DeepMind has run its AlphaGenome model across every single-letter change that can be made to the human genome — roughly nine billion of them — and released the results as a precomputed atlas. Each variant arrives with about 27,000 predicted values describing its likely effect on molecular processes across hundreds of cell types and tissues. The dataset runs to one petabyte, more than 30 times the size of the AlphaFold protein structure database, and it is available for noncommercial use through a web portal, an API and a skill in Google Antigravity. A commercial version is planned for Google Cloud.
The problem it addresses is one of scale in both directions. A human genome holds about three billion letters, and any individual carries millions of small differences from the reference sequence, usually a single substituted letter. Most are harmless. A few cause disease, and the sequence itself does not say which is which. Testing all nine billion possible substitutions in a laboratory is not a funding question, it is arithmetic.
AlphaGenome, introduced in 2025, reads stretches of DNA one million letters long and predicts how actively a gene is transcribed, whether regulatory proteins can bind to the DNA, and how segments get spliced out of a gene's transcript. Until now the model had to be run separately for each variant of interest. The atlas is the same model, run in advance, on everything.
That precomputation matters most for the roughly 98% of the genome that contains no protein blueprint. Noncoding stretches work as switches and regulators, determining when a gene turns on and in which tissue. That is where most disease-associated variants sit, and where their effect is hardest to interpret — the region where a researcher was least likely to have run the model themselves, because they had no candidate to run it on.
Thousands of numbers per variant are unusable in daily work, so DeepMind built a summary score: the AlphaGenome Variant Impact Score, or AVI. A small neural network combines AlphaGenome's output with the AlphaMissense protein model and two measures of how well a stretch of DNA has been conserved across millions of years of evolution. AVI uses 18 input features. CADD, the tool it is most often compared against, uses more than 150.
The training signal is the interesting compromise. For almost no variant is it known for certain whether it causes harm, so the team substituted a proxy: variants that are very rare in the population are treated as probably harmful, common ones as probably harmless, on the reasoning that damaging mutations are passed to the next generation less often and never spread. Despite that indirect supervision, AVI outperformed existing tools on benchmarks of variants with established clinical status, particularly in noncoding regions. On some tasks other tools did better. For each variant the atlas also reports which process drove the score — a change in splicing, say, or in the behaviour of a genetic switch.
The clearest demonstration comes from GREGoR, a consortium studying rare diseases with no identified cause. A child with severe epilepsy had no diagnosis despite having had their genome sequenced. AVI moved a variant in the gene DNM1, previously classified as uncertain, to the top of the candidate list. The AlphaGenome predictions also proposed a mechanism: the variant creates an incorrect splice site during transcript processing and lengthens the resulting protein by 13 building blocks — but only in a version of the gene that is read exclusively in the brain, which is why earlier analysis of blood samples had found nothing. A laboratory experiment confirmed the prediction, and the researchers recommended reclassifying the variant as likely disease-causing.
Run back over the consortium's already-solved cases, AVI placed the causal variant in the top 50 candidates 29.5% of the time. CADD managed 12.5%.
That pair of numbers is the honest measure of where this technology stands, and it is more modest than the petabyte implies. Better than double the previous tool is a real result. It also means that on cases whose answers were already known, the system failed to surface the right variant inside fifty guesses seven times out of ten. A clinical geneticist reading that will hear "a better ranking heuristic," not "a diagnostic." The line about other tools winning on some tasks is doing quiet work too: this is not a model that has replaced the category, it is one that is stronger where the category was weakest.
The proxy training deserves the same scepticism. Rarity is a reasonable stand-in for harm across a population, but it encodes population history — drift, founder effects, which populations are in the reference data — as biology. A score built that way will be most confident exactly where sampling has been deepest.
The population-genetics use may end up mattering more than the diagnostic one. To establish whether rare variants in a given region affect something like a blood composition measure, they have to be analysed together, because individually each is too rare to support a statistically reliable conclusion. Pool harmless variants with the ones that actually act on the trait and the signal disappears into noise. Gareth Hawkes at the University of Exeter used the atlas to group only variants the system predicted would act the same way, across genetic data from more than 54,000 UK Biobank participants, and found 22% more associations between noncoding variants and blood protein levels than conventional filters produced. From the same predictions the team also picked out 2,601 recurring short DNA motifs — effectively the words of the genome, the sequences regulatory proteins attach to.
DeepMind is explicit about the limits. The model does not know every cell type, and it does not account for effects that arise from changing the quantity of other regulatory proteins. The atlas and AVI are research instruments, and the company says they can be only one link in a chain of evidence behind a diagnosis.
Source: the-decoder.com
Which leaves the part the release is quietest about. A petabyte is not a file anyone downloads; it is a reason to compute where the data already lives. Free noncommercial access through a portal and an API, with the paid version on Google Cloud, makes the atlas both a genuine public good and the largest piece of demand generation DeepMind has shipped into its parent's cloud business. The science stands on its own. So does the fact that the most useful map of the human genome now has a landlord.