For more than fifty years, predicting a protein’s three-dimensional shape from its sequence remained an open problem. AlphaFold changed the scale of what was possible, reaching near experimental accuracy in many cases—even without a similar known structure.

Proteins are essential to life, and understanding their structure can facilitate a mechanistic understanding of their function. Yet structures have been determined for around 100,000 unique proteins, only a small fraction of the billions of known protein sequences.

Structural coverage is bottlenecked by the months to years of painstaking effort required to determine a single protein structure, so accurate computational approaches are needed for large-scale structural bioinformatics. Predicting a three-dimensional structure from only an amino acid sequence has been an important open research problem for more than fifty years, and existing methods fall far short of atomic accuracy when no homologous structure is available.

Computational methods have followed two complementary paths: one focused on physical interactions, and the other on evolutionary history. The physical interaction programme uses thermodynamic or kinetic simulation of protein physics, but molecular simulation is computationally intractable for even moderate-sized proteins, while protein stability is context-dependent and accurate physics models are difficult to produce.

The evolutionary programme instead derives constraints on protein structure from bioinformatics analysis of evolutionary history, homology to solved structures, and pairwise evolutionary correlations. The approach is presented as a computational method capable of predicting protein structures to near experimental accuracy in a majority of cases.

AlphaFold was entered into the CASP14 assessment from May through July 2020 under the team name AlphaFold2, using a completely different model from the CASP13 AlphaFold system. CASP is carried out biennially using recently solved structures that have not been deposited in the PDB or publicly disclosed, making it a blind test.

The AlphaFold network directly predicts the three-dimensional coordinates of all heavy atoms for a given protein using the primary amino acid sequence and aligned sequences of homologues as inputs. The paper provides the most important ideas and components in the main text, while the full network architecture and training procedure are provided in the Supplementary Methods.

The network comprises two main stages. First, the trunk processes the inputs through repeated layers of a neural-network block called Evoformer. The Evoformer processes its inputs to produce one array representing a processed multiple sequence alignment and another representing residue pairs.

The MSA representation is initialized with the raw MSA, providing its starting state for the network's subsequent processing of inputs. The key innovation is the exchange of information within the MSA and pair representations, enabling direct reasoning about spatial and evolutionary relationships.

Figure three lays out the network’s architecture from sequence information to three-dimensional structure. Panel a shows the Evoformer repeatedly mixing MSA and pair representations, while panels b and c interpret pair entries as directed residue-to-residue edges and update them through triangle operations, enforcing constraints involving three residues.

Panel d shows the structure module using invariant point attention to update residue features and backbone frames, then predict atom positions; together, these components connect evolutionary information with geometric structure. The key principle of Evoformer is to view protein-structure prediction as a graph inference problem in three-dimensional space, with graph edges defined by residues in proximity.

The elements of the pair representation encode information about the relation between residues, while MSA columns encode individual residues and rows represent the sequences in which those residues appear. Within this framework, update operations are applied in each block, with the different operations applied in series.

After the trunk, the structure module introduces an explicit three-dimensional structure as a rotation and translation for each residue, called global rigid body frames. These representations start with all rotations set to the identity and all positions set to the origin, but rapidly develop and refine a highly accurate protein structure with precise atomic details.

The structure module uses simultaneous local refinement, an equivariant transformer, and a loss term that places substantial weight on the orientational correctness of residues. The network repeatedly feeds outputs recursively into the same modules.

This whole-network iterative refinement, called recycling, contributes markedly to accuracy with minor extra training time. AlphaFold can train to high accuracy using only supervised learning on PDB data, but accuracy is enhanced using an approach similar to noisy student self-distillation.

A trained network predicts structures for around 350,000 diverse sequences from Uniclust30, producing a new dataset filtered to a high-confidence subset. The same architecture is then trained from scratch on a mixture of PDB data and the predicted-structure dataset, with cropping and MSA subsampling making it challenging to recapitulate the earlier predictions.

This self-distillation procedure makes effective use of unlabelled sequence data and considerably improves the accuracy of the resulting network. Figure four asks how different parts of AlphaFold contribute to accuracy and how predictions develop through the network.

Panel a reports ablations on eighty-seven CASP14 domains and two thousand two hundred sixty-one PDB chains, using GDT or lDDT-C-alpha differences from the three-seed baseline, with ninety-five percent bootstrap intervals. Panel b tracks GDT across one hundred ninety-two Evoformer blocks: the two domains of T1024 stabilize early, while T1064 changes repeatedly and continues refining near the network’s end.

In CASP14, AlphaFold structures were vastly more accurate than competing methods. AlphaFold had a median backbone accuracy of zero point nine six angstroms, compared with two point eight angstroms for the next best method.

As a comparison point, the width of a carbon atom is approximately one point four angstroms. AlphaFold also produced highly accurate side chains when the backbone was highly accurate and improved over template-based methods even when strong templates were available.

For all-atom accuracy, AlphaFold reached one point five angstroms, compared with three point five angstroms for the best alternative method. Figure one establishes AlphaFold’s accuracy through both benchmark scores and structural overlays.

In panel a, AlphaFold’s median backbone accuracy on eighty-seven CASP fourteen protein domains is shown alongside the top fifteen entries, with ninety-five percent confidence intervals; panels b through d display close agreement with experimental structures, including a zinc-binding site and the domain packing of a two-thousand-one-hundred-eighty-residue chain.

Panel e summarizes the network, which combines sequence alignments, templates, pair representations, and recycling to produce three-dimensional structures. In the recent-structure dataset, all structures were deposited in the PDB after the training-data cut-off and were analysed as full chains.

High side-chain accuracy was observed when backbone prediction was accurate, and pLDDT reliably predicted the lDDT-C-alpha accuracy of the corresponding prediction. The global superposition metric TM-score could also be accurately estimated.

Overall, the analyses showed that the high accuracy and reliability seen on CASP14 proteins transferred to an uncurated collection of recent PDB submissions. Figure two evaluates AlphaFold on recent PDB structures that postdate its training data.

Panel a shows the distribution of full-chain backbone r.m.s.d., with an overall median of one point four six angstroms across three thousand one hundred forty-four chains, while panel b links backbone accuracy to correct side-chain rotamers. Panels c and d test the model’s confidence estimates: pLDDT tracks measured lDDT-C-alpha, and pTM tracks full-chain TM-score, supporting their use as indicators of structural reliability.

Although AlphaFold has high accuracy across the vast majority of deposited PDB structures, factors still affect accuracy or limit the model’s applicability. Accuracy decreases substantially when the median alignment depth is less than around thirty sequences, while improvements in MSA depth above around one hundred sequences lead to small gains.

The model is much weaker for proteins with few intra-chain or homotypic contacts compared with heterotypic contacts, a situation that typically occurs for bridging domains within larger complexes. Conversely, AlphaFold is often able to give high-accuracy predictions for homomers, even when the chains are substantially intertwined.

AlphaFold has demonstrated utility to the experimental community for molecular replacement and for interpreting cryogenic electron microscopy maps. Because AlphaFold outputs protein coordinates directly, it produces predictions in GPU minutes to GPU hours depending on protein-sequence length, including around one GPU minute per model for three hundred eighty-four residues.

This creates the possibility of predicting structures at the proteome scale and beyond, including application to the entire human proteome in a companion paper. The explosion in available genomic sequencing techniques and data has revolutionized bioinformatics, but the intrinsic challenge of experimental structure determination has prevented a similar expansion in structural knowledge.

That contrast captures the broader opportunity: sequence data has expanded rapidly, while experimental structure determination remains the bottleneck. AlphaFold combines evolutionary information from multiple sequence alignments with geometric neural-network reasoning and iterative refinement.

Its results approach experimental accuracy broadly, while its weaknesses reveal where sequence information and molecular context still matter.