John Jumper · Nature 2021 · Computational model development and benchmark validation · n=?

Highly accurate protein structure prediction with AlphaFold

Cited 46735 times in the scientific literature.

Level 5 - mechanism / opinion, no new human data

Computational method development and benchmark validation (Level 5 by design analogy, not clinical CEBM)

OpenAlex W3177828909 · doi:10.1038/s41586-021-03819-2 · record verified 2026-08-27

What was done

The authors developed a redesigned neural network architecture, AlphaFold, designed to predict 3D protein structures directly from primary amino acid sequences. The model incorporates physical and biological structural constraints and utilizes multiple sequence alignments within a deep learning framework. Performance was evaluated in the blinded 14th Critical Assessment of protein Structure Prediction (CASP14) benchmark.

What was found

The abstract reports that AlphaFold regularly predicted protein structures with atomic accuracy, including targets with no known homologous structures. It demonstrated accuracy competitive with experimental methods in a majority of cases and significantly outperformed other computational approaches. The abstract does not provide specific numerical metrics or error bounds.

Why it matters

Predicting protein structures computationally at near-experimental accuracy helps bridge the massive gap between billions of known sequence entries and the relatively small number of experimentally determined structures, accelerating structural biology and functional characterization.

Limits

The abstract provides no quantitative error metrics, confidence intervals, or specific sample sizes (number of proteins tested in CASP14). It does not detail performance boundaries, computational resource requirements, or handling of multi-protein complexes and dynamic conformational states.

Cited by