Record: AlphaGenome Atlas precomputes effects for 9 billion human DNA SNVs

Claim of novelty

DeepMind claims AlphaGenome Atlas is a predictive map of molecular effects for every possible single-letter DNA change in the human genome: on the order of 9 billion single-nucleotide variants, released as a platform detailed in an accompanying paper (linked from the post as "our paper"; URL not separately extracted in this Record body).

Supporting product claims in the same post:

• Dataset size about 1 petabyte, described as more than 30× larger than the AlphaFold Database.
• AVI score: one number per variant combining AlphaGenome regulatory predictions and AlphaMissense protein-impact signal, aimed at coding and non-coding regions.
• AVI feature attributions that decompose the score into categories such as chromatin accessibility, splicing, and conservation.
• >2,500 recurrent DNA sequence motifs with locations.
• Delivery surfaces: intuitive academic website portal, AlphaGenome API, and a skill in Google Antigravity.

What was measured and on which data

The public post is a science communication layer. It is not a full methods table. Measured or reported items stated there:

• Precomputed AlphaGenome molecular-effect predictions genome-wide across "thousands" of molecular tracks spanning hundreds of human and mouse cell types and tissues (exact track count not enumerated in the post body we retrieved).
• Collaborator application: GREGoR Consortium / Broad Institute team (Laura Covill, Anne O'Donnell-Luria and colleagues) used AVI to prioritize variants in unsolved rare disease work and report a DNM1 splice-affecting variant tied to epileptic encephalopathy, with experimental screens said to validate the prediction and nearby similar variants.
• Collaborator application: Gareth Hawkes (University of Exeter) applied Atlas to whole-genome data from over 54,000 UK Biobank participants; DeepMind states this surfaced 22% more non-coding genetic associations than without the grouping approach, including regulatory variants affecting PLA2G7 and EGLN1 protein levels, and a BMI-focused pass over the top 1% AVI non-coding variants that identified 19 genetic regions.
• Motif resource used by Zeitlinger and Weilert (Stowers) to categorize transcription-factor effects on accessibility versus gene activation.

DeepMind also states internal testing shows AVI performance on "many variant pathogenicity and rare disease benchmarks." The post does not paste the benchmark names, splits, or numeric leaderboard in the HTML we retrieved.

Baseline fairness

Independent baselines are only partly specified in the blog:

• Size comparison is to the AlphaFold Database, a different modality (structures vs variant effects). Useful as a storage analogy. It is not a fairness baseline for pathogenicity ranking.
• Collaborator lifts (22% more associations; DNM1 find) are before/after or case narratives inside partner workflows. They are not a public, locked leaderboard with a named competing method run by DeepMind under identical filters.
• "Best-in-class" wording appears in the vendor post. Treat it as DeepMind's characterization until the paper's tables are checked.

What was not tested

Not disclosed in the blog post we retrieved:

• Clinical diagnostic claims or regulatory clearance.
• Prospective trial endpoints.
• Full per-benchmark numeric tables for AVI versus named competitors.
• Exact commercial SKU, price, or Cloud GA date ("soon").
• Whether every one of the 9 billion cells is equally calibrated in every tissue track.

Code / weights / data public?

Monday use or why not

• Academic rare-disease or statistical-genetics lab: Open the portal, rank a candidate list with AVI, and read attributions before wet-lab follow-up. That is the intended Monday path.
• Hospital diagnostic pipeline: Do not treat Atlas scores as a standalone diagnosis. Human genetics review still owns the call.
• Commercial biotech on Cloud: Wait for the stated Cloud commercial path and your counsel's DPA review; "soon" is not a date.
• General AI buyers: No general LLM SKU here. Ignore unless genomics is in scope.

Relation to the last paper in the line

Atlas sits on AlphaGenome (variant effect model already in research use) and folds in AlphaMissense for protein-altering impact. DeepMind frames Atlas as the precomputed, atlas-scale distribution layer (compare to how the AlphaFold Database redistributed structure predictions). The post positions Atlas as a baseline map that should improve as AlphaGenome weights improve, and as an ingredient for agentic science tools such as Google Antigravity.