chemical space “explore” as a source for drug discovery.ACS Medicinal Chemistry Letters, 14(4):466–472, 2023. doi: 10.1021/acsmedchemlett. 3c00021

    2023 doi

  33. [41]

    Direct k-ras inhibitors to treat cancers: progress, new insights, and approaches to treat resistance.Annual Review of Pharmacology and Toxicology, 64(1):231–253, 2024

    Ruth Nussinov and Hyunbum Jang. Direct k-ras inhibitors to treat cancers: progress, new insights, and approaches to treat resistance.Annual Review of Pharmacology and Toxicology, 64(1):231–253, 2024

    2024

  34. [42]

    The current understanding of kras protein structure and dynamics.Computational and structural biotechnology journal, 18:189–198, 2020

    Tatu Pantsar. The current understanding of kras protein structure and dynamics.Computational and structural biotechnology journal, 18:189–198, 2020

    2020

  35. [43]

    Pocket2Mol: Efficient Molecular Sampling Based on 3D Protein Pockets

    Xingang Peng, Shitong Luo, Jiaqi Guan, Qi Xie, Jian Peng, and Jianzhu Ma. Pocket2Mol: Efficient Molecular Sampling Based on 3D Protein Pockets. InProceedings of the 39th InternationalConference on MachineLearning, pages 17644–17655. PMLR, 2022. URLhttps://proceedings.mlr.press...

    2022

  36. [44]

    Decipher fundamental atomic interactions to unify generative molecular docking and design.bioRxiv, 2024

    Xingang Peng, Ruihan Guo, Yan Xu, Jiaqi Guan, Yinjun Jia, Yanwen Huang, Muhan Zhang, Jian Peng, Jiayu Sun, Chuanhui Han, Zihua Wang, and Jianzhu Ma. Decipher fundamental atomic interactions to unify generative molecular docking and design.bioRxiv, 2024. doi: 10.1101/2024.10.17...

    2024 doi

  37. [45]

    Pocket2mol: Efficient molecular sampling based on 3d protein pockets.ICML, 2022

    Xingang Peng et al. Pocket2mol: Efficient molecular sampling based on 3d protein pockets.ICML, 2022

    2022

  38. [46]

    Estimation of the size of drug-like chemical space based on 1d and 2d molecular descriptors.Journal of Computer-Aided Molecular Design, 27:675–679, 2013

    Pavel G Polishchuk, Timur I Madzhidov, and Alexandre Varnek. Estimation of the size of drug-like chemical space based on 1d and 2d molecular descriptors.Journal of Computer-Aided Molecular Design, 27:675–679, 2013. doi: 10.1007/s10822-013-9672-4

    2013 doi

  39. [47]

    Language models are unsupervised multitask learners

    Alec Radford, Jeffrey Wu, Rewon Child, et al. Language models are unsupervised multitask learners. 2019

    2019

  40. [48]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, et al. High-resolution image synthesis with latent diffusion models. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684–10695, 2022

    2022

  41. [49]

    The maximal and current accuracy of rigorous protein-ligand binding free energy calculations

    Gregory A Ross, Chao Lu, Guido Scarabelli, Steven K Albanese, Evelyne Houang, Robert Abel, Edward D Harder, and Lingle Wang. The maximal and current accuracy of rigorous protein-ligand binding free energy calculations. Communications Chemistry, 6(1):222, 2023

    2023

  42. [50]

    Accuracy or novelty: what can we gain from target-specific machine- learning-based scoring functions in virtual screening?Briefings in Bioinformatics, 22(5):bbaa410, 2021

    Chao Shen, Gaoqi Weng, Xujun Zhang, Elaine Lai-Han Leung, Xiaojun Yao, Jinping Pang, Xin Chai, Dan Li, Ercheng Wang, Dongsheng Cao, et al. Accuracy or novelty: what can we gain from target-specific machine- learning-based scoring functions in virtual screening?Briefings in Bio...

    2021

  43. [51]

    Ras proteins and their regulators in human disease

    Dhirendra K Simanshu, Dwight V Nissley, and Frank McCormick. Ras proteins and their regulators in human disease. Cell, 170(1):17–33, 2017. 17

    2017

  44. [52]

    Development and evaluation of a deep learning model for protein–ligand binding affinity prediction.Bioinformatics, 34(21):3666–3674, 2018

    Marta M Stepniewska-Dziubinska, Piotr Zielenkiewicz, and Pawel Siedlecki. Development and evaluation of a deep learning model for protein–ligand binding affinity prediction.Bioinformatics, 34(21):3666–3674, 2018

    2018

  45. [53]

    Zinc 15—ligand discovery for everyone.JCIM, 2015

    Teague Sterling and John J Irwin. Zinc 15—ligand discovery for everyone.JCIM, 2015

    2015

  46. [54]

    Comparative Assessment of Scoring Functions: The CASF-2016 Update.Journal of Chemical Information and Modeling, 59(2):895–913,

    Minyi Su, Qifan Yang, Yu Du, Guoqin Feng, Zhihai Liu, Yan Li, and Renxiao Wang. Comparative Assessment of Scoring Functions: The CASF-2016 Update.Journal of Chemical Information and Modeling, 59(2):895–913,

    2016

  47. [55]

    Oleg Trott and Arthur J Olson. Autodock vina: Improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading.Journal of Computational Chemistry, 31(2):455–461, 2010. doi: 10.1002/jcc.21334

    2010 doi

  48. [56]

    Wavenet: A generative model for raw audio.arXiv preprint arXiv:1609.03499, 2016

    Aaron Van den Oord, Sander Dieleman, Heiga Zen, et al. Wavenet: A generative model for raw audio.arXiv preprint arXiv:1609.03499, 2016

    2016 arXiv

  49. [57]

    Attention is all you need.Advances in neural information processing systems, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017

    2017

  50. [58]

    Token-Mol 1.0: tokenized drug design with large language models

    Jike Wang, Rui Qin, Mingyang Wang, Meijing Fang, Yangyang Zhang, Yuchen Zhu, Qun Su, Qiaolin Gou, Chao Shen, Odin Zhang, Zhenxing Wu, Dejun Jiang, Xujun Zhang, Huifeng Zhao, Jingxuan Ge, Zhourui Wu, Yu Kang, Chang-Yu Hsieh, and Tingjun Hou. Token-Mol 1.0: tokenized drug design...

    2025 doi

  51. [59]

    Accurate and reliable prediction of relative ligand binding potency in prospective drug discovery by way of a modern free-energy calculation protocol and force field

    Lingle Wang, Yujie Wu, Yuqing Deng, Byungchan Kim, Levi Pierce, Goran Krilov, Dmitry Lupyan, Shaughnessy Robinson, Markus K Dahlgren, Jeremy Greenwood, et al. Accurate and reliable prediction of relative ligand binding potency in prospective drug discovery by way of a modern f...

    2015

  52. [60]

    Identification of mrtx1133, a noncovalent, potent, and selective krasg12d inhibitor.Journal of medicinal chemistry, 65(4):3123–3133, 2021

    Xiaolun Wang, Shelley Allen, James F Blake, Vickie Bowcut, David M Briere, Andrew Calinisan, Joshua R Dahlke, Jay B Fell, John P Fischer, Robin J Gunn, et al. Identification of mrtx1133, a noncovalent, potent, and selective krasg12d inhibitor.Journal of medicinal chemistry, 65...

    2021

  53. [61]

    Tamgen: drug design with target- aware molecule generation through a chemical language model.Nature Communications, 15(1):9360, 2024

    Kehan Wu, Yingce Xia, Pan Deng, Renhe Liu, Yuan Zhang, Han Guo, Yumeng Cui, Qizhi Pei, Lijun Wu, Shufang Xie, Si Chen, Xi Lu, Song Hu, Jinzhi Wu, Chi-Kin Chan, Shawn Chen, Liangliang Zhou, Nenghai Yu, Enhong Chen, Haiguang Liu, Jinjiang Guo, Tao Qin, and Tie-Yan Liu. Tamgen: d...

    2024 doi

  54. [62]

    Target-aware Molecule Generation for Drug Design Using a Chemical Language Model*

    Yingce Xia, Kehan Wu, Pan Deng, Renhe Liu, Yuan Zhang, Han Guo, Yumeng Cui, Qizhi Pei, Lijun Wu, Shufang Xie, Si Chen, Xi Lu, Song Hu, Jinzhi Wu, Chi-Kin Chan, Shuo Chen, Liangliang Zhou, Nenghai Yu, Haiguang Liu, Jinjiang Guo, Tao Qin, and Tie-Yan Liu. Target-aware Molecule G...

    2024 doi

  55. [63]

    Accelerating discovery of bioactive ligands with pharmacophore-informed generative models.Nature Communications, 16(1):2391, 2025

    Weixin Xie, Jianhang Zhang, Qin Xie, Chaojun Gong, Yuhao Ren, Jin Xie, Qi Sun, Youjun Xu, Luhua Lai, and Jianfeng Pei. Accelerating discovery of bioactive ligands with pharmacophore-informed generative models.Nature Communications, 16(1):2391, 2025

    2025

  56. [64]

    Knowledge-guided diffusion model for 3D ligand-pharmacophore mapping

    Jun-Lin Yu, Cong Zhou, Xiang-Li Ning, Jun Mou, Fan-Bo Meng, Jing-Wei Wu, Yi-Ting Chen, Biao-Dan Tang, Xiang-Gen Liu, and Guo-Bo Li. Knowledge-guided diffusion model for 3D ligand-pharmacophore mapping. Nature Communications, 16(1):2269, 2025. ISSN 2041-1723. doi: 10.1038/s4146...

    2025 doi

  57. [65]

    ResGen is a pocket-aware 3D molecular generation model based on parallel multiscale modelling.Nature Machine Intelligence, 5(9):1020– 1030, 2023

    Odin Zhang, Jintu Zhang, Jieyu Jin, Xujun Zhang, RenLing Hu, Chao Shen, Hanqun Cao, Hongyan Du, Yu Kang, Yafeng Deng, Furui Liu, Guangyong Chen, Chang-Yu Hsieh, and Tingjun Hou. ResGen is a pocket-aware 3D molecular generation model based on parallel multiscale modelling.Natur...

    2023 doi

  58. [66]

    Data-driven parametrization of molecular mechanics force fields for expansive chemical space coverage.Chemical Science, 16:2730–2740, 2025

    Tianze Zheng, Ailun Wang, Xu Han, Yu Xia, Xingyuan Xu, Jiawei Zhan, Yu Liu, Yang Chen, Zhi Wang, Xiaojie Wu, Sheng Gong, and Wen Yan. Data-driven parametrization of molecular mechanics force fields for expansive chemical space coverage.Chemical Science, 16:2730–2740, 2025. doi...

    2025 doi

  59. [67]

    Hui Zhu, Jincai Yang, and Niu Huang. Assessment of the generalization abilities of machine-learning scoring functions for structure-based virtual screening.Journal of chemical information and modeling, 62(22):5485–5502, 2022

    2022

  60. [68]

    3D molecular generative framework for interaction-guided drug design

    Wonho Zhung, Hyeongwoo Kim, and Woo Youn Kim. 3D molecular generative framework for interaction-guided drug design. Nature Communications, 15(1):2688, 2024. ISSN 2041-1723. doi: 10.1038/s41467-024-47011-2. URL https://www.nature.com/articles/s41467-024-47011-2. 19

    2024 doi

  61. [2019]

    doi: 10.1021/acs.jcim.8b00545

    ISSN 1549-9596. doi: 10.1021/acs.jcim.8b00545. URLhttps://doi.org/10.1021/acs.jcim.8b00545

    doi

  62. [2023]

    URL https://enamine.net/compound-collections/real-compounds

Pith tools

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