9 SEP 2026 — Google DeepMind has published AlphaGenome Atlas, a petabyte of precomputed molecular predictions covering every one of the roughly nine billion single-letter substitutions possible in the human genome. Nine billion is not a discovery count. It is three billion base pairs multiplied by the three letters each one is not, which means the Atlas is exhaustive by construction.
What was released
About one petabyte of precomputed output, more than thirty times the size of the AlphaFold Database, covering all possible single-nucleotide variants. Alongside it comes the AlphaGenome Variant Impact score, which combines outputs from the AlphaGenome model released in 2025 and from AlphaMissense into one disruption signal per variant.
The practical change is in access. Where researchers previously requested predictions one at a time, they are now already computed and searchable, which turns a queue into a lookup.
It is free for non-commercial research, with commercial licensing left open.
Exhaustive is a computation, not a finding
Enumerating every possible substitution sounds like completeness and is really arithmetic. The genome has about three billion positions and each admits three other letters, so nine billion is what you get by multiplying, before anything is learned.
What the Atlas does not contain is which of those variants actually occur in people. The overwhelming majority never have and never will, because they are incompatible with development or simply have not arisen. A map of all possible changes is a different object from a map of observed human variation.
That does not make it useless. The point of precomputing is that you cannot know in advance which variant a clinician will look up. It does mean the nine billion is a measure of coverage rather than of knowledge.
The AlphaFold comparison is doing too much work
The framing everywhere is that this runs the AlphaFold Database playbook on human genetics, and the analogy breaks in a specific place.
AlphaFold predicted protein structures, and a predicted structure can be checked against crystallography or cryo-electron microscopy. The ground truth existed, was independent, and accumulated. That feedback loop is why confidence in AlphaFold's predictions could be calibrated.
Predicted regulatory effects of non-coding variants have no comparable reference at scale. There is no experiment that resolves nine billion variants, and the experiments that do exist — reporter assays, perturbation screens — cover a vanishing fraction and disagree with each other. A prediction of a variant's effect is harder to falsify, which makes the AlphaFold analogy point the wrong way.
A variant impact score is a marker
The AVI score ranks how disruptive a variant is predicted to be. It does not say whether a person carrying it becomes ill.
That distinction has been the recurring problem in this field, and we have hit it twice in a fortnight from different directions — methylation marks that record inflammation without anyone following carriers to an outcome, and a statin trial whose event endpoint moved while its independence endpoint did not.
A disruption score is further from a clinical outcome than either. It predicts a molecular consequence, which may or may not produce a cellular one, which may or may not produce a phenotype, in a genetic background and environment the model never saw.
Precomputation changes who can use it
The engineering decision worth crediting is the one that sounds least interesting. Moving from an on-demand service to a precomputed dataset removes a queue, and a queue is a gatekeeper.
A researcher at a well-funded institution could always get predictions; the constraint was patience rather than access. For a group in a country without the compute or the relationship, a downloadable table is a categorically different proposition from an API allocation.
A petabyte is its own barrier, of course, and few groups will hold a local copy. But querying a hosted table is cheap in a way that running a model is not, and that difference is what widens the population who can ask a question.
Where it helps
Variant interpretation in rare disease is the strongest case. A clinician sequencing a child with an undiagnosed condition typically finds several variants of uncertain significance, and the bottleneck is prioritising which to investigate. A ranked, precomputed score narrows that list. Even an imperfect ranking is a contribution when the alternative is an unordered one.
The same applies to research triage. A group with a genomic region of interest and no budget to assay every position can use the Atlas to choose what to test in the laboratory, which is prediction used as a hypothesis generator rather than as evidence.
Both uses follow a safe pattern: the prediction is followed by a human decision and, usually, an experiment.
What would make it checkable
Three things. A held-out validation against experimentally measured variant effects that the model was not trained on, reported as calibration rather than as a single accuracy figure.
Performance broken down by ancestry, because reference genomes and the variation catalogues behind them are heavily European and a score that degrades outside that population will degrade quietly. And a stated position on clinical use, since a freely available ranked score will be used clinically whether or not it was intended to be.
The release is a substantial engineering achievement and a convenience for researchers. It is a petabyte of what a model believes, and the difference between that and a petabyte of what has been measured is the whole of the interpretation.
View author profile → · Editorial policy
