For more than fifty years, protein folding resisted reliable computation. AlphaFold changed the benchmark by predicting structures at near experimental accuracy, even when no similar structure was known. Proteins are essential to life, and understanding their structure can facilitate a mechanistic understanding of their function.

Yet the structures of around 100,000 unique proteins represent 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 to address this gap and enable large-scale structural bioinformatics.

Predicting the three-dimensional structure from amino acid sequence has been an important open research problem for more than fifty years, and existing methods fall far short of atomic accuracy, especially when no homologous structure is available. AlphaFold is presented as the first computational method that can regularly predict protein structures with atomic accuracy even when no similar structure is known.

Computational prediction from protein sequence has 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 it remains highly challenging for moderate-sized proteins because molecular simulation is computationally intractable and models of protein physics are difficult to make sufficiently accurate.

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 designed to predict protein structures to near experimental accuracy in a majority of cases.

AlphaFold was entered into the CASP14 assessment from May to July twenty twenty, under the team name AlphaFold two. AlphaFold two was a completely different model from the CASP thirteen AlphaFold system. CASP is carried out biennially using recently solved structures that have not been deposited in the Protein Data Bank or publicly disclosed, making it a blind test.

AlphaFold greatly improves structure prediction by incorporating novel neural network architectures and training procedures based on evolutionary, physical, and geometric constraints of protein structures. The architecture jointly embeds multiple sequence alignments and pairwise features, and uses a new output representation and associated loss for accurate end-to-end structure prediction.

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

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

Its key innovations are mechanisms that exchange information within the multiple sequence alignment and pair representations, enabling direct reasoning about spatial and evolutionary relationships. The trunk is followed by a structure module that 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 includes a novel equivariant transformer for reasoning about unrepresented side-chain atoms and a loss term that places substantial weight on orientational correctness.

The network repeatedly feeds outputs recursively into the same modules; this whole-network iterative refinement is called recycling and contributes markedly to accuracy with minor extra training time. 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 pair representation encodes information about relations between residues, while multiple sequence alignment 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 update operations applied in series.

AlphaFold can train to high accuracy using only supervised learning on Protein Data Bank 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 Uniclust thirty, producing a new dataset filtered to a high-confidence subset.

The same architecture is then trained again from scratch using a mixture of Protein Data Bank data and this new dataset of predicted structures. This self-distillation procedure makes effective use of unlabelled sequence data and considerably improves the accuracy of the resulting network.

Figure one combines CASP fourteen benchmarking with structural examples and the AlphaFold architecture. In panel a, AlphaFold has a median C-alpha r.m.s.d. ninety-five of zero point nine angstroms, compared with two point eight angstroms for the next best method; panels b through d show close agreement for individual targets, including a zinc-binding site and correct domain packing in a two-thousand-one-hundred-eighty-residue chain.

Panel e traces the sequence, MSA, template, Evoformer, structure-module, and recycling information flow. In CASP fourteen, AlphaFold structures were vastly more accurate than competing methods. AlphaFold had a median backbone accuracy of zero point nine six angstroms r.m.s.d.

at ninety-five percent residue coverage. The next best-performing method had a median backbone accuracy of two point eight angstroms r.m.s.d. at ninety-five percent residue coverage. AlphaFold's all-atom accuracy was one point five angstroms r.m.s.d.

at ninety-five percent residue coverage, compared with three point five angstroms for the best alternative method. AlphaFold is scalable to very long proteins with accurate domains and domain packing, and the model provides precise per-residue estimates of reliability.

Figure two evaluates AlphaFold on recent PDB chains excluded from its training structures. Panel a reports a median full-chain backbone RMSD of one point four six ångströms across three thousand one hundred forty-four chains, while panel b links increasing lDDT-Cα with the fraction of correctly predicted side-chain rotamers.

Panels c and d show that predicted confidence tracks observed accuracy: pLDDT correlates with lDDT-Cα at Pearson’s r of zero point seven six, and pTM correlates with full-chain TM-score at zero point eight five. All structures in this dataset were deposited in the Protein Data Bank after the training data cut-off and were analysed as full chains.

High side-chain accuracy is observed when backbone prediction is accurate, and pLDDT reliably predicts the lDDT-C-alpha accuracy of the corresponding prediction. The analyses validate that AlphaFold's high accuracy and reliability on CASP fourteen proteins transfers to an uncurated collection of recent Protein Data Bank submissions.

Figure four disentangles which AlphaFold components contribute to accuracy across eighty-seven CASP14 domains and two thousand two hundred sixty-one difficult PDB chains, reporting mean score differences from three baseline seeds with ninety-five percent bootstrap intervals.

The ablations implicate templates, MSA processing, invariant point attention, recycling, and auxiliary objectives as distinct contributors. Panel b shows that the two T1024 domains stabilize early, whereas T1064 undergoes repeated structural changes and approaches its final domain GDT only near the end of four recycling iterations, spanning one hundred ninety-two Evoformer blocks.

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

AlphaFold 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.

The methodology combines bioinformatics and physical approaches by using a physical and geometric inductive bias in components that learn from Protein Data Bank data with minimal handcrafted features. AlphaFold can build hydrogen bonds effectively without a hydrogen bond score function.

This results in a network that learns more efficiently from the limited data in the Protein Data Bank while coping with the complexity and variety of structural data. AlphaFold has demonstrated utility for molecular replacement and for interpreting cryogenic electron microscopy maps.

Because AlphaFold outputs protein coordinates directly, it produces predictions in graphics processing unit minutes to graphics processing unit hours depending on protein sequence length; for example, around one graphics processing unit minute per model for three hundred eighty-four residues.

This opens the possibility of predicting structures at proteome scale and beyond, including application to the entire human proteome in a companion paper. AlphaFold combines evolutionary information, physical and geometric inductive bias, and iterative refinement to produce highly accurate structures at scale—while remaining weaker when alignments are shallow or proteins depend on heterotypic contacts.