The problem 01 · A 50-year grand challenge
From String to Shape
The same sequence always folds the same way — but no one could predict it.
A protein starts life as a long, floppy chain of tiny beads strung together in a fixed order — your body builds thousands of different ones, each from its own recipe. Let one go, and in the blink of an eye it collapses into a specific, intricate 3D lump, like a string that folds itself into the same piece of origami every single time.
That shape is the whole point. A protein's shape is what lets it do its job — grab oxygen, cut a molecule, form a muscle fibre. Same shape, same job. And here is the astonishing part: the same bead-string always folds into the same shape. Yet for fifty years, looking at just the string of beads, nobody could say what the folded shape would be. That was the protein folding problem.
The beads are amino acids — twenty kinds — joined into a chain called the backbone. The order of the amino acids is the protein's primary structure, written directly from a gene. As the chain forms, parts of the backbone coil into α-helices (springs) or line up side by side into β-sheets (the flat arrows), linked by floppy loops — the secondary structure — and these elements then pack together into one compact 3D fold.
Why is predicting that fold so hard? Count the possibilities. If each amino acid could sit in just a few orientations, a chain of length \(n\) has on the order of \(3^n\) possible shapes. For a small protein of 100 amino acids that is already about \(10^{47}\) shapes — so many that if the chain tried them one after another, faster than any atom moves, the search would outlast the age of the universe many times over. This is Levinthal's paradox.
And yet real proteins fold in microseconds to seconds. So folding cannot be a blind search through every shape — the chain is funnelled toward one special shape, its stable, lowest-energy native state. Knowing it has a definite answer, though, never told us how to compute it.
Two different problems hide under "protein folding." The first is the structure-prediction problem: given the primary sequence \(a_1 a_2 \cdots a_n\), output the native 3D coordinates of every atom. The second is the folding-process problem: by what physical mechanism, and along what pathway, does the chain reach that state so fast and so reliably? They are not the same question, and only the first has been cracked.
Anfinsen's thermodynamic hypothesis (1973) frames prediction: for many proteins the native state is determined by the sequence alone, as the global minimum of the free energy. Levinthal's paradox (1969) is the obstruction — the number of conformations grows exponentially, \(\Omega \sim k^{n}\), so an unbiased search is astronomically infeasible. Its resolution is the energy-landscape / funnel picture (Bryngelson–Wolynes, the principle of minimal frustration): the landscape is not flat and rugged but globally funnel-shaped, biasing the chain downhill toward the native basin without exhaustive search.
The funnel resolves the paradox in principle, but it does not hand us the mechanism. The detailed dynamics — folding pathways, intermediates, transition states, the kinetics of how a given sequence threads its funnel, and why folding sometimes fails into misfolded or aggregated states (amyloid, prion disease) — remain only partially understood.
Status: prediction solved; mechanism open. Computing the native shape from sequence is now routine (next panel). Explaining and simulating the folding process from physical principles, ab initio, is still an active, unsolved problem.
A bead-chain is released and folds to one specific compact shape — the same shape every time, looping on its own. Drag to turn it in 3D. Helices (coils) and sheets (arrows) packing into the fold, beside the count of possible shapes. As the chain lengthens, the count explodes past anything searchable. Drag to turn the fold. The folding funnel: the chain slides down a rugged energy landscape to its single lowest-energy native state — biased, not blind.
The solution 02 · Structure prediction · solved · AlphaFold2, 2020
How AlphaFold Cracked Prediction
A computer that learned the shape of life from examples.
The breakthrough was not a new law of physics. It was teaching a computer to recognise. Over decades, scientists had painstakingly measured the shapes of around two hundred thousand real proteins in the lab. AlphaFold — built by Google DeepMind — studied that whole library until it learned the hidden patterns linking a bead-string to the shape it folds into.
Then, shown a brand-new string it had never seen, it could predict the folded shape — in minutes, and often as accurately as a months-long lab experiment. Feed in the sequence, read out the structure. A problem that had resisted everyone for half a century was, for prediction at least, suddenly solved.
AlphaFold2 is a deep neural network trained on the experimentally-determined structures in the Protein Data Bank. Its cleverest ingredient is evolution. Line up the same protein from many species — a multiple sequence alignment — and you see which positions vary in lockstep. If two amino acids consistently change together across the tree of life, it is usually because they touch in the folded structure and must stay compatible.
So the network reads thousands of related sequences, turns that co-evolution signal into a contact / distance map — a grid predicting which pairs of amino acids sit close together in space — and then folds a single 3D structure consistent with the whole map. Earlier methods used pieces of this idea; AlphaFold2 made it work end to end.
The proof came at CASP14 in 2020 — a blind, biennial contest where teams predict structures not yet made public. AlphaFold2 reached accuracy close to the experiments themselves, far ahead of everyone else. DeepMind then ran it across nearly every known protein and released about 200 million predicted structures in a free public database.
AlphaFold2's pipeline takes two representations: a multiple sequence alignment (MSA) and a pairwise residue–residue tensor. The Evoformer — a stack of attention blocks — lets the MSA and pair representations exchange information repeatedly, enforcing geometric consistency (e.g. the triangle inequality on distances). A structure module then turns the refined pair representation into explicit 3D coordinates via a rotation-and-translation (\(\mathrm{SE}(3)\)) frame per residue, and the whole network is recycled several times and trained end-to-end against known structures. Each residue gets a confidence score, pLDDT, and a global pair-confidence, PAE.
The impact is real: the ~200M-structure database has accelerated drug discovery, enzyme design and basic biology. AlphaFold3 (2024) replaced the structure module with a diffusion-based generator and extended prediction beyond single proteins to complexes — proteins bound to DNA, RNA, small-molecule ligands and ions.
The honest close. AlphaFold predicts the native structure with remarkable accuracy. It does not reveal the folding mechanism, the pathway, or the physical rules of folding — it recognises the answer rather than simulating the process, and that process remains an open problem. Confidence also varies: low-pLDDT regions, intrinsically disordered proteins, points mutations and novel folds can be unreliable, and a prediction is a hypothesis, not a measurement.
Credit, precisely. AlphaFold's structure prediction was built at Google DeepMind, led by Demis Hassabis and John Jumper. They shared half of the 2024 Nobel Prize in Chemistry; the other half went to David Baker (University of Washington) — but for computational protein design, a separate strand of work, not for AlphaFold.
A sequence goes in; the machine — having studied a library of known shapes — predicts the fold that comes out, looping on its own. Drag to turn the predicted fold. Stacked relatives reveal positions that change together; the contact map of which pairs end up touching sharpens as more sequences are added, then folds to 3D. Drag to turn the fold. MSA + pair representation → the Evoformer's attention → the structure module's coordinates, recycled to refine. Colour = pLDDT confidence. Drag to turn the fold.