Google DeepMind Maps 9 Billion Possible DNA Variants

Google DeepMind announced the creation and public release of the AlphaGenome Atlas on 8 September, an online repository of precomputed predictions made using its AlphaGenome model. The Atlas covers all 9 billion possible single-letter changes to a reference human genome, saving researchers from selecting variants, writing code and running the computationally demanding model themselves.

The AlphaGenome model was originally announced in 2025, and a Nature paper in January provided more details; the model was released for public noncommercial use. It compares an original DNA sequence with an altered one and predicts how the change might affect gene expression and other regulatory activity.

The human genome contains roughly 3 billion base pairs, with three possible single-nucleotide substitutions at each position, producing 9 billion variants. The complete dataset is around 1 petabyte. Žiga Avsec, genomics lead at DeepMind, said early estimates indicated the team needed to improve calculation speed by a factor of 80 to compile the Atlas in a reasonable time. Techniques used included model distillation, GPU kernel optimization and elimination of redundant calculations.

The Atlas offers a more approachable interface, including a single-number impact score intended to show at a glance whether a variant is likely to be meaningful. Avsec said it could accelerate work in fundamental biology, disease research and treatment development. Pushmeet Kohli, VP of science at Google DeepMind, said understanding DNA is a grand challenge.

DeepMind's earlier work includes AlphaFold in 2020, which predicted protein three-dimensional structure from amino-acid sequences, and AlphaMissense in 2023, which predicted whether 71 million possible protein-altering variants were likely benign or pathogenic; results from both were placed in public databases.

Limitations exist: many diseases are associated with multiple genetic variants, and although AlphaGenome examines 1 million base pairs around a variant, some enhancers regulate genes over distances beyond the model's field of view, making their effects difficult to predict. Carl de Boer, a genomicist at the University of British Columbia who is not affiliated with DeepMind, said the resource seems useful, called AlphaGenome the field's leading model, and noted it is slow and computationally intensive. He said the impact score has a clear use but is probably easily misinterpreted. The Atlas is freely available for noncommercial research, with potential for commercial licensing.