Template Selection for Homology Modelling: A Guide

To choose a template for homology modelling, search the PDB with BLASTp (close homologs), HHpred (remote homologs) and SWISS-MODEL’s template search, then rank hits on sequence identity, query coverage, experimental quality (resolution and method), ligand state and oligomeric state. Above about 40% identity over a long alignment the choice is safe; at 20 to 35% identity, inspect carefully or compare with AlphaFold.

The template decides your model far more than the modelling program does. MODELLER, SWISS-MODEL and every other comparative modelling tool copy the backbone of the template into your target, so a poor template gives a poor model no matter how carefully you refine it afterwards. This guide shows how to find candidate templates, how to rank them, and how to tell when the best one is still not good enough. It was written by the StemSkills Lab team, who have spent more than ten years on sequence and structural bioinformatics, drug design and molecular modeling, and it sits in our computational biology skills roadmap between “learn sequence alignment” and “build your first model”.

Why does template choice matter so much in homology modelling?

Because structural accuracy falls as sequence identity falls, and the template sets the ceiling. In their classic analysis, Chothia and Lesk (EMBO J, 1986) fitted the main-chain RMSD of the conserved core between homologous proteins to the fraction of mutated residues H as delta = 0.40 e^1.87H Å. Plugging in their own formula:

• At 50% identity (H = 0.5), the shared core differs by about 1.0 Å.
• At 30% identity (H = 0.7), about 1.5 Å.
• At 20% identity (H = 0.8), about 1.8 Å, and the same paper notes that at roughly 20% identity the regions keeping the same fold may cover less than half of each molecule.

So a template at 50% identity gives you a model whose core is close to experimental quality, while a template at 20% identity may only be right for part of the chain. Loops, insertions and side chains are worse than the core in every case. That is why five minutes spent ranking templates saves days of refinement and simulation on a model that was never right.

What is the twilight zone, and what identity is “safe”?

The twilight zone is the identity range where sequence similarity stops being a reliable sign of shared structure. In Rost’s 1999 study (Protein Engineering 12:85-94) of more than a million alignments between proteins of known structure, alignments separated similar from dissimilar structures clearly above 40% identity for long alignments, and the signal blurred at 20 to 35%. Above a cut-off roughly matching 30% identity, 90% of pairs were homologous; below 25%, fewer than 10% were.

Two practical points follow:

• Identity must be read with alignment length. Rost showed that 10 similar residues in a 16-residue alignment (over 60%) still did not imply structural similarity. A 70% identity hit over 30 residues of a 400-residue protein is not a good template for the whole protein.
• Below about 30% identity, do not trust BLAST identity alone. Switch to profile methods (HHpred) and check secondary-structure agreement, or compare with an AlphaFold prediction.

How do you find candidate templates with BLASTp?

BLASTp against the PDB sequence database is the fastest way to find close homologs (roughly 30% identity and above). On the NCBI BLAST web page, paste your target sequence and set Database to “Protein Data Bank proteins (pdb)”. For a scripted search with BLAST+, download NCBI’s preformatted PDB database and run blastp with tabular output:

update_blastdb.pl --decompress pdbaa
blastp -query target.fasta -db pdbaa -evalue 1e-5 -max_target_seqs 50 \
  -outfmt "6 sseqid pident length qcovs evalue bitscore stitle" > hits.tsv

In that output, pident is percent identity over the aligned region, length is alignment length, and qcovs is query coverage per subject (how much of your target the hit covers). Sort on coverage and identity together, not identity alone. Option names are in the BLAST+ command-line user manual.

BLAST is quick and easy to read, but it is a pairwise sequence method. Once identity drops into the twilight zone it starts missing true homologs, which is where HHpred takes over.

How does HHpred find remote homologs that BLAST misses?

HHpred compares a profile hidden Markov model built from your sequence’s homologs against profile HMMs of PDB structures, so it detects relationships that pairwise identity hides. Run it on the MPI Bioinformatics Toolkit (described in Zimmermann et al., J Mol Biol 2018) with your sequence against the default PDB_mmCIF70 database.

Read the Probability column first, not the E-value. The HH-suite user guide calls probability a more sensitive measure than the E-value because it also accounts for secondary-structure similarity, and it states: “When it is larger than 95%, say, the homology is nearly certain.” The same guide suggests, as a rough rule, looking closely at any hit above 50% probability, or above 30% if it is among the top three hits.

A high-probability HHpred hit at 22% identity can be a far better template than a BLAST hit at 35% identity over a short fragment. HHpred also shows the predicted versus template secondary structure under each alignment; long stretches where a predicted helix aligns to a template strand are a warning sign.

How do you use SWISS-MODEL’s template search?

On SWISS-MODEL, paste your sequence and click Search For Templates instead of Build Model. According to the SWISS-MODEL documentation, the SWISS-MODEL Template Library (SMTL, derived from the PDB) is searched in parallel with BLAST and HHblits, which gives good alignments at both high and low identity. The method is described in Waterhouse et al., Nucleic Acids Research 2018.

The results table ranks templates by two scores you should understand:

• GMQE (Global Model Quality Estimate) combines properties of the target-template alignment and the template structure, using a neural network trained to predict the lDDT of the resulting model. Higher is better, on a 0 to 1 scale.
• QSQE (Quaternary Structure Quality Estimate), 0 to 1, estimates how accurate the interchain contacts would be for an oligomeric model. The documentation treats values above 0.7 as reliable enough to follow the template’s quaternary structure.

The same table shows identity, coverage, method and resolution, oligomeric state and bound ligands for each template, which makes it the easiest place to apply the criteria below. Our SWISS-MODEL homology modelling tutorial walks through building the model once you have chosen.

How do BLASTp, HHpred and SWISS-MODEL compare for template search?

A sensible workflow is to run SWISS-MODEL’s search and HHpred side by side. If they agree on the top template, you can proceed with confidence. If they disagree, the disagreement itself tells you the target needs closer inspection.

Which criteria should you use to rank templates?

Rank on these six criteria, roughly in this order. No single number decides it.

1. Homology confidence. HHpred probability near 100%, or BLAST identity well above 30% over a long alignment. If this fails, nothing else matters.
2. Query coverage. The template should cover the region you care about. For a single-domain question (an active site, a binding pocket), full-length coverage matters less than complete coverage of that domain.
3. Sequence identity. Higher identity means a closer core (see the Chothia and Lesk numbers above) and fewer alignment errors.
4. Experimental quality. SWISS-MODEL’s documentation says X-ray structures at high resolution are generally preferable and discourages averaged NMR structures. As a working rule, prefer crystal structures around 2.5 Å or better, and open the template’s wwPDB validation report to check R-free, clashscore and Ramachandran outliers. For cryo-EM templates, check the reported resolution and whether your region of interest is well resolved.
5. Ligand and conformational state. For docking, a template crystallised with a ligand in the same pocket (holo) usually gives a better-shaped site than an apo structure. Match the state you want to study: active or inactive kinase, open or closed loop.
6. Oligomeric state. If your protein works as a dimer, choose a template whose biological assembly is a dimer, and check QSQE in SWISS-MODEL. The SWISS-MODEL documentation warns when the predicted oligomeric state differs from the template’s biounit.

What does a template decision look like in practice?

Here is an illustrative case (made-up candidates to show the reasoning, not a real search). You want to dock inhibitors into a target kinase and find three templates:

A has the highest identity but misses half the domain, including part of the ATP pocket. C covers the domain but is an NMR structure without a ligand. B is the right choice: high enough identity for a reliable core, full coverage of the pocket, a good crystal structure, and a bound inhibitor that shapes the site you plan to dock into. Notice that the highest identity lost.

When should you use multiple templates or switch to AlphaFold?

Use multiple templates when no single structure covers your target, for example one template per domain, or one template for the core and another for a loop that is disordered in the first. MODELLER handles this directly; see our MODELLER homology modelling tutorial.

Switch to, or at least compare with, AlphaFold when your best template sits in the twilight zone (below roughly 30% identity), when HHpred probability is low, or when coverage is patchy. Check the AlphaFold Protein Structure Database for your UniProt accession before predicting anything yourself, and read its confidence scores with our guide to pLDDT and PAE. Homology modelling still has the edge when you need a specific state, such as a ligand-bound conformation from a holo template. Our comparison of AlphaFold, homology modelling and experimental structures covers that decision in detail.

What are the common template selection mistakes, and how do you fix them?

• Picking the top BLAST hit by identity alone. Fix: sort by coverage of your region of interest first, then identity.
• Ignoring the template’s missing residues. A loop that is unresolved in the template will be modelled from scratch. Fix: check REMARK 465 (PDB format) or the unobserved residues in the mmCIF file, and prefer a template where your key region is resolved.
• Modelling a monomer from a dimeric template without noticing. The interface residues end up exposed and can misbehave in MD. Fix: check the biological assembly on the RCSB entry page and QSQE in SWISS-MODEL.
• Trusting an alignment in the twilight zone without inspection. Fix: compare HHpred and SWISS-MODEL alignments; shifted gaps in helices or strands mean the alignment, not the template, needs correcting.
• Using an apo template for docking when a holo one exists. Fix: look for the same family crystallised with a ligand in your pocket, even at slightly lower identity.

After building, always validate and, if needed, refine the model; our guide on how to refine a homology model covers loop rebuilding and energy minimisation.

Frequently asked questions

What is the minimum sequence identity for homology modelling?

There is no single cut-off, but above about 40% identity over a long alignment the structural relationship is clear, and between 20 and 35% (the twilight zone) it becomes uncertain. Below roughly 30%, confirm homology with HHpred probability and compare against AlphaFold before trusting the model.

Is higher sequence identity always the better template?

No. A template with lower identity but full coverage of your region, better resolution and the right ligand or oligomeric state often gives a more useful model than a higher-identity template that covers only a fragment.

What GMQE value is good in SWISS-MODEL?

GMQE runs from 0 to 1 and higher means a more reliable expected model. Use it to compare templates for the same target rather than as an absolute pass mark, and check the global QMEANDisCo score of the built model as well.

Should I use BLAST or HHpred to find templates?

Use both. BLAST finds close homologs quickly and is easy to interpret; HHpred finds remote homologs that BLAST misses. If both point to the same template, your choice is well supported.

Can I use an NMR structure as a template?

You can, especially if it is the only option, but the SWISS-MODEL documentation discourages averaged NMR structures and generally prefers high-resolution X-ray structures. If you use an NMR ensemble, pick a representative model and validate the result carefully.

Template selection is the first real decision in any structure-based project, and it shapes the docking and MD work that follows. To see where it fits among the other skills, revisit the computational biology skills roadmap.