Running the model in advance
Google DeepMind published AlphaGenome Atlas on Monday, a database of AI-generated predictions for how single-letter changes to human DNA affect molecular biology.
The scale comes from simple arithmetic. The human reference genome is about three billion letters long, and at each position there are three possible substitutions. That is nine billion variants, and DeepMind has run its AlphaGenome model over all of them in advance.
The result is roughly one petabyte of data. DeepMind says that makes the Atlas more than thirty times the size of the AlphaFold Database, the protein structure resource the company released in 2021 and expanded to over 200 million predictions in 2022.
What it predicts, and the single number on top
AlphaGenome compares an original DNA sequence with an altered one and predicts how the change might affect gene expression and other regulatory activity. Precomputing that removes the step where a researcher has to run a model themselves to ask about a variant.
Alongside the raw predictions, DeepMind released the AlphaGenome Variant Impact score, or AVI. It condenses predictions from AlphaGenome and from AlphaMissense — the company’s model for variants that alter proteins — into a single number, so that variants in coding and non-coding regions can be ranked on the same scale.
That last point is the useful one. Most of the human genome does not code for proteins, and most tools for interpreting variants have been strongest exactly where AlphaMissense already worked. A regulatory variant sitting far from any gene has historically been much harder to score.
Access
The Atlas is free for academic research, through a website portal, the AlphaGenome API, and as a skill in Google Antigravity. Commercial access will run through Google Cloud.
DeepMind names research partners including the Broad Institute, the University of Exeter, the Stowers Institute for Medical Research, Boston Children’s Hospital and the GREGoR Consortium. Nature and IEEE Spectrum both covered the release.
What to watch next
These are predictions, not measurements. The value of the Atlas depends on how well AlphaGenome’s outputs hold up against experimental data for variants nobody has characterised yet, and that is a question the genetics community answers slowly, one benchmark at a time.
The concrete thing to watch is whether clinical variant-interpretation pipelines start citing AVI scores, and on what evidentiary footing. AlphaFold’s structures became infrastructure because people checked them; the same test applies here, and it has not been run yet.
