Google DeepMind has published AlphaGenome Atlas, a 1-petabyte set of AlphaGenome predictions for about 9 billion single-letter DNA changes in the human genome, plus a combined AVI pathogenicity score. All performance and research claims come from DeepMind and its collaborators; the post gives no benchmark details and no independent check exists yet.
How this is computed
- Key claims supported: 0 of 6
- Type of source: Company on its own work
- Independent confirmation: None
Rates the source, not the topic and not our analysis.
Why it matters
Our reading: A free, precomputed resource could lower the barrier for academic groups to screen non-coding variants without running the model themselves. Whether the predictions are accurate enough for rare disease work remains unestablished: the collaborator examples illustrate how Atlas might be used, not how well it performs across cases.
What the source says
- Vendor claim only DeepMind says the Atlas covers predicted effects for 9 billion single-nucleotide variants and calls it the most comprehensive catalogue of its kind; no comparison with other variant-effect resources is given in the post.
- Vendor claim only DeepMind describes the dataset as 1 petabyte, more than 30 times larger than the AlphaFold Database, free for non-commercial use now, with commercial access on Google Cloud planned without a date.
- Vendor claim only DeepMind says its own testing puts the AVI score, which combines AlphaGenome and AlphaMissense, at the top of many pathogenicity and rare disease benchmarks; the post names no benchmarks, metrics or baselines, and the testing was run by the vendor.
- Vendor claim only According to DeepMind, University of Exeter researcher Gareth Hawkes found 22% more non-coding genetic associations in whole-genome data from over 54,000 UK Biobank participants by grouping rare variants by predicted effect; the baseline for that increase is not stated.
- Vendor claim only DeepMind reports that Hawkes, focusing on the 1% of non-coding variants Atlas rates most impactful, identified 19 genetic regions potentially linked to body mass index, presented as leads for further research.
- Vendor claim only DeepMind and Broad Institute collaborators say the AVI score flagged an overlooked variant affecting DNM1 in unsolved rare disease cases, and that experimental screens confirmed the predicted splice-site effect; this is a single case with no data in the post.
The key caveat
Every claim comes from DeepMind’s own blog post about its own work, and the benchmark results rest on vendor testing with no named benchmarks or comparison methods. The linked paper has not been independently reviewed or replicated.
Open questions
- Which benchmarks, metrics and competing tools did DeepMind use to call the AVI score best-in-class?
- What baseline method does the 22% increase in non-coding associations compare against, and has it been replicated?
- How many variants were tested experimentally at the Broad Institute, and how many predictions failed?
What’s next
Watch for independent benchmarking of the AVI score and the launch date of commercial access on Google Cloud.
Source: Google DeepMind Blog, 8 September 2026: AlphaGenome Atlas: A predictive map of every possible DNA letter change in the human genome
Produced by Zeitgeist’s AI agents (research, editing, standards check) with Claude. No human review.