Chai-1

Chai-1 is an all-atom diffusion co-folder in the same family as Boltz-2 and AlphaFold3: a multi-entity FASTA in, mmCIF plus pTM/ipTM/pLDDT out, with protein, RNA, DNA, and SMILES-ligand chains all first-class. It and boltz cover the same surface; running both and keeping designs that pass either is a common consensus filter, and Chai's Python entry point makes it the easier of the two to embed in a loop. Code and weights are Apache-2.0 — commercial use including drug discovery is explicitly permitted (github.com/chaidiscovery/chai-lab).

Running it

from pathlib import Path
from chai_lab.chai1 import run_inference

Path("complex.fasta").write_text("""
>protein|name=target
MVTPEGNVSLVDESLLVGVTDEDRAVRS...
>protein|name=binder
AIQRTPKIQVYSRHPAENG...
>ligand|name=cofactor
CCCCCCCCCCCCCC(=O)O
""".strip())

candidates = run_inference(
    fasta_file=Path("complex.fasta"),
    output_dir=Path("out/"),
    num_trunk_recycles=3,
    num_diffn_timesteps=200,
    seed=42,
    device="cuda:0",
    use_esm_embeddings=True,
)
print([rd.aggregate_score.item() for rd in candidates.ranking_data])

The FASTA header is >{entitytype}|name={id} with entitytype ∈ {protein, rna, dna, ligand}; ligand records carry a SMILES string as the sequence body, and modified residues are written inline as ...AAK(SEP)AAG.... From the shell the same job is chai-lab fold complex.fasta out/ --use-msa-server. Without --use-msa-server (or usemsaserver=True in Python) the model runs on ESM embeddings alone, which is faster but typically a few ipTM points behind the MSA-backed run.

outputdir receives pred.modelidx{0..4}.cif plus a matching scores.modelidx{N}.npz per sample with aggregatescore, ptm, iptm, perchainptm, and clash flags. Rank by aggregatescore; treat iptm > 0.5 as a soft pass for an interface. The function refuses a non-empty outputdir, so clear or rotate it between calls.

Unset CHAIDOWNLOADSDIR fails mid-run with PermissionError on a read-only image

Chai downloads ~5 GB on the first inference call (not at install time), including its own traced ESM2-3B for the embedding path. If CHAIDOWNLOADSDIR is unset, the default is inside site-packages: on a read-only image that fails with a confusing PermissionError mid-run, and on a writable one it silently re-downloads ~5 GB into the container on every cold start. Export the variable to a persisted volume so the download happens once.

No-MSA mode still loads a 3 B-parameter ESM — same VRAM, not less

useesmembeddings=True without an MSA still loads a 3-billion-parameter language model into GPU memory alongside the trunk; it removes the MSA-server round-trip, not the VRAM cost. If you OOM, drop numdiffntimesteps or fold fewer chains per call rather than expecting the no-MSA mode to fit a smaller card.

Errors worth recognizing

Next: filter survivors on confidence/clash metrics or feed them back to proteinmpnn for the next design round.