DeepMind’s AlphaGenome Atlas Predicts the Impact of Every Possible Single-Letter DNA Change

Google DeepMind has used AlphaGenome Atlas to predict what roughly nine billion possible single-letter changes in the human genome could do inside the body. The atlas is designed to help researchers sort through millions of variants and focus on the few that may matter for disease. Its dataset spans one petabyte, more than 30 times the size of the AlphaFold database.

The human genome has about three billion DNA letters, and most individual variants are harmless. The hardest problem is identifying the small number that contribute to disease, especially in the roughly 98 percent of the genome that does not code for proteins. These noncoding regions act like switches and dials that influence when and where genes are active, making them central to many disease-linked variants.

DeepMind also built the AlphaGenome Variant Impact Score, or AVI, to condense thousands of predictions into one number. AVI combines AlphaGenome predictions with AlphaMissense and evolutionary conservation measures, using 18 input features. According to the article, AVI performed strongly against existing tools in tests on clinically classified variants, particularly in noncoding regions, while some tasks still favored competing approaches.

In one GREGoR consortium case, AVI elevated a previously unclear DNM1 variant in a child with severe epilepsy to the top of the candidate list. The AlphaGenome predictions suggested the variant created a faulty splice site in a brain-specific gene version, and a lab experiment confirmed the prediction. In population research, the atlas also helped group variants by predicted effect, turning up more links between noncoding variants and blood protein levels than conventional filters in work using UK Biobank data.

DeepMind describes AlphaGenome Atlas and AVI as research tools rather than standalone diagnostic systems. The article notes limits: AlphaGenome does not cover every cell type and can miss effects involving the amount of other regulatory proteins. The atlas is available for noncommercial use through a web portal, an API, and as a skill in Google Antigravity, with a commercial version planned through Google Cloud.