Chai1

Chai-1: Biomolecular Structure Prediction

Overview

Chai-1 is a multimodal structure prediction model that handles proteins, small molecules, RNA, DNA, and modifications in a single unified framework. It is a direct competitor to AlphaFold 3 and Boltz-2.

Key differentiators:

• Apache 2.0 license (commercial use explicitly permitted, including drug discovery)
• Simple pip install + FASTA-like input format
• Returns PAE, PDE, pLDDT, pTM, ipTM confidence metrics
• Optional MSA server integration (ColabFold MMseqs2)
• Supports templates, restraints, and covalent bonds

Supported entity types:

• protein — amino acid sequences
• ligand — SMILES-encoded small molecules
• rna — RNA sequences
• dna — DNA sequences
• Modified residues (e.g., phosphoserine: AAA(SEP)AAA)

Installation

pip install chai_lab==0.6.1

Requirements: Python ≥ 3.10, Linux, CUDA GPU with bfloat16 support.

Recommended GPUs: A100 (80GB), H100 (80GB), L40S (48GB). Also works on A10, A30, RTX 4090 for smaller complexes.

Model weights download automatically on first run to ~/.chai/ (or $CHAI_DOWNLOADS_DIR).

Input Format

Chai-1 uses a FASTA-like format with entity type headers:

>protein|name=receptor
AGSHSMRYFSTSVSRPGRGEPRFIAVGYVDDTQFVRFDSDAA...

>protein|name=peptide
GAAL

>ligand|name=inhibitor
CC(=O)Nc1ccc(O)cc1

>rna|name=guide_rna
AUGCUAGCUAGC

>dna|name=template
ATGCTAGCTAG

• Each entity needs a unique name= identifier
• All entities in one file form a complex
• Ligands use SMILES notation
• Modified residues use parenthetical notation: AAA(SEP)AAA (phosphoserine at position 4)

Core Usage

Python API

from pathlib import Path
from chai_lab.chai1 import run_inference

candidates = run_inference(
    fasta_file=Path("input.fasta"),
    output_dir=Path("output/"),
    num_trunk_recycles=3,
    num_diffn_timesteps=200,
    num_diffn_samples=5,
    seed=42,
)

# Access results
for i, (cif_path, ranking) in enumerate(zip(candidates.cif_paths, candidates.ranking_data)):
    score = ranking.aggregate_score.item()
    print(f"Sample {i}: {cif_path}  aggregate_score={score:.3f}")

# Confidence tensors
plddt = candidates.plddt   # (num_samples, num_tokens)
pae   = candidates.pae     # (num_samples, num_tokens, num_tokens)
pde   = candidates.pde     # (num_samples, num_tokens, num_tokens)

CLI

# Basic prediction (no MSA)
chai-lab fold input.fasta output_folder/

# With MSA server (more accurate)
chai-lab fold --use-msa-server input.fasta output_folder/

# With MSA + templates
chai-lab fold --use-msa-server --use-templates-server input.fasta output_folder/

# Convert a3m MSA to chai format
chai-lab a3m-to-pqt input_a3m_dir/ output.pqt

Key Parameters

Parameter	Default	Description
num_trunk_recycles	3	Recycling iterations (more = slower, better)
num_diffn_timesteps	200	Diffusion denoising steps
num_diffn_samples	5	Structures generated per trunk sample
num_trunk_samples	1	Independent trunk sampling runs
use_esm_embeddings	True	Use ESM protein language model embeddings
use_msa_server	False	Query ColabFold MMseqs2 server for MSA
msa_server_url	https://api.colabfold.com	MSA server endpoint
use_templates_server	False	Fetch PDB templates from server
low_memory	True	Offload tensors to CPU between operations
seed	None	Random seed for reproducibility
device	None	Torch device (default: cuda:0)
fasta_names_as_cif_chains	False	Use entity names as mmCIF chain IDs

Output Format

output_folder/
├── pred.model_idx_0.cif          ← best structure (mmCIF)
├── pred.model_idx_1.cif
├── ...
├── scores.model_idx_0.npz        ← confidence arrays
└── msa_coverage.png              ← MSA depth plot (if MSA used)

Scores NPZ arrays (per candidate):

• plddt — per-token pLDDT (0–100)
• pae — predicted aligned error matrix (N×N, Å)
• pde — predicted distance error matrix (N×N, Å)

SampleRanking fields (from candidates.ranking_data[i]):

• aggregate_score — primary ranking metric (higher = better)
• ptm_scores.complex_ptm — global fold quality (0–1)
• ptm_scores.interface_ptm — interface confidence (0–1)
• plddt_scores — per-chain pLDDT

Confidence Metrics

Metric	Range	Interpretation
aggregate_score	0–1	Primary ranking; combine pTM + clash + pLDDT
complex_ptm	0–1	Overall fold quality (>0.5 = good)
interface_ptm (ipTM)	0–1	Interface confidence (>0.6 = confident)
plddt per-residue	0–100	Local confidence (>70 = reliable)
PAE (Å)	0–31	Position error; low = confident relative placement

For complexes, filter by interface_ptm > 0.6 as the primary criterion.

MSA Handling

By default, Chai-1 runs without MSA (fast but less accurate). For best results:

# Option 1: use server (requires internet)
candidates = run_inference(
    fasta_file=Path("input.fasta"),
    output_dir=Path("output/"),
    use_msa_server=True,
)

# Option 2: precomputed MSA directory
candidates = run_inference(
    fasta_file=Path("input.fasta"),
    output_dir=Path("output/"),
    msa_directory=Path("msas/"),   # contains aligned.pqt files
)

Scripts

• scripts/predict.py — predict structures from FASTA input; see scripts/predict.py --help

Resources

• GitHub: https://github.com/chaidiscovery/chai-lab
• Paper: Chai Discovery, bioRxiv 2024 — https://doi.org/10.1101/2024.10.10.615955
• Web app: https://lab.chaidiscovery.com (no local setup required)

References

• references/api-reference.md — full Python API reference, confidence metric details, restraints, covalent bonds, template handling