Chai-1 results
One top-ranked mmCIF structure with aggregate_score, ptm, iptm and per-chain plddt_mean — how to read each, and why none of them is an affinity or an RMSD.
A finished Chai-1 job returns the top-ranked candidate structure as mmCIF text, three confidence numbers for the complex, and a per-chain breakdown. A chai_results JSON artifact carries the whole result object.
Result fields
| Field | Meaning |
|---|---|
structure_cif | The top-ranked candidate structure, as mmCIF text |
aggregate_score | The score the candidates were ranked by; higher is better. The returned structure is the highest-scoring one |
ptm | Predicted TM-score for the complex |
iptm | Interface predicted TM-score. null when there is only one entity, since there is no interface to score |
chains | One entry per chain: chain_id, sequence, n_residues, and plddt_mean |
num_chains | Number of protein chains folded. A ligand is not counted here, so a one-protein, one-ligand complex reports 1 |
num_models | The num_models value carried with the request |
plddt_mean is reported on a 0–100 scale, as a mean over that chain's atoms.
How to read them
plddt_meanjudges each chain's own fold. A low value on one chain means that chain's geometry is unreliable, whatever the complex-level scores say.iptmestimates how accurately the interface was modelled — not whether the chains bind. It is the number to read for a contact question, and it has real predictive power for pose correctness, but a highiptmis a statement about the model's own confidence in its geometry, not evidence of an interaction. Good chains with a weak interface is the common failure mode; so is good chains with a confident interface in the wrong place. See translatability.aggregate_scoreis a ranking quantity, so it is meaningful between candidates within one run, not across targets or across jobs.- None of these is an affinity. Chai-1 returns no binding-energy or affinity value. For a model-predicted affinity, Boltz-2 has an affinity head; for physics-based rescoring of a pose, use MM-GBSA.
Reliability guard
If inference scores the chains but no structure file is written, the job fails rather than returning an empty structure_cif alongside real-looking confidence numbers. A chai_predict job that reports success has a structure with atoms in it.
What is not returned
- No pose ensemble. One top-ranked structure per job.
- No RMSD or DockQ to a reference. There is no reference-structure input, so if you want an agreement number for your own target you have to compute it yourself against a deposited structure. Translatability covers what the published figures are and which metric they use.
- No physical-validity check. Inspect the geometry before building on it.
Read next
Was this page helpful?
Chai-1 inputs and parameters
Chain entries are chain_id, sequence_type and sequence; protein only. Three sampling parameters are not read — the run is fixed at 3 recycles, 200 steps.
How Chai-1 results compare to the bench
What a Chai-1 predicted complex has been shown to agree with experimentally, and why a no-MSA single-seed run sits at the weak end of the published range.