AlphaFold2 and the Protein-Folding Problem

In his 1972 Nobel lecture, Christian Anfinsen argued that a protein's amino-acid sequence should, in principle, fully determine its folded three-dimensional structure. That claim launched a fifty-year search for a computational shortcut: read the sequence, predict the shape, skip the months of X-ray crystallography or cryo-EM. The search stalled for decades. Then, at the fourteenth Critical Assessment of Structure Prediction (CASP14) in late 2020, DeepMind's AlphaFold2 scored a median GDT of 92.4 across all targets, a level of accuracy that had never before been reached routinely in the competition. On a scale where 100 is a perfect match to the experimental structure, a score in the low 90s means the prediction is often within the error of the experiment itself, on the order of an atom's width. Assessors described the problem, for a large class of proteins, as substantially solved.

It is worth being precise about what was solved. AlphaFold2 does not simulate folding. It does not integrate physical forces over time to watch a chain collapse into its native state. It is a supervised model that maps sequence-plus-evolutionary-context directly to a static set of atomic coordinates. Understanding how it does that, and where the "static" qualifier bites, is the whole lesson.

Why sequence-to-structure was hard

The naive framing (feed a string of amino acids to a network, read out coordinates) hides the difficulty. A modest 300-residue protein has thousands of degrees of freedom in its backbone and side chains; the space of physically plausible conformations is astronomically large; and the folded state is selected by a delicate balance of hydrophobic packing, hydrogen bonding, and electrostatics that no simple scoring function captures well. Direct physical simulation (molecular dynamics) can fold small proteins but is far too expensive at scale and drifts on longer chains.

The signal that eventually cracked the problem does not come from physics at all. It comes from evolution.

Co-evolution: the signal baked into the data

Take one protein and find every homologous sequence across the tree of life: the same protein in humans, mice, yeast, bacteria. Stack these aligned sequences into a multiple-sequence alignment (MSA), one row per organism, one column per residue position. Now read the alignment column by column.

Two residues that sit far apart in the sequence but touch in the folded 3D structure are under a shared constraint. If a mutation at position i would disrupt a contact, a compensating mutation at the touching position j tends to be selected alongside it, so the two positions vary together across species. That statistical coupling between columns is a fingerprint of spatial proximity. Extract enough of these coupled pairs and you have a noisy contact map, a strong prior on the fold.

This is the key idea for a machine-learning audience: the domain structure is in the input representation, not bolted on as a physics loss. Pre-AlphaFold methods already used co-evolution, but they read couplings with relatively shallow statistical models. AlphaFold2's advance was to make the extraction of co-evolutionary signal learnable and deep, and to couple it tightly to a geometry-aware output head, trained end to end.

The Evoformer

The heart of AlphaFold2 is the Evoformer, a stack of attention blocks that maintains and refines two representations in parallel:

• The MSA representation (rows = sequences, columns = residues). Attention runs along both axes. Row-wise attention lets a residue position attend across the other positions in the same sequence; column-wise attention lets a position attend across all the aligned sequences, which is precisely where the co-evolution signal lives.
• The pair representation (an L x L grid, one entry per residue pair). This is the model's evolving hypothesis about the relationship between every pair of residues, effectively a learnable, refined contact/distance map.

The two representations talk to each other every block. Information distilled from the MSA updates the pair grid; the pair grid, in turn, biases the MSA attention. A geometric consistency constraint (the triangle inequality on distances) is built into the pair-update operations, so the model's internal distance hypotheses are nudged toward something that could correspond to a real 3D arrangement. After many blocks, the pair representation encodes a rich, geometrically coherent picture of the fold, without any coordinates having been written yet.

The structure module and recycling

The structure module turns representations into geometry. It treats each residue as a rigid body (a local coordinate frame) and predicts a rotation and translation for each one, placing every residue in 3D space; side-chain torsion angles are predicted on top. Crucially it operates directly on 3D frames and is equivariant to global rotation and translation of the protein, so the network reasons about relative geometry rather than memorising absolute positions. The output is a full set of atomic coordinates.

Two training ideas make this work end to end. First, the whole pipeline (Evoformer plus structure module) is trained against experimental structures from the Protein Data Bank with a loss defined directly on atom positions, so the geometry head and the representation learner co-adapt rather than being trained in separate stages. Second, recycling: the model's output representations are fed back as additional input and the whole forward pass is run again, several times. Each pass refines the previous estimate, an iterative-improvement loop that costs compute but sharpens the final structure considerably.

pLDDT: the model tells you where to trust it

A prediction is only useful if you know where it is reliable. AlphaFold2 emits a per-residue confidence score, the pLDDT (predicted Local Distance Difference Test), on a 0 to 100 scale, where higher is more confident. It is a calibrated estimate of how well the local structure around each residue is likely to match the truth. In practice pLDDT is load-bearing: well-folded domain cores score high, and low-pLDDT stretches are a reliable flag for regions the model is unsure about (often flexible or intrinsically disordered ones). A separate predicted aligned error (PAE) score captures confidence in the relative positioning of different domains, which matters when a protein has multiple parts that could shift against each other.

Why this mattered for AI-for-science

The result landed as more than a benchmark win. The AlphaFold Protein Structure Database, built by DeepMind with EMBL-EBI, now hosts over 200 million predicted structures spanning essentially the known proteome, released openly. For a working biologist, a plausible structural model for almost any protein is now a lookup rather than a multi-year experimental campaign. The template is what generalises: a hard scientific inverse problem, an input representation that encodes the domain's own structure (here, evolutionary co-variation), a geometry-aware architecture, and end-to-end training against experimental ground truth. That recipe, not the specific network, is why AlphaFold2 is cited as a landmark for machine learning applied to the natural sciences.

When it falls down

The limits follow directly from how the model works, and an ML reader should be able to predict most of them.

• Shallow MSAs. The model's power comes from co-evolution. Orphan proteins with few known homologues, fast-evolving or engineered sequences, and many single-sequence inputs give the Evoformer little to read, and accuracy drops. Where the evolutionary signal is thin, so is the prediction.
• A single static structure, not dynamics. AlphaFold2 returns one conformation. It does not predict conformational ensembles, folding pathways, or the motion of a protein over time. Proteins that function by switching between states (open/closed, active/inactive) are represented by, at best, one of those states.
• Disordered and flexible regions. Intrinsically disordered regions have no single native fold; the model correctly signals this with low pLDDT, but a low-confidence coil is not a structure. Reading a confident shape into a disordered stretch is a common misuse.
• Binding-induced and context-dependent change. The base model predicts the isolated chain. It does not natively model how the structure shifts on binding a ligand, drug, ion, or another protein, nor the effect of the cellular environment.
• Point mutations. A single amino-acid substitution can destabilise or refold a protein, but it barely perturbs the MSA. The model often returns a near-identical structure for wild-type and mutant, so AlphaFold2 is not a reliable predictor of mutational effect on stability.
• Multimeric assemblies. The original AlphaFold2 predicts single chains. Complexes and their interfaces needed a separate model (AlphaFold-Multimer) and remain harder than single-domain monomers.

The honest one-line summary: AlphaFold2 predicts a static structure from evolutionary context with remarkable accuracy, and it is neither a folding simulator nor a substitute for experiment where the co-evolution signal runs out.

Further reading

• Highly accurate protein structure prediction with AlphaFold - Jumper et al, Nature 2021, the primary paper describing the Evoformer, structure module, recycling, and pLDDT.
• AlphaFold: a solution to a 50-year-old grand challenge in biology - DeepMind's CASP14 writeup, source of the 92.4 median GDT figure and the Anfinsen framing.
• AlphaFold Protein Structure Database - the open DeepMind and EMBL-EBI database of over 200 million predicted structures.
• google-deepmind/alphafold - the official AlphaFold2 inference implementation and model weights.