A predicted structure is cheap, fast and good enough for several decisions a small team makes every month: where to put a construct boundary, which surface loop to graft, whether two domains are likely to sit together, which residues to mutate first. It is not good enough for anything that depends on sub-angstrom geometry, and the common mistake is not using it at all rather than using it wrongly. This article is about which questions a model answers for a team with no structural biology group, and the five minutes of checking that should precede every answer.
Questions a predicted structure answers well
- Domain boundaries and construct design. Where ordered regions start and end, which linkers are flexible, which termini can be truncated without cutting into a fold. This is the highest-value routine use.
- Surface topology. Which residues are exposed, which patches are hydrophobic or charged, where a tag or a conjugation site can go without burying it.
- Fold assignment and family relationships. Whether an uncharacterised sequence adopts a known architecture, which is often the fastest route to a functional hypothesis.
- Mutation triage. Whether a residue is buried in the core, in a loop, or at an interface, which separates "likely destabilising" from "probably tolerated" before anyone clones anything.
- Interface hypotheses. Whether a complex is geometrically plausible, and which residues would be in contact if it is.
Questions it answers badly
- Binding affinity and the effect of a point mutation on it. A model gives geometry, not energetics, and small side-chain differences dominate affinity.
- Side-chain positions in a pocket at the precision docking needs. Rotamers in a predicted binding site are the least reliable part of the model, which is why a docking campaign on a predicted structure needs its own validation.
- Conformational states and dynamics. A prediction returns one structure. A protein with two functionally relevant states will be returned as one of them, or as an average of neither.
- Everything about disordered regions except that they are disordered, which is itself useful information.
The five-minute check before you use a model
1. Read the per-residue confidence over the region you care about, not the global mean. A high average hides a low-confidence loop, and the loop is usually the part the question is about. 2. Check how much evolutionary information was available. A sequence with few homologues produces a less reliable model, and the confidence score does not always make that obvious. 3. Look at the inter-domain arrangement separately from the domains. The individual folds are often right while their relative placement is not, and for a multi-domain protein that is the difference between a usable and a misleading model. 4. Compare with an experimental structure of a relative if one exists. Superimposing on a homologue takes minutes and catches gross errors immediately. 5. Sanity-check the chemistry. Disulfides paired as expected, a buried charged residue only where there is a partner, the active site residues in contact if the function is known. A model that violates known biochemistry is wrong whatever the score says.
Use the model to design the experiment, not to replace it
The productive pattern is a loop in which the model generates a cheap, falsifiable prediction and one experiment tests it. Examples that work:
| Model-based prediction | Cheap test |
|---|---|
| Residues 1 to 28 are a flexible extension | Express the truncated construct, compare expression and stability |
| This loop carries the specificity | Swap it with the equivalent loop from a relative, measure binding |
| These two residues form the interface | Mutate both to alanine, look for loss of complex formation |
| The construct boundary at 215 cuts into the fold | Make 215 and 240, compare yield |
Each of these turns a structural guess into a result, and a result into a better model of what is going on. The cost is one or two constructs.
When you do need an experimental structure
Three triggers, and they are worth naming in advance so the decision is not made late: a lead optimisation programme that will make dozens of compounds against one pocket, a mechanism claim that will go into a publication or a regulatory document, and any case where the predicted model and the data disagree in a way that matters. In those situations the prediction is the thing that makes the experimental work efficient, by telling you which construct to crystallise and which region to focus on.
Limits of the statement
- Confidence scores report the model's agreement with itself and with the evidence available, not correctness. A confidently wrong model is possible, particularly for folds unlike anything in the training data.
- A predicted complex is a hypothesis about an interaction, not evidence that the interaction occurs. That evidence is biochemical.
- Post-translational modifications, bound cofactors, metals and lipids are usually absent from the model, and for many proteins those are the functionally decisive parts.
- Nothing here substitutes for a structural biologist's judgement on a difficult case. It substitutes for having no structural information at all, which is the real alternative for most small teams.
Frequently asked questions
What is a predicted protein structure actually good for?
Construct design, domain boundaries, surface topology, fold assignment, mutation triage and interface hypotheses. These are geometric questions at a scale the prediction gets right, and they cover a large share of what a small team needs week to week.
Can a predicted structure be used for docking?
With care and with its own validation. Side-chain positions in the pocket are the least reliable part of a model, so a retrospective check on known actives for the target is the condition for trusting any ranking obtained from it.
How do you check a model without an experimental structure?
Read per-residue confidence over the region in question, check how many homologous sequences were available, assess the domain arrangement separately from the folds, superimpose on an experimental structure of a relative, and verify that known biochemistry such as disulfide pairing is respected.
Does a high confidence score mean the structure is correct?
No. It means the prediction is internally consistent and well supported by the available evidence. For an unusual fold, or a sequence with few relatives, a confident model can still be wrong.
When is an experimental structure unavoidable?
When a long optimisation campaign will target one pocket, when a mechanistic claim goes into a publication or a regulatory file, or when model and data disagree in a way that changes a decision.