Highly accurate protein structure prediction with AlphaFold

Despite decades of experimental effort, the structures of only around 100,000 unique proteins have been determined, a small fraction of the billions of known protein sequences, because structural coverage is bottlenecked by the months to years needed to solve a single protein structure. Predicting the three-dimensional structure a protein will adopt from its amino acid sequence alone, the structure prediction component of the 'protein folding problem,' has been an important open research problem for more than 50 years, and despite recent progress, existing methods fell far short of atomic accuracy, especially absent a homologous structure.

The paper presents an entirely redesigned version of AlphaFold, validated in the 14th Critical Assessment of protein Structure Prediction (CASP14), as the first computational method able to regularly predict protein structures with atomic accuracy even when no similar structure is known, demonstrating accuracy competitive with experimental structures in the majority of cases and greatly outperforming other methods. This is underpinned by a novel deep learning architecture that incorporates physical and biological knowledge about protein structure, leveraging multi-sequence alignments in its design.