AlphaGenome Atlas precomputes every possible human mutation, so you don't have to guess which ones matter

Google DeepMind has released AlphaGenome Atlas, a free, precomputed dataset covering the predicted molecular effects of every one of the roughly 9 billion possible single-nucleotide variants in the human genome. That number is the whole point: there are about 9 billion ways a single DNA letter can be substituted for another across the genome, and testing each one in a wet lab is not a scaling problem you solve with more grad students. It's simply not tractable. So DeepMind ran its AlphaGenome model — which predicts how a variant perturbs gene regulation — across the entire space up front, once, and published the result as a browsable resource.

I think of this the same way I think about a precomputed embedding index versus running a model at query time: the hard part is deciding what to precompute and how to compress it into something a person can actually query. AlphaGenome Atlas is DeepMind's answer to both halves of that problem, and the design choices are worth looking at closely.

From a model to an atlas

AlphaGenome itself already does useful work: given a variant, it predicts effects across gene expression, splicing, chromatin accessibility, and other regulatory tracks, for hundreds of human and mouse cell types. That's valuable for a researcher who already has a candidate variant in hand. It's less useful if your problem is the opposite — you have thousands of candidates and no idea which ones are worth a closer look.

Atlas closes that gap by inverting the workflow. Instead of running AlphaGenome on demand, DeepMind ran it on everything, producing what it describes as a 1-petabyte dataset — more than 30 times the size of the AlphaFold Database at its 2022 expansion, when that resource grew from roughly 190,000 experimentally solved protein structures to over 200 million predicted ones. The AlphaFold comparison is deliberate and, I think, apt: both projects took a model that was useful one query at a time and turned it into infrastructure a whole field could build on without needing to run inference themselves.

A single number, and the receipts behind it

The harder problem with exhaustive prediction is that thousands of regulatory-track values per variant, times 9 billion variants, is not something a researcher can page through. DeepMind's answer is the AlphaGenome Variant Impact (AVI) score — a single number per variant that combines AlphaGenome's regulatory predictions with AlphaMissense, DeepMind's existing model for scoring protein-altering variants. The fusion matters mechanically: it lets one score rank variants in the roughly 2% of the genome that codes for protein alongside variants in the other 98%, the non-coding regions that regulate when and how genes get expressed and that carry most trait-associated variation identified so far. Coding and non-coding variants have historically needed separate scoring frameworks; collapsing them into one ranked number is what makes rapid triage across a whole candidate list possible.

A ranking score you can't interrogate is a liability in a research setting — you need to know why a variant scored high before you act on it. DeepMind addresses this with AVI feature attributions, which decompose each score back into the regulatory categories driving it: splicing, chromatin accessibility, gene expression, conservation. Atlas also ships a catalogue of more than 2,500 recurrent DNA sequence motifs — the repeated "words" of the genome — mapped to their genomic locations, so a high attribution for splicing can be traced to the actual sequence feature it disrupts.

Where it's already found something real

The clearest evidence this isn't just a bigger spreadsheet comes from DeepMind's collaboration with the GREGoR Consortium. Researchers Laura Covill and Anne O'Donnell-Luria at the Broad Institute used the AVI score to re-rank candidate variants in an unsolved rare-disease case and surfaced one in the gene DNM1, which is strongly linked to epileptic encephalopathy — a variant earlier analysis had passed over. The attribution data didn't just flag the variant; it explained the mechanism, showing the change created an incorrect splice site that produced an abnormally extended protein. Experimental follow-up confirmed the prediction and turned up nearby variants with the same effect. That's the workflow Atlas is built for: cut a haystack of thousands of candidates down to the handful worth a wet-lab test, and hand back a testable hypothesis for why each one matters, not just a rank order.

A second collaboration, with Gareth Hawkes at the University of Exeter, is applying Atlas to a statistically harder problem: finding rare, non-coding variants associated with common traits at the population level, where the sheer number of harmless variants buries real signal in noise. DeepMind reports the AVI score performs at a best-in-class level across variant pathogenicity and rare-disease benchmarks, though as with any benchmark claim from the model's own developer, independent validation across more cohorts is what will actually establish how far that generalizes.

Why I'd use this over calling the model directly

AlphaGenome Atlas is available through a no-code web portal, the AlphaGenome API, and as a skill inside Google Antigravity — three access points aimed at three different users: the clinician who wants to look up one variant, the bioinformatician wiring it into a pipeline, and an agent doing the lookup as one step in a larger workflow. That last point is the one I'd flag for anyone building agentic research tools: a 9-billion-row precomputed index is a far better tool-use target than a model endpoint, because it turns a variant lookup into a deterministic read instead of a fresh inference call. Cheaper, faster, and reproducible in a way that matters when the answer feeds into a clinical or scientific claim.

The caveat is the same one that applies to AlphaFold's structures: a predicted score is a hypothesis, not a diagnosis. AlphaFold accelerated structural biology by giving researchers a starting point to test rather than a monopoly on ground truth, and the DNM1 case shows Atlas working the same way — the AVI score narrowed the search, but experimental validation is still what closed the case. Treat Atlas as the fastest way to decide what to test next, not as the test itself.