Protein Structure Prediction: What It Solved and What It Did Not

A protein is a chain of amino acids that folds into a specific shape, and the shape is what does the work. Going from the chain to the shape was one of biology’s defining open problems for fifty years. It is now, for a large class of proteins, something you do in minutes on a laptop. That is a real change, and it is narrower than the headline.

The problem, and why it was hard

Anfinsen’s experiments in the mid-twentieth century established that for many small globular proteins the sequence determines the fold: denature one, let it recover, and it returns to the same structure. So the information is in the sequence. Levinthal then pointed out the difficulty — a chain of any length has an astronomically large number of conformations, far too many to search, yet real proteins fold in milliseconds. Physics finds the answer without enumerating; a computer simulating the physics could not.

Meanwhile the practical gap widened. Sequencing became cheap, so the number of known protein sequences grew explosively. Determining a structure did not: X-ray crystallography needs a crystal, and plenty of proteins refuse to form one; NMR is limited by molecular size; cryo-EM removed some of those constraints but remains an instrument-heavy, expert-heavy process. Each structure was a project. The result was a catalogue of sequences vastly larger than the catalogue of shapes, with the shapes concentrated in whatever was easy to crystallise.

The field kept itself honest with CASP, a blind assessment run every two years. Organisers take proteins whose structures have been solved but not released, hand out the sequences, and score submitted predictions against the withheld answers. Nobody can tune to the test set, because the test set does not exist yet when you submit. In 2020 at CASP14, AlphaFold2 produced predictions for many targets at an accuracy that assessors described as competitive with experiment. The reason that result was believed immediately is entirely down to the format: it was a prospective, blind evaluation, which is rung 3 on the evidence ladder rather than another retrospective benchmark.

The signal: evolution as a contact map

The thing worth understanding is where the information comes from, because it explains every limitation later on. It is not physics. It is evolution.

Take a protein and collect its homologues from thousands of organisms — a multiple sequence alignment. Now look at two positions in that alignment. If those two residues touch in the folded structure, a mutation at one of them is destabilising unless something compensates, so over evolutionary time you see correlated substitution: position 40 changes and position 118 changes with it. Positions that do not touch have no reason to co-vary. So an alignment deep enough to show the pattern is, in effect, a noisy contact map of the structure, written down by natural selection over hundreds of millions of years.

Statistical methods exploited this before deep learning, with limited success — the correlations are heavily confounded by indirect and phylogenetic effects. What changed was learning the whole pipeline end to end: a network that refines an alignment representation and a pairwise residue-residue representation against each other, then a module that emits three-dimensional coordinates directly rather than a contact map for someone else to fold, with the output fed back through the network several times. Later systems replaced the explicit alignment with the internal representation of a protein language model trained on sequence databases, which is much faster because no alignment search is needed and is generally weaker precisely where you would expect: proteins with few known relatives.

Hold on to that dependency. The method works because the protein has well-populated evolutionary neighbours. Orphan proteins, fast-evolving viral proteins, and de novo designed sequences with no natural relatives are exactly the cases where it has least to go on.

The confidence score is half the tool

Predictions come with per-residue confidence — pLDDT, on a 0 to 100 scale — and a predicted aligned error matrix that says how reliable the relative placement of one region against another is. Using these properly is most of using the tool properly, and it is the part skipped in the summaries.

• Low confidence is often a result, not a failure. Regions predicted with low pLDDT correlate strongly with intrinsic disorder — parts of the protein that genuinely have no single structure in solution. The model saying “I do not know” frequently means “there is nothing to know.”
• High per-residue confidence does not mean the domains are arranged correctly. Two domains can each be predicted beautifully and be placed relative to each other with no justification. That is what the aligned-error matrix is for, and it is a separate question from pLDDT.
• A confidently predicted backbone does not mean the side chains in a pocket are right. Binding-site geometry depends on side-chain rotamers at a precision the backbone score says nothing about, which is why substituting a predicted structure for an experimental one in a docking campaign does not simply inherit the headline accuracy.

Four things it did not solve

• Function. A structure is not a mechanism. Knowing the shape of an enzyme does not tell you its substrate, its rate, its regulation or its role in a pathway. Structure narrows the hypothesis space; it does not close it.
• Dynamics. The output is one static conformation. Real proteins move: they open and close, they have allosteric states, they expose cryptic pockets that do not exist in the resting structure. Predicting the ensemble and the transitions between its members is a different and much less settled problem.
• The folding pathway. Predicting the endpoint is not the same as modelling the route. Misfolding kinetics, aggregation and chaperone dependence are questions about the trajectory, and the method that answers the first says nothing about them.
• Everything in context. Proteins work in complexes, in membranes, with cofactors, with post-translational modifications, at a particular pH. Systems extended to complexes, nucleic acids and small molecules exist and are used, but induced fit — where two partners change shape on binding — remains substantially harder than the single-chain case that got solved.

What came after prediction

The more consequential development is the inverse problem. Given a shape you want, generate a backbone that adopts it, then design an amino acid sequence that folds into that backbone. Diffusion models for backbone generation and sequence-design networks are now routine laboratory tools, and they are being used to make binders, enzymes and scaffolds that have no natural counterpart.

Note the difference in how those results are reported, because it is the right model for the whole field. A design pipeline emits many candidates. The paper reports how many were expressed, how many folded, and how many bound — laboratory numbers, measured after the fact, and typically a small fraction of what was generated. That fraction is the real output. The model made the search affordable; the experiment decided what was true.