Pith. sign in

REVIEW 3 major objections 4 minor 122 references

This review contends that geometric deep learning has moved multi-target drug design from serendipity to rational, structure-driven generation.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 05:28 UTC pith:BILPQOYL

load-bearing objection A useful but over-reaching survey: good map of GDL for multi-target drug design, but the 'rational automated generation era' claim outruns the evidence—even the authors later concede the lack of wet-lab confirmation. the 3 major comments →

arxiv 2607.20550 v1 pith:BILPQOYL submitted 2026-07-14 cs.LG cs.AI

Beyond SBDD: Geometric Deep Learning in Polypharmacology and Multi-target Drug Design

Tianming Han , Zhijie Pan , Wenchi Ge , Qi Zhao This is my paper
classification cs.LG cs.AI
keywords geometric deep learningpolypharmacologymulti-target drug designequivariant diffusionstructure-based drug designdrug combination synergytarget deconvolutiongenerative AI
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Geometric deep learning (GDL), by encoding 3D molecular structure directly, offers a path beyond the 'one drug, one target' paradigm that fails against cancer and neurodegenerative disease. The review's thesis is that GDL has converted polypharmacology from serendipitous discovery into a rational design problem: modern generative models can propose a single ligand that satisfies the 3D binding constraints of two distinct protein pockets at once. It maps the evolution from pocket-embedding and fragment assembly to end-to-end SE(3)-equivariant diffusion models, and claims these models resolve inter-pocket geometric conflicts that defeated sequential optimization. A sympathetic reader would care because if the thesis holds, multi-target drug design becomes programmable, and drug combinations with their dosing and toxicity complications could be replaced by single molecules.

Core claim

The review's central claim is that geometric deep learning has moved polypharmacology from serendipitous discovery ('dirty drugs') to a rational design paradigm: modern GDL generative models can construct a single ligand that satisfies the 3D binding constraints of two or more protein pockets simultaneously. It identifies three enabling capabilities — pocket embedding in latent space, multi-target bioactivity prediction via heterogeneous graph fusion, and de novo dual-target generation — and argues that the last is now realized by end-to-end architectures. DualDiff, for instance, resolves dual-target spatial conflicts by rigidly superimposing the two pockets and fusing geometric messages fro

What carries the argument

The load-bearing machinery is the SE(3)-equivariant diffusion model operating on a composite graph: two protein pockets are brought into a shared coordinate system (or, in FuseDiff, left unaligned), and a shared ligand node receives 'geometric pulls' from both binding sites through equivariant message passing during reverse diffusion. This is contrasted with the external-reward paradigm (AIxFuse, MDRL) in which reinforcement learning and docking feedback steer generation. The review also introduces a 'Weighted Target Interaction Graph' to deconvolve drug-combination synergy scores into target-pair supervision — the proposed bridge from macroscopic phenotypes to 3D geometric benchmarks.

Load-bearing premise

The entire pipeline assumes that aggregate synergy scores from two-drug screens can be decomposed into reliable, single-molecule geometric supervision for a specific target pair; if that deconvolution is not trustworthy, the benchmark-construction workflow collapses.

What would settle it

Take a reviewed dual-target generative model (e.g., DualDiff), generate 100 ligands for a target pair lacking co-crystal structures, and measure experimental binding (SPR or ITC) to both proteins; the central claim fails if fewer than a substantial fraction of computationally 'dual-active' hits show micromolar-or-better affinity for both targets, or if no generated molecule binds both targets with the intended affinity ratio.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If correct, dual-target ligands can be generated end-to-end without serial single-target screening or manual fragment splicing, bypassing the combinatorial explosion of multi-target screening.
  • Pre-trained single-target diffusion models can be repurposed zero-shot for new target pairs, as DualDiff demonstrates, making multi-target design a compositional extension of existing generative models.
  • The Weighted Target Interaction Graph workflow could turn large drug-combination databases (DrugComb, NCI-ALMANAC) into training data for 3D geometric models, converting phenotype-level synergy into structure-level supervision.
  • Alignment-free joint-density generation (FuseDiff) would allow multi-target design to handle pockets with different conformations, not just rigid superpositions, bringing generated ligands closer to true induced-fit binding.
  • The same equivariant machinery could incorporate 'negative design' — repulsive gradients that avoid anti-targets and off-target toxicity — turning selectivity into a tunable geometric constraint.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The deconvolution from synergy scores to target-pair supervision is the most fragile link in this entire program; if macroscopic synergy from drug-pair screens does not track single-molecule geometric affinity, the benchmark pipeline built on it will systematically mislead. The paper itself concedes this in its Section 4 challenge 3.
  • A concrete testable extension is to apply DualDiff-style layer-wise fusion to three-target combinations; if internal geometric coupling scales, the 'one-key-fits-two-locks' metaphor should generalize, and failure at three targets would reveal a capacity ceiling in current equivariant message passing.
  • Clinical translation hangs on wet-lab confirmation; until GDL-designed dual-target molecules are co-crystallized or otherwise experimentally shown to occupy both targets with the predicted occupancy ratio, the 'rational era' claim remains a computational promise rather than a therapeutic result.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. This review argues that geometric deep learning (GDL) can move polypharmacology from serendipitous multitarget screening to rational, structure-driven design. It surveys invariant and equivariant architectures, pocket representation and comparison, multi-target activity prediction, dual-target generative models (AIxFuse, MDRL, DualDiff, FuseDiff), multimodal data integration, validation strategies, and public data resources. It also proposes a workflow for constructing target–target geometric benchmarks from drug-combination synergy data, centered on a Weighted Target Interaction Graph with putative driver target pairs. The central thesis is that end-to-end GDL generative frameworks have already "formally transitioned" the field and "surmounted the computational barrier of combinatorial explosion."

Significance. If properly calibrated, the review fills a genuine gap: it assembles a broad and mostly accurate picture of GDL methods applied to polypharmacology, including useful comparative tables (Tables 2–4), a clear architectural taxonomy (invariant vs. equivariant, internal vs. external conditioning), and an updated list of public drug-combination resources. The proposed benchmark workflow in Section 3.3 is a constructive suggestion for an unmet need. The paper is also honest in places: Section 4 explicitly lists wet-lab validation as an unresolved challenge. However, the manuscript's headline claim overstates what the cited evidence supports, and the proposed benchmark is presented as a methodology although it is untested. The review would be valuable after the central claim is reframed as a roadmap aspiration rather than an accomplished transition, and after the proposed benchmark is clearly labeled a proposal with open validation questions.

major comments (3)
  1. [§2.4.2, §4 (Abstract and Conclusion)] The load-bearing claim that polypharmacology ligand design "has formally transitioned ... and enter[ed] an era of rational automated generation" is contradicted by the manuscript's own Section 4, Challenge 4: "GDL-generated multi-target candidates lack extensive wet-lab confirmation, and high virtual scores do not necessarily correspond to authentic intracellular target occupancy." The cited models (AIxFuse, MDRL, DualDiff) are evaluated with docking-derived or docking-in-the-loop metrics, and FuseDiff is an unreviewed 2026 arXiv preprint. No wet-lab dual-target engagement is reported for any model. Section 2.6.2 also places such validation in the future ("In the near future, it is expected..."). I request that the "formal transition" language be replaced with a calibrated claim — e.g., a promising but not yet experimentally validated route — and that the abstract and conclusion be revis
  2. [§3.3.2, §4 Challenge 3] The proposed Weighted Target Interaction Graph and "driver target pairs" depend on deconvolving macroscopic drug-pair synergy scores into geometric supervision for specific protein pocket pairs. The manuscript itself concedes in Section 4, Challenge 3, that combination therapies permit dose adjustment while single-molecule affinity ratios are fixed by structure, so in vitro synergy cannot be mapped to single-molecule affinities without strong assumptions. Since the benchmark workflow in Section 3.3.2 is presented as a "methodology" but contains no validation and no concrete algorithm for the deconvolution/credit assignment, it should be explicitly labeled a research proposal with testable steps, not an established infrastructure.
  3. [§2.4.2, claim of surmounting combinatorial explosion] The statement that end-to-end GDL frameworks "surmount the computational barrier of combinatorial explosion" is unsupported by the review. AIxFuse is an MCTS search and MDRL is an RL optimization; both are iterative search procedures, and no wall-clock, sample-complexity, or coverage comparison against fragment-linking or combinatorial docking baselines is provided. Either provide such evidence or soften the claim to indicate that these methods reformulate, rather than eliminate, the combinatorial search.
minor comments (4)
  1. [Table 3 / Ref [70]] FuseDiff is described as a peer-level "state-of-the-art" model but is an unreviewed 2026 arXiv preprint (arXiv:2603.05567). It should be marked as a preprint and its unreviewed status acknowledged.
  2. [Tables 1 and 4] There are minor numerical inconsistencies: DrugComb is described as "nearly 700,000" combinations in Section 3.1.1 but Table 4 lists ~740,000; NCI-ALMANAC text says ~100 drugs while Table 4 says 104 drugs. Please align these figures.
  3. [§2.6.2] The sentence "For instance, a single molecule inhibiting two kinases could achieve superior anti-tumor efficacy compared to individual inhibition" is a hypothetical, not a result. Rephrase to avoid implying experimental validation.
  4. [References] Reference [70] (2026) and several Journal of Pharmaceutical Analysis 2026 references should be checked for bibliographic completeness; if they are not yet published, use preprint or in-press labels consistently.

Circularity Check

0 steps flagged

No circular derivation found; the central review claim is a literature synthesis supported by external citations, with only peripheral self-citations.

full rationale

This is a review, not an original method or benchmark paper, so the usual input-output circularity patterns (self-definitional targets, fitted parameters renamed as predictions, uniqueness imported from authors) do not apply. The load-bearing statement in Section 2.4.2 that end-to-end generation frameworks signify that polypharmacology design has 'formally transitioned from serendipitous blind screening or multi-step splicing' is presented as a synthesis of externally cited methods [65,66,67,69], not as a result derived from the authors' own definitions or fitted quantities. The proposed Weighted Target Interaction Graph benchmark in Section 3.3.2 is explicitly a proposal ('we propose constructing a global Weighted Target Interaction Graph'), not a validation of the review's central claim. The two self-citations ([115] BIOEMU and [118] Ames model) appear only in suggested future directions and do not support the review's load-bearing claims. Section 4, challenge 4, does concede that 'GDL-generated multi-target candidates lack extensive wet-lab confirmation,' and challenge 3 concedes the deconvolution problem between phenotypic synergy and single-molecule affinity; these are serious evidence limitations for the review's enthusiastic framing, but they are correctness/risk concerns, not circularity. No circular step can be quoted, so the appropriate circularity score is low.

Axiom & Free-Parameter Ledger

0 free parameters · 4 axioms · 3 invented entities

This review contains no original derivation or fitted model, so there are no free parameters. The axioms listed are the domain assumptions that the surveyed and proposed pipelines depend on; the invented entities are conceptual constructs proposed in Section 3.3.2 and Section 4, none with independent evidence.

axioms (4)
  • domain assumption 3D structural complementarity is the primary driver of multi-target drug efficacy.
    The review assumes resolving geometric conflicts between binding pockets is the core bottleneck (Sections 1 and 2.4). Section 4, challenge 3 partially undercuts this by noting that phenotypic synergy and physical geometry are not directly equatable.
  • domain assumption Drug-combination synergy scores can be mapped to specific target pairs via deconvolution.
    Section 3.3.2's Weighted Target Interaction Graph depends on this mapping. The paper itself notes the risk of false positives from 'bystander targets' and combinatorial explosion.
  • domain assumption PDB and AlphaFold structures are adequate geometric ground truth for training multi-target models.
    Section 3.3.1 uses PDBbind plus AlphaFold DB with P2Rank and AutoDock Vina to build 12,917 target pairs. No experimental validation of these modeled complex conformations is provided.
  • domain assumption SE(3)-equivariance of a network implies its learned features correspond to physically meaningful interactions.
    Section 2.7 argues that equivariant architectures 'guard against spurious, frame-dependent features' and yield interpretability. This is a hopeful correspondence, not a proven property.
invented entities (3)
  • Weighted Target Interaction Graph no independent evidence
    purpose: Proposed benchmark construct to deconvolve drug-pair synergy scores into target-pair edge weights.
    Introduced in Section 3.3.2 as a recommendation; no implementation, data release, or validation is provided.
  • Driver target pairs no independent evidence
    purpose: Latent labels for high-confidence synergistic target pairs extracted from the weighted graph.
    Defined in Section 3.3.2; no experimental or structural evidence is supplied.
  • Occupancy-driven conditional generation no independent evidence
    purpose: Future framework to map QSP-derived target occupancy distributions into diffusion-model constraints.
    Proposed in Section 4, challenge 3; no implementation or proof of concept exists.

pith-pipeline@v1.3.0-alltime-deepseek · 29941 in / 9166 out tokens · 93829 ms · 2026-08-02T05:28:14.210361+00:00 · methodology

0 comments
read the original abstract

The traditional "one drug, one target" paradigm of structure-based drug design (SBDD) frequently proves inadequate for treating multifactorial diseases such as cancer and neurodegenerative disorders, owing to compensatory signaling pathways and the emergence of drug resistance. While polypharmacology offers a synergistic therapeutic strategy, the rational design of ligands capable of simultaneously satisfying the geometric constraints imposed by multiple targets remains a major computational bottleneck. This review positions geometric deep learning (GDL) as a powerful integrative approach to overcome these limitations. We systematically survey GDL architectures ranging from invariant graph neural networks to SE(3)-equivariant diffusion models that harness non-Euclidean molecular data to capture intrinsic three-dimensional (3D) structural interdependencies. We critically analyze GDL applications across three core dimensions, including the characterization of shared binding pockets via geometric embeddings, multi-target bioactivity prediction through heterogeneous graph fusion, and de novo generation of dual-target ligands. Particular emphasis is placed on emerging structure-conditioned generative algorithms that integrate diffusion models with reinforcement learning to autonomously resolve complex geometric conflicts between competing binding sites. Furthermore, we evaluate the pivotal role of multimodal omics integration and specialized geometric benchmarking infrastructures in validating these models. By synthesizing these methodological advances, this review elucidates the paradigm shift in drug discovery from serendipitous exploration to rational, structure-driven polypharmacological molecular engineering, thereby providing a clear, structured guide for navigating the complexities of next-generation therapeutics.

Figures

Figures reproduced from arXiv: 2607.20550 by Qi Zhao, Tianming Han, Wenchi Ge, Zhijie Pan.

Figure 1
Figure 1. Figure 1: Geometric deep learning (GDL) framework enabling the paradigm shift from traditional structure-based drug design (SBDD) to polypharmacology and multi-target drug design. (A) Transition from the conventional “one drug, one target” strategy to single-molecule polypharmacology, and further to the direct de novo generation of multi￾target drugs enabled by GDL, thereby overcoming drug resistance, harnessing syn… view at source ↗
Figure 2
Figure 2. Figure 2: Schematic illustrations of representative geometric deep learning architectures in molecular modeling and drug design. (A) Graph Convolutional Network (GCN). (B) Graph Attention Network (GAT). (C) Message Passing Neural Network (MPNN). (D) E(n)-Equivariant Graph Neural Network (EGNN). (E) SE(3)-Transformer. These panels depict the core computational mechanisms of neighborhood aggregation, attention weighti… view at source ↗
Figure 3
Figure 3. Figure 3: Schematic comparison of state-of-the-art structure-aware multi￾target ligand generation models using GDL. (A) AIxFuse. External reward [PITH_FULL_IMAGE:figures/full_fig_p078_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Proposed workflow for constructing target–target benchmarks for GDL in polypharmacology and multi-target drug design. (A) Primary data sources comprising high-throughput screening (HTS) synergy matrices (DrugComb, NCI-ALMANAC), clinical combination databases (DCDB 2.0, CDCDB), protein structures (PDBbind, AlphaFold DB), and target mapping resources (DrugBank, TTD). (B) Three-stage protocol: (1) target deco… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

122 extracted references · 55 canonical work pages

  1. [1]

    Polypharmacology: The science of multi -targeting molecules

    Kabir A, Muth A. Polypharmacology: The science of multi -targeting molecules. Pharmacological Research. 2022;176:106055. doi:10.1016/j.phrs.2021.106055

    arXiv 2022

  2. [2]

    Polypharmacology by Design: A Medicinal Chemist’s Perspective on Multitargeting Compounds

    Proschak E, Stark H, Merk D. Polypharmacology by Design: A Medicinal Chemist’s Perspective on Multitargeting Compounds. Journal of Medicinal Chemistry. 2019;62(2):420 –

    2019

  3. [3]

    one ligand – one pocket

    Data infrastructures for geometric polypharmacology To advance GDL beyond conventional single -target SBDD, models must be trained on datasets capable of capturing the complexity of multi - target interactions. Unlike classic affinity datasets such as PDBbind [100,101], which address only “one ligand – one pocket ” scenarios, polypharmacology requires dat...

  4. [4]

    one drug, one target

    Challenges and future perspectives The conventional “one drug, one target” paradigm in SBDD has revealed critical therapeutic limitations, particularly its propensity to elicit compensatory signaling pathways when addressing complex systemic diseases such as cancers exhibiting drug resistance, neurodegenerative disorders, and met abolic syndromes. In cont...

    2025

  5. [5]

    DNA -encoded library screening identifies CDK2 - targeting lead compounds with favorable drug -like properties for anticancer development

    Zhou L, Ju Y, Cao Z, Cai S, Su J, Lu J. DNA -encoded library screening identifies CDK2 - targeting lead compounds with favorable drug -like properties for anticancer development. Journal of Pharmaceutical Analysis. 2026;16(5):101498. doi:10.1016/j.jpha.2025.101498

    arXiv 2026

  6. [6]

    Pharmacological effects, classification, genetic and molecular studies of different chemotypes essential oil of Perilla frutescens (L.) Britt.: A review

    Wei W, Li Z, Wang B, Liu Y, Sun Y, Wen D, et al. Pharmacological effects, classification, genetic and molecular studies of different chemotypes essential oil of Perilla frutescens (L.) Britt.: A review. Journal of Pharmaceutical Analysis. 2026;16(5):10 1454. doi:10.1016/j.jpha.2025.101454

    arXiv 2026

  7. [7]

    Toward structure –multiple activity relationships (SMARts) using computational approaches: A polypharmacological perspective

    López -López E, Medina -Franco JL. Toward structure –multiple activity relationships (SMARts) using computational approaches: A polypharmacological perspective. Drug Discovery Today. 2024;29(7):104046. doi:10.1016/j.drudis.2024.104046

    arXiv 2024

  8. [8]

    De novo generation of multi -target compounds using deep generative chemistry

    Munson BP, Chen M, Bogosian A, Kreisberg JF, Licon K, Abagyan R, et al. De novo generation of multi -target compounds using deep generative chemistry. Nature Communications. 2024;15(1):3636. doi:10.1038/s41467-024-47120-y

    work page doi:10.1038/s41467-024-47120-y 2024

  9. [9]

    Cucurbitacin B mitigates Staphylococcus aureus pathogenicity and reprograms macrophage responses to restore host defense

    Sun L, Wang L, Guo D, Liu X, Wang B, Zhao Y, et al. Cucurbitacin B mitigates Staphylococcus aureus pathogenicity and reprograms macrophage responses to restore host defense. Journal of Pharmaceutical Analysis. 2026;16(5):101557. doi:10.1016/j.jpha.2026.101557

    arXiv 2026

  10. [10]

    AlphaFill: enriching AlphaFold models with ligands and cofactors

    Hekkelman ML, De Vries I, Joosten RP, Perrakis A. AlphaFill: enriching AlphaFold models with ligands and cofactors. Nature Methods. 2023;20(2):205 –213. doi:10.1038/s41592-022- 01685-y

    work page doi:10.1038/s41592-022- 2023

  11. [11]

    Accurate structure prediction of biomolecular interactions with AlphaFold 3

    Abramson J, Adler J, Dunger J, Evans R, Green T, Pritzel A, et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature. 2024;630(8016):493 –500. doi:10.1038/s41586-024-07487-w

    doi:10.1038/s41586-024-07487-w 2024

  12. [12]

    AlphaFold Protein Structure Database: massively expanding the structural coverage of protein-sequence space with high -accuracy models

    Varadi M, Anyango S, Deshpande M, Nair S, Natassia C, Yordanova G, et al. AlphaFold Protein Structure Database: massively expanding the structural coverage of protein-sequence space with high -accuracy models. Nucleic Acids Research. 2022;50(D1):D439 –D444. doi:10.1093/nar/gkab1061

    doi:10.1093/nar/gkab1061 2022

  13. [13]

    AI for targeted polypharmacology: The next frontier in drug discovery

    Cichońska A, Ravikumar B, Rahman R. AI for targeted polypharmacology: The next frontier in drug discovery. Current Opinion in Structural Biology. 2024;84:102771. doi:10.1016/j.sbi.2023.102771

    arXiv 2024

  14. [14]

    Structure -based drug design with geometric deep learning

    Isert C, Atz K, Schneider G. Structure -based drug design with geometric deep learning. Current Opinion in Structural Biology. 2023;79:102548. doi:10.1016/j.sbi.2023.102548

    arXiv 2023

  15. [15]

    Comprehensive detection and characterization of human druggable pockets through binding site descriptors

    Comajuncosa -Creus A, Jorba G, Barril X, Aloy P. Comprehensive detection and characterization of human druggable pockets through binding site descriptors. Nature Communications. 2024;15(1):7917. doi:10.1038/s41467-024-52146-3

    work page doi:10.1038/s41467-024-52146-3 2024

  16. [17]

    A geometric deep learning approach to predict binding conformations of bioactive molecules

    Méndez-Lucio O, Ahmad M, Del Rio-Chanona EA, Wegner JK. A geometric deep learning approach to predict binding conformations of bioactive molecules. Nature Machine Intelligence. 2021;3(12):1033–1039. doi:10.1038/s42256-021-00409-9

    work page doi:10.1038/s42256-021-00409-9 2021

  17. [18]

    Predicting locations of cryptic pockets from single protein structures using the PocketMiner graph neural network

    Meller A, Ward M, Borowsky J, Kshirsagar M, Lotthammer JM, Oviedo F, et al. Predicting locations of cryptic pockets from single protein structures using the PocketMiner graph neural network. Nature Communications. 2023;14(1):1177. doi:10.1038/s41467-023-36699-3

    doi:10.1038/s41467-023-36699-3 2023

  18. [19]

    Enabling late-stage drug diversification by high -throughput experimentation with geometric deep learning

    Nippa DF, Atz K, Hohler R, Müller AT, Marx A, Bartelmus C, et al. Enabling late-stage drug diversification by high -throughput experimentation with geometric deep learning. Nature Chemistry. 2024;16(2):239–248. doi:10.1038/s41557-023-01360-5

    work page doi:10.1038/s41557-023-01360-5 2024

  19. [20]

    Generating 3D small binding molecules using shape -conditioned diffusion models with guidance

    Chen Z, Peng B, Zhai T, Adu -Ampratwum D, Ning X. Generating 3D small binding molecules using shape -conditioned diffusion models with guidance. Nature Machine Intelligence. 2025;7(5):758–770. doi:10.1038/s42256-025-01030-w

    work page doi:10.1038/s42256-025-01030-w 2025

  20. [21]

    Magic shotguns versus magic bullets: selectively non - selective drugs for mood disorders and schizophrenia

    Roth BL, Sheffler DJ, Kroeze WK. Magic shotguns versus magic bullets: selectively non - selective drugs for mood disorders and schizophrenia. Nature Reviews Drug Discovery. 2004;3(4):353–359. doi:10.1038/nrd1346

    work page doi:10.1038/nrd1346 2004

  21. [22]

    Preclinical overview of sorafenib, a multikinase inhibitor that targets both Raf and VEGF and PDGF receptor tyrosine kinase signaling

    Wilhelm SM, Adnane L, Newell P, Villanueva A, Llovet JM, Lynch M. Preclinical overview of sorafenib, a multikinase inhibitor that targets both Raf and VEGF and PDGF receptor tyrosine kinase signaling. Molecular Cancer Therapeutics. 2008;7(10):3129 –3140. doi:10.1158/1535-7163.MCT-08-0013

    work page doi:10.1158/1535-7163.mct-08-0013 2008

  22. [23]

    Machine learning-aided generative molecular design

    Du Y, Jamasb AR, Guo J, Fu T, Harris C, Wang Y, et al. Machine learning-aided generative molecular design. Nature Machine Intelligence. 2024;6(6):589–604. doi:10.1038/s42256-024- 00843-5

    doi:10.1038/s42256-024- 2024

  23. [24]

    Improving de novo molecular design with curriculum learning

    Guo J, Fialková V, Arango JD, Margreitter C, Janet JP, Papadopoulos K, et al. Improving de novo molecular design with curriculum learning. Nature Machine Intelligence. 2022;4(6):555–563. doi:10.1038/s42256-022-00494-4

    work page doi:10.1038/s42256-022-00494-4 2022

  24. [25]

    PPARα: An emerging target of metabolic syndrome, neurodegenerative and cardiovascular diseases

    Lin Y, Wang Y, Li P. PPARα: An emerging target of metabolic syndrome, neurodegenerative and cardiovascular diseases. Frontiers in Endocrinology. 2022;13:1074911. doi:10.3389/fendo.2022.1074911

    arXiv 2022

  25. [26]

    Semi-Supervised Classification with Graph Convolutional Networks

    Kipf TN, Welling M. Semi-Supervised Classification with Graph Convolutional Networks. Proceedings of the International Conference on Learning Representations. 2017

    2017

  26. [27]

    Graph Attention Networks

    Veli č kovič P, Cucurull G, Casanova A, Romero A, Li ò P, Bengio Y. Graph Attention Networks. Proceedings of the International Conference on Learning Representations. 2018

    2018

  27. [28]

    Kinase drug discovery 20 years after imatinib: progress and future directions

    Cohen P, Cross D, Jänne PA. Kinase drug discovery 20 years after imatinib: progress and future directions. Nature Reviews Drug Discovery. 2021;20(7):551 –569. doi:10.1038/s41573- 021-00195-4

    work page doi:10.1038/s41573- 2021

  28. [29]

    Kaul U, Parmar D, Manjunath K, Shah M, Parmar K, Patil KP, et al. New dual peroxisome proliferator activated receptor agonist —Saroglitazar in diabetic dyslipidemia and non - alcoholic fatty liver disease: integrated analysis of the real world evidence. Cardiovascular Diabetology. 2019;18(1):80. doi:10.1186/s12933-019-0884-3

    work page doi:10.1186/s12933-019-0884-3 2019

  29. [30]

    SchNet – A deep learning architecture for molecules and materials

    Schütt KT, Sauceda HE, Kindermans P -J, Tkatchenko A, Müller K -R. SchNet – A deep learning architecture for molecules and materials. The Journal of Chemical Physics. 2018;148(24):241722. doi:10.1063/1.5019779

    doi:10.1063/1.5019779 2018

  30. [31]

    Directional Message Passing for Molecular Graphs

    Gasteiger J, Groß J, Günnemann S. Directional Message Passing for Molecular Graphs. Proceedings of the International Conference on Learning Representations. 2020

    2020

  31. [32]

    E(n) Equivariant Graph Neural Networks

    Satorras VG, Hoogeboom E, Welling M. E(n) Equivariant Graph Neural Networks. Proceedings of the 38th International Conference on Machine Learning. 2021;139:9323–9332

    2021

  32. [33]

    Neural Message Passing for Quantum Chemistry

    Gilmer J, Schoenholz SS, Riley PF, Vinyals O, Dahl GE. Neural Message Passing for Quantum Chemistry. Proceedings of the 34th International Conference on Machine Learning. 2017;70:1263–1272

    2017

  33. [34]

    Graph Attention Site Prediction (GrASP): Identifying Druggable Binding Sites Using Graph Neural Networks with Attention

    Smith Z, Strobel M, Vani BP, Tiwary P. Graph Attention Site Prediction (GrASP): Identifying Druggable Binding Sites Using Graph Neural Networks with Attention. Journal of Chemical Information and Modeling. 2024;64(7):2637–2644. doi:10.1021/acs.jcim.3c01698

    work page doi:10.1021/acs.jcim.3c01698 2024

  34. [35]

    Learning from Protein Structure with Geometric Vector Perceptrons

    Jing B, Eismann S, Suriana P, Townshend RJL, Dror R. Learning from Protein Structure with Geometric Vector Perceptrons. Proceedings of the International Conference on Learning Representations. 2021

    2021

  35. [36]

    CASTER -DTA: equivariant graph neural networks for predicting drug –target affinity

    Kumar R, Romano JD, Ritchie MD. CASTER -DTA: equivariant graph neural networks for predicting drug –target affinity. Briefings in Bioinformatics. 2025;26(5):bbaf554. doi:10.1093/bib/bbaf554

    work page doi:10.1093/bib/bbaf554 2025

  36. [37]

    EquiBind: Geometric Deep Learning for Drug Binding Structure Prediction

    Stärk H, Ganea O, Pattanaik L, Barzilay DR, Jaakkola T. EquiBind: Geometric Deep Learning for Drug Binding Structure Prediction. Proceedings of the 39th International Conference on Machine Learning. 2022;162:20503–20521

    2022

  37. [38]

    SE(3) -Transformers: 3D Roto -Translation Equivariant Attention Networks

    Fuchs F, Worrall D, Fischer V, Welling M. SE(3) -Transformers: 3D Roto -Translation Equivariant Attention Networks. Advances in Neural Information Processing Systems. 2020;33:1970–1981

    2020

  38. [39]

    Tensor field networks: Rotation- and translation -equivariant neural networks for 3D point clouds

    Thomas N, Smidt T, Kearnes S, Yang L, Li L, Kohlhoff K, et al. Tensor field networks: Rotation- and translation -equivariant neural networks for 3D point clouds. arXiv. 2018. doi:10.48550/arXiv.1802.08219

    Pith/arXiv arXiv doi:10.48550/arxiv.1802.08219 2018

  39. [40]

    Structure -based drug design with equivariant diffusion models

    Schneuing A, Harris C, Du Y, Didi K, Jamasb A, Igashov I, et al. Structure -based drug design with equivariant diffusion models. Nature Computational Science. 2024;4(12):899–909. doi:10.1038/s43588-024-00737-x

    doi:10.1038/s43588-024-00737-x 2024

  40. [41]

    PILOT: equivariant diffusion for pocket - conditioned de novo ligand generation with multi -objective guidance via importance sampling

    Cremer J, Le T, Noé F, Clevert D -A, Schütt KT. PILOT: equivariant diffusion for pocket - conditioned de novo ligand generation with multi -objective guidance via importance sampling. Chemical Science. 2024;15(36):14954–14967. doi:10.1039/D4SC03523B

    doi:10.1039/d4sc03523b 2024

  41. [42]

    Equivariant diffusion for structure-based de novo ligand generation with latent -conditioning

    Le T, Cremer J, Clevert D-A, Schütt KT. Equivariant diffusion for structure-based de novo ligand generation with latent -conditioning. Journal of Cheminformatics. 2025;17(1):90. doi:10.1186/s13321-025-01028-x

    work page doi:10.1186/s13321-025-01028-x 2025

  42. [43]

    Geometric Deep Learning for Structure -Based Ligand Design

    Powers AS, Yu HH, Suriana P, Koodli RV, Lu T, Paggi JM, et al. Geometric Deep Learning for Structure -Based Ligand Design. ACS Central Science. 2023;9(12):2257 –2267. doi:10.1021/acscentsci.3c00572

    doi:10.1021/acscentsci.3c00572 2023

  43. [44]

    GeoDiff: a Geometric Diffusion Model for Molecular Conformation Generation

    Xu M, Yu L, Song Y, Shi C, Ermon S, Tang J. GeoDiff: a Geometric Diffusion Model for Molecular Conformation Generation. Proceedings of the International Conference on Learning Representations. 2022

    2022

  44. [45]

    SE(3)-equivariant ternary complex prediction towards target protein degradation

    Xue F, Zhang M, Li S, Gao X, Wohlschlegel JA, Huang W, et al. SE(3)-equivariant ternary complex prediction towards target protein degradation. Nature Communications. 2025;16(1):5514. doi:10.1038/s41467-025-61272-5

    work page doi:10.1038/s41467-025-61272-5 2025

  45. [46]

    Electron -density-informed effective and reliable de novo molecular design and optimization with ED2Mol

    Li M, Song K, He J, Zhao M, You G, Zhong J, et al. Electron -density-informed effective and reliable de novo molecular design and optimization with ED2Mol. Nature Machine Intelligence. 2025;7(8):1355–1368. doi:10.1038/s42256-025-01095-7

    work page doi:10.1038/s42256-025-01095-7 2025

  46. [47]

    DeepPocket: Ligand Binding Site Detection and Segmentation using 3D Convolutional Neural Networks

    Aggarwal R, Gupta A, Chelur V, Jawahar CV, Priyakumar UD. DeepPocket: Ligand Binding Site Detection and Segmentation using 3D Convolutional Neural Networks. Journal of Chemical Information and Modeling. 2022;62(21):5069–5079. doi:10.1021/acs.jcim.1c00799

    work page doi:10.1021/acs.jcim.1c00799 2022

  47. [48]

    GraphBind: protein structural context embedded rules learned by hierarchical graph neural networks for recognizing nucleic-acid-binding residues

    Xia Y, Xia C-Q, Pan X, Shen H-B. GraphBind: protein structural context embedded rules learned by hierarchical graph neural networks for recognizing nucleic-acid-binding residues. Nucleic Acids Research. 2021;49(9):e51–e51. doi:10.1093/nar/gkab044

    work page doi:10.1093/nar/gkab044 2021

  48. [49]

    Deciphering interaction fingerprints from protein molecular surfaces using geometric deep learning

    Gainza P, Sverrisson F, Monti F, Rodolà E, Boscaini D, Bronstein MM, et al. Deciphering interaction fingerprints from protein molecular surfaces using geometric deep learning. Nature Methods. 2020;17(2):184–192. doi:10.1038/s41592-019-0666-6

    work page doi:10.1038/s41592-019-0666-6 2020

  49. [50]

    GPSFun: geometry -aware protein sequence function predictions with language models

    Yuan Q, Tian C, Song Y, Ou P, Zhu M, Zhao H, et al. GPSFun: geometry -aware protein sequence function predictions with language models. Nucleic Acids Research. 2024;52(W1):W248–W255. doi:10.1093/nar/gkae381

    work page doi:10.1093/nar/gkae381 2024

  50. [51]

    Unified protein –small molecule graph neural networks for binding site prediction

    Wang J, Dokholyan NV. Unified protein –small molecule graph neural networks for binding site prediction. Proceedings of the National Academy of Sciences. 2026;123(10):e2524913123. doi:10.1073/pnas.2524913123

    work page doi:10.1073/pnas.2524913123 2026

  51. [52]

    Spatiotemporal identification of druggable binding sites using deep learning

    Kozlovskii I, Popov P. Spatiotemporal identification of druggable binding sites using deep learning. Communications Biology. 2020;3(1):618. doi:10.1038/s42003-020-01350-0

    work page doi:10.1038/s42003-020-01350-0 2020

  52. [53]

    Targeting protein– ligand neosurfaces with a generalizable deep learning tool

    Marchand A, Buckley S, Schneuing A, Pacesa M, Elia M, Gainza P, et al. Targeting protein– ligand neosurfaces with a generalizable deep learning tool. Nature. 2025;639(8054):522–531. doi:10.1038/s41586-024-08435-4

    work page doi:10.1038/s41586-024-08435-4 2025

  53. [54]

    Protein Binding Site Representation in Latent Space

    Lohmann F, Allenspach S, Atz K, Schiebroek CCG, Hiss JA, Schneider G. Protein Binding Site Representation in Latent Space. Molecular Informatics. 2025;44(1):e202400205. doi:10.1002/minf.202400205

    work page doi:10.1002/minf.202400205 2025

  54. [55]

    Applications of machine learning in drug discovery and development

    Vamathevan J, Clark D, Czodrowski P, Dunham I, Ferran E, Lee G, et al. Applications of machine learning in drug discovery and development. Nature Reviews Drug Discovery. 2019;18(6):463–477. doi:10.1038/s41573-019-0024-5

    doi:10.1038/s41573-019-0024-5 2019

  55. [56]

    Rethinking drug design in the artificial intelligence era

    Schneider P, Walters WP, Plowright AT, Sieroka N, Listgarten J, Goodnow RA, et al. Rethinking drug design in the artificial intelligence era. Nature Reviews Drug Discovery. 2020;19(5):353–364. doi:10.1038/s41573-019-0050-3

    work page doi:10.1038/s41573-019-0050-3 2020

  56. [57]

    RELATION: A Deep Generative Model for Structure-Based De Novo Drug Design

    Wang M, Hsieh C -Y, Wang J, Wang D, Weng G, Shen C, et al. RELATION: A Deep Generative Model for Structure-Based De Novo Drug Design. Journal of Medicinal Chemistry. 2022;65(13):9478–9492. doi:10.1021/acs.jmedchem.2c00732

    work page doi:10.1021/acs.jmedchem.2c00732 2022

  57. [58]

    Navigating structure-based drug discovery with emerging innovations in physics - and knowledge -based approaches

    Lam JH, Katritch V. Navigating structure-based drug discovery with emerging innovations in physics - and knowledge -based approaches. npj Drug Discovery. 2025;2(1):29. doi:10.1038/s44386-025-00031-4

    work page doi:10.1038/s44386-025-00031-4 2025

  58. [59]

    Generative Deep Learning for de Novo Drug Design─A Chemical Space Odyssey

    Özçelik R, Brinkmann H, Criscuolo E, Grisoni F. Generative Deep Learning for de Novo Drug Design─A Chemical Space Odyssey. Journal of Chemical Information and Modeling. 2025;65(14):7352–7372. doi:10.1021/acs.jcim.5c00641

    doi:10.1021/acs.jcim.5c00641 2025

  59. [60]

    De novo design of protein structure and function with RFdiffusion

    Watson JL, Juergens D, Bennett NR, Trippe BL, Yim J, Eisenach HE, et al. De novo design of protein structure and function with RFdiffusion. Nature. 2023;620(7976):1089 –1100. doi:10.1038/s41586-023-06415-8

    doi:10.1038/s41586-023-06415-8 2023

  60. [61]

    Guided multi -objective generative AI to enhance structure -based drug design

    Kadan A, Ryczko K, Lloyd E, Roitberg A, Yamazaki T. Guided multi -objective generative AI to enhance structure -based drug design. Chemical Science. 2025;16(29):13196 –13210. doi:10.1039/D5SC01778E

    doi:10.1039/d5sc01778e 2025

  61. [62]

    Multi-Objective Molecule Generation using Interpretable Substructures

    Jin W, Barzilay DR, Jaakkola T. Multi-Objective Molecule Generation using Interpretable Substructures. Proceedings of the 37th International Conference on Machine Learning. 2020;119:4849–4859

    2020

  62. [63]

    FragGen: towards 3D geometry reliable fragment-based molecular generation

    Zhang O, Huang Y, Cheng S, Yu M, Zhang X, Lin H, et al. FragGen: towards 3D geometry reliable fragment-based molecular generation. Chemical Science. 2024;15(46):19452–19465. doi:10.1039/D4SC04620J

    work page doi:10.1039/d4sc04620j 2024

  63. [64]

    Target -aware 3D molecular generation based on guided equivariant diffusion

    Hu Q, Sun C, He H, Xu J, Liu D, Zhang W, et al. Target -aware 3D molecular generation based on guided equivariant diffusion. Nature Communications. 2025;16(1):7928. doi:10.1038/s41467-025-63245-0

    doi:10.1038/s41467-025-63245-0 2025

  64. [65]

    Structure -aware dual-target drug design through collaborative learning of pharmacophore combination and molecular simulation

    Chen S, Xie J, Ye R, Xu DD, Yang Y. Structure -aware dual-target drug design through collaborative learning of pharmacophore combination and molecular simulation. Chemical Science. 2024;15(27):10366–10380. doi:10.1039/D4SC00094C

    work page doi:10.1039/d4sc00094c 2024

  65. [66]

    A 3D generation framework using diffusion model and reinforcement learning to generate multi -target compounds with desired properties

    Yuan Y, Pan X, Li X, Zhang R, Su W. A 3D generation framework using diffusion model and reinforcement learning to generate multi -target compounds with desired properties. Journal of Cheminformatics. 2025;17(1):93. doi:10.1186/s13321-025-01035-y

    work page doi:10.1186/s13321-025-01035-y 2025

  66. [67]

    bystander targets

    made foundational contributions to bridging this phenotype–structure gap. Starting from DrugCombDB, they screened drug combinations exhibiting positive synergistic effects using four mainstream synergy scoring systems, namely ZIP, Bliss, Loewe, and HSA. They subsequently performed target deconvolution to map the drugs to specific protein targets, and inte...

  67. [68]

    MARS: Markov Molecular Sampling for Multi-objective Drug Discovery

    Xie Y, Shi C, Zhou H, Yang Y, Zhang W, Yu Y, et al. MARS: Markov Molecular Sampling for Multi-objective Drug Discovery. Proceedings of the International Conference on Learning Representations. 2021

    2021

  68. [69]

    Automated design of multi -target ligands by generative deep learning

    Isigkeit L, Hörmann T, Schallmayer E, Scholz K, Lillich FF, Ehrler JHM, et al. Automated design of multi -target ligands by generative deep learning. Nature Communications. 2024;15(1):7946. doi:10.1038/s41467-024-52060-8

    work page doi:10.1038/s41467-024-52060-8 2024

  69. [70]

    FuseDiff: Symmetry-Preserving Joint Diffusion for Dual-Target Structure-Based Drug Design

    Wu J, Qiao A, Wang Z, Wei Z, Chen S. FuseDiff: Symmetry-Preserving Joint Diffusion for Dual-Target Structure-Based Drug Design. arXiv. 2026. doi: 10.48550/arXiv.2603.05567

    work page doi:10.48550/arxiv.2603.05567 2026

  70. [71]

    Building a knowledge graph to enable precision medicine

    Chandak P, Huang K, Zitnik M. Building a knowledge graph to enable precision medicine. Scientific Data. 2023;10(1):67. doi:10.1038/s41597-023-01960-3

    doi:10.1038/s41597-023-01960-3 2023

  71. [72]

    Reprogramming Pretrained Target - Specific Diffusion Models for Dual -Target Drug Design

    Zhou X, Guan J, Zhang Y, Peng X, Wang L, Ma J. Reprogramming Pretrained Target - Specific Diffusion Models for Dual -Target Drug Design. Advances in Neural Information Processing Systems. 2024;37:87255–87281. doi: 10.52202/079017-2769

    work page doi:10.52202/079017-2769 2024

  72. [73]

    DeepDTAGen: a multitask deep learning framework for drug -target affinity prediction and target -aware drugs generation

    Shah PM, Zhu H, Lu Z, Wang K, Tang J, Li M. DeepDTAGen: a multitask deep learning framework for drug -target affinity prediction and target -aware drugs generation. Nature Communications. 2025;16(1):5021. doi:10.1038/s41467-025-59917-6

    doi:10.1038/s41467-025-59917-6 2025

  73. [74]

    3D molecular generative framework for interaction -guided drug design

    Zhung W, Kim H, Kim WY. 3D molecular generative framework for interaction -guided drug design. Nature Communications. 2024;15(1):2688. doi:10.1038/s41467-024-47011-2

    work page doi:10.1038/s41467-024-47011-2 2024

  74. [75]

    Systematic identification of genomic markers of drug sensitivity in cancer cells

    Garnett MJ, Edelman EJ, Heidorn SJ, Greenman CD, Dastur A, Lau KW, et al. Systematic identification of genomic markers of drug sensitivity in cancer cells. Nature. 2012;483(7391):570–575. doi:10.1038/nature11005

    doi:10.1038/nature11005 2012

  75. [76]

    A Landscape of Pharmacogenomic Interactions in Cancer

    Iorio F, Knijnenburg TA, Vis DJ, Bignell GR, Menden MP, Schubert M, et al. A Landscape of Pharmacogenomic Interactions in Cancer. Cell. 2016;166(3):740 –754. doi:10.1016/j.cell.2016.06.017

    doi:10.1016/j.cell.2016.06.017 2016

  76. [77]

    A knowledge graph to interpret clinical proteomics data

    Santos A, Colaço AR, Nielsen AB, Niu L, Strauss M, Geyer PE, et al. A knowledge graph to interpret clinical proteomics data. Nature Biotechnology. 2022;40(5):692 –702. doi:10.1038/s41587-021-01145-6

    doi:10.1038/s41587-021-01145-6 2022

  77. [78]

    GeOKG: geometry -aware knowledge graph embedding for Gene Ontology and genes

    Jeong C -U, Kim J, Kim D, Sohn K -A. GeOKG: geometry -aware knowledge graph embedding for Gene Ontology and genes. Bioinformatics. 2025;41(4):btaf160. doi:10.1093/bioinformatics/btaf160

    work page doi:10.1093/bioinformatics/btaf160 2025

  78. [79]

    Hi -GeoMVP: a hierarchical geometry-enhanced deep learning model for drug response prediction

    Chen Y, Zhang L. Hi -GeoMVP: a hierarchical geometry-enhanced deep learning model for drug response prediction. Bioinformatics. 2024;40(4):btae204. doi:10.1093/bioinformatics/btae204

    work page doi:10.1093/bioinformatics/btae204 2024

  79. [80]

    Generalized biomolecular modeling and design with RoseTTAFold All -Atom

    Krishna R, Wang J, Ahern W, Sturmfels P, Venkatesh P, Kalvet I, et al. Generalized biomolecular modeling and design with RoseTTAFold All -Atom. Science. 2024;384(6693):eadl2528. doi:10.1126/science.adl2528

    doi:10.1126/science.adl2528 2024

  80. [81]

    Evolutionary -scale prediction of atomic- level protein structure with a language model

    Lin Z, Akin H, Rao R, Hie B, Zhu Z, Lu W, et al. Evolutionary -scale prediction of atomic- level protein structure with a language model. Science. 2023;379(6637):1123 –1130. doi:10.1126/science.ade2574

    doi:10.1126/science.ade2574 2023

Showing first 80 references.