DeepMind's AlphaMissense AI Tool Predicts Pathogenicity of 71 Million Missense Mutations

On September 20, DeepMind, a subsidiary of Google, announced its new AI tool called AlphaMissense, which has evaluated all 71 million possible missense mutations; the tool is now on GitHub and the results have been published in the journal Science.

Human experts have so far identified clinical effects for only about 0.1% of missense variants, which alter the structure of proteins, the main work molecules of the body. Among these variants, 32% are classified as possibly pathogenic, 57% as benign, and the remainder as uncertain.

AlphaMissense predictions display two protein structures with mutations, where red indicates harmful, blue indicates benign, and gray indicates uncertain.

Project researcher Žiga Avsec from DeepMind’s London office said: "Previously, discovering pathogenic variants required expensive and labor‑intensive experiments, each protein being unique and each assay needing to be designed separately, often taking months. Now, using AI predictions, researchers can preview results for thousands of proteins at once, helping prioritize resources and accelerate more complex studies."

Avsec also noted that AlphaMissense has limitations; its pathogenicity predictions do not reveal the biophysical properties of the variants, but he remains optimistic about its prospects.

Researcher Cheng Jun added, "These AI predictions were never intended for standalone clinical diagnosis; they should always be used alongside other evidence, yet they will still help improve rare‑disease diagnosis rates and assist researchers in discovering new disease‑causing genes."