lamm-mit/chai

Chai-1 Structure Prediction

Use when the user needs to predict molecular structures — proteins, nucleic acids, small molecules, or multi-chain complexes — using the Chai-1 foundation model. Supports both local GPU inference and the Chai Discovery API for remote execution.

Requirements

• Python 3.10+
• 16 GB GPU VRAM (A10G sufficient; A100 for large complexes)
• Or: use Chai Discovery API (no local GPU needed)

Installation

pip install chai-lab

Local Usage

Python API

from chai_lab.chai1 import run_inference
import torch
from pathlib import Path

# Single protein
results = run_inference(
    fasta_file=Path("input.fasta"),
    output_dir=Path("results/"),
    num_trunk_recycles=3,
    num_diffn_timesteps=200,
    seed=42,
    device=torch.device("cuda:0"),
    use_esm_embeddings=True,
)

# Access results
for i, result in enumerate(results):
    print(f"Model {i}: pTM={result.ptm:.3f}, ipTM={result.iptm:.3f}")

FASTA Input Format

# Single chain
>protein|A
MTEYKLVVVGAGGVGKSALTIQLIQNHFVDEYDPT

# Complex: separate chains with different headers
>protein|A
EVQLVESGGGLVQPGGSLRLSCAASGFTFSDYYMSWVRQAP
>protein|B
MTEYKLVVVGAGGVGKSALTIQLIQNHFVDE

# With small molecule (SMILES)
>protein|A
MTEYKLVVVGAGGVGKS...
>ligand|L
CC1=CC=C(C=C1)NC(=O)C2=CC=C(C=C2)CN3CCN(CC3)C

# RNA
>rna|R
GCGGAUUUAGCUCAGUUGGGAGAGCGCCAGACUGAAGAUCUGGAGGUCCUGUGUUCGAUCCACAGAAUUCGCACCA

Chai Discovery API (No Local GPU)

import requests

# Submit prediction job
response = requests.post(
    "https://api.chaidiscovery.com/v1/predictions",
    headers={"Authorization": f"Bearer {CHAI_API_KEY}"},
    json={
        "sequences": [
            {"type": "protein", "chain_id": "A", "sequence": "MTEYKLVV..."},
            {"type": "protein", "chain_id": "B", "sequence": "EVQLVES..."}
        ],
        "num_diffn_timesteps": 200,
        "num_trunk_recycles": 3,
    }
)
job_id = response.json()["job_id"]

# Poll for results
import time
while True:
    status = requests.get(
        f"https://api.chaidiscovery.com/v1/predictions/{job_id}",
        headers={"Authorization": f"Bearer {CHAI_API_KEY}"}
    ).json()
    if status["status"] == "completed":
        break
    time.sleep(30)

# Download structure
structure_url = status["results"]["structure_url"]

Output Files

| File | Contents | |------|----------| | pred.model_idx_0.cif | Top-ranked structure (CIF format) | | pred.model_idx_0.npz | Confidence arrays (pLDDT, PAE, pDE) | | scores.json | Aggregate scores per model |

Parsing Confidence Scores

import numpy as np

data = np.load("pred.model_idx_0.npz")
plddt = data["plddt"]                    # Per-residue, shape (N,)
pae = data["pae"]                        # N×N matrix, Angstroms
pde = data.get("pde")                    # Predicted Distance Error

# Interface residues (chain A = target, chain B = binder)
chain_a_len = 150  # length of chain A
interface_pae = pae[:chain_a_len, chain_a_len:].mean()
print(f"Interface PAE: {interface_pae:.2f} Å (< 10 = good)")

Chai vs. Other Predictors

| Feature | Chai-1 | Boltz | AF2 | |---------|--------|-------|-----| | Speed (complex) | Fast | Medium | Slow | | Small molecules | ✓ | ✓ | ✗ | | RNA/DNA | ✓ | ✓ | ✗ | | API available | ✓ | ✗ | ✗ | | Open weights | ✓ | ✓ | ✓ | | GPU VRAM | 16 GB | 24 GB | 32 GB |

Quality Thresholds

| Metric | Marginal | Good | Excellent | |--------|----------|------|-----------| | Mean pLDDT | <60 | 60–80 | >80 | | ipTM (complex) | <0.5 | 0.5–0.75 | >0.75 | | Interface PAE | >20 Å | 10–20 Å | <10 Å |

Use Cases

• Fast validation: Predicts binder-target complexes quickly before committing to expensive MD simulations.
• Ligand complexes: Predicts protein-small molecule binding poses from SMILES input.
• Ensemble scoring: Generates multiple models and ranks them by ipTM for design selection.
• Nucleic acid interactions: Predicts protein-DNA/RNA complex structures.