🧬 Genomics
DeepMind Precomputed All 9 Billion Human DNA Mutations. We Checked the Petabyte Math.
The AlphaGenome Atlas predicts the molecular effect of every possible single-letter DNA change. We audited the arithmetic behind the 1-petabyte headline, priced the storage bill, and measured the improvement over the field's standard ranking tool.
Nine billion.
That is how many single-letter changes are possible in the roughly three billion letters of the human reference genome: three alternatives at each position. On Tuesday, Google DeepMind announced that it has precomputed a prediction for the molecular effect of every one of them. Its result, the AlphaGenome Atlas, is a 1-petabyte dataset that the company says is more than 30 times larger than the AlphaFold Database. Big numbers in press releases deserve auditing, so we ran the math DeepMind did not show its work for, starting with the only question that matters about a 1-petabyte claim: does the arithmetic underneath it actually close.
Auditing the Petabyte
For each of the 9 billion variants, the Atlas carries what reporting on DeepMind's materials describes as roughly 27,000 predictions across eleven categories, from RNA splicing to transcription-factor binding across hundreds of human and mouse cell types, computed from a window of one million surrounding base pairs. Assume each prediction is a 32-bit float, the standard currency of machine-learning output. Four bytes.
Now multiply: 9 × 109 variants, times 27,000 predictions each, times 4 bytes. You get 9.72 × 1014 bytes, or about 0.97 petabytes, which means DeepMind's "1 petabyte" headline lands within a rounding error of the per-variant prediction count, a consistency check that takes ninety seconds with a calculator and that nobody, apparently, had bothered to run before this article. Bonus finding: roughly 108 kilobytes per DNA letter change, about the size of a medium web image, spent on describing what one altered letter of DNA might do to a cell.
| Input | Value | Source |
|---|---|---|
| Possible single-letter variants | 9 × 109 | DeepMind announcement, Sep 8 2026 |
| Predictions per variant | ~27,000 (11 categories) | Reporting on DeepMind materials |
| Bytes per prediction | 4 (float32) | Standard ML storage; stated assumption |
| Computed total | 0.97 PB | This article |
| Claimed total | ~1 PB | DeepMind announcement |
| Size vs AlphaFold Database | >30× | DeepMind announcement |
What Lives Inside
Raw predictions are only half the release. DeepMind also shipped the AlphaGenome Variant Impact (AVI) score, a single number per variant folding AlphaGenome's regulatory predictions together with AlphaMissense (the 2023 protein-altering model) plus evolutionary conservation and loss-of-function signals. Around 9,000 researchers had already been hitting the AlphaGenome API since the model shipped, but using it meant writing code and running a model one independent genomicist calls very slow and computationally heavy. This release answers the standing request: one score that says whether you should care. One number. Rank everything. The underlying model, published in Nature in January 2026, had already beaten the previous best on expression-effect benchmarks, lifting eQTL effect-magnitude correlation from 0.39 to 0.49 against the prior state of the art; the Atlas is what happens when you stop making researchers run it themselves.
In benchmarks described in a companion preprint, AVI ranked a known causal variant among the top 50 candidates 29.5 percent of the time, versus 12.5 percent for CADD, the ranking method it is meant to replace. Divide 29.5 by 12.5 and the answer is 2.36, which means that for every 100 causal variants, AVI surfaces about 17 more in a researcher's top-50 shortlist than the field's standard tool, and that is the kind of margin that decides whether a rare-disease lab finds the needle or re-reads the haystack.
Beyond single letters, the Atlas scores more than 100 million insertions and deletions and maps over 2,500 recurring regulatory motifs, the short DNA "words" transcription factors bind. Stowers Institute researchers used them to sort transcription factors into those that merely open DNA up versus those that also switch genes on, a distinction nobody had catalogued genome-wide before.
Two Labs Already Cashed It In
Working with the GREGoR Consortium, Broad Institute researchers applied the AVI score to unsolved rare-disease cases and prioritized a non-coding variant nobody had flagged, in DNM1, a gene strongly linked to epileptic encephalopathy. AlphaGenome's prediction came with a mechanism: the variant created an incorrect splice site that produced an abnormally extended protein. Experimental screens confirmed it. Prediction, mechanism, validation: the full chain.
At population scale, Gareth Hawkes at the University of Exeter ran the Atlas against whole-genome data from over 54,000 UK Biobank participants. Grouping rare variants by predicted molecular effect uncovered 22 percent more non-coding genetic associations than the noise would otherwise allow, pinning regulatory variants behind PLA2G7 (linked to aging) and EGLN1 (a cellular oxygen sensor). In a field where non-coding variants are famously where association signals go to die unexplained, a 22 percent lift from data that already existed is a real number.
An 80x Feat Nobody Mentions
Here is the part that matters more than the petabyte. Running a model that reads a million base pairs around each of 9 billion variants "seemed impossible" at the outset, in the words of DeepMind genomics lead Žiga Avsec. His team got there by cutting inference cost 80-fold with model distillation plus hand-tuned GPU kernels, turning an impossible computation into a project with a launch date. Without the 80×, this atlas is not a bigger project. It is not a project at all.
Pricing the Storage Bill
Now the part nobody prices. This Atlas is free for academic and non-commercial research. Free to use, that is; somebody still has to store it, and that somebody is Google. Its own Cloud Storage pricing lists Standard storage at $0.020 per GB per month: one million gigabytes times $0.020 is $20,000 per month, about $240,000 per year for raw storage, before serving, API traffic, or the GPU farm that generated it. A lab that wants the whole thing locally pays egress at $0.12 per GB, roughly $120,000 for one full download.
Google is paying at least a quarter-million a year to keep this searchable as a public good, while commercial access comes to Google Cloud "soon," so every academic lab that builds a workflow on the free portal is a future enterprise customer. Useful. Generous. Priced in petabytes.
Steel-Manning the Skeptics
At full strength, the case against celebrating this is simple: a prediction is not a measurement, and this petabyte is 9 billion guesses about cells, not facts about people. AlphaGenome predicts molecular effects in cell lines. Human disease happens in whole organisms, and most of it is polygenic, driven by many variants acting together, while the Atlas scores single letters one at a time. Its million-base-pair window still misses long-range enhancers beyond its field of view, and precomputation does not fix uncertainty; it just serves it faster. And the 29.5-versus-12.5 benchmark, the number that anchors the "worth it" case, comes from a preprint that has not yet survived peer review. Celebrate the map. But it is a map drawn by a model, and models are wrong in bulk, confidently, at petabyte scale, and the only thing more dangerous than a wrong prediction is a wrong prediction served instantly, for free, behind a beautiful interface, which is exactly what this release is.
What This Analysis Did Not Prove
Honest boundaries: the ~27,000 figure comes from secondary reporting on DeepMind's materials, not a published data dictionary, so the audit's headline check rests on that attribution. The $240,000-a-year figure is raw storage alone, so the true carrying cost is higher. The CADD comparison is arithmetic on preprint figures, and top-50 ranking is a proxy for usefulness, not a measurement of it. We did not test the portal or independently reproduce any benchmark. DeepMind's own disclaimer stands: the model "has not been validated for, and is not approved for, any clinical use."
What You Can Do
If you run variant analysis anywhere: stop re-running AlphaGenome inference yourself. The model is deliberately heavy, the API already serves ~9,000 researchers, and the Atlas exists so you never pay that compute bill twice. Query the portal or the API first, and use the AVI score as triage before spending a dollar on experiments. For core facilities, treat these predictions like reference data: pull once, annotate your pipelines, and do not download the petabyte over egress unless you have $120,000 and a reason.
If you are a clinician, a patient, or just someone with a sequenced genome: do not treat the AVI score as a diagnosis. It ranks variants by predicted impact; it knows nothing of your history, environment, or other variants. A high score is a research lead, not a verdict.
If you fund science or build research infrastructure, the lesson generalizes: precomputation is now the cheapest way to democratize an expensive model, because one heroic engineering effort plus a storage subscription beats ten thousand labs each re-running the same inference.
Bottom Line
DeepMind turned its hardest genomics model into a public lookup table: 9 billion possible DNA letter changes, 1 petabyte of predictions, free to any academic with a browser. Our audit says the petabyte arithmetic checks out to within a rounding error, the new impact score ranks causal variants 2.36 times better than the standard tool, and the "free" part costs Google roughly $240,000 a year in storage before anyone downloads a byte. But the pattern is the point: take an expensive model, engineer it 80 times cheaper, precompute once, then hand out free access to everyone. That is how you turn a capability only 9,000 coders could touch into infrastructure for all of biology, and when the same formula lands in protein design, materials, and climate, as it will, remember that the math checked out here first.