How AlphaFold Solved Protein Folding

01The shape is the function

Proteins are chains of amino acids, but biology does not operate on a chain in a line. The chain folds into a three-dimensional structure, and that shape controls which molecules it can bind, what reactions it can catalyse, and how it interacts with the cell.

For decades, researchers could determine structures with X-ray crystallography, cryo-electron microscopy, and NMR. These methods are powerful, but expensive and slow compared with the number of sequences nature contains.

02Why folding looked impossible

A protein with hundreds of amino acids has an astronomical number of possible conformations. The naive strategy—try every shape and choose the lowest-energy one—quickly becomes unusable. Biology’s physical constraints narrow the search, but they do not make the brute-force version practical.

The problem is also inverse: the sequence is available, while the folded structure is the hidden object. Evolution supplies clues because related proteins often preserve structural patterns even when their exact sequences diverge.

03CASP turned prediction into a contest

The Critical Assessment of Structure Prediction, or CASP, gave the field a recurring test against experimentally determined structures. AlphaFold 1 placed first at CASP13 in 2018. AlphaFold 2 then produced a much larger jump at CASP14 in 2020.

The score that made the result legible was GDT: a measure of how closely predicted coordinates align with the experimental structure. A score of 100 is a complete match; AlphaFold 2 exceeded 90 for roughly two-thirds of the proteins in the cited CASP14 result.

04The breakthrough was information fusion

AlphaFold 2 did not discover a single magic folding rule. It combined several sources of evidence: the amino-acid sequence, multiple sequence alignments that expose evolutionary constraints, pairwise relationships between residues, and a structure module that iteratively refines a three-dimensional hypothesis.

The training corpus was enormous. Wikipedia describes a custom database with more than 65 million protein families represented as alignments and hidden Markov models, covering more than two billion sequences from reference databases and metagenomes.

05Attention made distant residues talk

Transformers are useful here because residues that are far apart in the sequence may be close together in the folded structure. Attention mechanisms let the model update relationships across the whole sequence rather than treating each position as an isolated local patch.

The output is a probability-shaped prediction, not a microscope image. Confidence varies across a protein: rigid domains can be clear while flexible loops, disordered regions, or alternate conformations remain uncertain.

06AlphaFold 3 widened the target

AlphaFold 2 focused on protein structures. AlphaFold 3, announced in 2024, extended the task to complexes involving proteins with DNA, RNA, ions, and ligands. Wikipedia reports a minimum 50% improvement in accuracy for protein interactions with other molecules compared with existing methods.

That matters because drug discovery depends on interactions, not isolated shapes. A binding pocket, a nucleic-acid interface, and a transient complex are different scientific questions.

07Solved prediction is not solved biology

The phrase “solved protein folding” needs a boundary around it. AlphaFold made static structure prediction dramatically more useful; it did not reveal the complete physical mechanism of folding, guarantee every prediction, or replace experiments. Researchers noted that accuracy was insufficient for about one-third of predictions in the CASP14 summary.

The durable contribution is a new research loop: predict a structure, choose the experiment that would distinguish competing hypotheses, and use the result to improve the next question. AI did not end structural biology. It changed which experiments are worth doing first.