DeepMind maps all 9B human DNA variants

Human DNA contains around 3 billion base pairs, but scientists only really understand the 2% that codes for proteins. The other 98% has long been a biological black box, even though changes hidden inside can trigger devastating diseases.

Google DeepMind just released AlphaGenome Atlas, a massive predictive database mapping the molecular consequences of all 9 billion possible single-letter genetic variations across the entire human genome.

Precomputing 9 billion mutations into 1PB of data

DNA is written in four letters: A, C, G, and T. Swapping a single letter can be completely harmless, or it can cause severe illness. Testing all 9 billion single-letter substitutions in a wet lab is practically impossible.

While DeepMind had already introduced the AlphaGenome model to predict molecular impacts, running the model from scratch for individual variants was a computational hurdle. Ziga Avsec, genomics lead at DeepMind, explained that it took significant time to precompute this vast search space.

The result is a massive 1-petabyte dataset, more than 30 times larger than the historic AlphaFold Database. Instead of running heavy models, researchers can now simply look up any mutation like an entry in an encyclopedia.

To make sense of the ocean of numbers, DeepMind introduced the AlphaGenome Variant Impact (AVI) score. It combines AlphaGenome's non-coding insights with AlphaMissense, their existing tool for protein-altering mutations, condensing everything into a single priority score.

Solving cold cases in clinical genetics

External teams testing early access are already seeing concrete results.

At the Broad Institute, Laura Covill and her team used the AVI score to track down an elusive mutation in the DNM1 gene. The model flagged that the change would create an incorrect splice site, handing researchers the vital piece of evidence needed to crack an unsolved rare disease case.

In another study, Dr. Gareth Hawkes applied the atlas to genetic records from more than 54,000 UK Biobank participants. By grouping mutations by predicted molecular function, his team found 22% more non-coding genetic associations, isolating 19 genetic regions tied to body mass index (BMI) within the top 1% of impactful variants.

The atlas is accessible through a free, no-code web portal for academic researchers, alongside an API and integration into Google's Antigravity platform. Google confirmed that commercial access will follow shortly via Google Cloud.

Why it matters

We think DeepMind's smartest move here was turning expensive AI inference into ready-to-use biological infrastructure.

Most medical researchers and small genetics labs do not have the compute clusters or engineering teams required to run massive deep learning models across full genomes. Precomputing all 9 billion single-letter swaps takes the compute barrier off the table. A geneticist looking at a patient's mystery mutation can simply query the portal and get instant actionable priorities.

For the AI landscape, this shows DeepMind doubling down on its biology moat. Following Demis Hassabis and John Jumper winning the 2024 Nobel Prize in Chemistry for AlphaFold, DeepMind is clearly positioning itself as the foundational operating system of digital biology, feeding upstream discoveries directly into Hassabis's drug discovery vehicle, Isomorphic Labs.

That said, researchers must treat these numbers as predictions rather than ground truth. An algorithmic score is an educated roadmap, not clinical proof. Lab benches will still have to run the physical experiments, and how well the model handles the subtle complexities of non-coding regulatory loops across diverse patient populations remains to be seen.