DeepMind Maps 9 Billion DNA Variants with AlphaGenome Atlas

Google DeepMind has publicly released the AlphaGenome Atlas, a public database of precalculated predictions for all 9 billion possible single-letter changes to the human genome. Built using the company's AlphaGenome AI model—originally announced in 2025 and detailed in a January Nature paper—the new database is freely available for noncommercial research, with options for commercial licensing. By mapping these variations across the 3 billion base pairs of human DNA, the tool aims to show how mutations in noncoding regions affect gene regulation and expression.

Compiling this 1-petabyte dataset required massive computational optimization. The human genome contains three potential nucleotide swaps at every position, totaling 9 billion variants. To make the project feasible, DeepMind engineers had to accelerate their calculation speeds by a factor of 80. They achieved this benchmark by distilling the model, optimizing GPU kernels, and cutting out repetitive computations. The underlying AlphaGenome model analyzes a substantial window of 1 million base pairs surrounding each genetic variant to make its predictions.

For genomic practitioners, the Atlas eliminates the need to write complex code or run computationally demanding simulations themselves. Instead, scientists can access 11 different output types through an approachable web interface. The platform also introduces a highly anticipated one-number impact score to help researchers quickly identify meaningful variants. While Carl de Boer, a genomicist at the University of British Columbia, warned that this simplified metric is "going to be easily misinterpreted" given the complexity of the genome, he noted that the Atlas provides a highly useful resource, especially for researchers lacking access to advanced hardware.

The project builds on DeepMind's previous biological AI milestones, including the 2020 protein-folding model AlphaFold and the 2023 AlphaMissense tool, which predicted the effects of 71 million protein-altering variants. Although the AlphaGenome Atlas will help scientists prioritize laboratory experiments, it has limitations. It cannot easily predict long-distance genetic interactions beyond its 1-million-base-pair window, nor can it easily map complex diseases associated with multiple simultaneous mutations. Nonetheless, DeepMind genomics lead Žiga Avsec believes the resource will significantly accelerate disease research and treatment development.