You can't reset your genome

Google DeepMind has run its AlphaGenome model across every possible single-letter change to the human genome and published the results. All 9 billion single nucleotide variants, each scored for its predicted regulatory impact, add up to a one-petabyte dataset called AlphaGenome Atlas.

The human genome is roughly 3 billion base pairs. About 2% codes for proteins and is reasonably well understood; the other 98% is non-coding and far less so. AlphaGenome, the model underneath the Atlas, predicts how a single change in those non-coding regions disrupts molecular processes such as protein production. The Atlas runs that model exhaustively, so instead of running inference on a variant of interest, a researcher queries a precomputed answer. The useful engineering move is turning the entire variant space into a lookup rather than a per-query model call.

On top of the raw predictions sits the AlphaGenome Variant Impact score, AVI. It combines coding and non-coding predictions into a single number meant for triage, so a researcher can rank variants, look at the top of the list, and skip the thousands below.

Two cases show how it is being used. At the Broad Institute, Laura Covill’s team used the AVI score to prioritize variants in unsolved rare disease cases. It flagged a variant in the DNM1 gene, predicting that the change created an incorrect splice site, and that became supporting evidence to solve the case. Separately, Gareth Hawkes applied the Atlas to data from more than 54,000 UK Biobank participants. Grouping variants by their predicted molecular effect, he found 22% more non-coding genetic associations than before, and from the top 1% of impactful variants he identified 19 genetic regions linked to body mass index.

How those results are described matters. In the rare-disease case the score provided supporting evidence; in the Biobank case it directed the next stage of targeted research. In both, the Atlas prioritized and suggested, and something else did the confirming. The AVI score is a ranking signal computed over model predictions, and model predictions are not measurements, so the sensible way to use it is as a triage layer that points at where to spend wet-lab or statistical effort.

Access is through a website portal that DeepMind says needs no coding, aimed at clinical researchers and biologists rather than people set up to run the model themselves.