Story
September 9, 2026
DeepMind Puts 9 Billion DNA Mutations in a Browser—But the Lab Still Has the Last Word
DeepMind presents AlphaGenome Atlas as a way to turn an overwhelming genetic search into a usable research tool, while outside researchers see potential to speed up rare-disease and non-coding DNA studies. The shared caveat is crucial: AI can prioritize the clues, but it cannot replace biological proof.
On Tuesday, Google DeepMind unveiled AlphaGenome Atlas, a searchable database that predicts the molecular impact of all nine billion possible single-letter substitutions in human DNA. The release targets a longstanding problem in genetics: scientists can read the genome, but often cannot tell which tiny changes matter for health and disease.1
DeepMind built the atlas by running its AlphaGenome model across a reference human genome and comparing each DNA letter with its three possible replacements. The result is meant to give researchers an immediate starting point for questions that previously demanded slow, variant-by-variant model runs or laboratory testing. Pushmeet Kohli, the company’s AI-for-science chief, cast the advance as a next chapter after the Human Genome Project: “We bought the book, but we did not understand how to read it.”2
The company says the tool can flag changes that may alter gene expression, splicing or protein production, including in the vast non-coding part of the genome that has been especially difficult to interpret. It also released an AlphaGenome Variant Impact score to help scientists rank which mutations deserve closer scrutiny. Google chief executive Sundar Pichai said the free academic resource works in an ordinary browser, “without any coding required.”
3
Early tests suggest why that ranking could matter. In one unresolved epilepsy case, researchers used the score to identify a DNM1 variant predicted to create an abnormal splice site; laboratory work later confirmed the prediction and the variant was reclassified as likely pathogenic. In a separate analysis of more than 54,000 UK Biobank participants, filtering rare non-coding variants by predicted molecular effects produced 22% more associations than an analysis without Atlas.2
But DeepMind’s own researchers draw a line between a promising map and a medical verdict. Genomics lead Žiga Avsec said the system can miss some variants, particularly in enhancers, and warned that its outputs are not “the universal truth.” Atlas is therefore a powerful triage tool—not a substitute for experiments or the full evidence needed for a clinical diagnosis.2