Story
September 11, 2026
DeepMind’s Atlas Makes 9 Billion DNA Mutations Searchable—But Not Proven
DeepMind presents AlphaGenome Atlas as a way to turn an overwhelming genetic search problem into a practical research tool, while outside users see promise in its ability to prioritize rare-disease clues. Both the company and its testers draw the same boundary: predictions can guide experiments, not replace them.
Google DeepMind on Tuesday released AlphaGenome Atlas, a searchable database of AI predictions covering all nine billion possible single-letter substitutions in the human genome. The ambition is straightforward: help scientists identify which tiny DNA changes may disrupt gene regulation, protein production or other biological processes—and which can safely be ignored.1
The release builds on AlphaGenome, DeepMind’s earlier model for estimating the effects of genetic variants. Instead of making researchers run predictions one mutation at a time, Atlas precomputes them across a reference genome, pairing each variant with an average of roughly 27,000 molecular predictions across human and mouse cell and tissue types. It also introduces the AlphaGenome Variant Impact, or AVI, score to rank mutations by likely significance.2
Google’s pitch is accessibility as much as scale. Sundar Pichai said the interactive resource works “in a regular web browser, without any coding required,” and is free for academic researchers.
3 Commercial access, however, is expected later through Google Cloud licensing—a split that leaves the most immediate benefit with non-commercial science.
Early tests suggest the system can cut through the noise. In one retrospective analysis of unsolved rare-disease cases, researchers used Atlas to flag a DNM1 variant in a patient with epileptic encephalopathy; laboratory work later confirmed the prediction and the variant was reclassified as likely pathogenic. In a separate U.K. Biobank analysis, filtering variants by predicted molecular effect produced 22% more associations than an analysis without Atlas, according to the company’s account.2
Still, DeepMind’s own researchers caution against treating the map as a diagnosis engine. Genomics lead Žiga Avsec said the predictions are “accurate enough to really point us in the right direction,” but not “the universal truth.”2 The model performs better on some variant classes, including splicing and promoters, than on others such as enhancers. Atlas may shrink the haystack; scientists must still verify what they find.