AlphaGenome Atlas precomputes predictions for 9 billion single-nucleotide variants in 1 PB dataset

DeepMind released AlphaGenome Atlas on 8 September 2026 as a 1-petabyte resource containing thousands of regulatory predictions per variant across hundreds of human and mouse cell types. The release bundles an AVI score that merges AlphaGenome outputs with AlphaMissense protein-impact scores and supplies feature attributions for each variant. External collaborators have already used the atlas to flag and experimentally confirm variants in unsolved rare-disease cases.

The dataset exceeds the AlphaFold Database scale by more than 30 times. Where AlphaFold expanded coverage from 190 000 experimental structures to 200 million predictions, AlphaGenome Atlas shifts from per-variant inference to exhaustive pre-computation. No independent benchmark numbers for variant-effect accuracy appear in the announcement; only internal consistency checks and limited collaborator verifications are cited.

Operationally the portal, API and Antigravity skill lower the barrier for non-computational labs to rank variants by predicted impact. The absence of disclosed test-set performance or prospective validation cohorts means downstream users must still generate their own ground truth. Next milestones will be measured by the number of peer-reviewed studies that both cite the AVI score and report orthogonal experimental confirmation within 24 months.