AlphaFold vs Homology Modelling vs Experimental Structures: Which to Use, When

Check the RCSB Protein Data Bank first: an experimental structure covering your region of interest at good resolution is the safest starting point. If none exists, build a homology model with SWISS-MODEL when a template above roughly 30% sequence identity is available, and use AlphaFold when no usable template exists. Judge every option on coverage, confidence and fitness for your next step.

You have a protein sequence and a plan to run docking or a molecular dynamics simulation. Before any of that, you need coordinates: a three-dimensional structure to start from. There are three ways to get one, and picking the wrong source quietly weakens everything downstream. This guide is a decision framework for an MSc student, so you choose a defensible structure instead of grabbing whatever appears first in a search. It sits inside the wider computational biology skills roadmap, one step before you validate the structure and prepare it for a pipeline.

What are the three sources of a protein structure?

Every starting structure comes from one of three routes, and they differ in what they are: measured, borrowed, or predicted.

• Experimental structures are determined by X-ray crystallography, cryo-electron microscopy, or NMR, then deposited in the RCSB Protein Data Bank. They are real measurements of a real molecule, and the PDB holds more than 200,000 of them. This is your first stop.
• Homology models are built by copying the fold of a related protein whose structure is already known (the template), then threading your sequence onto it. Tools such as SWISS-MODEL automate this. The quality depends almost entirely on how similar the template is.
• AlphaFold predictions come from a deep-learning model that predicts a structure directly from sequence, with no template required. Predictions for over 200 million proteins are available through the AlphaFold Protein Structure Database, and you can generate your own with ColabFold.

The rest of this article is how to choose between them, and how to read the confidence signal each one gives you.

Should you always start with an experimental structure?

Yes, look there first, but do not accept the first hit blindly. Search the PDB by your protein name, gene, or UniProt accession and check three things about any candidate.

Does it cover the region you care about? A structure may be deposited for only one domain, or it may have missing loops exactly where your binding site sits. Compare the modelled residues against the full sequence before trusting it.

What is the resolution? For an X-ray structure, resolution is the single most useful quality number. As a working guide, structures better than about 2.0 Angstrom are high quality with well-defined side chains, structures between 2.0 and 2.5 Angstrom are usually fine for docking, and anything worse than roughly 3.0 Angstrom means side-chain positions are uncertain. Also read the R-free value, which reports how well the model fits data it was not fitted against. Cryo-EM structures report a resolution too, and the same logic applies: higher resolution means more reliable atomic detail.

What state was it captured in? An apo (ligand-free) structure and a holo (ligand-bound) structure of the same protein can differ at the active site because binding reshapes the pocket. If you are docking, a holo structure of a similar complex often gives a more realistic pocket than an apo one.

When a good experimental structure covers your region, use it. You are starting from a measurement, not a prediction, and that is the strongest possible foundation for docking or MD.

When should you use homology modelling instead?

Use homology modelling when there is no experimental structure of your protein, but there is a solved structure of a close relative to act as a template. The method rests on a well-established observation in structural biology: proteins with similar sequences fold into similar shapes, so a good template transfers its fold reliably.

The deciding number is the sequence identity between your target and the best available template. A common rule of thumb is that above about 30% identity you can expect a usable model, and the higher the identity climbs, the more accurate the model becomes, especially in the core. Below roughly 20 to 30% you enter what Burkhard Rost named the “twilight zone”, where an alignment can look plausible yet place residues wrongly, and models become unreliable. SWISS-MODEL reports its own quality estimates, the GMQE and the per-model QMEANDisCo score, so you are not left guessing. The SWISS-MODEL workspace is described by Waterhouse and colleagues in their 2018 paper in Nucleic Acids Research, and our own SWISS-MODEL tutorial walks through a build end to end.

The catch is that a homology model is only as good as its template near the region you care about. Insertions and deletions relative to the template, and loops that differ between the two proteins, are modelled with far less confidence than the conserved core, so treat those regions cautiously.

When is AlphaFold the right choice?

Reach for AlphaFold when there is no experimental structure and no template close enough for homology modelling. AlphaFold predicts a structure from sequence alone, and its accuracy changed the field. In the CASP14 blind assessment its predictions reached a median domain GDT_TS of 92.4, and a median backbone accuracy of 0.96 Angstrom root-mean-square deviation, which its authors reported as “competitive with experimental structures in a majority of cases”. The method is set out by Jumper and colleagues in their 2021 Nature paper, and for many human proteins a ready-made model already exists in the AlphaFold database, so you may not need to run anything. Our AlphaFold walkthrough covers generating your own with ColabFold.

Two cautions matter. First, AlphaFold is superb at the folded core but weak at intrinsically disordered regions, which it draws as low-confidence spaghetti because they have no single structure to predict. Second, a standard AlphaFold model is apo and ligand-free, and it does not model bound cofactors, metals, or the conformational change a ligand induces, so a pocket may be modelled in a resting state that is not the one you want to dock into.

How do you read the confidence of each option?

Every source gives you a different confidence signal, and comparing them fairly is the heart of a good decision. Do not compare a resolution number directly to a pLDDT score. They measure different things.

For AlphaFold, the per-residue confidence is the pLDDT, reported on a 0 to 100 scale. Per the EBI AlphaFold documentation, a pLDDT above 90 is very high confidence, 70 to 90 is confident with a generally good backbone, 50 to 70 is low and should be treated with caution, and below 50 usually marks a disordered or unreliable region. There is a second metric, the Predicted Aligned Error (PAE), which tells you whether the relative position of two domains is trustworthy: a low PAE between domains means their arrangement is confident, a high PAE means you should not read anything into how they are placed against each other.

Whatever you choose, run the final structure through a standard stereochemistry check (a Ramachandran plot and clash analysis) before you commit. A high pLDDT or a low resolution number does not excuse you from validation, and our guide on how to validate a protein structure covers the checks that apply to all three sources.

Side-by-side: which structure for which situation?

Here is the decision compressed into one table. Read it top to bottom: prefer the earlier rows when they are available.

A practical order of operations: search the PDB first; if nothing fits, check the AlphaFold database for a ready model and note its pLDDT; if you have a strong template and want a build you control, run SWISS-MODEL. Often the smartest move is to obtain two options, such as an AlphaFold model and a homology model, and compare them before you start molecular dynamics.

Frequently asked questions

Is AlphaFold now better than experimental structures?

No. An experimental structure of your exact protein, at good resolution and covering your region, is still the strongest starting point because it is a measurement of the real molecule in a real state. AlphaFold is remarkable when no structure exists, but it predicts an apo model and can be wrong at flexible regions and induced-fit pockets.

Can I dock a ligand into an AlphaFold model?

You can, but with care. AlphaFold models are ligand-free, so the pocket may be in a resting conformation rather than the bound one. Check the pLDDT across the binding site (aim for values above 70), consider refining or minimising the pocket, and validate any pose against known binders where possible.

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

As a rule of thumb, above about 30% identity to the template gives a usable model, and accuracy rises with identity. Below roughly 20 to 30% you are in the twilight zone, where the alignment itself becomes unreliable. Always read the SWISS-MODEL GMQE and QMEANDisCo scores rather than trusting identity alone.

How do I compare a pLDDT score with a crystal resolution?

You do not compare them directly, because they measure different things. Resolution describes the experimental data quality; pLDDT describes AlphaFold’s confidence in its own prediction. Use each metric within its own source, then judge all candidates on the shared criteria of coverage, confidence and fitness for your next step.

Where do I get each type of structure?

Experimental structures come from the RCSB PDB. Ready-made AlphaFold models come from the AlphaFold Protein Structure Database, or you can generate your own with ColabFold. Homology models are built on the SWISS-MODEL server from a template of your choice.

This guide was written by the StemSkills Lab team, whose members bring more than ten years of combined work in sequence and structural bioinformatics, drug discovery and design, and multiscale molecular modeling. For the full learning path, start with the computational biology skills roadmap and follow it into molecular docking.