Modern Structure Prediction

Predict protein structures using state-of-the-art machine learning models. This covers cloud APIs, local installations, and interpretation of results.

Model Comparison

Model	Complexes	Ligands	Speed	Access
AlphaFold3	Yes	Yes	Slow	Server only (2025)
ESMFold	No	No	Fast	API or local
Chai-1	Yes	Yes	Moderate	Local or API
Boltz-1	Yes	Yes	Moderate	Local
ColabFold	No*	No	Moderate	Colab/local

*ColabFold can predict complexes with AlphaFold-Multimer.

ESMFold (Fastest Single-Chain)

Via ESM Atlas API

import requests

def predict_esmfold(sequence):
    '''Predict structure using ESMFold API'''
    url = 'https://api.esmatlas.com/foldSequence/v1/pdb/'
    response = requests.post(url, data=sequence, timeout=300)
    if response.status_code == 200:
        return response.text
    raise Exception(f'ESMFold failed: {response.status_code}')

sequence = 'MVLSPADKTNVKAAWGKVGAHAGEYGAEALERMFLSFPTTKTYFPHFDLSH'
pdb_text = predict_esmfold(sequence)
with open('predicted.pdb', 'w') as f:
    f.write(pdb_text)

Local ESMFold

import torch
import esm

def predict_esmfold_local(sequence, device='cuda'):
    '''Run ESMFold locally (requires ~16GB GPU memory)'''
    model = esm.pretrained.esmfold_v1()
    model = model.eval().to(device)

    with torch.no_grad():
        output = model.infer_pdb(sequence)
    return output

# Extract pLDDT from ESMFold output
def extract_esmfold_plddt(pdb_text):
    plddt = {}
    for line in pdb_text.split('\n'):
        if line.startswith('ATOM') and line[12:16].strip() == 'CA':
            resnum = int(line[22:26])
            bfactor = float(line[60:66])
            plddt[resnum] = bfactor
    return plddt

AlphaFold3 (Server)

AlphaFold3 predictions via the server at alphafoldserver.com.

Prepare Input JSON

import json

def create_af3_input(sequences, job_name='prediction'):
    '''Create AlphaFold3 server input JSON'''
    entities = []
    for i, seq in enumerate(sequences):
        entities.append({
            'type': 'protein',
            'sequence': seq,
            'count': 1
        })

    job = {
        'name': job_name,
        'modelSeeds': [1],
        'sequences': entities
    }
    return json.dumps(job, indent=2)

# Single protein
input_json = create_af3_input(['MVLSPADKTNVKAAWGKVGAHAGEYGAEALERMFLSFPTTKTYFPHFDLSH'])

# Protein complex
input_json = create_af3_input([
    'MVLSPADKTNVKAAWGKVGAHAGEYGAEALERMFLSFPTTKTYFPHFDLSH',
    'MGHFTEEDKATITSLWGKVNVEDAGGETLGRLLVVYPWTQRFFDSFGNLSS'
])

Process AF3 Results

import json
from Bio.PDB import PDBParser
import numpy as np

def analyze_af3_result(result_dir):
    '''Analyze AlphaFold3 prediction results'''
    # Load summary
    with open(f'{result_dir}/summary_confidences.json') as f:
        summary = json.load(f)

    # Extract confidence metrics
    iptm = summary.get('iptm', None)  # Interface pTM (complexes)
    ptm = summary.get('ptm', None)    # Predicted TM-score
    ranking = summary.get('ranking_score', None)

    print(f'pTM: {ptm:.3f}' if ptm else 'pTM: N/A')
    print(f'ipTM: {iptm:.3f}' if iptm else 'ipTM: N/A')

    return summary

AF3 Confidence Interpretation

Metric	Range	Interpretation
pTM	0-1	Overall structure confidence
ipTM	0-1	Interface prediction quality
pLDDT	0-100	Per-residue confidence
PAE	0-30A	Position error between residue pairs

Chai-1 (Local Open-Source)

Installation

pip install chai-lab

Basic Prediction

from chai_lab.chai1 import run_inference
import numpy as np
from pathlib import Path

def predict_chai1(fasta_path, output_dir='chai_output'):
    '''Run Chai-1 structure prediction'''
    Path(output_dir).mkdir(exist_ok=True)

    candidates = run_inference(
        fasta_file=Path(fasta_path),
        output_dir=Path(output_dir),
        num_trunk_recycles=3,        # 3: Standard. Use 5+ for difficult targets.
        num_diffn_timesteps=200,     # 200: Standard. 500 for higher quality.
        seed=42,
        device='cuda:0'
    )
    return candidates

# Candidates are sorted by confidence
# candidates.cif files contain predicted structures

Chai-1 with Ligands

# Chai-1 supports protein-ligand complexes
# Include ligand SMILES in input FASTA with special format

def create_chai_fasta_with_ligand(protein_seq, ligand_smiles, output_file):
    '''Create Chai-1 input with protein and ligand'''
    with open(output_file, 'w') as f:
        f.write('>protein|chain_A\n')
        f.write(f'{protein_seq}\n')
        f.write('>ligand|chain_B\n')
        f.write(f'{ligand_smiles}\n')

Boltz-1 (Open-Source Complex Prediction)

Installation

pip install boltz

Basic Prediction

from boltz import Boltz1

def predict_boltz1(sequences, output_dir='boltz_output'):
    '''Run Boltz-1 structure prediction'''
    model = Boltz1()

    result = model.predict(
        sequences=sequences,
        output_dir=output_dir,
        recycling_steps=3,   # 3: Standard. Increase for difficult targets.
        sampling_steps=200   # 200: Standard. 500 for publication quality.
    )
    return result

Boltz-1 for Complexes

# Boltz-1 handles heteromeric complexes
def predict_complex_boltz(chain_sequences):
    '''Predict protein complex with Boltz-1'''
    model = Boltz1()

    result = model.predict(
        sequences=chain_sequences,  # List of sequences for each chain
        output_dir='complex_output'
    )

    # Extract interface metrics
    return result

ColabFold (AlphaFold2 + MMseqs2)

Command Line

# Install ColabFold
pip install colabfold

# Run prediction
colabfold_batch input.fasta output_dir/

# With custom templates
colabfold_batch input.fasta output_dir/ --templates

# For complexes (use : to separate chains)
# Create FASTA like: >complex\nSEQUENCE1:SEQUENCE2

Python API

from colabfold.batch import run_colabfold

def predict_colabfold(fasta_file, output_dir, use_templates=False):
    '''Run ColabFold prediction'''
    run_colabfold(
        input_path=fasta_file,
        result_dir=output_dir,
        use_templates=use_templates,
        num_models=5,           # 5: Standard. Use 1 for quick predictions.
        num_recycles=3,         # 3: Standard. Increase for multimers.
        model_order=[1,2,3,4,5]
    )

Comparing Predictions

from Bio.PDB import PDBParser, Superimposer
import numpy as np

def compare_predictions(pdb_files, labels=None):
    '''Compare multiple structure predictions'''
    parser = PDBParser(QUIET=True)
    structures = [parser.get_structure(f'model_{i}', f) for i, f in enumerate(pdb_files)]

    # Extract CA atoms from first chain
    def get_ca_atoms(struct):
        return [r['CA'] for r in struct[0].get_residues() if 'CA' in r]

    all_atoms = [get_ca_atoms(s) for s in structures]

    # Pairwise RMSD
    n = len(structures)
    rmsd_matrix = np.zeros((n, n))

    for i in range(n):
        for j in range(i+1, n):
            min_len = min(len(all_atoms[i]), len(all_atoms[j]))
            super_imposer = Superimposer()
            super_imposer.set_atoms(all_atoms[i][:min_len], all_atoms[j][:min_len])
            rmsd_matrix[i,j] = rmsd_matrix[j,i] = super_imposer.rms

    return rmsd_matrix

# Compare ESMFold vs AlphaFold3 vs Chai-1
rmsd = compare_predictions(['esmfold.pdb', 'af3.pdb', 'chai1.pdb'])
print('RMSD matrix:')
print(rmsd)

When to Use Each Model

Scenario	Recommended Model
Quick single-chain prediction	ESMFold (API)
Highest accuracy single chain	AlphaFold3 or ColabFold
Protein-protein complex	AlphaFold3, Chai-1, or Boltz-1
Protein-ligand complex	AlphaFold3 or Chai-1
No GPU available	ESMFold API or AlphaFold3 server
Large-scale screening	ESMFold (local)
Open-source requirement	Chai-1 or Boltz-1

Memory Requirements

Model	GPU Memory	Notes
ESMFold	~16 GB	Sequence length dependent
ColabFold	~8-16 GB	Model size dependent
Chai-1	~24 GB	Complex size dependent
Boltz-1	~24 GB	Complex size dependent

Related Skills

• alphafold-predictions - Download pre-computed AlphaFold structures
• structure-io - Parse and write structure files
• geometric-analysis - RMSD, superimposition, distance calculations
• structure-navigation - Navigate predicted structure hierarchy