Homology modeling: principles and the pipeline

Comparative or homology modeling predicts a target protein's three dimensional structure from the experimentally solved structure of one or more homologs. It rests on one robust observation: over evolutionary time protein structure diverges far more slowly than sequence, so proteins that share a common ancestor almost always share a fold. The task therefore reduces to finding a solved homolog and copying its coordinates onto the target sequence as faithfully as possible.

Nearly every tool runs the same five stages in the same order: template search, target template alignment, backbone transfer for the conserved core, loop modeling for insertions and deletions, and side chain packing. MODELLER, SWISS-MODEL, and Rosetta differ in the algorithm inside each stage, not in the ordering of the stages themselves.

Template search

Template search finds which solved structures are detectably related to the target. BLAST and PSI-BLAST compare a sequence or a sequence profile against a database and are reliable when identity is high. HHsearch instead aligns a hidden Markov model built from the target's multiple sequence alignment against HMMs of known structures, which is far more sensitive for remote homologs; the usual recipe is to build a target profile with HHblits and then search a structural set such as PDB70.

# 1. Build a target profile: 3 HHblits iterations over a large sequence DB
hhblits -i target.fasta -d UniRef30_2023_02 -oa3m target.a3m -n 3 -cpu 4

# 2. Search the PDB70 template HMMs with that profile
hhsearch -i target.a3m -d pdb70 -o target.hhr -cpu 4

# 3. Read the ranked hit table
head -n 12 target.hhr

Query         sp|P0XXXX|TARGET
Match_columns 248
No_of_seqs    112 out of 1543
Neff          8.7
Searched_HMMs 34891
Date          Wed Jul  8 14:22:07 2026
Command       hhsearch -i target.a3m -d pdb70 -o target.hhr -cpu 4

 No Hit                              Prob E-value P-value  Score    SS Cols Query HMM  Template HMM
  1 3f7g_A Alpha/beta-hydrolase fold 99.9 8.1E-31 2.3E-35  214.6   0.0  231    4-241     2-233 (236)
  2 1uxz_B Carboxylesterase          99.6 3.4E-24 9.7E-29  182.1   0.0  224    9-240     1-227 (229)
  3 5and_A Putative lipase           97.2 2.6E-11 7.4E-16  118.3   0.0  205   18-238    12-221 (240)

Backbone, loops, side chains

For every target residue aligned to a template residue, backbone transfer copies the main chain atom coordinates directly from the template, reconstructing the structurally conserved core. Insertions and deletions in the alignment fall into loops, which have no template to copy and are rebuilt by loop modeling using fragment databases or spatial restraint satisfaction. Side chain packing then places the side chains, typically with backbone dependent rotamer libraries as in SCWRL4, retaining template rotamers at conserved positions where possible.

Alignment and the twilight zone

Target to template sequence identity is the strongest single predictor of model accuracy. Above roughly 50 percent identity, models usually fall within about 1 Angstrom backbone RMSD and errors concentrate in loops and side chains. Below roughly 30 percent identity you enter the twilight zone, where similarity no longer reliably signals homology and, crucially, a correct alignment becomes hard to obtain even when the fold is genuinely shared.

The alignment decides which template residue donates coordinates to each target residue, so alignment quality is the dominant error source in comparative modeling. A single register shift displaces every downstream residue by one position and corrupts the local fold, an error larger than anything loop or side chain modeling introduces. Spend your effort here: prefer profile probabilities over raw percent identity, and manually verify that conserved catalytic and structural motifs align correctly before building.