Skip to content

Chai-1

Summary

Chai-1 is a multi-modal foundation model for joint structure prediction of proteins, antibodies, small molecules, DNA, RNA and covalent modifications, following the general design of AlphaFold 3. Its defining feature is a protein-language-model input track that lets it retain most of its accuracy in a "single-sequence" mode with no multiple sequence alignment (MSA). It can additionally be prompted with experimental restraints (contacts, pockets, docking distances) to boost accuracy on hard targets.[1]

Architecture

Chai-1 follows the trunk/diffusion design of AlphaFold 3 but trains one model across all evaluations (data cutoff 2021-01-12) rather than separate specialist models.[1] Two additions extend it beyond AF3:

  • Protein language model embeddings: per-residue embeddings from a 3-billion-parameter protein language model are added as an input track alongside (or instead of) MSAs, enabling accurate single-sequence prediction.[1]
  • Constraint features: optional pocket, contact, and docking-distance features that mimic experimental restraints (e.g. from cross-linking mass spectrometry or epitope mapping), each included independently with 10% probability during training and randomly dropped out to force robustness.[1]

Modified residues are mapped to their canonical parent residue or an unknown ("X") token; non-protein-chain residues (DNA, RNA, ligands) receive a mask token. Chai-1 was trained on 128 Nvidia A100 GPUs (batch size 128, 30 days) on PDB and AlphaFoldDB structures, and inference is limited to 2,048 tokens.[1]

Performance

  • Protein-ligand (PoseBusters): 77% ligand-RMSD (<2 Å) success rate given only sequence and ligand composition, comparable to AlphaFold 3's reported 76%; conditioning on the apo protein structure raises this to 81%.[1]
  • Protein multimers (low-homology PDB set, n=929 interface clusters): average DockQ success rate of 0.751 vs. 0.677 for AlphaFold-Multimer 2.3 (two-sided Wilcoxon p=6.24×10⁻¹⁰); in single-sequence mode (no MSA), Chai-1 still reaches 0.698, comparable to AF-Multimer 2.3 run with MSAs.[1]
  • Antibody-protein interfaces: Chai-1 outperforms AF-Multimer 2.3 (p=3.25×10⁻⁵), and its single-sequence mode performs similarly to its own MSA-based mode here — the authors attribute this to antibody sequences carrying less exploitable evolutionary signal in the first place.[1]
  • Protein monomers (CASP15, n=69): average LDDT of 0.849 vs. 0.843 for AF2.3; on the subset where AF2.3 struggles (LDDT<0.75), Chai-1 reaches 0.643 vs. 0.552 (p=3.66×10⁻⁴). ESM3's largest (98B-parameter) model scores 0.801 on the full 70-target set.[1]
  • Restraint conditioning: on antibody-antigen interfaces, a single distance restraint (≤15 Å) raises the fraction of acceptable DockQ predictions from a 35% (unconstrained) baseline to 57%; conditioning on four epitope residues more than doubles DockQ success across quality cutoffs.[1]

Nucleic-acid structure prediction (run without RNA MSAs) is comparable to RoseTTAFold2NA on both a low-homology protein-nucleotide set and 9 CASP15 RNA targets, despite RoseTTAFold2NA having full evolutionary information available.[1]

Availability

Model weights and inference code are released for non-commercial use at github.com/chaidiscovery/chai-lab; a web server (lab.chaidiscovery.com) is available free of charge, including for commercial drug-discovery use.[1] This is a preprint, not yet peer-reviewed.

See Also

  • AlphaFold 3 — the architecture Chai-1's design most closely follows.
  • Protenix — an open-source reproduction of AlphaFold 3 with a similar scope.
  • Boltz-2 — a related co-folding foundation model that adds binding-affinity regression.

Citations

[1] Chai Discovery team (2024), "Chai-1: Decoding the molecular interactions of life", bioRxiv Supports: architecture, single-sequence/restraint capabilities, all benchmark figures above. Location: Results (Sections 2.1–2.4), Methods.