DeepMind's new genome atlas grades all 9 billion possible DNA typos

Key points

• Google DeepMind launched AlphaGenome Atlas on 8 September 2026, a free academic database of predicted effects for all 9 billion possible single-letter DNA changes in the human genome.
• The 1-petabyte dataset is more than 30 times the size of the AlphaFold protein database DeepMind expanded in 2022.
• A new AlphaGenome Variant Impact (AVI) score condenses predictions into one number for each variant, covering protein-coding and non-coding DNA alike.
• Broad Institute researchers used the AVI score to flag a variant in the DNM1 gene linked to epileptic encephalopathy, later confirmed in the lab.
• A University of Exeter fellow applied Atlas to whole-genome data from over 54,000 UK Biobank participants to find rare variants tied to common traits.

Google DeepMind has published what it calls the most complete map ever made of how tiny DNA typos might affect the human body. AlphaGenome Atlas went live on 8 September 2026 as a free tool for academic researchers, and it's the kind of resource that tends to quietly reshape a field over years rather than make headlines for a week.

The numbers are hard to absorb. There are roughly 9 billion single-letter changes possible in human DNA. Testing each one in a lab would take lifetimes. DeepMind ran them all through its AI model instead and saved the answers.

What is AlphaGenome Atlas, in plain English?

It's a giant lookup table of predictions. For every possible one-letter change in the human genome, the atlas estimates what that change might do to nearby biology: which genes turn on or off, how proteins get built, how DNA is packaged inside cells.

Think of DNA as a four-letter alphabet (A, C, G, T) spelling out instructions for the body. Change one letter and sometimes nothing happens. Sometimes a child is born with a serious disease. Sorting the harmless typos from the dangerous ones was mostly guesswork for anything outside the small slice of DNA that directly codes for proteins. Atlas is an attempt to end that guesswork at scale.

The atlas is built on AlphaGenome, DeepMind's earlier AI model that reads stretches of DNA and predicts their molecular effects. Precomputing every answer turns a research tool into a reference book.

How big is the dataset?

One petabyte, roughly a million gigabytes. That's more than 30 times the size of the AlphaFold protein structure database. DeepMind wants Atlas to play the same role for genetics. Our earlier genomics coverage from 11 August showed how AI reading genetic data as a connected system, rather than a list of parts, already uncovered 641 genes linked to schizophrenia. Atlas extends that logic to the whole genome.

Researchers can reach it three ways: a website portal for scientists who don't code, an API for those who do, and a skill inside Google Antigravity.

What is the AVI score?

The AlphaGenome Variant Impact score is a single number summarising how disruptive a given DNA change is likely to be. It blends predictions from AlphaGenome with those from AlphaMissense, DeepMind's older model for protein-altering variants.

The score works for the 98% of the genome that doesn't code for proteins. That non-coding stretch controls when and where genes switch on, and it houses most of the DNA changes linked to common traits and diseases. It's also been the hardest to interpret. Each score comes with feature attributions, a plain readout of which biological processes (splicing, gene expression, DNA accessibility) are driving the number, so researchers get the ranking and the reason at once.

Has it found anything useful yet?

Yes, in two early studies with academic partners.

At the Broad Institute, Laura Covill and Anne O'Donnell-Luria, working with colleagues through the GREGoR Consortium, re-examined unsolved rare disease cases. They flagged a variant in a gene called DNM1, strongly associated with epileptic encephalopathy, a severe seizure disorder in children. The prediction said the variant created a faulty splice site, causing the cell to build an abnormally long protein. Lab experiments confirmed it.

At the University of Exeter, Medical Research Council fellow Gareth Hawkes ran Atlas against whole-genome data from more than 54,000 UK Biobank participants. Grouping rare variants by their predicted molecular effects lifted signals for common traits out of the statistical noise.

What does this mean for patients?

Not much directly, and not soon. Atlas is a research tool, not a diagnostic. Families waiting on a rare disease diagnosis won't log in themselves. But the labs that do this detective work now have a faster way to shortlist suspect variants, which could shorten the years-long odyssey many families face before getting an answer. That's the real bet DeepMind is making: not one dramatic cure, but thousands of smaller searches that finally end.