Google DeepMind maps the predicted impact of 9 billion DNA variants

Google DeepMind has introduced AlphaGenome Atlas, a platform that predicts the molecular effects of all 9 billion possible single-nucleotide variants in the human genome. The resource is available free for academic research through a web portal, API, and Google Antigravity skill.

The release tackles a core challenge in genetics: testing every possible DNA change experimentally is effectively impossible. AlphaGenome Atlas precomputes thousands of predictions per variant across gene regulation, including data from hundreds of human and mouse cell types and tissues.

A ranking system for genetic variants

The platform also introduces the AlphaGenome Variant Impact (AVI) score. It combines predictions from AlphaGenome with AlphaMissense, Google DeepMind’s model for protein-altering variants, producing one score researchers can use to rank variants by likely biological impact. Feature attributions show whether predicted effects are linked to processes such as RNA splicing, gene expression, chromatin accessibility, or conservation.

That matters especially outside protein-coding DNA. Coding regions account for only about 2% of the genome, while the remaining 98% contains regulatory sequences where many trait-associated variants are found. Researchers can also explore more than 2,500 recurring DNA motifs and connect them to affected functional sequences.

The underlying dataset is approximately 1 petabyte—over 30 times larger than the AlphaFold Database. Google DeepMind says early collaborators have used it to prioritize variants in unsolved rare-disease cases, experimentally verify candidates, and identify rare variants linked to common traits. For AI and biology builders, the project demonstrates how precomputed model outputs can turn an otherwise inaccessible, genome-scale inference task into a practical discovery and prototyping tool.

Source: Google DeepMind Blog