DeepMind scores all 9 billion human DNA variants in AlphaGenome Atlas

Google DeepMind has published a prediction for every single-letter change that can be made to human DNA. The company introduced AlphaGenome Atlas on 8 September, a repository covering all 9 billion possible single-nucleotide variants in the genome. It’s free for academic research through a website portal, so a biologist doesn’t have to write code to query it.

The Atlas is what you get by running an existing model to exhaustion rather than one query at a time. DeepMind precomputed AlphaGenome’s output at every position in the reference genome, producing roughly 1 petabyte of data. The company says that’s more than 30 times the size of the AlphaFold Database. The accompanying preprint adds that each variant carries an average of 27,000 experiment-specific scalar predictions, alongside more than 100 million indels scored from gnomAD, UK Biobank and All of Us.

On top of that sits the AlphaGenome Variant Impact score, or AVI, one number per variant. It folds AlphaGenome’s regulatory predictions together with AlphaMissense, DeepMind’s protein-variant model, then converts the result to a PHRED scale where 10 marks the top 10% of predictions and 20 marks the top 1%. That matters because ranking is exactly what researchers kept asking for.

DeepMind prediction database	Released	Scale
AlphaFold Database	Expanded 2022	More than 200 million protein structure predictions
AlphaMissense	2023	71 million protein-altering variants classified
AlphaGenome Atlas	8 September 2026	9 billion single-nucleotide variants, about 1 petabyte

Demand for the underlying model was already there. Around 9,000 researchers had reached AlphaGenome’s predictions through its API, DeepMind product manager Dhavi Hariharan told Nature, but doing so meant writing software. Compute was the second barrier, which is why the Atlas took engineering rather than just patience. IEEE Spectrum reports that early estimates put the job out of reach until the team raised its calculation speed by a factor of 80, using model distillation, GPU kernel optimisation and the removal of redundant calculations.

If you remove the friction, you also increase the curiosity for people to dive in. Instant access is something that feels magical.

Žiga Avsec, AlphaGenome team lead, Google DeepMind, via Nature

What it has already turned up

Working with the Broad Institute, researchers used the AVI score to re-examine rare disease cases that earlier work had left unsolved. They flagged a variant affecting DNM1, a gene strongly linked to epileptic encephalopathy, and the predictions showed the mechanism: the variant created an incorrect splice site that abnormally extended the resulting protein. Experimental screens validated it. That’s triage before anyone touches a bench, the same shape of win as when an LLM literature search surfaced 17 cohorts four catalogues had missed.

A second result came from population genetics, and it scales further. Gareth Hawkes, a Medical Research Council fellow at the University of Exeter, applied the Atlas to whole-genome data from over 54,000 UK Biobank participants. Grouping rare variants by predicted molecular effect uncovered 22% more non-coding genetic associations than were otherwise detectable in the statistical noise. Narrowing to the 1% of non-coding variants the Atlas rated most impactful, he identified 19 genetic regions relevant to body mass index.

What the single score hides

Carl de Boer, a genomicist at the University of British Columbia who isn’t affiliated with DeepMind, calls AlphaGenome the field’s leading model while also describing it as “very slow and computationally intensive”. His reservation about the Atlas is the top-line number itself. “It has a clear use, but it also is probably going to be easily misinterpreted,” he told IEEE Spectrum. “We’re talking about a very complex system, and there’s a lot of moving parts.”

Field of view is the other limit. AlphaGenome reads 1 million base pairs around the variant in question, and some enhancers regulate genes from further away than that. Martin Kircher, a bioinformatician at the Max Delbrück Centre for Molecular Medicine in Berlin, told Nature the Atlas won’t replace experiments, or replace accounting for differences specific to individuals when diagnosing disease, though he called it “a useful and generous way to scale up access to a strong model”.

DeepMind’s own disclaimer states that AlphaGenome hasn’t been validated or approved for any clinical use. Commercial access is the next thing to watch, since the company says paid use is coming to Google Cloud while the non-commercial portal is live today. Downstream, validation in the lab is still the bottleneck, which is where work on robots running long scientific protocols meets predictions like these.