Homology Modeling — Protein Structure Prediction

MODELLER 10.x · ColabFold · ESMFold · Biopython · pdbfixer · ProDy. For building protein 3D models from sequence when no experimental structure is available.

When to Use This Skill

• No X-ray/CryoEM structure for your target protein (or coverage is partial)
• Building a receptor model for docking or MD when AlphaFold DB lacks your variant/mutant
• Constructing chimeric or engineered proteins not in existing databases
• Validating or improving an AI-predicted structure with experimental template data
• Generating a starting conformation for free-energy calculations (→ force-fields skill)

Decision Tree — Which Method to Use

Target sequence available?
  NO → retrieve from UniProt / NCBI first

Do you have a homologous template (sequence identity > 25%)?
  YES + identity > 50%  → MODELLER (comparative, references/modeller-basics.md)
  YES + identity 25–50% → MODELLER multi-template or AlphaFold2 with template
  NO / < 25%             → AlphaFold2 / ColabFold (references/alphafold-esm.md)

Throughput?
  Single target           → ColabFold interactive / MODELLER script
  Batch (>10 proteins)    → ColabFold batch CLI or ESMFold API
  No MSA / fast screen    → ESMFold (references/alphafold-esm.md)

After modeling:
  → Validate model        → references/structure-quality.md
  → Prepare for MD        → references/structure-prep.md
  → Prepare for docking   → references/structure-prep.md + docking skill

Quick Start

# --- Option A: MODELLER comparative modeling (single template) ---
from modeller import Environ
from modeller.automodel import AutoModel

env = Environ()
env.io.atom_files_directory = ['.', '../templates']

a = AutoModel(env,
              alnfile  = 'alignment.pir',   # PIR format — see modeller-basics.md
              knowns   = '5HT2A_template',  # template code (PDB ID, no extension)
              sequence = 'TARGET_SEQ')      # target sequence ID in .pir file

a.starting_model = 1
a.ending_model   = 5    # generate 5 models, pick best by DOPE score

a.make()

# Select best model
results = [(m.molpdf, m.name) for m in a.outputs
           if m['failure'] is None]
results.sort()
print(f"Best model: {results[0][1]}  DOPE: {results[0][0]:.1f}")

# --- Option B: ColabFold (AlphaFold2 engine, local CLI) ---
colabfold_batch target.fasta output_dir/ \
    --num-models 5 \
    --num-recycle 3 \
    --amber \
    --use-gpu-relax

# Best model: output_dir/target_relaxed_rank_001_*.pdb
# Scores:     output_dir/target_scores_rank_001_*.json

# --- Option C: ESMFold (single-sequence, no MSA, fastest) ---
import torch, esm

model = esm.pretrained.esmfold_v1()
model = model.eval().cuda()

sequence = "MKTAYIAKQRQISFVKSHFSRQ..."   # full amino acid sequence

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

with open("esmfold_model.pdb", "w") as f:
    f.write(output)
print("Model saved to esmfold_model.pdb")

Router — What to Read

Task	Reference
MODELLER automodel, PIR format, loop refinement, multi-template, DOPE ranking	references/modeller-basics.md
ColabFold CLI, AlphaFold2 output parsing, ESMFold API, pLDDT/PAE interpretation	references/alphafold-esm.md
DOPE scores, Ramachandran analysis, MolProbity, ProDy, RMSD to experiment	references/structure-quality.md
pdbfixer, propka3, disulfide bonds, protonation states, ACE/NME capping	references/structure-prep.md
HHblits/HHpred template search, BLAST, Biopython alignments, multi-template selection	references/template-search.md

Method Comparison

Method	Best for	Seq. ID required	Speed	Accuracy
MODELLER (automodel)	Close homologs, custom restraints	> 30%	Medium	★★★★ (with good template)
MODELLER (multi-template)	Coverage gaps, divergent regions	> 25%	Medium	★★★★
ColabFold / AlphaFold2	Any target, captures remote homologs	None	Slow (GPU)	★★★★★
ESMFold	Fast screen, no MSA, single sequence	None	Fast (GPU)	★★★

Key Tools

Tool	Install	Role
modeller	conda install -c salilab modeller (requires license key)	Comparative modeling
colabfold	pip install colabfold[alphafold] or conda	AF2-based prediction
esm	pip install fair-esm	ESMFold single-sequence prediction
biopython	pip install biopython	PDB I/O, BLAST, alignments, Ramachandran
pdbfixer	conda install -c conda-forge pdbfixer	Missing residues, H addition
propka	pip install propka	pKa prediction, protonation states
prody	pip install prody	Structural analysis, NMA, chain alignment

Installation

# MODELLER (requires free academic license from https://salilab.org/modeller/)
conda install -c salilab modeller
# Set MODELLER license key:
export KEY_MODELLER="XXXXXXXX"  # add to ~/.bashrc

# ColabFold (local, GPU recommended)
pip install "colabfold[alphafold]"
# OR via conda (recommended for reproducibility):
conda install -c conda-forge -c bioconda colabfold

# ESMFold
pip install fair-esm
# ESMFold also requires torch >= 2.0 and ~15 GB VRAM for full model

# Biopython + ProDy
pip install biopython prody

# pdbfixer + propka (structure prep)
conda install -c conda-forge pdbfixer
pip install propka

# Verify
python -c "from modeller import Environ; print('MODELLER OK')"
colabfold_batch --help
python -c "import esm; print('ESM OK')"

Related Skills

• docking → use homology model as receptor for virtual screening (check pLDDT > 80 in pocket)
• force-fields → MD simulation of the built model (OpenMM, AMBER, GROMACS)
• mdanalysis → trajectory analysis after MD equilibration of the model
• qm-dft → QM refinement of active-site geometry (xTB/ORCA)
• free-energy → FEP/TI relative binding free energies using model receptor
• ase → QM/MM or GFN2-xTB optimization of small binding-site models
• PDB: use PDB MCP or pdb_database skill → download template PDB