AlphaGenome Atlas: DeepMind scores every possible human DNA change

A free API and 1PB dataset rank all 9 billion single-letter variants — coding and non-coding — by one AVI score, now a skill inside Google Antigravity.

Nowline SEP 10 2:00 AM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • One AVI score for any variant, coding or not

    The AlphaGenome Variant Impact score fuses AlphaGenome's regulatory predictions with AlphaMissense into a single number, so you can rank a non-coding tweak right next to a missense mutation. Feature attributions decompose each score into splicing, chromatin, conservation and protein effects.

  • Free today: web portal, GitHub API, Antigravity skill

    Query alphagenome.google/atlas free for non-commercial use, call the AlphaGenome API on GitHub, or invoke it as a scientific skill inside Google Antigravity. Commercial access via Google Cloud Model Garden is 'coming soon,' and clinical use is explicitly not approved.

  • 9 billion variants, precomputed — no GPU on you

    The Atlas covers every possible single-letter change in the human genome in a 1-petabyte dataset, ~30x the AlphaFold Database, with thousands of predictions per variant across hundreds of cell types and tissues. It's precomputed, so a lookup costs you nothing in inference.

  • Build this weekend: a variant-triage app

    Wire the API into a tool that ranks candidate mutations by AVI and shows what each one disrupts — the same move researchers used to pin a DNM1 variant behind an epileptic encephalopathy. A bundled compendium of 2,500+ DNA motifs gives regulatory-element tooling a running start.

  • The proof it works: 22% more UK Biobank hits

    Rescoring 54,000 UK Biobank genomes with the Atlas surfaced 22% more non-coding associations and 19 fresh genome regions tied to BMI — signal that variant-effect ranking now catches things older annotation missed.

  • Elsewhere: OpenAI claims a Millennium Prize proof

    OpenAI says an unreleased internal model ran ~10,000 coordinating agents for 88 hours to produce a Navier–Stokes solution with a Lean-formalized proof, though a mathematician disputes how it learned of concurrent work. Nothing shippable yet, but a marker for where agent swarms are heading.