AI can predict a protein's shape in minutes. That is not the same as knowing what it does

A protein is manufactured as a string. The cell reads a gene and links amino acids in order, and what comes off the ribosome is a floppy chain with no useful properties at all. Within microseconds it collapses into a specific three-dimensional shape, and only then does it become an enzyme, a receptor, an antibody, a motor. The sequence is written in the DNA. The shape is not — it is a consequence, and for fifty years nobody could reliably work out that consequence from the string.

That changed abruptly. Between 2020 and 2022 structure prediction went from an open problem to something a student can run in a browser over lunch, and in 2024 the Nobel Prize in Chemistry went half to David Baker for computational protein design and half, jointly, to Demis Hassabis and John Jumper for solving the prediction problem. What is much less widely understood is how it was solved — because the method does not simulate folding, and that single fact explains both its power and every one of its remaining blind spots.

Why the shape is the whole story

Function in biology is geometry. An enzyme works because a pocket on its surface fits one molecule and not another, holding it in a strained arrangement that makes a reaction easier. A receptor works because a groove matches the shape of a signalling molecule. An antibody works because a loop at its tip is complementary to a patch on a pathogen.

Get the shape wrong and you lose the function entirely, which is why misfolding is its own category of disease — the amyloid aggregates of Alzheimer's and the prion disorders are proteins that have adopted a stable arrangement they were not supposed to have, and then recruited their neighbours into it. A drug, likewise, is usually a molecule shaped to occupy a particular pocket. Knowing the pocket is where the design work starts.

For most of the field's history that shape came from experiment: X-ray crystallography, nuclear magnetic resonance, and more recently cryo-electron microscopy. These work, and they remain the ground truth, but a single structure could take months to years — and crystallography needs the protein to crystallise, which many important proteins simply refuse to do. The result is a permanent imbalance. The Protein Data Bank holds a couple of hundred thousand experimentally determined structures. The sequence databases hold hundreds of millions of proteins. Almost everything we knew about was known only as a string.

Why brute force was never going to work

The obvious approach — simulate the physics and let the chain fold — runs into a counting problem that Cyrus Levinthal pointed out in 1969. Each bond in the backbone can take several rotational states, so a chain of a hundred amino acids has an astronomical number of possible conformations. Searching them one at a time at atomic timescales would take longer than the age of the universe.

And yet real proteins fold in microseconds to milliseconds. That is Levinthal's paradox, and its resolution — that folding follows funnelled energy landscapes rather than a random search — was good physics but did not deliver a practical predictor. Molecular dynamics simulation can fold small proteins given enormous computing time. It could not be pointed at a genome.

What the model is actually reading

The breakthrough came from abandoning the physics question and asking an evolutionary one instead.

Take the protein you care about and collect every related sequence from every organism where evolution has kept a version of it — a multiple sequence alignment, often thousands of rows deep. Now look down the columns. Most positions vary independently. But some pairs of positions vary together: whenever one mutates, the other mutates in a compensating way. That pattern has an obvious structural explanation. Those two residues are touching. A mutation that would break the contact only survives if its partner changes to restore the fit.

Co-evolution as a contact signal was known before AlphaFold. What deep learning added was the ability to use it properly — a transformer architecture that passes information back and forth between the sequence alignment and a pairwise map of residue-residue relationships, refining both together, and then a final module that turns that abstract map into actual atomic coordinates in one shot rather than by simulated folding. At the CASP14 blind assessment in 2020 — a biennial competition running since 1994, in which teams predict structures that have been solved but not yet published — this approach returned a median accuracy score above 90 on a 100-point scale, a level generally regarded as comparable with experiment. The field's reaction was not that a competition had been won but that a problem had closed.

In 2022 the resulting database was expanded to cover essentially every protein in the reference sequence databases — over two hundred million predicted structures, released openly. A 2024 successor generation extended the same machinery to complexes: proteins bound to DNA, RNA, ions and small molecules, which is the form most drug discovery questions actually take.

It is not simulating a molecule. It is reading what evolution has already tested and discarded over billions of years. That is why it is superb where the evolutionary record is deep, and unreliable exactly where the record is thin.

The four things it still gets wrong

Sequences with no relatives. If the alignment is shallow — an orphan gene, a rapidly evolving viral protein, a molecule designed in a lab — the co-evolution signal is weak or absent, and accuracy drops. Newer protein language models such as ESMFold soften this by learning general sequence regularities without needing an alignment, but they do not remove the dependence on having seen something like it before.

Proteins that move. A prediction is a single static picture, usually of the most stable arrangement. Real proteins are machines: transporters open and close, kinases switch between active and inactive states, motors ratchet. A structure that is right about the resting state can be silent about the conformational change that constitutes the actual function.

Point mutations. This one surprises people. The models are excellent at predicting the fold of a sequence and often poor at predicting how a single amino-acid substitution destabilises it — because the fold is inferred from the family, and one changed letter barely moves the family signal. For clinical variant interpretation, that gap matters.

Regions with no fixed shape. A significant fraction of the human proteome is intrinsically disordered — genuinely floppy, functional precisely because it is not fixed. Here the model's own uncertainty is the useful output. The per-residue confidence score, pLDDT, runs from 0 to 100, and stretches below about 50 correlate strongly with disorder. A low-confidence region is often not a failed prediction but a correct statement that there is nothing rigid to predict.

Behind all four sits one discipline: a predicted structure is a hypothesis. It is an extremely good one, good enough to design experiments around, choose crystallisation targets with, and interpret cryo-EM maps faster. It is not evidence about mechanism, and the experimental methods it was said to have replaced are now busier, because there are far more specific questions worth asking.

Running the idea backwards

The more consequential half of the 2024 prize was the other one. If a network can map sequence to shape, the reverse becomes tractable: specify a shape you want and generate a sequence that folds into it. That is de novo protein design, and it has produced binders that grip a chosen target, novel enzymes catalysing reactions with no natural counterpart, and self-assembling nanoparticle scaffolds for vaccines.

This is where the discipline stops being descriptive and becomes engineering — and it is why the same community that celebrated the achievement moved quickly to set screening norms for synthesis providers, because a technology that designs binding proteins to order deserves that attention early rather than late.

Why it matters for students and researchers

The bottleneck has moved, exactly as it did in genomics. Structures are no longer scarce; two hundred million of them are a free download. What is scarce is the judgement to use them — reading a confidence score correctly, recognising when a prediction reflects a family average rather than your specific variant, knowing which questions still require a beamline or a microscope.

That judgement sits at the join of two fields that used to be separate degrees. The person who understands both what a transformer is doing to a sequence alignment and what a hydrogen bond does in an active site is not a biologist who dabbles in code or an engineer who read a biology paper — that combination is now the standard job description in drug discovery, enzyme engineering and synthetic biology, and it is what applied AI and bioengineering coursework should be building toward.

Frequently asked questions

What is the protein folding problem?

It is the problem of predicting a protein's three-dimensional shape from its amino-acid sequence alone. The shape determines the protein's function, but it is not written explicitly in the gene, and for about fifty years there was no reliable way to compute it.

How does AlphaFold predict protein structure?

It does not simulate folding. It collects related sequences from other organisms and detects positions that mutate together — a signature that those residues touch in the folded protein — then uses a deep neural network to turn that co-evolutionary pattern into atomic coordinates.

Is a predicted structure as good as an experimental one?

For well-represented protein families the accuracy is often comparable to experiment, but a prediction remains a hypothesis. Crystallography, NMR and cryo-EM are still needed to confirm mechanism, capture alternative states and resolve how a protein behaves with a drug bound.

What does the pLDDT confidence score mean?

It is a per-residue confidence estimate from 0 to 100. High values indicate a well-determined local structure, while values below roughly 50 usually indicate an intrinsically disordered region — which is genuine information about the protein, not merely a failure of the model.

What is de novo protein design?

It is the reverse of prediction: specifying a desired shape or function and computationally generating an amino-acid sequence that folds into it. It has produced novel binders, enzymes and vaccine scaffolds that do not exist in nature, and it was recognised in the 2024 Nobel Prize in Chemistry.