Google DeepMind's genome prediction tool could unlock new disease treatments

Google DeepMind has released AlphaGenome Atlas, an AI-driven resource that predicts how nearly every possible single-letter change in human DNA could influence the body's molecular machinery. The tool aims to accelerate the discovery of disease-causing mutations and help researchers understand which genetic variations are most likely to affect health.

A predictive map of DNA variation#

The human genome operates with a four-letter chemical alphabet—A, C, G, and T—organized across roughly three billion base pairs that carry the instructions for making proteins and controlling gene activity. Within this vast sequence lie approximately nine billion potential single-letter substitutions, each one a candidate for study. The challenge researchers face is determining which changes are harmless, which drive human diversity, and which contribute to disease.

AlphaGenome Atlas contains predictions for all nine billion variants, estimating how each might reshape molecular biology—for instance, by altering the quantity of a protein produced. The company describes it as "the most comprehensive catalogue of how genetic mutations affect molecular biology."

Tools for navigating genetic complexity#

To help scientists identify the variants most worthy of investigation, Google is introducing a Variant Impact Score (AVI) that synthesizes predictions from multiple models to rank mutations by their potential biological significance. Researchers can access the atlas through a web portal, integrate it into Google's Antigravity agentic platform, or use the AlphaGenome interface directly.

Building on earlier breakthroughs#

This release extends work from prior DeepMind projects. AlphaGenome, unveiled last year, helped identify genetic disease drivers, while AlphaMissense focused on predicting how small mutations might damage proteins. The Atlas broadens that scope dramatically, covering non-coding regions that regulate gene behavior—regions far larger than the protein-coding sequences themselves.

Ziga Avsec, DeepMind's genomics lead, noted that "it took us some time to really precompute and also analyze this many variants because the space is so big." The resulting dataset spans roughly one petabyte of information, generated by training AlphaGenome on public human and mouse genome databases.