henhaobio
Sign in

ScienceBioinformaticsPeer-reviewed journal article

Deep contrastive learning enables genome-wide virtual screening

Yinjun JiaiD, Bowen GaoiD, Jiaxin Tan, Jiqing ZhengiD, Xin HongiD, Wenyu ZhuiD, Haichuan Tan, Yuan XiaoiD, Liping TaniD, Hongyi Cai, Yanwen HuangiD, Zhiheng DengiD, Xiangwei WuiD, Yue Jin, Yafei YuaniD, J. Tian, Wei He, Wei‐Ying MaiD, Ya-Qin ZhangiD, Wei ZhangiD, Lei Liu, Chuangye YaniD, Yanyan LaniD

PDFDOIPreprintCodeData

In-depth summary

Claude Sonnet 4.6 · xhigh effort · 8 Nov 2026 · from the full text of version 3 (29 Apr 2025) · CC BY 4.0

DrugCLIP, a contrastive dense-retrieval model for pockets and molecules, beats docking on DUD-E and LIT-PCBA and screened about 10,000 human AlphaFold2 proteins in 24 hours.

Docking takes seconds to minutes per protein-ligand pair, and one campaign of 1 billion molecules against a single target took two weeks on 10,000 CPU cores, so genome-wide screening is impractical. Affinity-regression deep models suffer from inconsistent labels and train/test distribution shift; DrugCLIP instead aligns pocket and molecule embeddings and ranks by cosine similarity.

Key results

  1. ShownDrugCLIP gave higher early enrichment (EF1%) than every baseline on both standard virtual screening benchmarks, including physics-based docking tools and regression-based machine learning models.

    EF1% above Glide-SP, AutoDock Vina, Surflex, NNscore, RFscore, Pafnucy, OnionNet, PLANET, Gnina and BigBind on DUD-E (102 targets, 22,886 actives) and LIT-PCBA (15 targets). Numeric values are in Tables S4 and S5, not in the text. · Evaluating DrugCLIP performance with benchmarks…; In silico validation with DUD-E and LIT-PCBA…

  2. ShownDrugCLIP kept outperforming docking baselines when training data excluded similar scaffolds (versus Glide-SP) or whole protein families (versus AutoDock Vina), and with pocket side-chain errors up to 3 Å RMSD (versus Vina).

    After Murcko-scaffold removal it still outperformed Glide-SP; after protein-family removal (HMMER/Pfam) it still outperformed AutoDock Vina; it outperformed Vina even with 3 Å side-chain RMSD error (Fig. 2C, 2D). · Evaluating DrugCLIP performance with benchmarks…; Figure legends

  3. ShownScreening cost scales with the number of targets plus molecules, which made a genome-wide campaign over 500 million molecules feasible on one 8-GPU node.

    LIT-PCBA in 38 s (DrugCLIP-S) vs Glide 3 days, Uni-Dock 22 hours, PLANET 3 hours; more than 10 trillion pair scorings in about 24 hours on 8 A100 GPUs; database spans 9,908 targets vs 4,810 in ChEMBL 34. · Evaluating DrugCLIP performance with benchmarks…; Genome-wide virtual screening with DrugCLIP; Figure legends

  4. ShownWet-lab testing of DrugCLIP picks for 5HT2AR and NET yielded diverse binders, including two NET inhibitors more potent than bupropion that were solved by cryo-EM.

    5HT2AR: 78 compounds, 8 agonists in calcium flux, best affinity 21.0 nM, EC50 60.3 nM. NET: 100 compounds, 15% over 60% inhibition, 12 beat bupropion; IC50 0.31 μM and 1.14 μM vs 1.5 μM; cryo-EM 2.87 Å and 2.98 Å. · Evaluating DrugCLIP performance with benchmarks…; The structure determination of NET and its…

  5. ShownGenPack refinement raised DrugCLIP screening enrichment on AlphaFold2 and apo structures and docking success on AlphaFold2 structures; for TRIP12, which lacks reference structures or inhibitors, 10 of 57 tested compounds showed SPR Kd below 50 μM.

    DUD-E subset EF1%: Fpocket 19.0% vs 24.1% with GenPack (AF2); apo 11.5% to 20.4%; AF2 docking success 19.1% to 38.7%. TRIP12: 10 of 57 compounds with SPR Kd under 50 μM (17.5%); best 10.8 μM and 11.9 μM. · Applying DrugCLIP to AlphaFold-predicted…; Evaluating the effectiveness of GenPack model

Methods, data and reproducibility

Pocket and molecule encoders (Uni-Mol architecture) are aligned with a batch-softmax contrastive loss: the pocket encoder is pretrained on 5.5 million ProFSA pseudo pocket-fragment pairs, then both are fine-tuned on BioLip2 complexes with RDKit conformers. Screening ranks molecules by cosine similarity.

  • ProFSA pretraining set: pseudo-ligand/pseudo-pocket pairs extracted from protein-only PDB structures · n = 5.5 million pairs
  • BioLip2 protein-ligand complexes for fine-tuning (redundancy-filtered) · n = around 43,980 complexes
  • DUD-E benchmark (actives plus decoys) for enrichment and generalization tests · n = 102 targets; 22,886 actives
  • LIT-PCBA benchmark derived from PubChem BioAssay · n = 15 targets
  • GenPack training/validation from PDBbind, tested on DUD-E pairs · n = 13,137 train, 1,479 validation, 101 test pairs
  • AlphaFold DB human predicted proteins used for genome-wide pocket detection · human · n = 20,296 proteins; 24,692 super-domains
  • ZINC20 and Enamine REAL chemical libraries, property- and alert-filtered · n = over 500 million drug-like molecules
  • Cryo-EM structures of human NET bound to two DrugCLIP hits · human · n = 2,918 and 4,687 micrographs · PDB 9JEL, 9JF3; EMD-61420, EMD-61426
Methods & tools
Uni-Mol, RDKit, Fpocket, GenPack, Glide-SP, AutoDock Vina, Uni-Dock, PLANET, IFD-MD, TM-align, MMseqs2, cryoSPARC
Software & environment
PyTorch (no version stated) for similarity scoring; Schrodinger Suite for Glide-SP and Prime; GraphPad Prism 8 for curve fitting. No containers or version pins stated.
Assay conditions & compute
Pretraining and fine-tuning on 4 NVIDIA A100 GPUs; GenPack training on a single A100; genome-wide screen in about 24 hours on one node with 8 A100 GPUs.
Reproducibility gaps
Authors note human evaluation of candidates affects screening outcomes, so the released pipeline is only a reference for wet-lab validation. Per-target numeric results are in supplementary tables not in the text.
Limitations (5)
  • Pocket definition matters for predicted structures: using Fpocket instead of pockets from structural alignment cut the DUD-E subset EF1% from 29.3% to 19.0%, so automated pocket detection remains a weak point. · stated by the authors
  • Genome-wide pocket detection used only 20,296 of the 20,504 AlphaFold human proteins, because 208 proteins above 2500 amino acids lack accessible PAE data, so the largest proteins were excluded. · stated by the authors
  • Candidate molecules for wet-lab tests were chosen with human expert review, clustering, docking and filters, which can influence outcomes and limits fully automatic reproducibility. · stated by the authors
  • Benchmark baseline results are taken from earlier studies and some baseline speeds are converted or estimated, so comparisons were not run under uniform conditions. · noted in this summary
  • The genome-wide hits (over 2 million) are filtered by DrugCLIP Z-score and Glide score and were experimentally tested for only three targets, so hit rates for most of the 9,908 targets are unknown. · noted in this summary

Read this if you want a fast contrastive-retrieval alternative to docking for ultra-large library screening, including AlphaFold2 pockets, plus a released genome-wide hit database.

Citations
67
FWCI
241.63
Field percentile
top 1%

Abstract

Numerous protein-coding genes are associated with human diseases, yet approximately 90% of them lack targeted therapeutic intervention. While conventional computational methods, such as molecular docking, have facilitated the discovery of potential hit compounds, the development of genome-wide virtual screening against the expansive chemical space remains a formidable challenge. Here we introduce DrugCLIP, a novel framework that combines contrastive learning and dense retrieval to achieve rapid and accurate virtual screening. Compared to traditional docking methods, DrugCLIP improves the speed of virtual screening by up to seven orders of magnitude. In terms of performance, DrugCLIP not only surpasses docking and other deep learning-based methods across two standard benchmark datasets, but also demonstrates high efficacy in wet-lab experiments. Specifically, DrugCLIP successfully identified agonists with < 100 nM affinities for 5HT2AR, a key target in psychiatric diseases. For another target NET, whose structure is newly solved and not included in the training set, our method achieved a hit rate of 15%, with 12 diverse molecules exhibiting affinities better than bupropion. Additionally, two chemically novel inhibitors were validated by structure determination with Cryo-EM. Finally, a novel potential drug target TRIP12, with no experimental structures and inhibitors for reference, was used to challenge DrugCLIP. DrugCLIP achieved a hit rate of 17.5% by screening a pocket identified on an AlphaFold2-predicted structure, verified with multi-cycle SPR assays. Molecules with the highest affinities also showed a dose-dependent inhibition to the enzymatic function of TRIP12. Building on this foundation, we present the results of a pioneering trillion-scale genome-wide virtual screening, encompassing approximately 10,000 AlphaFold2 predicted proteins within the human genome and 500 million molecules from the ZINC and Enamine REAL database. This work provides an innovative perspective on drug discovery in the post-AlphaFold era, where comprehensive targeting of all disease-related proteins is within reach.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISUtgés '26Hong '26Wang '24Oliveira '26Zhou '24Gaiser '25
68% semantic matchBioinformatics
Javier S. Utgés, David T. Jones, Christine Anne Orengo·bioRxiv·Sep 18, 2026