On 8 September 2026, Google DeepMind released AlphaGenome Atlas and ranked all 9 billion possible single-letter DNA mutations in the human genome. The free research portal and API hand biologists a phone-readable prioritization score — the AlphaGenome Variant Impact score, or AVI — instead of asking them to run the model one variant at a time. Atlas is a predictive catalogue for non-commercial research, not a diagnostic test and not a substitute for lab experiments.
Why it matters
Most of the genome that shapes disease risk sits outside the 2% that codes for proteins. That “dark matter” — promoters, splices, enhancers, and other regulatory DNA — is where most trait-associated variants live, and it has been brutally hard to read at scale. Until now, a researcher who wanted AlphaGenome’s molecular predictions either queried variants one by one or spent years in wet-lab assays that could never exhaust the search space.
Atlas collapses that bottleneck. DeepMind precomputed the effects of every possible single-nucleotide swap, packaged them into a 1-petabyte atlas more than 30× larger than the AlphaFold Database, and gave collaborators a single ranking number that works for both coding and non-coding DNA. Early partners at the Broad Institute, the University of Exeter, and the Stowers Institute already used it to reopen rare-disease cases, pull rare non-coding signals out of UK Biobank noise, and map regulatory DNA “words” across cell types. For anyone hunting why a mutation matters — or which of thousands of candidates to test next — that is the difference between a haystack and a shortlist.
Key numbers
| Quantity | Value (DeepMind / collaborator reports, 8 Sep 2026) |
|---|---|
| Peg | Released 8 September 2026 |
| Single-letter variants scored | ~9 billion SNVs (every possible base swap) |
| Dataset size | 1 petabyte (>30× AlphaFold Database) |
| Predictions per variant (avg.) | ~27,000 molecular-effect scores |
| Genome coverage focus | 98% non-coding + 2% coding |
| AVI score guide | AVI 10 ≈ top 10%; AVI 30 ≈ 1 in 1,000 |
| Motifs catalogued | >2,500 recurrent DNA sequence motifs |
| UK Biobank rare-variant lift | 22% more non-coding associations (n>54,000) |
| BMI follow-up | Top 1% impactful variants → 19 regions |
| Rare-disease ranking (GREGoR retrospective) | Causal in top 50: 29.5% (AVI) vs 12.5% (CADD) |
| Extra population edits scored | >100 million insertions/deletions (UK Biobank, All of Us) |
Those figures come from DeepMind’s launch materials and Fortune’s reported briefing with the team — not from invented press math. The phone-readable headline remains 9 billion ranked mutations; the table is what pays that click.
How they built the ranking
DeepMind’s AlphaGenome model takes long DNA sequences and predicts thousands of regulatory properties across hundreds of human and mouse cell types and tissues. For Atlas, the team ran that model across a human reference genome, comparing each reference base with each of the three possible alternatives. The resulting catalogue links every SNV to a thick stack of molecular predictions — gene expression, splicing, chromatin accessibility, and more — then compresses the signal into the AVI score.
AVI blends AlphaGenome’s regulatory predictions with AlphaMissense, DeepMind’s earlier model for protein-altering changes. That is why one number can rank both the 2% coding genome and the 98% that switches genes on and off. Each AVI score also comes with feature attributions: how much of the flag comes from splicing versus expression versus protein impact. An AVI of 10 means the variant sits among the most impactful tenth of the genome; an AVI of 30 puts it among roughly the strongest one in a thousand.
Atlas also ships a genome-wide motif map — more than 2,500 recurring short DNA sequences that transcription factors tend to bind — so researchers can ask not only how strong a variant looks, but which regulatory word it may break.
What partners already found
Rare disease, reopened. With the GREGoR Consortium, Laura Covill and Anne O’Donnell-Luria at the Broad Institute used AVI to prioritize variants in unsolved rare-disease cases. In a patient with epileptic encephalopathy, Atlas pointed to a DNM1 variant: 69% of its score came from splicing. The model predicted a spurious splice site in a brain-expressed isoform that lengthens the protein by 13 amino acids — a change blood RNA sequencing had missed because that segment is barely expressed in blood. Lab screens confirmed the prediction, and the variant was reclassified as likely pathogenic. In a retrospective GREGoR test on previously solved cases, AVI put the known causal variant among a patient’s top 50 candidates 29.5% of the time, versus 12.5% for CADD.
Population genetics, less noise. Gareth Hawkes (MRC fellow, University of Exeter) applied Atlas to whole-genome data from more than 54,000 UK Biobank participants. Grouping rare variants by predicted molecular effect yielded 22% more non-coding associations than the same analysis without Atlas, and in one locus narrowed 526 candidates to four. Focusing on the top 1% of non-coding variants by Atlas impact, he flagged 19 genetic regions linked to body-mass index — a shortlist for the next round of targeted work, not a finished obesity map. The same approach surfaced regulatory variants tied to circulating proteins including PLA2G7 (linked to aging biology) and EGLN1 (an oxygen sensor).
Motifs, sorted. Julia Zeitlinger and Melanie Weilert at the Stowers Institute used the motif resource to separate transcription factors that mainly open DNA from those that also turn genes on or off — the kind of catalogue that would take decades of manual assays to approximate.
What this is not
Atlas is not a clinical diagnosis. DeepMind is explicit: predictions can only be part of an evidence chain, and they are not “the universal truth.” AlphaGenome is stronger on some classes (splicing, promoters) and weaker on others (notably some enhancers). It can miss indirect effects that act through changing levels of regulatory proteins. AVI is a prioritization score — which variants deserve wet-lab attention — not a doctor’s label.
It is also not a finished commercial product for clinics. Non-commercial access opened on 8 September 2026 via the website portal and API; commercial licensing on Google Cloud is promised “soon.” Sister company Isomorphic Labs will still need a commercial license. EMBL-EBI’s Ewan Birney has said Ensembl’s Variant Effect Predictor is working to integrate AVI — useful plumbing, not a regulatory clearance.
Do not read “9 billion” as “we now understand every disease mutation.” Read it as: the ranking layer exists, partners already used it, and the hedges are written into the release.
What to watch
Three near-term checks will show whether Atlas becomes daily infrastructure or a splashy demo.
- Independent replication. Outside labs should reproduce the UK Biobank lift and GREGoR-style prioritization gains on their own cohorts, not only DeepMind-supplied case studies.
- Hard classes. Watch enhancer and long-range regulatory performance as the base model improves — Avsec has already flagged those as weaker spots.
- Ecosystem plug-ins. Ensembl VEP integration, Cloud commercial terms, and whether rare-disease consortia adopt AVI as a default filter will decide whether clinicians and genetic counselors ever see these ranks in their ordinary tools.
For a progress scoreboard that fits a phone screen, keep the pair DeepMind actually shipped: 9 billion single-letter mutations ranked on day one, and a 22% lift in rare non-coding associations when Exeter filtered UK Biobank with those ranks. Everything else — drug targets, finished diagnoses, “solved the genome” slogans — has to wait for the experiments Atlas was built to prioritize.
Sources
- Google DeepMind blog, 8 September 2026: AlphaGenome Atlas: Molecular predictions for 9 Billion human DNA variants
- Google Keyword blog, 8 September 2026: Introducing AlphaGenome Atlas
- Fortune, 8 September 2026: Google DeepMind publishes AI-powered predictions for effect of all 9B mutations to human DNA (briefing numbers: ~27,000 predictions/variant; AVI 10 / 30; GREGoR 29.5% vs 12.5%; DNM1 splicing 69% / +13 aa; >100M indels)
