Structure prediction

Chai-1

Molecular structure prediction, including proteins. Similar to AlphaFold3.

A multi-modal foundation model for molecular structure prediction that performs at the state-of-the-art across a variety of benchmarks. Chai-1 enables unified prediction of proteins, small molecules, DNA, RNA, glycosylations, and more.

Documentation ↗ Input parameters ↗ Examples ↗ Home page ↗ Paper ↗ Source code ↗ License ↗

No jobs before this; it will run immediately.

Not logged in. Register an account to keep track of your jobs and results, and add presets and bookmarks. Email results to this address when complete:

Ready

Example preset

Choose an example to fill in its settings and input structures.

Input mode

Input

Official documentation ↗

Job name Names the run and its result archive.

Molecules

One box per entity. Its name becomes the name of its FASTA record, which is what restraints refer to it by; left blank, entities are named A, B, C… in order. Proteins, DNA, RNA and SMILES ligands are supported. Use the FASTA mode for glycans.

Chai FASTA file Upload a native Chai FASTA file.

Evolution and templates

Official documentation ↗

Use ESM embeddings Upstream default. Disable for a sequence-only ablation.

Use public MSA server Generate protein MSAs with the ColabFold MMseqs2 server below. This sends the protein sequences to that server and needs network access. A local aligned.pqt directory or the supplied example MSAs take precedence. Turn off for private sequences or offline runs; Chai-1 then relies on its ESM embeddings alone.

MSA server URL ColabFold-compatible MMseqs2 endpoint; used only when the MSA server is on. The public server asks for serial queries from a single IP, so host your own for large batches.

Local aligned.pqt directory Existing compute-node directory with one <sequence SHA256>.aligned.pqt per protein, as documented in the official MSA example. Takes precedence over the MSA server.

Use supplied example MSAs Use the two aligned.pqt files from Chai’s official examples, downloaded when Chai-1 is installed or first used. Only suitable for the sequences in the official protein-complex preset.

Search public templates Search pdb100 templates with the MSA server; requires the public MSA server.

Local template hits .m8 Existing compute-node MMseqs2 .m8 template hits file. Exclusive with the template server.

Restraints

Official documentation ↗

Native restraints CSV Paste a Chai restraint table with the official header. Supports contact, pocket and covalent rows. Use either this table or the guided fields below.

Restraints CSV file Upload the same native restraint table instead of pasting it.

Pocket restraints (JSON) Guided format: [{"chainA":"A","res_idxA":387,"chainB":"B"}]. First chain has the specific residue. Chain letters follow FASTA entity order.

Contact restraints (JSON) Guided format: [{"chainA":"A","res_idxA":387,"chainB":"B","res_idxB":101}].

Covalent restraints (JSON) Guided format: [{"chainA":"A","res_idxA":217,"covalentAtomA":"SG","chainB":"B","covalentAtomB":"S1"}]. Ligand/glycan partners need an atom name but no residue index.

Guided minimum distance (Å)

Guided maximum distance (Å)

Sampling

Official documentation ↗

Trunk samples

Trunk recycles

Diffusion timesteps

Diffusion samples per trunk Number of candidate structures to generate per trunk. More samples use more GPU memory; try 1 for large complexes or 16 GB GPUs.

MSA subsample per recycle Upstream recycle_msa_subsample argument; 0 disables subsampling.

Random seed

Low-memory inference Move model components as needed to lower peak GPU memory.

CUDA device CUDA device such as cuda:0.

Ready

Endpoint POST /api/chai1/

Send JSON or standard form data. A GET request returns this tool's field schema.

API reference for Chai-1 →