DeepMind's AlphaGenome Atlas scores every possible human DNA change

Google DeepMind released AlphaGenome Atlas, a free database that predicts the molecular effect of every possible single-letter change to human DNA. That works out to about 9 billion predictions stored across a petabyte of data, roughly thirty times the size of the AlphaFold Database. For each variant the Atlas gives thousands of predicted effects on gene regulation across hundreds of human and mouse cell types, then rolls them into a single AVI score that ranks how much a change is likely to matter. The point that sets it apart is coverage of the 98 percent of the genome that does not code for proteins, where most disease-linked variants sit and where interpretation has always been hardest.

DeepMind says the Atlas has already surfaced results researchers had missed. It flagged overlooked epilepsy-linked variants in the DNM1 gene, and across more than 54,000 UK Biobank participants it found 22 percent more non-coding associations affecting protein levels. It also pointed to 19 genomic regions that may influence body mass index. The scores are precomputed AlphaGenome outputs combined with AlphaMissense, reachable through a web portal, an API, or Google Antigravity. DeepMind is clear that none of this is medical advice: the Atlas has no clinical validation, so a high score is a lead to chase, not a diagnosis.

Why it matters

For genetics researchers, the slow step is often guessing which of thousands of candidate variants deserves lab time. A precomputed impact ranking that also covers non-coding DNA gives a concrete place to start, as long as you treat the scores as hypotheses to test rather than answers.