Homology Modelling Tutorial for Beginners: Build a Protein Structure with SWISS-MODEL (Step by Step)

Homology modelling builds a 3D structure for a protein that has no experimental structure by copying the fold of a related protein whose structure is already known. On the SWISS-MODEL web server you paste your sequence, choose a template with good sequence identity and coverage, build the model, then read the GMQE and QMEANDisCo scores to judge whether it is accurate enough to dock or simulate.

You have a protein sequence, your project needs a 3D structure, and a search of the Protein Data Bank turns up nothing for your exact protein. That is the normal case, not the exception. This tutorial walks an MSc student through building a model on SWISS-MODEL, a free web server that needs no install, and then reading its quality scores so you know whether the result is trustworthy. It sits inside the wider computational biology skills roadmap, where getting a structure is usually the gate before docking and molecular dynamics.

What is homology modelling, and when should you use it?

Homology modelling (also called comparative modelling) predicts the 3D structure of a target protein from its sequence by using an experimentally solved structure of a related protein, the template, as a scaffold. The core assumption is simple and well supported: proteins with similar sequences tend to fold into similar shapes, so if you can align your sequence to a known structure, you can borrow its backbone and rebuild the side chains and loops.

The reason this matters is the size of the gap between known sequences and known structures. The RCSB Protein Data Bank holds more than 200,000 experimentally determined structures, while UniProt catalogues hundreds of millions of protein sequences. For the vast majority of proteins you will meet in a project, no one has ever solved the structure. Homology modelling is one of the oldest and most reliable ways to close that gap when a suitable template exists.

The canonical SWISS-MODEL paper by Waterhouse and colleagues (2018), published in Nucleic Acids Research, opens by framing exactly this role:

“Homology modelling has matured into an important technique in structural biology, significantly contributing to narrowing the gap between known protein sequences and experimentally determined structures.”

Use homology modelling when a related structure exists and you want a fast, interpretable model tied to a specific experimental template, for example a particular ligand-bound or functional state. Later in this guide we compare it directly with AlphaFold and with MODELLER so you can pick the right tool for your case.

What do you need before you start?

You need three things, and only the first is mandatory: your protein’s amino acid sequence, a web browser, and a rough idea of which organism and function your protein has. There is nothing to install and no GPU to rent.

Start at UniProt, the standard reference for protein sequences. Search for your protein by name and organism, open the correct entry, and copy either the UniProt accession (for example, a code like P00533) or the FASTA sequence from the “Sequences” section. Confirm you have the right isoform and that the length matches the construct you actually care about. A wrong or truncated sequence is the most common reason a beginner’s model comes out useless, so spend a minute here before anything else.

How do you build a model in SWISS-MODEL, step by step?

The whole workflow happens in the browser and usually finishes in a few minutes. Here is the sequence of actions for a single protein chain.

1. Open the server and start a project. Go to swissmodel.expasy.org and click “Start Modelling”.
2. Enter your target. Paste the amino acid sequence (or the UniProt accession) into the “Target Sequence” box. Give the project a name so you can find it later.
3. Search for templates. Click “Search For Templates”. SWISS-MODEL scans its template library, which is derived from the PDB, using both a fast BLAST search and the more sensitive HHblits profile search, then returns a ranked list of candidate structures.
4. Choose a template. Read the template table (covered in detail below) and tick the box next to the best candidate. For a first model, pick one good template rather than several.
5. Build the model. Click “Build Models”. SWISS-MODEL aligns your sequence to the template, transfers the conserved backbone, rebuilds side chains and loops with its ProMod3 engine, and returns a results page.
6. Download and inspect. Download the model as a PDB file, then open it in a viewer such as PyMOL, ChimeraX, or Mol* to look at it before you trust it.

How do you choose a good template by sequence identity and coverage?

The template list shows several columns, but two decide most of the outcome: sequence identity (how much of your sequence matches the template) and coverage (how much of your sequence the template spans). Higher is better on both. Below roughly 30% sequence identity, alignments enter what Rost (1999) named the “twilight zone” in Protein Engineering, where alignment errors climb sharply and models become unreliable. Above about 50%, you can usually expect a good backbone.

Use this rough guide when you scan the candidates.

Also weigh the practical columns: prefer an X-ray template with high resolution (a lower number in angstroms is better), check that the template’s oligomeric state matches how your protein behaves, and prefer a template that covers your region of interest even if another has slightly higher identity elsewhere. If two templates each cover a different domain, model the domains separately.

How do you read GMQE and QMEANDisCo to judge the model?

SWISS-MODEL does not just hand you a structure. It hands you two estimates of how much to believe it, and reading them is the single most important skill in this tutorial. A pretty ribbon means nothing if the scores are poor.

GMQE (Global Model Quality Estimate) is reported before you build, next to each template, on a scale from 0 to 1. It predicts the expected quality of the final model by combining the target-template alignment with properties of the template. Higher is better. Use it to compare templates: a template with a higher GMQE will, on average, give a better model.

QMEANDisCo is reported after you build, and it scores the model you actually got. It is a composite quality estimate that adds distance constraints from homologous structures, described by Studer and colleagues (2020) in Bioinformatics. The global QMEANDisCo score runs from 0 to 1, higher is better, and per the SWISS-MODEL results interpretation, models scoring above roughly 0.6 are generally considered reliable. Just as useful is the per-residue local quality plot: it shows which parts of the model are trustworthy (typically the core) and which are not (often loops and the flexible N and C termini).

Practical rule for downstream work: if your binding site sits in a high-scoring region and the global QMEANDisCo is healthy, the model is a reasonable starting point for docking. If the active-site loops score poorly, fix or refine them first, or treat any docking result there as a hypothesis rather than a finding.

Homology modelling vs AlphaFold: which should you use?

This is the question every student asks now, and the honest answer is that they solve overlapping but different problems. For a single-domain protein with no obvious template, AlphaFold usually gives the stronger baseline, and our companion guide on how to predict a protein structure with AlphaFold covers that route. Homology modelling still wins in specific, common situations.

When a homology model still beats AlphaFold: when you need the structure in the conformation of a particular template, such as an inhibitor-bound or otherwise functional state that AlphaFold will not reproduce; when you are modelling a point mutant relative to a known wild-type structure and want the change measured against that exact scaffold; or when your docking study must match an experimental complex. Homology models are also transparent, because you know precisely which template and alignment produced them.

When to reach for MODELLER instead: SWISS-MODEL is the right first tool because it is automated and fast, but it trades away control. When you need to combine several templates, refine a difficult loop, script hundreds of models, or hand-edit the alignment, move to MODELLER, the command-line program by Šali and Blundell (1993) that builds models by satisfaction of spatial restraints. It has a steeper learning curve and expects Python, so treat it as your second step, not your first.

What are the common problems, and how do you fix them?

Most beginner failures fall into a handful of buckets. Here are the ones you are most likely to hit and what to do about each.

• No usable template found, or all below 30% identity. Your protein has no close relative in the PDB. Do not force a twilight-zone model. Switch to AlphaFold through ColabFold or check the AlphaFold Protein Structure Database first.
• Low coverage. The template only spans part of your sequence, so the model covers only that region. Split your protein into domains and model each against its best template, then study them separately.
• Good GMQE but poor QMEANDisCo. The expected quality and the delivered quality disagree, usually because of a shaky alignment or a template mismatch. Rebuild with a different template, or inspect the alignment on the results page for gaps in important regions.
• Low local scores at the active site. Loops near the binding site are unreliable. Refine those loops (this is where MODELLER helps) or clearly flag any docking there as tentative.
• Wrong oligomeric state. Your protein is a dimer but you modelled a monomer, or the reverse. Pick a template whose quaternary structure matches your protein so the interface is modelled too.

Frequently asked questions

Is SWISS-MODEL free to use?

Yes. SWISS-MODEL is a free academic web server hosted by the SIB Swiss Institute of Bioinformatics and the Biozentrum of the University of Basel. You can build models in a browser without an account, and creating a free account lets you save projects.

What sequence identity do I need for a reliable homology model?

As a rule of thumb, above 50% identity gives a reliable backbone, 30% to 50% gives a generally correct fold with uncertain loops, and below 30% is the twilight zone where the fold itself may be wrong. Always confirm quality with the QMEANDisCo score rather than identity alone.

Can I dock a ligand into a homology model?

Yes, if the model is good enough. Check that the binding site sits in a high-scoring region of the QMEANDisCo local plot and that global quality is healthy. Docking into low-confidence loops produces results you cannot trust, so validate the site before you commit to it.

Is homology modelling obsolete now that AlphaFold exists?

No. AlphaFold is often the stronger baseline for a template-free single domain, but homology modelling remains the better choice when you need a model in a specific ligand-bound or functional state, when modelling a mutant against a known structure, or when you want a fast model tied to a named experimental template.

How long does a SWISS-MODEL run take?

For a single chain the template search and model build usually finish within a few minutes. Larger proteins, complexes, or busy server periods can take longer, but you never wait hours as you might with a local install.

Written by the StemSkills Lab team, structural and computational scientists with more than 10 years of combined experience in sequence and structural bioinformatics, drug discovery and design, and multiscale molecular modelling.