alphafold

AlphaFold2 Structure Prediction

Use when the agent needs to run AlphaFold2 for protein structure prediction and complex modeling. Covers validating designed sequences, predicting binder-target complexes, and calculating confidence metrics (pLDDT, pTM, ipTM).

Distinct from alphafold-database (which retrieves pre-computed structures) — this skill covers running AF2 predictions on custom sequences.

Requirements

• Python 3.8+
• CUDA 11.0+, 32 GB GPU VRAM minimum (A100 recommended)
• For multimers: ColabFold recommended over local install

Deployment Options

1. ColabFold (Recommended for Multimers)

pip install colabfold[alphafold]

# Single chain
colabfold_batch input.fasta output_dir/ \
    --model-type alphafold2_ptm \
    --num-recycles 3

# Complex (multimer) — comma-separate chains in FASTA header
# >complex:ChainA,ChainB
colabfold_batch complex.fasta output_dir/ \
    --model-type alphafold2_multimer_v3 \
    --num-recycles 20 \
    --num-models 5

2. LocalColabFold

# Install
wget https://raw.githubusercontent.com/YoshitakaMo/localcolabfold/main/install_colabbatch_linux.sh
bash install_colabbatch_linux.sh

# Run offline
colabfold_batch sequences.fasta results/ \
    --model-type alphafold2_multimer_v3 \
    --num-recycles 3 \
    --use-gpu-relax

3. OpenFold (PyTorch reimplementation)

pip install openfold
python run_pretrained_openfold.py \
    --fasta_paths input.fasta \
    --output_dir results/ \
    --model_device cuda:0

Key Parameters

Confidence Metrics

import numpy as np
import json

# Load result JSON
with open("result_model_1.json") as f:
    result = json.load(f)

plddt = np.array(result["plddt"])           # Per-residue confidence 0-100
ptm = result["ptm"]                          # Global TM-score estimate 0-1
iptm = result.get("iptm", None)             # Interface TM-score (multimer only)
pae = np.array(result.get("pae", []))       # Predicted Aligned Error matrix

# Quality thresholds
print(f"Mean pLDDT: {plddt.mean():.1f}")    # >70 = good, >90 = excellent
print(f"pTM: {ptm:.3f}")                    # >0.5 = confident fold
if iptm:
    print(f"ipTM: {iptm:.3f}")              # >0.6 = reliable complex, >0.8 = high confidence

Self-Consistency Validation for Designed Sequences

# Design → predict → measure similarity to input backbone
# 1. Generate sequences with ProteinMPNN
# 2. Predict structure of each sequence with AF2
# 3. Calculate TM-score / RMSD vs. design backbone

python3 -c "
from Bio.PDB import PDBParser, Superimposer
# Compare predicted vs. designed structure
# High TM-score (>0.8) = sequence encodes target fold
"

Output Files

Quality Thresholds

Common Issues