Chai-1 on the platform

Chai-1 co-folds a complex from sequence: you give it the chains and it predicts the assembled structure in one pass, ranked by its own confidence scores. On the platform it serves as a second co-folding model alongside Boltz-2, built on a different architecture — though two co-folding models agreeing is weaker evidence than it sounds, because their errors are correlated. Read how Chai-1 results compare to the bench before you act on a predicted interface.

Running chai_predict needs a signed-in account; it is not part of the guest tool set.

What Chai-1 does here

• Predicts one complex from protein sequences and returns the top-ranked candidate.
• Scores the whole complex and its interface: ptm for the complex, iptm for the interface, an aggregate_score used for ranking, and a mean pLDDT per chain.
• Runs on GPU as an async job, typically 5–30 minutes.
• Fails loudly rather than returning a hollow result. If scoring succeeds but no structure was written, the job fails instead of shipping confidence numbers next to an empty structure.

When to use it

• You want a second opinion on a predicted complex. Co-fold the same chains with Chai-1 and Boltz-2. Treat a disagreement as a warning; treat an agreement as weak support, not confirmation — published pose errors correlate strongly between co-folding methods and much less between a co-folding method and a physics-based one.
• You are working on a protein–protein assembly and have sequences rather than a structure, and the chains have close structural relatives in the PDB.
• You want an interface hypothesis to guide an experiment — mutagenesis, epitope mapping, or a crystallography or cryo-EM request.

For an antibody–antigen or nanobody–antigen complex, read translatability first. That is the hardest published case for every co-folding model, and the configuration here — no MSAs, a fixed seed, one returned structure — has none of the mitigations the literature uses.

When to use something else

Limits to know before you run

• Protein chains and one ligand. A chain entry typed protein is folded as a protein chain and an entry typed ligand as a small molecule from its SMILES, at most one per job. An entry typed dna, rna or glycan is refused rather than folded as a protein — use Boltz-2 for a complex that contains DNA or RNA.
• Sampling is yours; the output format is not. num_trunk_recycles (1–20, default 3) and num_diffusion_steps (10–500, default 200) are applied and reported back, and both scale what the run costs. The seed is fixed. output_format is not applied: the structure always comes back as mmCIF — see inputs.
• No MSAs, no templates, no restraints. The run uses ESM embeddings only. There is no way to supply a multiple sequence alignment, a structural template, or an epitope or distance restraint — all three are documented levers on Chai-1's accuracy, and none is reachable here. See translatability.
• One structure per job. The top-ranked candidate comes back, not a pose ensemble.
• A confidence score is not an affinity. Chai-1 returns no binding-energy or affinity value at all.

Citation

Chai Discovery. Chai-1: Decoding the molecular interactions of life. bioRxiv (2024). doi: 10.1101/2024.10.10.615955

Source code: chaidiscovery/chai-lab.

Read next