AlphaMissense catalogues 71 million genetic variants for disease risk

Google DeepMind published AlphaMissense, a model adapted from AlphaFold that predicts whether “missense” mutations — single-letter substitutions that change one amino acid in a protein — are likely to cause disease or are harmless. The team used it to score all 71 million possible missense variants in the human genome and released the full catalogue freely for research and commercial use, alongside a companion paper in Science.

Missense variants are common but usually of unknown significance: a given substitution might disrupt a protein’s function and contribute to disease, or might have no effect at all, and determining which requires either laboratory experiments or clinical case data that exist for only a small fraction of possible mutations. DeepMind said human experts had definitively classified roughly 0.1% of the 71 million variants; AlphaMissense’s model, fine-tuned on AlphaFold’s protein-structure predictions plus population-frequency data treating rare variants as more likely pathogenic, categorised 89% of them as either likely pathogenic or likely benign, and the company reported it outperformed prior computational predictors on clinical validation sets.

Because AlphaMissense produces predictions rather than confirmed diagnoses, its outputs were positioned as a triage and research tool — narrowing the enormous space of uncertain variants down to a manageable shortlist for clinical geneticists and researchers to investigate, rather than a diagnostic instrument to be used on its own. The release extended DeepMind’s earlier AlphaFold work into a second concrete biomedical application within three years of AlphaFold’s 2021 release, and it became one of the most cited examples, alongside AlphaFold itself, of AI research producing an openly released scientific resource with direct clinical relevance rather than a narrow benchmark result.