henhaobio
Sign in

bioRxivSynthetic BiologyPreprint

Chai-1: Decoding the molecular interactions of life

Chai Discovery, Jacques BoitreaudiD, Jack Dent, Matt McPartloniD, Joshua MeieriD, Vinicius Reis, Alex RogozhnikoviD, Kevin WuiD

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

Abstract

We introduce Chai-1, a multi-modal foundation model for molecular structure prediction that performs at the state-of-the-art across a variety of tasks relevant to drug discovery. Chai-1 can optionally be prompted with experimental restraints (e.g. derived from wet-lab data) which boosts performance by double-digit percentage points. Chai-1 can also be run in single-sequence mode with-out MSAs while preserving most of its performance. We release Chai-1 model weights and inference code as a Python package for non-commercial use and via a web interface where it can be used for free including for commercial drug discovery purposes.

In-depth summary

Claude Opus 5.5 · xhigh effort · 17 Oct 2026 · from the full text of version 2 (15 Oct 2024) · CC BY-NC 4.0

Chai-1 matches AlphaFold3 on PoseBusters ligand poses (77% vs 76%), beats AlphaFold-Multimer 2.3 on low-homology interfaces, and stays competitive without MSAs on multimers and antibodies.

All-atom models such as RoseTTAFold All-Atom and AlphaFold3 predict complexes of proteins, nucleic acids and small molecules, yet AlphaFold3's commercial-use restrictions kept it out of the authors' direct benchmarks. Chai-1 adds protein language model embeddings for MSA-free prediction and trainable pocket, contact and docking restraint inputs, uses one model with a 2021-01-12 data cutoff for every evaluation, and releases weights and inference code for non-commercial use plus a free web server.

Key results

  1. ShownGiven only the protein sequence and ligand chemistry, Chai-1 places ligands about as well as AlphaFold3 on PoseBusters; supplying the apo protein structure raises its own success rate further.

    Ligand RMSD < 2 Å success 77% vs 76% reported for AlphaFold3; supplying the apo protein structure (docking setting) raised it to 81%, which the authors use mainly as a test of prompt following. All PoseBusters structures post-date the 2021-01-12 training cutoff; 4/428 were cropped to 2048 tokens. · 2.2 Protein-ligand prediction

  2. ShownOn post-cutoff, low-homology protein-protein interfaces Chai-1 outperforms AlphaFold-Multimer 2.3 (AF2.3), including on antibody-protein interfaces and at stricter DockQ cutoffs.

    Cluster-averaged DockQ success (DockQ > 0.23) 0.751 vs 0.677 for AF2.3 over 2362 interfaces in 929 clusters (two-sided Wilcoxon p = 6.24 × 10^-10); 0.743 without templates, not significantly lower. Antibody-protein subset (268 interfaces, 122 clusters): Chai-1 ahead with p = 3.25 × 10^-5. · 2.3 Multimeric protein prediction

  3. ShownWithout MSAs, Chai-1 stays close to MSA-using AF2.3 on complexes and to RoseTTAFold2NA on nucleic acids, but falls behind on single-chain protein folding.

    Single-sequence multimer DockQ success 0.698 vs 0.677 for AF2.3 with MSAs (not significant); on antibody-protein interfaces it matches full-MSA Chai-1 and beats AF2.3 (p = 4.04 × 10^-4). Low-homology monomers: worse than AF2.3 (p = 4.46 × 10^-16). MSA-free nucleic acid predictions were comparable to RoseTTAFold2NA (interface LDDT; 9 CASP15 RNA targets). · 2.3 Multimeric protein prediction; 2.4 Protein monomer prediction; 2.5 Nucleic acid structure prediction

  4. SuggestedPrompting with restraints improves antibody-antigen docking, but high-quality predictions remain rare.

    One antibody-antigen distance restraint (θ ≤ 15 Å) raised DockQ-acceptable predictions from 35% to 57%; four randomly sampled epitope residues more than doubled success at every DockQ cutoff compared with blind prediction, yet high-quality predictions stayed at 4-8%. The restraints were simulated, not measured. · 2.3 Multimeric protein prediction

  5. ShownWith MSAs, Chai-1 folds protein monomers slightly better than AF2.3, with a clearer gain on CASP15 targets that AF2.3 finds hard.

    Low-homology set of 447 monomers (271 clusters): small but significant Cα-LDDT gain (p = 7.21 × 10^-10). CASP15 (69 targets): mean LDDT 0.849 vs 0.843; on the 14 targets with AF2.3 LDDT < 0.75, 0.643 vs 0.552 (p = 3.66 × 10^-4). ESM3's own report gives 0.801 on all 70 targets with its 98-billion-parameter model. · 2.4 Protein monomer prediction

Methods, data and reproducibility

Chai-1 largely follows the Abramson et al. architecture (pair-bias self-attention, diffusion sampling), adding residue embeddings from a 3 billion parameter protein language model and pocket, contact and docking restraint features trained with dropout. One model trained on PDB (cutoff 2021-01-12) and AlphaFoldDB structures was benchmarked on PoseBusters, CASP15 and a temporal low-homology PDB set.

  • PoseBusters protein-ligand benchmark, all released on or after 2021-01-13 (7D6O excluded as obsolete) · n = 428 structures
  • Low-homology multimer evaluation set from PDB entries released 2022-05-01 to 2023-01-12 (non-NMR, resolution better than 4.5 Å) · n = 2362 interfaces in 1054 structures, 929 interface clusters
  • Antibody-protein subset of the multimer evaluation set · n = 268 interfaces, 129 structures, 122 clusters
  • Low-homology protein monomer set from the same temporal split · n = 447 monomers in 271 clusters
  • CASP15 protein monomer targets (T1169 excluded) · n = 69 of 70 targets
  • CASP15 RNA targets and low-homology protein-nucleic acid complexes, compared with RoseTTAFold2NA · n = 9 RNA targets
  • Training data: PDB structures before the 2021-01-12 cutoff plus AlphaFoldDB structures; templates from PDB70; MSAs from OpenProteinSet or jackhmmer
Methods & tools
Chai-1, AlphaFold-Multimer 2.3 via ColabFold 1.5.5, RoseTTAFold2NA (Tamarind Bio API), jackhmmer v3.4, MMseqs2, OpenProteinSet, DockQ, PoseBusters python package, UniRef90, UniProt, MGnify, UniClust30+BFD, PDB70, AlphaFoldDB
Reagents & environment
Released as a Python package; MSAs from jackhmmer v3.4 (the web server runs one iteration and skips Uniclust30+BFD and templates); AF2.3 baseline via ColabFold version 1.5.5 with MMseqs2 against UniRef30 (version 2303) and colabfold_envdb (version 202108).
Assay conditions & compute
Trained on 128 Nvidia A100 GPUs with batch size 128 for 30 days; inference capped at 2048 tokens, by default with 4 recycles and 5 trunk × 5 diffusion samples (25 structures).
Reproducibility gaps
Weights and code (non-commercial licence), the GitHub repository, the web server and the PoseBusters predictions are all mentioned, but no links survive in this text; server and paper results may differ slightly.
Limitations (6)
  • Individual chains can be predicted correctly but placed in the wrong relative orientation; in the authors' example the complex is poor without added contact information. · stated by the authors
  • Removing modified residues, or swapping them for standard analogues, can change predicted structures substantially. · stated by the authors
  • AF3 and ESM3 could not be run because of commercial-use restrictions; their numbers are quoted from their own papers, and interface results are not directly comparable to AF3's because the evaluation sets differ. · stated by the authors
  • Nucleic acids are modelled without MSAs; AF3 reports LDDT 0.473 on 8 CASP15 RNA targets, which the authors read as better RNA performance than theirs. · stated by the authors
  • High-quality antibody-antigen predictions stay at 4-8% even with restraints. · stated by the authors
  • Restraint gains were measured with simulated epitope and distance restraints rather than real HDX-MS or cross-linking data, so benefits with noisy experimental restraints are untested here. · noted in this summary

Read this if you predict protein-ligand, protein-protein or antibody complex structures and want an openly available model that can run without MSAs or be steered with pocket and contact restraints.

Citations
443

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
THISBen-Joseph '23Liang '24Peng '26Jung '26Kosonocky '26Yang '26
62% semantic matchBioinformatics
JJ Ben-Joseph, Tim Oates·bioRxiv·Nov 11, 2023· 1 citations