AlphaGenome Atlas maps nine billion DNA changes

The AlphaGenome Atlas builds directly on the AlphaGenome model introduced in 2025, which reads DNA segments one million letters long. Previously, each variant had to be queried individually; the new release precomputes roughly 27,000 prediction values per variant, turning a slow process into instant lookup. This shift matters most for the 98 percent of the genome that does not code for proteins, where hidden regulatory functions have been difficult to analyze.

The atlas's companion tool, AVI, deliberately uses fewer features than CADD but achieved stronger rankings in retrospective clinical tests. The epilepsy case from the GREGoR consortium illustrates the diagnostic pathway: an atlas prediction identified a splice-site error in a brain-specific gene version, explaining why earlier blood-based analysis missed it. A lab experiment confirmed the finding, leading researchers to recommend reclassifying the variant as likely disease-causing.

The real test will be clinical adoption. DeepMind acknowledges the atlas is only one link in an evidence chain, not a diagnostic itself. For population studies, grouping variants by predicted effect produced 22 percent more links to blood protein levels, suggesting datasets like UK Biobank may yield more discoveries with this approach. Whether hospitals and research institutions integrate these tools into routine workflows, and how the planned commercial version through Google Cloud performs, will determine its broader impact beyond the initial research community.