DeepMind precomputed all nine billion single-letter mutations in the human genome
The Atlas attaches one damage score to every variant, and a line at the bottom of the page says it has not been approved for any clinical use.
Nobody at DeepMind had to decide which mutations were worth predicting. They did all of them.
AlphaGenome Atlas, published on 8 September, holds predicted molecular effects for every possible single-letter change in the human genome. That is 9 billion of them, which is three alternative letters at each of roughly three billion positions, so the word "all" is doing literal work. All of it runs to a petabyte, more than thirty times the AlphaFold Database, and it is free for academic research through a web portal, an API and a skill inside Google Antigravity.
One number per variant is the part that will travel
One score is the piece that will matter. The AlphaGenome Variant Impact score folds AlphaGenome's regulatory predictions and AlphaMissense's protein predictions into one figure per variant, so a researcher can rank nine billion mutations by how much damage they probably do. DeepMind says it is best in class across variant pathogenicity and rare disease benchmarks, and pairs each score with attributions showing which process drove it (splicing, chromatin accessibility, conservation).
And it has already found things. Working with the GREGoR Consortium, researchers at the Broad Institute used the score to prioritise variants in unsolved rare disease cases and surfaced one in DNM1, a gene strongly linked to epileptic encephalopathy. Its effect is an incorrect splice site that extends the resulting protein, and experimental screens confirmed it.
That last clause is what separates this from a demo.
The real test ran on 54,000 people
At Exeter, Gareth Hawkes ran the Atlas against whole-genome data from more than 54,000 UK Biobank participants and pulled out 22 percent more non-coding associations than the statistics alone surfaced, among them regulatory variants driving PLA2G7 and EGLN1.
Hold on to that percentage. And most trait-associated variation sits in the 98 percent of the genome that does not code for protein, where real signal drowns in harmless noise. A ranking that lifts a fifth more genuine associations out of that noise is an instrument rather than an announcement.
The line at the bottom of the page
"AlphaGenome has not been validated for, and is not approved for, any clinical use."
We would not bet on that sentence surviving contact with the score above it. A single interpretable number attached to every possible mutation is exactly the artefact that leaks out of research and into decisions, the way a p-value did. Divide the petabyte by the count and you get about 111 kilobytes of prediction per variant, and precisely one digit that anyone will read.
Would you want your own variant of uncertain significance ranked by a model that has never been near a clinical trial? We would want the false-negative rate first (it is not on the page).
So it rarely is.
A petabyte is a fact about access, not only size
Nobody is downloading this. Every researcher who uses the Atlas queries it where Google keeps it, free for non-commercial work today and on Google Cloud for commercial use soon, which makes the most comprehensive map of human genetic variation a hosted service.
AlphaFold's database went the same way and the field decided that was fine. But this one is thirty times larger, and it predicts what a stretch of DNA does rather than what a protein looks like. Function is the thing people act on.
