AlphaGenome Atlas
Vintage: 2026-09. Primary evidence is the DeepMind issuer blog AlphaGenome Atlas: A predictive map of every possible DNA letter change in the human genome (dated September 8, 2026) plus official @GoogleDeepMind X text, as synthesized in 2026-09-08-x-10am-deepmind-alphagenome-atlas-9b-variants-addy-osmani. Lead X post has untranscribed video — flagged for
/transcribe-clipping; do not treat video as grain. Atlas is a precomputed map + AVI layer on existing alphagenome, not a first-time model invent today. Distinct from world-labs-atlas.
AlphaGenome Atlas
One-line summary: DeepMind’s 8 Sep 2026 1-petabyte precomputed catalogue of alphagenome molecular-effect predictions for ~9 billion single-nucleotide variants (every single-letter change in the human genome), plus an AVI score. Academic portal + API + google-antigravity skill today; Google Cloud commercial path “soon.” Not validated for clinical use.
What it is
A genome-wide map layer: issuer says it was built by precomputing existing AlphaGenome predictions at scale, not by announcing a new base model. Full grain: alphagenome-atlas-precomputed-map. Scoring layer: alphagenome-variant-impact.
Why it matters to this thread
Frontier lab science releases and AI applied beyond coding (healthcare / molecular biology) are in-scope. This is the first dated AlphaGenome / Atlas page in the thread. Keep issuer “best-in-class” and collaborator vignettes attributed.
Key facts (from 2026-09-08-x-10am-deepmind-alphagenome-atlas-9b-variants-addy-osmani)
Issuer blog (September 8, 2026)
- From 2026-09-08-x-10am-deepmind-alphagenome-atlas-9b-variants-addy-osmani (AlphaGenome Atlas blog, dated September 8, 2026): platform with predictions for effects of 9 billion single-nucleotide variants — “every single-letter change possible” in the human genome; framed as the most comprehensive catalogue of how genetic mutations affect molecular biology; free academic website portal.
- From the same source (same blog): built by precomputing AlphaGenome predictions at scale. AlphaGenome itself already existed as a per-variant model / API / Nature paper path — Atlas is the genome-wide map layer.
- From the same source (same blog): dataset size ~1 petabyte, claimed >30× larger than the AlphaFold Database.
- From the same source (same blog): resources include thousands of molecular-effect predictions per variant across gene-regulation modalities and hundreds of human/mouse cell types/tissues; an AVI score (AlphaGenome + AlphaMissense); AVI feature attributions (e.g. splicing, chromatin accessibility, conservation); >2,500 recurrent DNA sequence motifs / locations.
- From the same source (same blog): access via website portal + AlphaGenome API + skill in Google Antigravity today; commercial on Google Cloud soon. Base AlphaGenome already academic on GitHub/API and commercial on Cloud Model Garden.
- From the same source (same blog): explicit clinical disclaimer — not a substitute for medical advice; not validated/approved for clinical use.
Official X thread (text only; lead post has video)
- From the same source (@GoogleDeepMind, 2026-09-08 14:03:52Z): launching Atlas — AI searchable database mapping predicted impact of all 9 billion possible single-letter DNA changes. Video attached — flagged; post text only.
- From the same source (@GoogleDeepMind): Atlas >30× AlphaFold Database; 1-petabyte dataset; link variants to molecular mechanisms they disrupt.
- From the same source (@GoogleDeepMind): AVI combines AlphaGenome, AlphaMissense, other features — rank low→high impact; reveal damage modes (gene switches, RNA splicing).
- From the same source (@GoogleDeepMind): resources via Atlas website, AlphaGenome API, Antigravity skill, Google Cloud soon — shortlink resolves to the issuer blog above.
Collaborator vignettes (issuer-reported, not independently re-verified)
Keep attributed to the DeepMind blog as synthesized in the source. No person or institute pages created this pass.
- From the same source (issuer blog): GREGoR / Broad (Covill, O’Donnell-Luria) — AVI prioritized a DNM1 splice-site variant in epileptic encephalopathy, later experimentally supported.
- From the same source (issuer blog): Exeter (Hawkes) — UK Biobank WGS (~54k) grouping by predicted molecular effects → +22% more non-coding associations; BMI screen focusing on top 1% impactful non-coding variants → 19 genetic regions.
- From the same source (issuer blog): Stowers (Zeitlinger / Weilert) — motif categorization for TF accessibility vs gene on/off.
What this source does not establish
- Not a first-time AlphaGenome model launch. Nature / model / API story is earlier. Today is the precomputed Atlas + AVI. See alphagenome, alphagenome-atlas-precomputed-map.
- “Best-in-class” AVI and collaborator wins are issuer claims. Keep attributed until independent corroboration is filed.
- Not validated for clinical use. DeepMind disclaims medical advice / clinical approval.
- Lead X video not transcribed (2097325048109384166). Do not cite the video file.
- Not world-labs-atlas (World Labs omni world model, Sep 1).
- Did not re-file same-day Buckmaster / Bubeck / Anandkumar / Rumik material (morning clip).
- No ticker.