On Tuesday, Google DeepMind unveiled AlphaGenome Atlas, a database built by running its AlphaGenome model across every possible single-letter substitution in the human reference genome: roughly 9 billion potential variants. The company calls it the most comprehensive catalogue yet of how mutations may affect molecular biology, with predictions spanning gene regulation, splicing and protein effects.¹

The pitch is speed. Instead of testing variants one at a time in a laboratory—or waiting for a model to calculate them—researchers can search precomputed results, including an AlphaGenome Variant Impact score designed to rank the most consequential leads. The atlas is free to academic researchers through a browser-based portal, API and Google’s Antigravity platform.¹

Google chief executive Sundar Pichai framed that access as central to the release, saying the tool works “in a regular web browser, without any coding required,” and is free for academics.Sundar Pichai (sundarpichai)² The commercial path is less clear: DeepMind says licensing through Google Cloud is coming, while its drug-discovery sibling Isomorphic Labs will also need a commercial licence.³

Early demonstrations give the company’s argument some weight. In rare-disease work with the GREGoR Consortium, Atlas helped flag a DNM1 variant linked to epileptic encephalopathy; laboratory experiments later supported the prediction. In a separate UK Biobank analysis, filtering by predicted molecular effect produced 22% more associations than an analysis without Atlas.³

But the atlas is a compass, not a diagnosis. DeepMind’s own genomics lead, Žiga Avsec, said the predictions can point downstream studies “in the right direction” but are not “the universal truth.”³ Independent scrutiny is sharper: because the model is trained on limited cell types and existing datasets, it remains unclear whether it delivers insight beyond those inputs. As one assessment put it, “At the moment, it’s not clear whether we’re there yet.”⁴