Google DeepMind Maps 9 Billion Possible DNA Variants

Google DeepMind has turned one of genomics’ most tedious tasks into a database. On 8 September, it announced the AlphaGenome Atlas, a public repository of precomputed predictions for all 9 billion possible single-letter changes in a reference human genome.

The work builds on AlphaGenome, the DNA model DeepMind announced in 2025 and described in a Nature paper in January. The model compares an original DNA sequence with an altered one and predicts how that change may affect gene expression and other regulatory activity. Until now, researchers had to pick variants themselves, write code, and run the model on their own hardware. The Atlas removes that friction.

That matters because most of the genome is non-coding, and the regulatory parts are messy. Some segments act close to a gene, others work from far away, and their effects can change across cell types and tissues. Carl de Boer, a genomicist at the University of British Columbia, calls understanding these DNA changes “fundamental to understanding most disease.” DeepMind is betting that easier access to predictions will help scientists narrow down what to test in the lab.

The scale is absurdly large. The human genome has roughly three billion base pairs, and each position can be swapped in three different ways, which gets you to the Atlas’ 9 billion variants. DeepMind says the full dataset is around 1 petabyte. The company also says early estimates suggested the team would need an 80x speedup to make the project practical, so it used model distillation, GPU kernel optimization, and other cuts to redundant computation.

The Atlas also adds a single-number impact score, meant to give researchers a quick sense of whether a variant might matter. That should make the tool easier to use, but de Boer warns it can be misleading because the biology is still complicated. AlphaGenome examines 1 million base pairs around each variant, yet some enhancers work over distances beyond that window. The Atlas is freely available for noncommercial research, with commercial licensing possible, and DeepMind says it can save people from repeating the same simulations over and over.