AlphaGenome Atlas Maps All 9 Billion Human DNA Variants

What Google DeepMind Released on September 8

Google DeepMind has published the AlphaGenome Atlas, a free repository of precomputed predictions covering every possible single-letter change in human DNA. The human genome runs to roughly 3 billion base pairs, and each position admits three possible substitutions, which is how the Atlas arrives at 9 billion variants. Instead of asking researchers to run the AlphaGenome model themselves, DeepMind has run it across the whole genome in advance and put the results behind a web interface.

• 9 billion single-nucleotide variants covered — every possible one-letter change in the human genome
• Roughly 1 petabyte of predictions, which DeepMind says is about 30 times the size of the AlphaFold Database
• Free web access for academic and non-commercial research, with commercial access via Google Cloud coming later
• An AVI score (AlphaGenome Variant Impact) condenses each variant into a single number, paired with feature attributions

The predictions span hundreds of human and mouse cell types and tissues, and the release also catalogues more than 2,500 recurrent sequence motifs — the repeating "words" that regulatory machinery reads.

Why Precomputing Matters More Than It Sounds

The interesting engineering story here is not the model, which has been available for a while. It is the decision to run it exhaustively. AlphaGenome is a sequence-to-function model, and it is expensive. Carl de Boer of the University of British Columbia, quoted by IEEE Spectrum, called it the field's leading model while noting it is very slow and computationally intensive — precisely the kind of tool that quietly excludes labs without a GPU cluster.

Doing the work once, centrally, removes that barrier. DeepMind genomics lead Žiga Avsec told IEEE Spectrum the team had to improve calculation speed by a factor of 80 to make a full sweep tractable, using model distillation and GPU optimisation. That is the same shape of story as the AlphaFold Database in 2022: the capability existed beforehand, but the precomputed resource is what actually changed day-to-day practice for biologists.

What Does the AVI Score Actually Tell You?

The AVI score is the Atlas's headline usability feature. It folds AlphaGenome and AlphaMissense predictions together with conservation and protein loss-of-function signals into one number per variant, so a researcher scanning a list can see at a glance which changes are worth a second look. Each score comes with feature attributions that decompose it into the molecular processes driving it — RNA splicing, gene expression, and so on — which is a meaningful difference from a bare confidence value.

DeepMind is clear that this is a ranking aid rather than a verdict, and outside researchers have made the same point. De Boer cautioned that a single-number summary of something as tangled as gene regulation is easy to misread. Several genuine limits apply: many diseases involve combinations of variants, some enhancers act from beyond the model's roughly 1-million-base-pair window, and every prediction still needs laboratory validation.

Where This Fits in the AI-for-Science Trend

The Atlas lands in a stretch where the most consequential AI research releases have been infrastructure rather than chatbots. It follows WeatherNext 3's move into hourly 5km forecasts and sits alongside clinical work like ChatGPT Health's Epic records integration. The pattern is consistent: take a capable model, run it against a domain exhaustively, and hand the output to the specialists who know what to do with it. More on that shift is in our AI coverage.

For rare-disease work in particular, the practical value is triage. A clinical genomics team looking at a patient with an unexplained condition often faces thousands of candidate variants and no principled way to order them. A ranked list with molecular reasoning attached does not diagnose anyone, but it decides what gets tested first — and in a field where a single functional assay can take weeks, that ordering is most of the work.

Sources: Google DeepMind — September 8, 2026; IEEE Spectrum — September 8, 2026.