Google DeepMind has released AlphaGenome Atlas, a precomputed set of molecular effect predictions covering roughly 9 billion single-nucleotide variants across the human genome. Every position, every possible substitution, scored in advance and packaged with a summary number the team calls the AlphaGenome Variant Impact (AVI) score. One value per variant.
That framing matters more than the model architecture does. AlphaGenome itself arrived earlier as something you query. The Atlas is the same capability turned into a static artifact you download and join against, which is a very different kind of object to put in a lab’s hands.
Precomputation is the actual release
Running a deep genomic model over a variant list is not hard for a well-funded group. It is annoying, expensive, and slightly different every time someone does it, which means results across labs are never quite comparable. Freezing one exhaustive pass and shipping it removes that variance. Everyone scoring a variant against the Atlas is scoring against the same numbers, from the same model, at the same version.
That is a standards play dressed up as a data release. Whoever owns the reference table that clinical and research pipelines join against gets to define what “predicted impact” means in practice, and DeepMind now has a strong claim on that slot for non-coding regions where existing annotation is thinnest.
One number is a compression, not a verdict
The AVI score is the part I would treat carefully. Collapsing a model’s multi-track molecular predictions (expression, splicing, chromatin, whatever the model outputs across tissues) into a single scalar makes the data usable at genome scale. It also throws away the reason a variant scored high. A splice disruption and a distal regulatory shift can land on the same number and mean entirely different things at the bench.
The risk is not that the score is wrong. It is that a convenient scalar with a big name attached becomes a filter threshold in a pipeline somewhere, and then a paragraph in a paper, and then a soft prior in someone’s clinical interpretation, long before anyone has established what its false positive rate looks like in the regions where it is being trusted most. Prediction is not evidence of function. Nine billion predictions are not nine billion pieces of evidence.
Where this actually gets used
The near-term winners are variant-of-uncertain-significance triage and GWAS follow-up, the two places where researchers are drowning in candidate positions and have no principled way to rank the non-coding ones. Having a precomputed prior for every position is genuinely useful there, because the alternative is arbitrary distance-to-gene heuristics.
The open question is calibration. An exhaustive map is only as good as its behaviour on the variants nobody has characterized yet, and by definition those are the ones the model was not trained to get right. DeepMind has shipped the map. What it costs to check the map against wet-lab reality, variant by variant, has not changed at all.