Google DeepMind Maps Every Possible DNA Change in the Human Genome
A new tool called AlphaGenome Atlas holds predictions for nine billion genetic variants. Researchers can use it free today, and here is what that actually means.
Key points
- Google DeepMind released AlphaGenome Atlas on Tuesday, a catalogue of predictions for all nine billion possible single-letter changes in the human genome.
- The dataset is roughly 1 petabyte in size, about a million gigabytes.
- Atlas is free for non-commercial research now, with commercial access on Google Cloud coming soon.
- A companion scoring tool called the Variant Impact Score lets scientists quickly rank which DNA changes are most likely to matter.
- The project builds on AlphaGenome, a model related to AlphaFold, which won Demis Hassabis and John Jumper the Nobel Prize in Chemistry in 2024.
Your DNA is a set of instructions written in a four-letter chemical alphabet: A, C, G and T. The human genome contains roughly three billion of these letter pairs, governing everything from eye colour to how your body fights infection.
Swap one letter in the wrong place and you might get a disease. Swap it somewhere harmless and nothing happens. The trouble is there are about nine billion possible single-letter swaps, and figuring out which ones matter has always been brutally slow work.
Google DeepMind thinks it has a shortcut.
What does AlphaGenome Atlas actually do?
Atlas holds a pre-computed prediction for every one of those nine billion variants, showing researchers how each change might alter biology at a molecular level, such as shifting how much of a particular protein the body produces.
Think of it like a giant reference book. Instead of running an experiment on one DNA change at a time, a scientist can look up a variant in seconds and get an immediate read on whether it looks dangerous or worth investigating further.
The underlying model, AlphaGenome, was trained on public databases of human and mouse genomes. It learned to spot patterns between DNA changes and biological processes, much the way a spell-checker learns which letter combinations look wrong. Applying those patterns to every possible variant produced a dataset roughly 1 petabyte in size.
DeepMind's genomics lead Ziga Avsec explained, in a press briefing reported by The Verge AI, that pre-computing predictions for a space this large simply took time: "Basically it took us some time to really precompute and also analyze this many variants because the space is so big."
We covered the Atlas launch on 8 September 2026 in "DeepMind's new genome atlas grades all 9 billion possible DNA typos", where early users reported wins in rare disease searches.
What does the Variant Impact Score add?
Atlas does the look-up. The Variant Impact Score, or AVI, does the ranking.
It draws on several of Google's other prediction models to give each variant a single score, helping researchers sort a long list of suspects and focus on the ones most likely to drive disease. Ranking first, then digging into the molecular detail, is a faster workflow than the other way round.
Should researchers trust it?
Atlas is a prediction tool, not a diagnostic one. Its conclusions come from patterns in training data, so results still need experimental confirmation before anyone draws firm medical conclusions.
Atlas extends into the parts of the genome that don't code for proteins directly, but instead control when and how genes switch on or off. Earlier tools like AlphaMissense focused mainly on protein-coding regions, so this is a genuinely broader view. The honest watch-out: broader coverage also means more predictions that nobody has yet tested in a lab.
Researchers can explore Atlas through a web portal or through Google's agentic development platform. It's free for non-commercial use now, with commercial access on Google Cloud arriving soon.
Common questions
Will this lead to new medical treatments?
Possibly, but not quickly. Atlas helps scientists identify which genetic variants are worth studying. Turning that knowledge into a drug or therapy still takes years of lab work and clinical trials.
Is my personal genetic data involved?
No. AlphaGenome was trained on publicly available databases of human and mouse genomes, not on data from consumer DNA-testing services or individual patients.
