henhaobio
Sign in

bioRxivPharmacology & ToxicologyPreprint

AlphaFold3 for Structure-guided Ligand Discovery

Kartikeya M. MenoniD, Aakash DavasamiD, Guo Chen, Claire Bryant, Zifang Deng, Brandon Lam, Chao YangiD, Dylan Barcelos, Fangyu LiuiD, Assaf AloniD, Jiankun LyuiD

PDFDOICodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

Deep-learning methods for protein structure prediction, such as AlphaFold2 (AF2) and RosettaFold (RF), have transformed structural biology and accelerated downstream biological discovery. More recent models, including AlphaFold3 (AF3) and RosettaFold All-Atom (RFAA), extend this capability to protein-ligand co-folding, enabling direct prediction of bound complexes from sequence and ligand inputs. This advance has raised the possibility that such models could function not only as structure predictors but also as AI-based molecular docking engines for virtual screening. Yet their true impact on ligand discovery remains unclear and, in many cases, controversial. Here, we systematically assess AF3 co-folding across tasks central to small-molecule discovery compared to conventional molecular docking. First, retrospective enrichment of actives over decoys showed that AF3 outperforms the physics-based docking program DOCK3 across 43 drug targets in DUDE-Z; however, this advantage is largely driven by hidden ligand-only biases inherent to computational decoy sets. In contrast, in three large experimental datasets (the sigma-2 receptor ({sigma}2), the D4 dopamine receptor (D4), AmpC {beta}-lactamase (AmpC)) with over 2,500 tested molecules that lack such biases, DOCK3 achieved stronger overall enrichment, while AF3 contributed mainly to early enrichment. Second, out-of-sample pose reproduction on >8,000 protein-ligand complexes deposited after the AF3 training and validation cutoff showed that AF3 accuracy is strongly dependent on training-set similarity, indicating that the model memorizes atomic positions more than learning general principles of molecular recognition. Finally, in the first prospective head-to-head screen against the {sigma}2 receptor, novel to AF3s training set, AF3 achieved a 13% hit rate and identified a 13 nM binder directly from the screen. However, compared with the parallel DOCK3 campaign, AF3 delivered a two-fold lower hit rate, despite yielding a similar affinity distribution among the top five hits. A crystal structure of the {sigma}2 receptor bound to the most potent AF3-derived hit confirmed that AF3 produced a near-native ligand pose. AF3 therefore represents the beginning of deep-learning-based structure-guided ligand discovery: a complementary tool rather than a replacement for conventional docking, with practical applications both as a screening engine and as a post-docking filter that improves hit rates. More broadly, this work establishes a framework for evaluating next-generation deep-learning co-folding models and quantifying their impact on small-molecule discovery.

In-depth summary

Claude Opus 5.5 · xhigh effort · 2 Nov 2026 · from the full text of version 2 (12 Jun 2026) · CC BY-NC-ND 4.0

Retrospective and prospective benchmarks show AlphaFold3 co-folding memorizes training poses and achieves half the prospective hit rate of DOCK3 at 6,700-fold higher cost, but excels as a post-docking filter.

Co-folding models like AlphaFold3 (AF3) predict protein-ligand complexes directly from sequence and ligand SMILES, raising the question of whether they can replace physics-based molecular docking in virtual screening. Benchmarking AF3 against DOCK3 across 43 DUDE-Z targets, >2,500 experimentally tested compounds on three targets, >8,000 post-cutoff PDB complexes, and a prospective 690,000-compound screen against the sigma-2 receptor defines its memorization limits and value as a post-docking filter.

Key results

  1. ShownApparent enrichment gains of AlphaFold3 over DOCK3 on the DUDE-Z decoy benchmark are largely driven by hidden ligand-only biases that vanish in experimental screening datasets.

    On 43 DUDE-Z targets, template-guided AF3 recovered poses (<2 Å RMSD) in 86% (37/43) and beat DOCK3 logAUC in 93% (40/43), but ligand-only pTM (no protein input) gave non-random enrichment in 84% (36/43) and matched or beat DOCK3 in 47% (20/43). On experimental D4, sigma-2, and AmpC sets (>2,500 compounds), ligand-only logAUC was near zero (-4.08 to 2.15) and DOCK3 achieved higher overall AUC. · Abstract; Retrospective Benchmarking with the DUDE-Z Set; Benchmark on large experimental testing sets…

  2. ShownAcross three unbiased experimental ultra-large docking datasets (sigma-2, D4, and AmpC), DOCK3 achieves higher overall enrichment, while AF3 ipTM rescoring of top docking tranches nearly doubles experimental hit rates.

    On sigma-2, DOCK3 reached an AUC of 77.28 vs 65.66 (AF3 ipTM) and 67.01 (AF3 mPAE), though AF3 improved early enrichment (logAUC and EF10) on sigma-2 and D4; applying an AF3 ipTM filter (>0.80 or >0.83) to the top docking-score bin increased experimental hit rates from 34% to 68% (2.0-fold) for sigma-2, 39% to 66% (1.7-fold) for D4, and 26% to 47% (1.8-fold) for AmpC. · Abstract; Benchmark on large experimental testing sets…; Figure legends

  3. ShownOut-of-sample evaluation on 8,076 post-cutoff PDB complexes shows that AF3 pose accuracy depends strongly on 3D ligand-pocket similarity to the training set and frequently fails to recover native binding interactions.

    AF3 predicted binding pockets (<2 Å RMSD) in 98.2% of 854 chemically novel complexes (2D Morgan Tc <= 0.35), but joint pocket and ligand (<2 Å) success dropped from 49.6% (424/854) under 2D novelty to 17.8% (173/970) when combined 3D ligand-pocket similarity (ComboROCS * PS-Score) was <= 0.35; 62% (2,622/4,235) of poses with ligand RMSD < 2 Å failed to recover >=80% of native interactions. · Abstract; Out-of-Sample Performance; Figure legends

  4. ShownIn a prospective head-to-head screen of 692,689 monocations against the sigma-2 receptor, AF3 achieved half the hit rate of DOCK3 at a 6,700-fold higher compute cost, but identified a 13 nM binder with a crystallographically validated pose.

    Testing 108 DOCK3 and 114 AF3 candidates at 1 µM yielded a 25% (27/108) hit rate for DOCK3 (top 5 Ki = 39-199 nM) vs 13% (15/114) for AF3 (top 5 Ki = 13-131 nM) with 0 shared scaffolds; AF3 took 23,000 A100 GPU hours vs 673 CPU hours for DOCK3 (6,700-fold efficiency gap), and the crystal structure of top AF3 hit ZINCkH000002xsk4 bound to sigma-2 (PDB 36EW) showed 0.67 Å C-alpha and 1.8 Å ligand RMSD. · Abstract; Prospective Screen of σ2 Receptor; Prospective Screens on the σ2 Receptor; Data availability

Methods, data and reproducibility

Template-accelerated AlphaFold3 (v3.0.0, 5 seeds x 5 samples) was benchmarked against DOCK3 (v3.8) on DUDE-Z, three experimental docking sets, and 8,076 post-cutoff PDB complexes using APoc, ComboROCS, PoseBusters, and LUNA interaction fingerprints. Both engines prospectively screened 692,689 Enamine monocations against the sigma-2 receptor, validated by radioligand assays, Boltz-2, and crystallography.

  • DUDE-Z docking benchmark of protein targets with actives and property-matched decoys (plus AF3 training set ligands before Sep 30, 2021) · n = 43 targets (77,396 molecules in 15 full sets; 50 actives + 500 decoys/target)
  • Large experimental compound sets from ultra-large library docking campaigns against AmpC, D4 dopamine receptor, and sigma-2 receptor · n = AmpC n=1,520 (+127), D4 n=553, sigma-2 n=506; 9 post-cutoff crystal structures · PDB: 7MFI, 7M93, 7M94, 7M95, 7M96, 9C81, 9C84, 9DHL, 9C6P
  • Out-of-sample protein-ligand complexes from RCSB PDB (deposited Jan 13, 2023 to Dec 2024) matched to BioLip2 · multiple species · n = 8,076 analyzed complexes (854 with 2D Tc <= 0.35; 970 with 3D*pocket <= 0.35)
  • Prospective Enamine in-stock monocation virtual screen, radioligand binding assays, and LCP crystal structure of sigma-2 receptor · human (Expi293 membranes) and Bos taurus (bovine sigma-2 in Sf9 insect cells) · n = 692,689 monocations screened; 108 DOCK3 and 114 AF3 tested; 1 crystal structure · PDB: 36EW
Methods & tools
AlphaFold3, DOCK3, Boltz-2, PoseBusters, OpenEye ROCS (ComboROCS), APoc, LUNA, DockRMSD, RDKit, HMMER (Jackhmmer), DIALS, PHENIX
Software & environment
AlphaFold3 v3.0.0, DOCK3.8, HMMER v3.4, APoc v1.0, DockRMSD v1.1, LUNA v0.13.1, RDKit v2024.03.5 and v2025.03.3, GraphPad Prism v10.6.1.
Assay conditions & compute
Using PDB templates reduced AF3 runtime from 20 to 2 minutes per complex; prospective screen of ~690,000 compounds required 23,000 A100 GPU hours for AF3 versus 673 CPU hours for DOCK3 (a 34-fold time and 6,700-fold hardware-adjusted efficiency gap).
Reproducibility gaps
AlphaFold3 model parameters require permission from Google DeepMind, OpenEye ROCS is commercial software, and computational poses of the top-ranked prospective AF3 and DOCK3 hits will only be released upon publication.
Limitations (4)
  • Prospective virtual screening was restricted to 692,689 monocationic molecules due to AF3 GPU compute constraints, preventing evaluation on ultra-large (>10^7) make-on-demand libraries or neutral sigma-2 ligands. · stated by the authors
  • AF3 exhibited a strong ranking bias toward positively charged molecules on the sigma-2 receptor, placing no neutral ligands in the top 1% or 5% of ipTM scores and only 1% in the top 10% across the 506-compound experimental set. · stated by the authors
  • Neither AF3 complex confidence scores (ipTM, mPAE) nor Boltz-2 predicted IC50 values correlated strongly with experimental binding affinities (Ki) among active hits. · stated by the authors
  • Prospective head-to-head screening between AF3 and DOCK3 was performed on a single target (the sigma-2 receptor) using a rigid 7MFI PDB receptor template. · noted in this summary

Read this if you use AlphaFold3, Boltz-2, or physics-based molecular docking for structure-based virtual screening and want to know when co-folding generalizes or improves hit rates.

Citations
10

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
THISKim '26Yasumitsu '25Maidanik '26Lyu '23Zheng '25Akabari '26
79% semantic matchBiophysics
Jongbin Kim, G.J. Correy, Brendan W. Hall+17 more·eLife·Oct 12, 2026· 7 citations