DeepMind’s Atlas Turns 9 Billion DNA Guesses Into a Searchable Map

DeepMind’s Atlas Turns 9 Billion DNA Guesses Into a Searchable Map
On Tuesday, Google DeepMind released AlphaGenome Atlas, a searchable database predicting the molecular consequences of roughly 9 billion possible single-letter substitutions in human DNA. The launch targets a central problem in genetics: researchers can read the genome, but often cannot tell which tiny changes alter biology and which are harmless.

DeepMind built the Atlas by running its AlphaGenome model across a reference genome and comparing every DNA letter with its three alternatives. The result is a precomputed catalogue of predicted effects on gene regulation, expression and protein production—work that previously required laborious experiments or model runs one variant at a time. As the company’s Pushmeet Kohli put it, the Human Genome Project meant “we bought the book,” but “we did not understand how to read it.”

The company has made the tool free for non-commercial academic use through a browser-based portal, alongside a Variant Impact score intended to rank the mutations most worth pursuing. Google chief executive Sundar Pichai emphasized the accessibility pitch: the resource works “in a regular web browser, without any coding required.”

Early users argue the ranking system can make a practical difference. In work with the GREGoR rare-disease consortium, researchers used the Atlas to revisit an epilepsy case and flag a DNM1 variant whose predicted splicing effect was later confirmed in the laboratory. In a separate analysis of more than 54,000 UK Biobank genomes, filtering variants by their predicted molecular effect produced 22% more associations than an analysis without Atlas; researcher Gareth Hawkes said the tool could “shrink the haystack.”

Still, DeepMind draws a firm line between a useful lead and a clinical answer. Its scientists say the model performs better on some variant classes, including splicing and promoters, than on enhancers, and acknowledge that indirect effects and training-data gaps remain. AlphaGenome Atlas, in other words, may accelerate the first pass through genetic uncertainty—but laboratory evidence remains the final judge.

Continue reading https://foxvector.com/stories/01a09047-50dd-32d7-73a4-12dbbb63cef2