On September 19, 2023, DeepMind released AlphaMissense, a model that predicts whether a missense mutation - a single-letter DNA change that swaps one amino acid for another in a protein - is likely to be benign or to cause disease. The work was published the same week in Science as “Accurate proteome-wide missense variant effect prediction with AlphaMissense.”
AlphaMissense was adapted from AlphaFold, the protein-structure model, and fine-tuned on databases of human and primate population genetics. By combining structural context with evolutionary conservation, it scored every possible single amino-acid substitution in the human proteome without being explicitly trained on clinical labels.
The accompanying catalogue covered all 71 million possible human missense variants. AlphaMissense classified 89 percent of them as either likely benign or likely pathogenic - 57 percent likely benign and 32 percent likely pathogenic. DeepMind noted that before this work only about 0.1 percent of those variants had been clinically confirmed by human experts, making the predictions a potentially large resource for diagnosing rare genetic disease.
Like AlphaFold itself, AlphaMissense is a demonstration that a model built for one biological task can be repurposed to attack a related one, here moving from protein shape to the medical consequences of mutations.