Your genome has roughly 3.1 billion letter positions, and each can be swapped for any of three alternatives — about 9 billion possible single-letter edits. As of September 8, every one of those "what if" questions has a precomputed predicted answer, freely browsable online.
This is DeepMind's AlphaGenome Atlas. First, a common misconception to clear up: there is no new model here. The engine is still AlphaGenome, released as a preprint in June 2025 and published in *Nature* on January 28, 2026. Atlas simply ran that model across the entire variant space once and stored the outputs — about 1 PB. DeepMind's own analogy is apt: this is AlphaFold's "database moment." AlphaFold 2 was the model; AlphaFold DB released predictions for 200 million protein structures, letting biologists bypass the "can call an API" barrier. Atlas aims to do the same for regulatory genetics.
- June 2025 — AlphaGenome preprint: a regulatory prediction model at single-base resolution
- January 2026 — Published in *Nature*; ~9,000 researchers have used its API
- September 8, 2026 — Atlas launches: a precomputed database covering the full variant space, ~1 PB
What's in the 1 PB
Three layers:
1. Molecular scores (the heaviest): ~27,000 scores per variant on average — transcription, splicing, chromatin accessibility, and more — across hundreds of human and mouse biosamples. 2. AVI, an aggregate pathogenicity score using only 18 features (compared with 150+ for its benchmark rival CADD), with SHAP decomposition into readable contributions. 3. 2,601 clustered cis-regulatory motif clusters, backed by 253 billion motif instances.
Scope caveat: coverage is complete only for single-nucleotide variants (SNVs), ~9 billion of them. Indels include only the ~100 million observed ones, sourced from gnomAD, UK Biobank, and All of Us. The "every possible DNA change" claim is discounted for indels.
Where it wins
Benchmark wins concentrate in noncoding regions — the darkest continent of the human genome. Retrieval precision (AUPRC) on pathogenic ClinVar variants: intronic 0.76 vs. 0.44 for the best alternative; synonymous 0.57 vs. 0.35; 3' UTR 0.50 vs. 0.18. The rare-disease gap is starker: on GREGoR consortium cases, recall of the true causative variant among the top 50 candidates is 29.5%, versus 12.5% for CADD.
The most telling example is a concrete case: DNM1-related epileptic encephalopathy, unsolved for a decade because the causative variant created a cryptic splice site that extended exon 10a in-frame — a position not covered by SpliceAI's precomputed score tables. Atlas caught it at 0.943, matching the specialized tool SpliceAI, and experiments in five cell lines validated the mechanism. That is the release's core message: precomputed breadth fills the coverage blind spots of specialized tools. On the population-genetics side, an Exeter team tested 54,000 UK Biobank genomes: 22% more noncoding associations found, and the PLA2G7 gene narrowed from 526 intractable variants to 4 candidates (p = 1.74×10⁻¹³).
Where it loses
The other side of the ledger: 5' UTR lost (0.26 vs. 0.27); the Mendelian noncoding benchmark TraitGym was a tie (0.76 vs. 0.77); fine-mapping of complex traits saw near-zero gain (0.28 vs. 0.27). The paper benchmarks only scoring methods, not sequence-to-function models like Borzoi or Enformer. Scoring variants independently inherently ignores gene–gene interactions — a popular HN counterexample is the full saturation-mutagenesis study of the ΦX174 phage, where a mere 5,000-base genome already defeats AI variant-effect predictors.
There is also a formal reservation: the companion Atlas paper is a DeepMind self-hosted preprint, not peer-reviewed. The model itself passed review; the Atlas did not. The "not for any clinical use" disclaimer on the site is not boilerplate. The harshest Hacker News comment called it "just announcing a cache" — unkind, but technically hard to dispute.
Why a cache is worth something
Because someone paid the cost of 9 billion model queries once, for the whole field. Variant annotation at this scale used to require code, API access, and a compute budget; now the barrier is "can open a webpage," and filling the institution field with "None" is allowed. A commercial version will run on Google Cloud, pricing unannounced — that will signal how DeepMind plans to charge pharma.
DeepMind positions Atlas as "a baseline, not an endpoint." Two watchpoints ahead: when full indel and structural-variant coverage arrives, and whether the 1 PB academia gets for free still deserves today's cheers once commercial pricing is published.
---
Sources: Google DeepMind official blog on AlphaGenome Atlas and the alphagenome.google/atlas portal; the Atlas companion preprint; the AlphaGenome paper, Nature 649:1206–1218 (2026-01-28); Nature News coverage (Ewen Callaway); GREGoR and UK Biobank case data; Hacker News discussion thread (2026-09-08).