biotender-max/chai
> Structure prediction using Chai-1, a foundation model for molecular structure. (2) Validating designed binders, (3) Predicting protein-ligand complexes, (4) Using the Chai API for high-throughput prediction, (5) Need an alternative to AlphaFold2. For QC thresholds, use protein-qc. For AlphaFold2 prediction, use alphafold. For ESM-based analysis, use esm.
npx skills add https://github.com/BioTender-max/awesome-bio-agent-skills --skill chai| Requirement | Minimum | Recommended |
|-------------|---------|-------------|
| Python | 3.10+ | 3.11 |
| CUDA | 12.0+ | 12.1+ |
| GPU VRAM | 24GB | 40GB (A100) |
| RAM | 32GB | 64GB |
> First time? See Installation Guide to set up Modal and biomodals.
cd biomodals
modal run modal_chai1.py \
--input-faa complex.fasta \
--out-dir predictions/
GPU: A100 (40GB) | Timeout: 30min default
pip install chai_lab
python -c "
import chai_lab
from chai_lab.chai1 import run_inference
# Run prediction
run_inference(
fasta_file='complex.fasta',
output_dir='predictions/',
num_trunk_recycles=3
)
"
git clone https://github.com/chaidiscovery/chai-lab.git
cd chai-lab
pip install -e .
chai-lab predict \
--fasta complex.fasta \
--output predictions/
>binder
MKTAYIAKQRQISFVKSHFSRQLE...
>target
MVLSPADKTNVKAAWGKVGAHAGE...
>protein
MKTAYIAKQRQISFVKSHFSRQLE...
>ligand|smiles
CCO
>protein
MKTAYIAKQRQISFVKSHFSRQLE...
>dna
ATCGATCGATCG
| Parameter | Default | Range | Description |
|-----------|---------|-------|-------------|
| num_trunk_recycles | 3 | 1-10 | Recycles (more = better) |
| num_diffn_timesteps | 200 | 50-500 | Diffusion steps |
| seed | 0 | int | Random seed |
predictions/
├── pred.model_idx_0.cif # Best model (CIF format)
├── pred.model_idx_1.cif # Second model
├── scores.json # Confidence scores
├── pae.npy # PAE matrix
└── plddt.npy # pLDDT values
Note: Chai-1 outputs CIF format. Convert to PDB if needed:
from Bio.PDB import MMCIFParser, PDBIO
parser = MMCIFParser()
structure = parser.get_structure("pred", "pred.model_idx_0.cif")
io = PDBIO()
io.set_structure(structure)
io.save("pred.model_idx_0.pdb")
import numpy as np
import json
# Load scores
with open('predictions/scores.json') as f:
scores = json.load(f)
plddt = np.load('predictions/plddt.npy')
pae = np.load('predictions/pae.npy')
print(f"pLDDT: {plddt.mean():.3f}")
print(f"pTM: {scores['ptm']:.3f}")
print(f"ipTM: {scores.get('iptm', 'N/A')}")
# Predict complex with Chai
chai-lab predict --fasta binder_target.fasta --output val/
# Check ipTM > 0.5
scores = json.load(open('val/scores.json'))
if scores['iptm'] > 0.5:
print("Design passes validation")
# FASTA with SMILES
fasta = """
>protein
MKTA...
>ligand|smiles
CCO
"""
# Chai handles both protein and small molecules
# Multiple sequences
for fasta in sequences/*.fasta; do
chai-lab predict \
--fasta "$fasta" \
--output "predictions/$(basename $fasta .fasta)"
done
| Aspect | Chai-1 | AlphaFold2 |
|--------|--------|------------|
| MSA required | No | Yes |
| Small molecules | Yes | No |
| DNA/RNA | Yes | Limited |
| Speed | Faster | Slower |
| Accuracy | Comparable | Reference |
$ chai-lab predict --fasta complex.fasta --output predictions/
[INFO] Loading Chai-1 model...
[INFO] Running inference...
[INFO] Saved 5 models to predictions/
predictions/scores.json:
{
"ptm": 0.82,
"iptm": 0.71,
"ranking_score": 0.76
}
What good output looks like:
Should I use Chai?
│
├─ What are you predicting?
│ ├─ Protein-protein complex → Chai ✓ or ColabFold
│ ├─ Protein + small molecule → Chai ✓
│ ├─ Protein + DNA/RNA → Chai ✓
│ └─ Single protein only → Use ESMFold (faster)
│
├─ Need MSA?
│ ├─ No / want speed → Chai ✓
│ └─ Yes / want accuracy → ColabFold
│
└─ Priority?
├─ Highest accuracy → ColabFold with MSA
├─ Speed / no MSA → Chai ✓
└─ Ligand binding → Chai ✓
| Campaign Size | Time (A100) | Cost (Modal) | Notes |
|---------------|-------------|--------------|-------|
| 100 complexes | 30-60 min | ~$10 | Standard validation |
| 500 complexes | 2-4h | ~$45 | Large campaign |
| 1000 complexes | 5-8h | ~$90 | Comprehensive |
Per-complex: ~20-40s for typical binder-target complex.
find predictions -name "*.cif" | wc -l # Should match input count
Low pLDDT: Increase num_trunk_recycles
Low ipTM: Check chain order, interface region
OOM errors: Use A100-80GB or reduce batch
Slow prediction: Reduce num_diffn_timesteps
| Error | Cause | Fix |
|-------|-------|-----|
| RuntimeError: CUDA out of memory | Complex too large | Use A100-80GB or split prediction |
| KeyError: 'iptm' | Single chain predicted | Ensure FASTA has multiple chains |
| ValueError: invalid SMILES | Malformed ligand | Validate SMILES with RDKit |
| torch.cuda.OutOfMemoryError | GPU exhausted | Reduce num_diffn_timesteps to 100 |
Next: protein-qc for filtering and ranking.
Take biotender-max/chai from the repository into ~/.claude/skills for personal
use, or into .claude/skills inside a project.
The agent identifies a skill by the name field in its header. Two skills with the
same name cannot sit side by side — one of them will be ignored.
The instructions reference pip.
Without those the skill loads but fails at the first command.