Google DeepMind Precomputed Every Possible Single-Letter DNA Change in the Human Genome: 9 Billion Variants, 1 Petabyte, 30 Times AlphaFold DB

Google DeepMind released AlphaGenome Atlas on September 8, 2026, a 1-petabyte dataset holding precomputed molecular effect predictions for all 9 billion possible single-nucleotide variants in the human genome, more than 30 times the size of the AlphaFold Database. Alongside it the company published the AlphaGenome Variant Impact (AVI) score, a single number summarizing each variant's impact that works across both the 2% of the genome that codes for proteins and the remaining 98% of non-coding sequence. The resource is free for academic research through a website portal. ASAP works from DeepMind's own announcement as the primary source to set out what this changes and where it stops.

Nine billion is the number that makes laboratory testing impossible

The problem DeepMind frames is straightforward: there are roughly 9 billion possible single-letter mutations in the human genome, and testing each one in the lab is practically impossible. That is why interpreting how genetic variation affects biology at the molecular level has been the bottleneck in genomics.

AlphaGenome, the model DeepMind released previously, predicts how genetic variants affect biological processes. Atlas is not a new model but the result of running that model's predictions across the entire genome in advance. As the announcement puts it, just as an atlas is a collection of maps linking features of the land, AlphaGenome Atlas charts the molecular effects of DNA variants across the genome.

The resource comes in four layers. First, thousands of molecular effect predictions per variant, spanning multiple aspects of gene regulation across hundreds of human and mouse cell types and tissues. Second, the AVI score, which condenses those predictions into a single number. Third, AVI feature attributions, which decompose each score into interpretable categories such as chromatin accessibility, splicing and conservation. Fourth, a compendium of more than 2,500 recurrent DNA sequence motifs, the "words" of the genome, together with their locations.

What the AVI score targets is the 98% of the genome that does not code

The AVI score merges AlphaGenome's regulatory effect predictions with the output of AlphaMissense, DeepMind's model for protein-altering variants. The point DeepMind emphasizes is that the score works for both the 2% of the genome that codes for proteins and the remaining 98% that orchestrates gene activity. According to the announcement, most trait-associated variants live in that 98%.

The design aims squarely at the gap existing tools left open. Coding variants came with relatively clear criteria, such as whether an amino acid changes or a stop codon appears, and prediction tools including AlphaMissense concentrated there. Non-coding variants were harder to rank on a single scale, because which gene they regulate, in which cell type, and by how much all depend on context. Atlas does not sidestep that difficulty; it precomputes cell-type-specific predictions in bulk, derives a single score from them, and then hands that score back decomposed into which processes are predicted to be disrupted.

That decomposition matters in practice. A score alone lets you rank candidate variants but not learn why one is dangerous, which means it does not translate into an experiment. With feature attributions attached, a ranking converts directly into a validation plan. DeepMind's claim that the AVI score gives best-in-class performance across pathogenicity and rare disease benchmarks is the headline, but the choice to return an interpretable breakdown alongside the score looks like the design decision with the longer life.

Three completed studies show what the resource is for

DeepMind's announcement presents three results that external academic collaborators produced with AlphaGenome Atlas before the September 8, 2026 release. The first is in rare disease. Laura Covill and Anne O'Donnell-Luria of the Broad Institute, working with the GREGoR Consortium, used the AVI score to prioritize candidate variants and surfaced one that earlier research had overlooked. The gene was DNM1, strongly linked to epileptic encephalopathy, and the underlying AlphaGenome predictions showed the mechanism as well: the variant created an incorrect splice site that led to an abnormal extension of the resulting protein. Experimental screens validated the prediction and turned up nearby variants with similar effects.

The second is in population genetics. Gareth Hawkes of the University of Exeter applied Atlas to whole-genome data from over 54,000 UK Biobank participants. Grouping rare variants by their predicted molecular effects uncovered 22% more non-coding genetic associations, signals that would otherwise have stayed buried in statistical noise. That work pinpointed regulatory variants driving circulating levels of PLA2G7, linked to aging, and EGLN1, a cellular oxygen sensor. Hawkes extended the same approach to body mass index, focusing on the 1% of non-coding variants Atlas predicts to be most impactful, and identified 19 genetic regions.

The third is in basic biology. Julia Zeitlinger and Melanie Weilert at the Stowers Institute for Medical Research used the motif resource to categorize which transcription factors affect only DNA accessibility and which also turn genes on and off.

Precomputation, not a new model, is the substance of this release

There is no new model here. AlphaGenome and AlphaMissense both already existed; what Atlas did was run their predictions genome-wide ahead of time and freeze the result into a petabyte. Why that choice matters becomes clear from the user's side.

Querying a model through an API one variant at a time requires that the researcher has already narrowed the field. You have to decide where to look before you can look. A fully computed table inverts that. It becomes possible to sort the whole genome by score and sweep the top of the list, which is exactly the shape of Hawkes's body mass index analysis, where filtering to the top 1% yielded 19 regions. Precomputation is not a compute optimization; it changes the direction of the research question, shifting weight from hypothesis testing toward hypothesis generation.

DeepMind's extended discussion of the AlphaFold Database precedent points the same way. The 2022 expansion grew coverage from roughly 190,000 experimental structures to more than 200 million predicted ones, and pairing that with a portal usable by researchers with no coding experience was decisive for adoption. Atlas opens three routes at once: the website portal, the AlphaGenome API, and a skill in Google Antigravity.

Why the AlphaFold precedent does not transfer cleanly

The difference between AlphaFold's structure predictions and AlphaGenome Atlas's variant effect predictions deserves attention, because only one of the two has a single ground truth. A predicted protein structure has a single ground truth that experiment can confirm, and AlphaFold reports its own confidence per prediction. A variant's molecular effect, by contrast, is a conditional quantity that shifts with cell type, tissue, developmental timing, and combination with other variants. Covering hundreds of human and mouse cell types is an attempt to span that conditionality broadly, not to eliminate it.

That difference sets the scale of the misuse risk. A wrong structure prediction shows up quickly in the lab, but a wrong variant effect prediction puts a false candidate at the top of a ranking, burns experimental resources, and leaves the real causal variant buried. It is telling that all three cases in the announcement pair prediction with either experimental validation or statistical association analysis rather than standing on prediction alone. In the Broad Institute case, the decisive step was that experimental screens confirmed what the model said.

The notice at the end of the announcement points the same direction. The information AlphaGenome Atlas provides is not intended to substitute for professional medical advice, diagnosis or treatment, and AlphaGenome has not been validated for, and is not approved for, any clinical use. DeepMind itself describes the resource as a baseline rather than an endpoint.

What a lab should check this week

Access terms are the first thing to settle, and the AlphaGenome Atlas portal is open for non-commercial use as of its September 8, 2026 launch day. The base AlphaGenome model is already available for academic use on GitHub and via the API, and the base model is available commercially on Google Cloud through Model Garden. Commercial access to Atlas itself is described only as coming soon on Google Cloud, so anyone planning industrial use needs to confirm timing and terms separately.

The most practical first test is to layer AVI scores onto data you have already analyzed and see whether the ranking changes. Hawkes's 22% gain and the Broad Institute's DNM1 finding both came from re-sorting existing data rather than collecting new cohorts, and reanalysis costs far less than new experiments. Two caveats belong in the study design: the predictions are built on human and mouse cell types, and population-specific allele frequency information is not among the items Atlas provides.

What remains unconfirmed

Four items remain unconfirmed in DeepMind's announcement, beginning with the review status of the paper it links. The announcement links a paper but does not state whether peer review is complete as of publication. Second, benchmark detail behind the AVI score. The claim of best-in-class performance across pathogenicity and rare disease benchmarks appears without figures naming which benchmarks, which metrics, or how large the margins were.

Third, the conditions attached to the 22% figure. That number is the increase in detectable non-coding associations obtained by grouping rare variants by predicted molecular effect across more than 54,000 UK Biobank participants, and whether the same margin reproduces in other cohorts or other traits is a separate question. Fourth, the start date and licensing terms for commercial use of Atlas, which the announcement describes only as coming soon.

Source: ASAP summary based on Google DeepMind's announcement "AlphaGenome Atlas: A predictive map of every possible DNA letter change in the human genome" (September 8, 2026)