AlphaGenome Atlas: DeepMind precomputes the molecular effect of all 9 billion possible human DNA variants — 1 petabyte of free answers
🔎 Genomics shifts from computing to looking up
On September 8, 2026, Google DeepMind released AlphaGenome Atlas: the genomic equivalent of its famous AlphaFold database. Concretely, predictions of the molecular effect of the 9 billion possible single-nucleotide variants in the human genome — the 3 possible substitutions for each of the ~3 billion positions — were precomputed, then stored in a dataset of roughly 1 petabyte.
The shift is brutal. Until now, querying a genomic model about a variant required picking your candidates, writing code, and paying for the compute — every single time, for every team, for every question. From now on, the answer already exists, freely accessible from a simple web browser for academic research.
The announcement, picked up notably by Nature, was summed up by Forbes with a phrase that stings: "the AlphaFold playbook, with a cash register." The science is given away; the infrastructure will be sold.
And no, this Atlas has nothing to do with Boston Dynamics' humanoid robot — even though the principle is, curiously, the same: a machine that does everything, all by itself, while humans decide what to do with it.
The essentials
- 9 billion precomputed variants: the 3 possible substitutions for each position of the human genome (~3 billion base pairs), published on September 8, 2026 by Google DeepMind.
- ~1 petabyte of data, more than 30 times the size of the AlphaFold database — the most complete catalog of the effect of mutations on molecular biology.
- Free for academic research via the web portal alphagenome.google/atlas, without a single line of code. Also available via the AlphaGenome API (GitHub) and as a skill in Google Antigravity.
- New AVI score (AlphaGenome Variant Impact): a single number per variant, combining AlphaGenome and AlphaMissense, covering both coding and non-coding regions.
- Results already in: a variant of the DNM1 gene predicted to disrupt RNA splicing has been experimentally validated; +22% more non-coding associations detected across more than 54,000 genomes from the UK Biobank; 19 genomic regions potentially linked to BMI.
- The money: commercial access will go through Google Cloud "soon", pricing not announced (September 2026).
- The acknowledged limitation: a starting point for research, not clinical proof.
Recommended Tools
| Tool | Primary use | Price (September 2026) | Ideal for |
|---|---|---|---|
| AlphaGenome Atlas | Look up the predicted effect of a variant, no code required | Free (academic research) | Biologists, geneticists, teachers |
| API AlphaGenome (GitHub) | Integrate predictions into research pipelines | Free (academic use) | Bioinformaticians, genomics teams |
| Google Antigravity Skill | Have AI agents query the Atlas | Included in Antigravity | Research agent developers |
| AlphaGenome on Google Cloud Model Garden | Commercial use of the base model | Already paid; Atlas access "soon", pricing not announced | Biotechs, pharma |
One honest caveat: the web portal covers single-letter substitutions. For anything beyond that scope, the API and the model remain the right tool — and that's precisely where Google intends to charge.
Why 9 billion variants, and why it's a mini-revolution
Because the question "what does this mutation actually do?" is now being asked at the scale of the entire genome — and no laboratory can answer it by hand.
The human genome contains about 3 billion base pairs. At each position, three substitutions are possible: that makes 9 billion single-letter changes (IEEE Spectrum, 2026). Testing them in the lab, one by one? It's physically impossible, Gigazine points out — it would take centuries at the bench.
The worst part? About 98% of this genome doesn't code for proteins. These non-coding regions regulate gene expression and splicing, and the effect of mutations there remains largely misunderstood (SiliconAngle, September 2026). The gap to fill is enormous — and it's regulatory, not coding.
Before the Atlas, a researcher's daily routine looked like this: pick a few candidate variants, write code, run the AlphaGenome model — a costly computation — and start over. "Initially, it seemed impossible to do this computationally," Avsec, a member of the team, told IEEE Spectrum.
To make the precomputation possible, DeepMind had to speed up the computation by a factor of 80, using three levers: model distillation, GPU kernel optimization, and elimination of redundant calculations. Then the company ran the model once. For everyone.
This is the heart of the shift, summed up in a table:
| Before the Atlas | With the Atlas | |
|---|---|---|
| Access | A model you have to run yourself | A table of answers to consult |
| Skills | Code + GPU access | A browser |
| Coverage | Variants chosen case by case | All 9 billion substitutions |
| Turnaround | Hours to days per batch of variants | Immediate |
| Marginal cost | Computation on every query | Zero (academic use) |
A model you query a hundred times is research. The same model run 9 billion times for everyone is infrastructure. DeepMind understood the difference — and that's the entire announcement.
The AVI score: a single number to sort 9 billion mutations
AVI (AlphaGenome Variant Impact) is a single score per variant, obtained by combining AlphaGenome and AlphaMissense, which covers both the coding and non-coding regions of the genome.
Why add a score where detailed predictions already exist? Because with 9 billion rows, the number one problem becomes sorting. A single number per variant makes prioritization possible at a glance — this was the most requested feature by researchers, according to IEEE Spectrum.
Concretely, AVI ranks best-in-class on pathogenicity and rare disease benchmarks (AlphaSignal, 2026). And it isn't limited to the 2% of the genome that codes for proteins, where legacy tools stopped.
That said, be careful not to over-interpret. The Atlas isn't reducible to this score: it also predicts the molecular processes that are disrupted — gene expression, RNA splicing, protein function. That's the point emphasized by the Stowers Institute, a partner in the project: until now, no single resource allowed researchers to both rank variants at genome scale and understand what they disrupt.
A single score is a sorting tool. The molecular details, for their part, remain the raw material of any serious biological hypothesis.
DNM1, UK Biobank, BMI: concrete results from day one
Yes — and that's what sets this announcement apart from a mere press release: the Atlas has already produced verifiable results, one of which has been experimentally confirmed.
The most striking case comes from the GREGoR consortium. The AVI score surfaced a previously overlooked variant in the DNM1 gene — strongly linked to epileptic encephalopathy. The predictions indicated an incorrect splice site, lengthening the protein. Laboratory confirmation: it was exactly right (AlphaSignal, 2026).
At the population scale, the University of Exeter applied the Atlas to whole-genome data from more than 54,000 UK Biobank participants. The result: rare associations between non-coding variants and protein levels — an additional +22% non-coding associations compared to standard analyses. Signals that were previously invisible become detectable, simply because the machine has already been running.
The same logic applies to BMI: by filtering the top 1% of non-coding variants by predicted impact, the team identified 19 genomic regions potentially linked to body mass index. Not a diagnosis here — leads, ranked and ready to be tested.
At the Broad Institute, the AVI score made it possible to prioritize a previously ignored non-coding variant, associated with an unsolved rare disease case. The scientific consortium behind the project sets the tone: Google DeepMind, Stowers Institute, Broad Institute, University of Exeter, Memorial Sloan Kettering Cancer Center, and Stanford.
"AlphaGenome Atlas is a powerful example of how AI can extend human knowledge and advance scientific discovery," summarizes Pushmeet Kohli, VP of Science at DeepMind and Chief Scientist at Google Cloud.
What I see in this: AI doesn't replace experimentation, it targets it. The lab no longer searches for a needle in a haystack; it receives a haystack that has already been sorted.
"The AlphaFold playbook, with a cash register": the real issue is economic
Google is applying the exact AlphaFold recipe: give away the science, sell the place where it lives.
The grid is crystal clear (Forbes, September 10, 2026). Free for non-commercial use. GitHub API for academia. Google Antigravity skill for AI agents. And for businesses: Google Cloud Model Garden, availability "soon", price not announced. The base AlphaGenome model is already commercial there, for that matter — the precedent exists.
Jon Markman's analysis deserves to be quoted as is: Google ran the model once for everyone, then sells what keeps the science alive — hosting, licensing, compute. The price Google Cloud sets will say, in black and white, what a petabyte of answers is worth.
My reading: it's the same mechanics as the free AI APIs from Groq, Google, or OpenRouter. Academic free access isn't a gift, it's bait. It installs a standard, creates a soft dependency, and the bill arrives at the precise moment the value is demonstrated — when a biotech wants to industrialize.
The open question is fascinating: how much is a petabyte of biological answers worth on the market? Google Cloud's answer, in the coming months, will set a reference price for all the precomputed science to come.
The acknowledged limitation: a starting point, not proof
DeepMind says so itself: the Atlas's predictions are a starting point — promising variants still require experimentation.
Let's recall what the Atlas predicts: molecular regulatory effects. Not diseases, not diagnoses, not treatments (Gigazine, 2026). Between "this variant disrupts a splice site" and "this variant causes this disease," there lies all the remaining biological work.
The DNM1 case, incidentally, shows the proper use: prediction, then experimental validation. The "predict-verify" loop isn't short-circuited; it's accelerated. It's the same tension as with the AI avatar in customer service: automate the triage without eliminating the human — we replace the routine, never the responsibility.
Second limitation, technical: the base model accepts 1 million base pairs as input (SiliconAngle, 2026). Effects that exceed this window — very long-range genomic interactions, structural rearrangements — remain outside the precomputed scope.
Third limitation, epistemological: a single score per variant is by construction a compression. The AVI says "look at me" or "move along"; it doesn't say why. For the why, you have to go back down to the detailed predictions, then back up to the test bench.
My take: this candor is rare and welcome. Too many AI announcements sell the end of human work. Here, the positioning is clear — the Atlas is an amplifier of that work, not its substitute.
Massive precomputation: the next default strategy for AI science?
Yes — "pay for the computation once, serve free answers" is a repeatable recipe, and it will be copied.
The AlphaFold precedent had sketched it out: hundreds of millions of predicted structures, freely browsable. The Atlas takes it up a notch: exhaustive coverage of a finite combinatorial space. And the result weighs "only" 1 petabyte — run-of-the-mill storage by 2026, once the computation has been paid for.
The condition can be summed up in one sentence: when the number of future queries far exceeds the cost of an exhaustive computation, precomputation wins. Human genome, 9 billion substitutions: a textbook case. But the list of candidates is long — protein-ligand interactions, mutation effects on protein stability, climate ensembles, materials screening.
The movement extends far beyond DeepMind, for that matter: OpenAI solved the Navier-Stokes problem in 88 hours with 10,000 AI agents, formal proof included, published openly. AI-assisted science is no longer content with merely accelerating computation: it is industrializing the production of public results.
What remains to be seen is where all this leads. When a quantum physicist claims that AGI is already here, an Atlas containing answers to questions no one had yet asked looks less like a tool than a taste of things to come.
The bottleneck is shifting. Yesterday: computation. Today: storage and distribution. Tomorrow: experimental validation — and humans' ability to read 9 billion answers. Guess where the cloud providers want to position themselves.
❌ Common Mistakes
Mistake 1: treating the AVI score as a diagnosis
The AVI is a research prioritization score, not a clinical test. No medical decision can be based on it. The solution: use it for triage, then go through experimental validation and the usual clinical pathways — no shortcuts.
Mistake 2: querying the model when the answer is already in the Atlas
If you're looking for the effect of a simple single-letter substitution, the answer most likely already exists in the Atlas — writing code and paying for compute would be a waste. The solution: check the web portal first, and reserve the API for out-of-scope cases.
Mistake 3: ignoring non-coding regions on the grounds that they don't "code" anything
98% of the genome is non-coding and regulates gene expression. Focusing only on coding variants means reading just 2% of the book. The solution: leverage the Atlas's non-coding coverage, which is precisely its main added value over previous generations of tools.
❓ Frequently Asked Questions
Is AlphaGenome Atlas really free?
Yes, for academic research, via the web portal alphagenome.google/atlas — no coding skills required. The GitHub API is also free for academic use. Commercial use, meanwhile, will go through Google Cloud, with a price not yet announced as of now (September 2026).
What's the difference between AlphaGenome and AlphaGenome Atlas?
AlphaGenome is the AI model, which accepts up to 1 million base pairs as input. The Atlas is the product of its massive execution: the precomputed predictions of the 9 billion possible single-nucleotide substitutions, directly browsable. In short: the model computes, the Atlas answers.
What is the AVI score, in one sentence?
A single number per variant, combining AlphaGenome and AlphaMissense, that lets you judge at a glance whether a mutation — coding or non-coding — deserves attention, with best-in-class performance on pathogenicity and rare disease benchmarks (2026).
Can you use the Atlas with an AI agent?
Yes: the Atlas is available as a skill in Google Antigravity, Google's agentic environment. An agent powered by Gemini 3.1 Pro or GPT-5.5 can therefore integrate the Atlas's predictions into automated research workflows, without any genomics-specific development.
Does the Atlas replace laboratory experiments?
No, and DeepMind says so explicitly: it's a starting point. The DNM1 variant, predicted by the AVI and then experimentally confirmed, illustrates the right workflow — the AI proposes and prioritizes, the lab decides. Validation biology remains indispensable.
✅ Conclusion
By precomputing the molecular effect of the 9 billion possible variants of human DNA and offering it to science for free, DeepMind is changing the question posed to researchers: it's no longer "how much compute am I willing to pay for," but "what question will I ask of these 9 billion answers" — and the upcoming Google Cloud bill will tell us exactly what a petabyte of knowledge is worth. If the AI-science pairing intrigues you, revisit the OpenAI and Navier-Stokes case mentioned earlier: research is no longer improving — it's shifting gears.