Chai-1

What is Chai-1?

Chai-1 is a multi-modal foundation model for molecular structure prediction developed by Chai Discovery. Released in October 2024, it can predict 3D structures for proteins, small molecules, DNA, RNA, and multi-component complexes containing any combination of these molecular types.

Chai-1 addresses a key limitation of earlier structure prediction tools by supporting truly unified multi-modal predictions. Rather than requiring separate models for different molecular types, Chai-1 handles everything in a single framework, including protein-ligand complexes, protein-DNA interactions, and systems with post-translational modifications.

ProteinIQ runs chai_lab 0.6.1. See the Chai source documentation for the model and native input formats.

How does Chai-1 work?

Chai-1 uses a diffusion-based architecture to generate molecular structures. Like image generation models, it starts from random noise and iteratively refines toward physically plausible 3D coordinates through learned denoising steps.

Diffusion-based structure generation

The model learns to reverse a diffusion process that gradually corrupts molecular structures with noise. During inference, it starts from random atomic coordinates and progressively denoises them through hundreds of steps to produce a coherent structure.

Each diffusion step adjusts atomic positions based on the model's learned understanding of molecular geometry, chemical bonding, and inter-molecular interactions. More diffusion steps generally produce higher-quality structures at the cost of increased computation time.

Evolutionary information from MSA

For protein structure prediction, Multiple Sequence Alignment (MSA) provides crucial evolutionary information. MSA aligns your target protein against thousands of homologous sequences from different organisms.

Coevolution patterns in the MSA reveal which residues are spatially close in 3D, if two positions consistently mutate together across evolution, they're likely in contact. This signal is one of the most powerful inputs for modern structure prediction.

Chai-1 can automatically generate MSAs via the ColabFold server, searching sequence databases to find homologs for each protein chain.

Recycling for iterative refinement

Chai-1 uses a recycling mechanism where the model feeds its prediction back through the network multiple times. Each recycling step allows the model to refine its prediction based on the previous iteration, progressively improving accuracy.

This is similar to how AlphaFold2 works, the model essentially "thinks twice" about difficult regions, using information from earlier predictions to resolve ambiguities.

Multi-modal representation

Unlike models trained exclusively on proteins, Chai-1 was designed from the ground up to handle diverse molecular types. Proteins, nucleic acids, and small molecules are all represented in a unified framework that captures their distinct chemical properties while enabling cross-modal interactions.

This unified architecture is what enables Chai-1 to predict complex multi-component systems like protein-ligand complexes, protein-DNA interactions, or systems with covalently attached modifications.

Inputs & settings

Molecule inputs

Chai-1 accepts multiple molecule types that you can combine into a single prediction job. Chain IDs (A, B, C...) are assigned automatically and displayed in the UI, you'll need these when defining constraints.

• Protein: Upload PDB/CIF files, paste FASTA sequences, or fetch directly from RCSB using a PDB ID. Up to 10 chains.
• Ligand (SMILES): Enter SMILES strings, upload SDF/MOL/MOL2 files, or fetch from PubChem by compound ID. Up to 10 ligands.
• DNA: Paste sequences in FASTA format or upload structure files. Up to 10 chains.
• RNA: Paste sequences in FASTA format or upload structure files. Up to 10 chains.

At least one protein, DNA or RNA chain is required. Chai supports up to 2,048 model tokens; modified residues and ligand atoms can consume more tokens than ordinary polymer residues. PDB/CIF molecule uploads supply the sequence from the first observed chain. Their coordinates are not used as template conditioning. Multiple FASTA records remain separate chains. Chain copies also become separate chains, in input order.

Native Chai FASTA

Select Native Chai FASTA to describe the complete complex in one file. Each record uses a unique type|name header, where type is protein, dna, rna, ligand or glycan. This mode supports Chai's modified polymer and glycan syntax. Do not add separate molecule inputs in this mode.

>protein|receptor
AC(SEP)EFGHIKLMNPQRSTVWY
>ligand|ethanol
CCO

Parenthesized CCD codes describe modified protein, DNA or RNA residues. Put modifications directly in this file instead of using Modified residues. Native names identify template queries; restraint chain IDs follow FASTA order: A, B, C and so on. The prepared FASTA and chain map are included with the downloads.

Prediction parameters

• Number of samples: Diffusion samples per trunk, an integer from 1 to 10. Default: 5.
• Recycling steps: Trunk refinement iterations, an integer from 1 to 10. Default: 3.
• Diffusion steps: Denoising steps, an integer from 50 to 500. Default: 200. Increasing this value increases runtime and does not guarantee better accuracy.
• Random seed: Nonnegative integer up to 9,007,199,254,740,991. Empty by default, which uses native random sampling. Chai's default GPU execution can produce different results even with the same seed and runtime; a fixed seed does not guarantee identical structures. See the Chai reproducibility report.
• Use ESM embeddings: Enabled by default, as in Chai's native API.
• Trunk samples: Independent trunk samples, an integer from 1 to 10. Default: 1. This value multiplied by Number of samples must be at most 10. Every resulting prediction is retained.
• MSA rows sampled per recycle: Nonnegative integer. Default: 0, which disables per-recycle subsampling.

MSA & templates

MSA (Multiple Sequence Alignment) provides evolutionary context that substantially improves protein structure prediction accuracy. Templates are known experimental structures from homologous proteins that can guide the prediction.

• Generate MSA: Disabled by default. Searches for homologs through Chai's public ColabFold endpoint when enabled. Submitted protein sequences are sent to that service. ProteinIQ does not accept a custom server URL.
• Use templates: Disabled by default. Requires Generate MSA because Chai performs template search through the same server request.
• Supplied MSAs: Upload one A3M or native aligned.pqt file per unique protein query sequence, with Generate MSA disabled. The query sequence selects the matching chain or identical chain copies. Native Parquet files retain their source databases, pairing keys, comments and bytes. Missing chains use Chai's single-sequence fallback.
• A3M source database: uniref90 by default; also accepts uniprot, bfd_uniclust and mgnify. This is a model feature, so choose the database that produced the alignment.
• A3M pairing keys: Defaults to native UniProt species pairing. Alternatives are species keys for all selected sources or no pairing keys. Chai's converter reads identifiers in the headers; it does not infer pairing from row order. Supply native aligned.pqt for arbitrary pairing keys or alignments combining databases.
• Custom template hits: Upload Chai's tab-separated .m8 table with template search disabled. Query IDs must match native FASTA names, or chain_A, chain_B and so on in molecule mode. Chai retrieves the referenced PDB templates and performs its native alignment. This input does not accept arbitrary local template CIF files. See the native template example.

Constraints

Constraints guide the prediction toward specific structural features when you have prior knowledge about the system. This is particularly useful when you know the binding site from experimental data or have crosslinking mass spectrometry data.

• Contact constraints: Guide two residues toward a maximum distance in angstroms. Format: chain:residue,chain:residue|max_distance. Example: A:10,B:5|8.0. Residue positions start at 1; the default maximum is 8 angstroms when omitted.
• Pocket constraints: Guide a chain toward specified residues. Format: ligand_chain|protein:res1,protein:res2|max_distance. Example: C|A:45,A:46|6.0. The default maximum is 6 angstroms when omitted.
• Native restraints: Upload Chai's CSV for contact, pocket or covalent restraints, including atom selectors such as C10@SG and @C1. Use the CSV or convenience controls, not both. Keep the native CSV columns: chainA,res_idxA,chainB,res_idxB,connection_type,confidence,min_distance_angstrom,max_distance_angstrom,comment,restraint_id. Chai 0.6.1 does not use the confidence and minimum-distance fields to weight these restraints. See its covalent bond examples.

Constraints guide inference and do not guarantee the requested geometry. Inspect the resulting structure. Malformed annotations and references to missing residues are rejected instead of omitted.

Modifications

Post-translational modifications (PTMs) can significantly affect protein structure. Chai-1 supports common modifications using standard CCD (Chemical Component Dictionary) codes from the PDB.

• Modified residues: Specifies modified residues using chain IDs shown in the molecule labels. Format: chain:position:CCD_code. Common codes: SEP (phosphoserine), TPO (phosphothreonine), PTR (phosphotyrosine), MSE (selenomethionine), MLY (methylated lysine).

Understanding the results

Chai-1 returns multiple ranked predictions, each with a structure file and associated confidence information.

Structure files

Each prediction produces a CIF structure file representing one possible conformation. We recommend examining the top 2-3 structures rather than relying solely on the highest-ranked prediction, especially for flexible complexes or multi-domain proteins.

Confidence interpretation

Predictions are sorted by Chai's aggregate confidence, highest first. Original model filenames and sample indices retain the native ordering. The table includes pTM, ipTM, per-chain pTM, the chain-pair ipTM matrix, and Chai's inter-chain clash indicators and counts.

Mean token pLDDT (0-1) is calculated from the returned per-token array. CIF atom B-factors contain Chai's per-atom pLDDT scaled to 0-100, so these are different summaries. Native score NPZ files, full PAE/PDE/pLDDT arrays, their sample index, generated MSAs, prepared inputs and complete run logs remain downloadable. PAE and PDE are in angstroms. Confidence is not a measurement of binding affinity or proof that a pose is correct.

If predictions differ substantially across samples, this may indicate:

• The target is intrinsically flexible
• The model is uncertain about the conformation
• MSA depth is insufficient for confident prediction

In such cases, consider enabling MSA generation or increasing the number of samples to better explore conformational space.

Pricing and runtime

Chai-1 costs 50 credits per minute on an A100 80 GB GPU. Runtime is prorated by the second and the final charge is rounded up to a whole credit. A 10-minute run costs 500 credits.

Credits are reserved in five-minute windows as the job runs. Set a spending limit before submission; the maximum is 3,000 credits for one hour. Unused reserved credits are returned. Failed or timed-out jobs are refunded. Recorded runtime from restarted containers counts toward the charge.

If the spending limit or available credits run out, prediction stops and the reserved runtime is charged. Stopping before prediction finishes may leave no completed structures. Metered runs require a paid plan.

Best practices

Getting good predictions

Enable MSA for proteins. The evolutionary information from MSA substantially improves prediction accuracy. Only disable it when working with synthetic proteins that lack natural homologs, or when you need a quick rough estimate.

Use template hits for structural conditioning. A PDB/CIF uploaded as a molecule supplies its sequence only. Use the template-search option or a native template-hit table when the prediction should use template information.

Check your SMILES stereochemistry. Ligand binding is highly stereospecific. Make sure your SMILES strings correctly represent the enantiomer you intend to model.

Choosing the right settings

When to use constraints

Pocket constraints are valuable when:

• You know the binding site from crystallography
• Mutagenesis data indicates specific binding residues
• You want to dock a ligand to a specific location

Contact constraints are useful for:

• Incorporating crosslinking mass spectrometry data
• Modeling known protein-protein interfaces
• Guiding specific domain orientations