The Change Isn't Accuracy — It's Turning a Model Into a Lookup Table
The model itself isn't new; AlphaGenome shipped last year. What happened here is that someone ran it to completion: rather than waiting for researchers to ask, compute every possible question in advance and store the answers. DeepMind has walked this path before. The AlphaFold Database turned protein structure prediction from "you have to run it yourself" into "you look it up," and usage patterns changed afterward. On the genome side the friction is even more visible: one outside researcher's assessment is that AlphaGenome is the leading model in the field but "very slow and computationally intensive." In other words, whether you could use it depended largely on whether you had the compute. The Atlas moves that threshold onto Google's side, and incidentally removes the waste of everyone worldwide re-running the same simulations. Each variant comes with thousands of predictions rather than a single number — from which tissues a nearby gene is expressed in to the shape of folded chromatin.
The AVI Score Exists Because Nobody Can Read That Much
More predictions create a new problem: with thousands of data points per variant, researchers cannot triage. The AVI score is built for that — a single disruption signal fusing AlphaGenome and AlphaMissense outputs, putting coding and non-coding regions on one scale, with feature attributions and 2,500+ motifs to say why a given variant ranks high. The non-coding part is the point. Variant-interpretation tools have been reasonably mature in coding regions and consistently weak outside them, and the UK Biobank result — 22% more non-coding associations across 54,000+ genomes — lands squarely on that weakness. The DNM1 splice variant is a different kind of evidence: a validated rare-disease variant, showing it can also land on something specific in an individual case.
How to Use It, and How Not To
The boundary first, and DeepMind states it plainly: the output is not a substitute for professional medical advice and is not approved for clinical use. That is not boilerplate. What the Atlas provides is **prediction** and **ranking**, not experimental evidence and not diagnosis. The sensible use is triage — putting limited wet-lab capacity behind the variants most likely to yield something, with validation still to follow. Then the other half of accessibility. The non-commercial web portal is free, the API is on GitHub, and it is also exposed as a Google Antigravity agent skill; but commercial use goes through Google Cloud, "coming soon," with no price announced. For an academic lab this is a directly usable public resource. For a company that wants to build it into a product, the number that matters most has not been published yet. That is the familiar shape left over from the AlphaFold Database playbook: the lookup is free, and what gets charged for is wiring the lookup into your own pipeline.
via: Google DeepMind blog, Nature, IEEE Spectrum, MarkTechPost technical write-up