Protein Structure · Chai Discovery
Chai-1
Predicts 3D structures of multimodal biochemical complexes, including proteins, nucleic acids, and small molecules.
What Chai-1 is
Chai-1 is a multimodal molecular structure prediction foundation model developed by Chai Discovery. It provides unified structure determination for assemblies containing proteins, DNA, RNA, small molecules, modified residues, and glycosylations in a single predictive framework.
The model addresses the challenge of predicting heterogeneous biomolecular interactions at all-atom resolution. It is designed as an open-weights tool for researchers studying complex multi-entity assemblies, offering an alternative architecture capable of handling multi-protein interactions alongside bound chemical entities.
Outputs include all-atom three-dimensional coordinates in PDB format accompanied by per-residue predicted Local Distance Difference Test (pLDDT) confidence scores spanning 0 to 100, where scores above 70 generally indicate reliable coordinate placement.
How it works
Chai-1 generates all-atom structure predictions from primary sequence data, incorporating nucleotide sequences, protein sequences, and small molecules provided as SMILES strings. By default, the model uses sequence embeddings and can sample multiple conformations (defaulting to five samples) without requiring multiple sequence alignments (MSAs) or structural templates.
For enhanced prediction accuracy, inference can incorporate MSAs and structural templates via external servers or custom files. Users can also supply custom experimental restraints to guide the folding process, including inter-chain contact boundaries and explicit covalent bonds for structures such as branched ligands.
Key facts
- Version
- 1.0
- Supported modalities
- Protein sequence, nucleotide sequence (DNA and RNA), small molecules (SMILES)
- Outputs
- All-atom PDB coordinates and residue-level pLDDT confidence scores
- Hardware requirement
- CUDA GPU with bfloat16 support (recommended: A100 80GB, H100 80GB, or L40S 48GB; minimum 40 GB GPU memory for typical complexes)
What you can use it for
- Protein–Ligand Complex Prediction
- Determining all-atom structures of small molecules bound to target proteins by supplying ligand structures as SMILES strings alongside amino acid sequences.
- Protein–Nucleic Acid Assembly
- Predicting multi-chain assemblies containing protein subunits together with DNA or RNA strands in unified coordinate spaces.
- Restraint-Guided Molecular Folding
- Incorporating known inter-chain contact restraints or user-specified covalent bonds to steer folding in modified residues and branched ligand configurations.
- Rapid Embedding-Only Complex Sampling
- Generating initial conformational ensembles using default embedding-based inference without compiling MSAs or structural templates.
Run it here
No installation and no GPU of your own: fill in the inputs and it runs on our servers. Results stay in your browser unless you sign in.
Runs on On-demand GPU
Runs on our servers with on-demand compute. A first run needs time to load the model; active capacity can be reused and scales down when idle.Run this model
Chai-1
Strengths and limits
Good at
- Operates across diverse molecular entities within a single unified model, including proteins, nucleic acids, small molecules, and post-translational glycosylations.
- Can perform inference using sequence embeddings alone without mandatory MSA generation or structural templates.
- Accommodates user-defined physical restraints, such as inter-chain contacts and specific covalent bonds.
- Supports integration with external ColabFold MMseqs2 servers for automated MSA generation as well as custom local MSAs and template CIF files.
Watch out for
- Requires substantial GPU hardware with bfloat16 support, typically needing 40 GB to 80 GB of VRAM for larger complexes.
- Default MSA generation through the public ColabFold server relies on a shared external resource that uses a different search strategy than the paper's benchmarks.
- Custom template ingestion requires specific formatting and manual configuration of m8 hit files and corresponding CIF structure files.
Inputs and outputs
You provide
Entity sequences
One entity per line; SMILES for ligands. Chai-1 FASTA headers are understood too.
Search for homologues (MSA)optional
Recyclesoptional
More passes, more accuracy, proportionally more time.
Structures to sampleoptional
You get back
All-atom coordinates in PDB format, viewable and downloadable.
pLDDT score per residue (0–100). Above 70 is generally reliable.
Licence
Check the licence before commercial use · Chai Discovery Community License
The Chai Discovery Community License licence carries terms worth reading before you depend on it.
Questions
Can I use Chai-1 commercially?
Check the licence before commercial use. The Chai Discovery Community License licence carries terms worth reading before you depend on it.
How do I run Chai-1 without installing anything?
Use the form under “Run it here” on this page: fill in the inputs and press Run model. It runs on our servers on 2× A100, typically in 4.0min, and a run costs about 1,260 credits. Visitors without an account get 1,000 credits a month.
Can Chai-1 fold complexes without multiple sequence alignments (MSAs)?
Yes. By default, Chai-1 runs inference using sequence embeddings without MSAs or structural templates, though supplying MSAs and templates is recommended for higher accuracy.
How are small-molecule ligands supplied to the model?
Ligands are specified directly in the input FASTA sequence file as SMILES strings alongside the protein or nucleic acid sequences.
What confidence metrics does Chai-1 provide?
The model outputs per-residue pLDDT scores on a scale from 0 to 100, where values above 70 are considered generally reliable.
Can I enforce specific covalent bonds between components?
Yes. Chai-1 supports input feature contexts that allow users to explicitly specify covalent connections, which is useful for branched ligands and post-translational modifications.
How to cite it
Chai Discovery. Chai-1: Decoding the molecular interactions of life. bioRxiv (2024). doi: 10.1101/2024.10.10.615955.
Written from Chai-1's own documentation (code). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.