Highly accurate protein structure prediction with AlphaFold

Written for the Research library (claude-opus-5-5) from the paper text. Original — nature.com/articles/s41586-021-03819-2.pdf

Summary

Experimentally determining a protein's 3D shape can take months or years. As a result, only about 100,000 unique proteins have solved structures, while billions of sequences are known. Predicting structure from the amino acid sequence alone has been an open problem for roughly half a century. Earlier computational methods fell well short of atomic accuracy whenever no homologous structure existed. The authors present a fully redesigned AlphaFold, a deep network that takes a sequence together with multiple sequence alignments (MSAs) and structural templates and directly outputs coordinates for all heavy atoms. Its core is the Evoformer, which jointly refines an MSA representation and a residue-pair representation before a structure module builds the 3D model. Training adds several ideas, including self-distillation on unlabelled sequences and a masked MSA loss. In the blind CASP14 assessment, AlphaFold matched experimental quality for most targets and far outperformed competing methods. This accuracy also held on recently released PDB chains. The model reports per-residue confidence, which makes its predictions usable for large-scale structural biology.

Key points

• CASP14 results (entered as "AlphaFold2"): median backbone error was 0.96 Å r.m.s.d.95, compared with 2.8 Å for the next-best method. All-atom error was 1.5 Å versus 3.5 Å. For scale, a carbon atom is about 1.4 Å wide.
• Architecture: a 48-block Evoformer feeds an 8-block structure module, and the whole network is recycled three times. Inputs come from genetic database searches (UniRef90, BFD, Uniclust30, MGnify) and template search (PDB70).
• Generalisation: on 3,144 recent PDB chains with low similarity to the training data, the median full-chain Cα r.m.s.d.95 was 1.46 Å. Training used the PDB snapshot of 28/08/2019.
• Confidence estimates: pLDDT tracks true lDDT-Cα (Pearson's r = 0.76), and pTM tracks TM-score (r = 0.85), both measured on 10,795 chains.
• Scalability: the model correctly predicted domain packing for T1044, a 2,180-residue single chain. It also produced accurate side chains, for example at a zinc-binding site, even though it does not model the ion explicitly.