Google DeepMind Maps Every Possible DNA Change in the Human Genome

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, equivalent to about 200,000 DVDs worth of information.
• Atlas is free for non-commercial research starting now, with a commercial version 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 extends Google DeepMind's earlier work on AlphaFold, the protein-prediction model that won 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, and together they govern everything from eye colour to how your body fights infection.

Change a single letter in the wrong place, and you might get a disease. Change it somewhere harmless, and nothing happens at all. The problem is that there are about nine billion possible single-letter swaps across the genome, 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, telling 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 to test 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, neutral, 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, the same 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. That is about a million gigabytes.

DeepMind's genomics lead Ziga Avsec explained, as 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."

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 quickly 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 much faster workflow than the other way around.

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.

Crucially, Atlas extends into the parts of the genome that do not 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. This is a broader view.

Researchers can explore Atlas through a web portal or through Google's agentic development platform. It is 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.