Google DeepMind launched AlphaGenome Atlas on Tuesday, putting AI-generated predictions for every possible single-letter change in the human genome into a searchable database. The company’s premise is blunt: with roughly 9 billion possible substitutions, testing each mutation in a laboratory is “practically impossible.”¹

Atlas precomputes how variants may alter gene regulation, protein production and other molecular processes across hundreds of human and mouse cell and tissue types. DeepMind says its new AlphaGenome Variant Impact, or AVI, score condenses those predictions with results from its AlphaMissense protein model, allowing researchers to rank mutations and see whether splicing, gene expression or protein change is driving the result.¹

The release is also a bid to make that computational firepower less exclusive. Sundar Pichai said the browser-based tool requires no coding and is free to academic researchers, calling it “an interactive resource” for mapping the predicted impact of all 9 billion variants.Sundar Pichai (sundarpichai)² Commercial access, however, is expected through a Google Cloud licensing arrangement, with terms still unspecified.³

The most compelling early examples come from disease research. In work with the GREGoR Consortium, researchers used AVI to revisit unsolved rare-disease cases and identified a likely pathogenic DNM1 variant in a patient with epileptic encephalopathy; laboratory experiments then confirmed the model’s splicing prediction.³ In a separate UK Biobank analysis of more than 54,000 people, filtering by predicted molecular effects produced 22% more associations than an analysis without Atlas.³

Still, DeepMind’s own researchers draw a firm line between prioritisation and diagnosis. Genomics lead Žiga Avsec said the outputs are “accurate enough to really point us in the right direction,” but researchers should not treat them as “the universal truth.”³ The atlas may shrink the haystack; it does not eliminate the need to find the needle in the lab.