AlphaGenome Atlas: 9 Billion Human DNA Variants

Google DeepMind has unveiled the AlphaGenome Atlas, a massive database housing precomputed molecular effect predictions and AVI scores for 9 billion human single-nucleotide variants, allowing researchers to instantly access complex genetic insights without on-demand processing bottlenecks.

The underlying AlphaGenome model, released in June 2025, predicts how DNA variants alter molecular processes such as gene expression and RNA splicing. While widely utilized, previous applications were restricted to analyzing one variant or region at a time.

The Atlas fundamentally changes this unit of work. The DeepMind team executed AlphaGenome across all 9 billion single-nucleotide variants and stored the outputs, resulting in a staggering 1-petabyte dataset—surpassing the AlphaFold Database, which holds over 200 million protein structures, by more than 30 times.

Testing 9 billion mutations experimentally is entirely unfeasible, and running large models on demand for individual candidate variants remains far too slow for genome-scale investigations. A comprehensive lookup table with integrated interpretation effectively removes both barriers.

Stock photo for illustration only, not from the actual event

According to the DeepMind team, the AVI score delivers best-in-class performance across numerous variant pathogenicity and rare disease benchmarks, with extensive metric evaluations detailed in their technical report.

Regarding deployment, the Atlas is queryable today for non-commercial research through an online portal and API, with commercial access on Google Cloud listed as coming soon. Meanwhile, the core AlphaGenome model remains accessible for academic use on GitHub and commercial use via Model Garden on Google Cloud.

Source: MarkTechPost