AlphaGenome Atlas: Google DeepMind Maps 9 Billion Genetic Variants at Petabyte Scale
Google DeepMind has released AlphaGenome Atlas, a precomputed dataset covering 9 billion single-nucleotide variants across the human genome, built to give researchers a shared starting point for interpreting genetic variation instead of running predictions one variant at a time.
Google DeepMind has released AlphaGenome Atlas, a precomputed dataset covering 9 billion single-nucleotide variants across the human genome, built to give researchers a shared starting point for interpreting genetic variation instead of running predictions one variant at a time.
A petabyte of precomputed biology
AlphaGenome Atlas is built by running AlphaGenome predictions across the entire genome in advance, rather than requiring researchers to query the model variant by variant. The result is a 1-petabyte dataset spanning 9 billion single-nucleotide variants, more than 30 times the size of the AlphaFold Database. For each variant, the atlas stores thousands of molecular effect predictions, drawn from hundreds of human and mouse cell types and tissues, along with over 2,500 recurrent DNA sequence motifs that appear throughout the genome. That scale changes what a research question can look like: instead of asking a model to evaluate one candidate variant, a lab can now search across billions of precomputed results for patterns that would otherwise take enormous compute time to generate on demand.
A single score for variants that used to be hard to compare
Coding regions make up only 2% of the genome, with the remaining 98% classified as non-coding. Historically, tools for scoring the functional impact of a variant have treated these two categories differently, making it hard to compare a coding mutation against a non-coding one on common ground. The AlphaGenome Variant Impact score, or AVI, combines predictions from AlphaGenome and AlphaMissense specifically to support ranking across both coding and non-coding regions. Alongside the score itself, AVI produces feature attributions that point to the molecular processes a given variant is predicted to disrupt, giving researchers a starting explanation rather than just a number. That combination of a unified score and an interpretable attribution is what turns the atlas from a static reference into a tool for active hypothesis generation.
Where the score has already been put to work
DeepMind points to two concrete uses of AVI in ongoing research. In an unsolved rare disease study, researchers used the score to prioritize candidate variants and identified a DNM1 variant associated with epileptic encephalopathy; the predicted molecular effects of that variant were subsequently confirmed through experimental screens. Separately, an analysis of UK Biobank data covering more than 54,000 participants used AlphaGenome Atlas to identify 22% more non-coding genetic associations than were found without it. A related analysis narrowed in on the 1% of non-coding variants predicted to be most impactful and, from that narrowed set, identified 19 genetic regions linked to body mass index. Both cases show the atlas functioning as a filter that turns an intractably large search space into a shortlist worth testing further.
What the atlas is not, and who can use it now
DeepMind is explicit that AlphaGenome has not been validated or approved for clinical use, so none of these results translate directly into diagnostic or treatment decisions. What it does offer is a research tool: the atlas is available now for academic research through a free website portal, and separately for non-commercial use through that same website. Beyond the browser interface, the data can also be reached through the AlphaGenome API and through Google Antigravity, though DeepMind has not specified pricing or eligibility terms for those two routes. A commercial access path on Google Cloud is planned but not yet live. For a lab weighing whether to explore the atlas, the practical entry point today is the free website portal, with the API and Antigravity worth watching as those access terms become clearer.
Specs
- Dataset size: 1 petabyte
- Variants covered: 9 billion single-nucleotide variants
- DNA motifs: Over 2,500 recurrent sequences
- Model score: AlphaGenome Variant Impact (AVI)
- Coverage: Human and mouse cell types and tissues
- Access: Website portal, AlphaGenome API, and Google Antigravity
Sources
- AlphaGenome Atlas — Google DeepMind