From the source
Google DeepMind is explaining how its AlphaGenome model and AlphaGenome Atlas dataset help researchers narrow down which DNA changes cause disease.
- The genome holds 3 billion base pairs, and any single letter can be swapped for three others, leaving 9 billion possible single-letter changes that no lab could test one by one.
- DeepMind says it used AlphaGenome to predict the outcomes of all 9 billion changes and built AlphaGenome Atlas so scientists can decide what to research faster.
- Atlas returns a single AVI (AlphaGenome Variant Impact) score showing a variant's potential impact across the full genome, which helps researchers prioritize the most disruptive changes.
- AlphaGenome launched in June 2025 and Atlas in September, and DeepMind has since launched AlphaGenome 2, which tackles cell type specificity by moving beyond tissue-level averages to predict how variants disrupt individual cell types.
- Gareth Hawkes at the University of Exeter used Atlas to narrow 1 billion variants, focusing on the most impactful 1% to identify 19 variant groups linked to body mass index.
- Laura Covill at the Broad Institute used the AVI score on unsolved rare diseases, where Atlas flagged a critical variant in a gene linked to neurodevelopmental disorders and helped solve the case.





