Google DeepMind has released AlphaGenome Atlas. The database predicts what every possible single-nucleotide variant in the human genome does. All nine billion of them. The result runs to a petabyte.
There’s a score attached, AVI — AlphaGenome Variant Impact — that folds predictions for coding and non-coding regions into one number. Access is through a free web portal, no programming needed.
Two examples that didn’t come from the lab
At the Broad Institute, researchers used the Atlas to pin down a critical variant in the DNM1 gene and closed a rare disease case with it.
In the UK Biobank, the Atlas turned up 22 percent more non-coding associations for complex traits such as body mass index.
The second one is the interesting one. Non-coding DNA — once dismissed as junk — is by far the larger part of the genome, and mapping variant to effect there has been the hard part. That’s where the gain sits.
The announcement is signed by Pushmeet Kohli, VP Science at Google DeepMind, and Žiga Avsec, who leads the Genomics Initiative there. Their framing: the Atlas offers “grounded genomic insights that will accelerate the pace of biological discovery.”
Reading it
The actual news here isn’t the model, it’s the precomputation. AlphaGenome already existed. What’s new is that somebody ran the whole thing once and left the answers lying around as a reference. Checking a variant no longer means standing up a model — you look it up.
That’s the same move AlphaFold made with its structure database, and that changed more than the model alone did. A tool you have to install gets used by a few hundred people. A web portal gets used by tens of thousands.
The usual caveat still applies: these are predictions, not measurements. A petabyte of plausible numbers is still a petabyte of hypotheses. Worth a lot for deciding what to test next, worth nothing as a finding.
Sources: Google DeepMind: AlphaGenome Atlas