The Ribbon Is Not the Result: A 15-Minute AI Protein-Model Audit for UK Students

If you have opened a protein-structure model in the last term at a UK university, you are not alone. AlphaFold, ColabFold, ESMFold and a growing shelf of web tools have made confident-looking structures feel like a normal part of coursework. That is broadly good news. The problem is that a predicted structure is not a result in the same way a gel image or a crystal structure is a result. It is a hypothesis with a colour scheme.

At Byte Size Bio, we like a simple rule: treat the ribbon diagram as the start of the argument, not the end of it. Before you cite a model, drop it into a slide, or use it to design a mutant, spend fifteen minutes on an audit. It will not make you a structural biologist. It will stop you making the most common mistakes.

The picture is not the proof

Prediction tools are excellent at some things and quietly unreliable at others. They are strongest for globular proteins with lots of homologues in the Protein Data Bank. They are weaker for flexible loops, disordered regions, membrane proteins, rare folds, ligand-bound states and anything involving post-translational modifications. They also predict a single static conformation, which is not how most proteins behave in a cell.

That does not mean you should avoid them. It means you should read the confidence data before you read the cartoon.

Trap 1: the low-confidence tail

Most tools give you a per-residue confidence score. AlphaFold uses pLDDT, where scores above 70 are generally decent, above 90 are strong, and below 50 mean the model is guessing. The trap is that the pretty part of the picture draws your eye, while the important part of your protein sits in a low-confidence loop.

Ask a blunt question: is the bit I care about in a confident region? If you are studying an active site, a binding interface or a catalytic residue, and that region has low confidence, the model is not evidence for your claim. It is a placeholder. The same goes for domain-to-domain orientation. pLDDT tells you about local confidence; PAE tells you about how confident the model is that two domains sit in the right relative position. A model can have excellent local scores and still get the overall architecture wrong.

Trap 2: training-data déjà vu

The model learned from known structures. That means well-studied protein families get strong predictions partly because the answer is already in the training set. Orphan proteins, recently discovered families and proteins from under-studied organisms are harder. This is a bias problem, not a bug, but it matters for UK labs working on non-model systems.

Before you trust a model, check what it was trained on and when. Look at the database version. Check whether your sequence is the canonical isoform or a splice variant. If you have a disease mutation, a truncation or a tag, make sure the model actually used that sequence. It is surprisingly easy to audit a wild-type model while writing about a mutant.

Also check ligands, cofactors and metals. Many prediction tools return an apo-like model. If your argument depends on a bound ATP or a zinc ion, the model may not support it. A ribbon diagram without the ligand can make a binding site look more solved than it is.

Trap 3: ribbon theatre

Rainbow colouring is not confidence colouring. A ribbon diagram coloured by spectrum can make a disordered tail look as solid as a barrel. A cropped image can hide missing residues. A cartoon with no scale, no confidence map and no sequence coverage is closer to illustration than analysis.

For your own work, use confidence colouring by default. Show the full chain, including low-confidence regions. Note missing residues. If you are comparing two models, show the same view and the same colour scale. If you are putting a model in a figure, caption it as a prediction and state the tool, version and date.

The 15-minute audit

Here is a routine you can run before a supervisor meeting, a lab report or a presentation.

0-2 minutes: provenance. Write down the tool, version, database release, date, sequence length, isoform and any mutations or tags. If you cannot fill this in, you are not ready to interpret the model.

2-5 minutes: confidence map. Find the pLDDT plot. Note the percentage of residues above 70. Identify low-confidence regions. Check PAE if you are making claims about domain arrangement or complexes. Circle the region you actually care about.

5-7 minutes: biological sanity. Does the model include the expected secondary structure? Are the termini where you expect them? Is the active site accessible? Are disulfides plausible? If it is a membrane protein, is there a hydrophobic band? If it is an enzyme, is the cofactor site present?

7-10 minutes: comparison. BLAST your sequence or search for close homologues with experimental structures. Compare the predicted fold to a PDB entry if one exists. Look for conserved residues in a multiple sequence alignment. If the model disagrees with a well-supported experimental structure, do not ignore it.

10-12 minutes: red flags. Check for long low-confidence tails, unlikely knots, clashes, missing loops near the site of interest, and domains packed in a way that contradicts known biochemistry. You do not need a full validation suite for a student report, but you do need to notice obvious nonsense.

12-15 minutes: write the caveat. Add one sentence to your notes or figure caption: what the model predicts, which regions are confident, and what it cannot tell you. Save the session or job ID. Future you will be grateful.

What to write in your report

A good caption sounds like this: 'Predicted structure of X generated with Tool Y in 2026; residues 1-180 have pLDDT above 70, while the C-terminal tail is low confidence and should not be interpreted.' That is honest, useful and quick.

A bad caption sounds like: 'Structure of X.' That implies you solved it. You did not.

None of this is about distrusting AI. It is about using it like any other lab instrument: check the settings, know the limitations, and do not over-read the output. In 2026, that is a basic UK life-science skill, not a specialist one. Fifteen minutes now saves a lot of awkward questions later.