AlphaGenome Atlas launches predictive map of every possible human DNA single‑letter variant

TL;DR

Google DeepMind announced AlphaGenome Atlas, a free 1‑petabyte platform that provides pre‑computed molecular effect predictions and a unified AlphaGenome Variant Impact (AVI) score for every one of the ~9 billion possible single‑nucleotide variants in the human genome, dramatically expanding the reach of its AlphaGenome AI model for research and discovery.

What AlphaGenome Atlas Is

AlphaGenome Atlas is a pre‑computed dataset that covers every possible single‑letter DNA change (≈9 billion SNVs) across the human genome. It stores:

  • Molecular effect predictions for each variant across dozens of regulatory dimensions (e.g., chromatin accessibility, RNA splicing, gene expression) in hundreds of human and mouse cell types.
  • AlphaGenome Variant Impact (AVI) score – a single scalar that combines AlphaGenome and AlphaMissense predictions to rank variant impact in both coding and non‑coding regions.
  • AVI feature attributions – additive contributions from interpretable biological categories that explain why a variant receives its AVI score.
  • DNA sequence motif catalogue – >2,500 recurrent motifs with genomic coordinates, supporting motif‑level analyses.

The entire resource occupies ~1 petabyte, more than 30 × the size of the AlphaFold Database, and is accessible via a web portal, an open‑source API, and as a skill in Google Antigravity.

How the AVI Score Works

The AVI score condenses two deep‑learning models:

  1. AlphaGenome – predicts variant effects on regulatory processes.
  2. AlphaMissense – predicts protein‑altering impact of missense variants.

By integrating these signals, the AVI provides a best‑in‑class ranking metric on a wide range of pathogenicity and rare‑disease benchmarks. Feature attributions accompany each score, highlighting the most disrupted molecular processes (e.g., splicing, transcription, conservation).

Demonstrated Scientific Impact

Rare‑Disease Variant Prioritization

In collaboration with the GREGoR Consortium, Broad Institute researchers used the AVI score to prioritize candidate variants in unsolved rare‑disease cases. They identified a splice‑disrupting SNV in DNM1 linked to epileptic encephalopathy, and experimental screens validated the predicted functional consequence.

Enhancing Population‑Genetics Analyses

Gareth Hawkes applied AlphaGenome Atlas to whole‑genome data from >54 k UK Biobank participants. By grouping rare variants according to predicted molecular effects, he uncovered 22 % more non‑coding associations than standard GWAS, revealing regulatory variants influencing protein levels (e.g., PLA2G7, EGLN1) and body‑mass‑index‑related loci.

Motif‑Level Functional Annotation

Researchers at the Stowers Institute leveraged the motif catalogue to distinguish transcription‑factor binding motifs that affect only chromatin accessibility from those that also modulate transcription, providing finer‑grained regulatory insight.

Access and Availability

The underlying AlphaGenome model is already available for academic use on GitHub and via the AlphaGenome API.

Future Directions

AlphaGenome Atlas is positioned as a baseline resource; as AlphaGenome models improve, future releases will increase prediction accuracy and coverage. The team envisions tighter integration with agentic systems (e.g., Google Antigravity) to automate end‑to‑end scientific pipelines, from hypothesis generation to experimental design.

Community and Acknowledgements

The atlas was co‑designed with input from collaborators at the University of Exeter, Broad Institute, Boston Children’s Hospital, Stowers Institute, Harvard, MSKCC, MGH, and the University of Kansas Medical Center. A long list of contributors and technical supporters is acknowledged in the original blog post.

The AlphaGenome Atlas predictions are not intended for clinical diagnosis or treatment and have not been validated for clinical use.

Sources