DeepMind's AlphaGenome Atlas maps the impact of all 9B single-letter DNA changes

Google DeepMind has published AlphaGenome Atlas, a precomputed database that scores the predicted regulatory effect of every possible single-nucleotide variant in the human genome. Running its AlphaGenome model across all roughly 9 billion single-letter changes produced a one-petabyte dataset, with the stated goal of illuminating the ~98% of non-coding DNA whose function is still poorly understood. Rather than forcing researchers to interpret thousands of raw predictions, the Atlas condenses coding and non-coding effects into a single AlphaGenome Variant Impact (AVI) score for prioritizing which variants merit follow-up.

Early collaborators point to concrete payoffs. A Broad Institute team used AVI scores to flag a DNM1 variant predicted to introduce a faulty splice site, adding evidence that helped close an unsolved rare-disease case. Separately, a researcher applying the Atlas to more than 54,000 UK Biobank participants reported 22% more non-coding associations by grouping variants on predicted molecular effect, and narrowed the top 1% of impactful variants down to 19 genomic regions tied to BMI.

DeepMind is releasing the Atlas through a no-code web portal aimed at clinicians and biologists, framing it as a way to democratize access to genome-scale variant interpretation. The significance is less a single discovery than the shift to precomputed, queryable AI predictions as standard infrastructure for genomics — though, as with any model-derived catalog, the outputs are prioritization hypotheses that still require experimental validation.