DeepMind's AlphaGenome predicts what a single change in DNA does to gene activity
AlphaGenome reads up to a million DNA letters at once. DeepMind reports it matched or beat other models on 24 of 26 tests.
Only a small part of DNA spells out proteins. Much of the rest acts as switches that decide which genes are busy in which cells, and researchers often cannot say what a change in that region does. That is the gap AlphaGenome aims at. Google DeepMind announced it in June 2025 and published the paper in Nature on 28 January 2026.
The model reads up to 1 million DNA letters at once and predicts what is happening along that stretch letter by letter, including where the cell cuts and rejoins a gene’s message before using it. Change one letter and it predicts how those readings shift, which is the question a geneticist looking at a patient’s DNA actually asks.
DeepMind reports that it beat the best rival models on 22 of 24 tests of describing a stretch of DNA, and matched or beat them on 24 of 26 tests of what changing a letter does. Researchers can use it free for non-commercial work.
What it means
This is a research tool, aimed at the scientists who study why a DNA difference raises someone’s risk of a disease. If its predictions hold up in other hands, it narrows the list of suspects they have to test in a lab. It is not a test for patients: DeepMind says it is not validated for clinical or personal genome use, and the scores so far come from DeepMind itself.
Caveats
- The test results are reported by DeepMind, the model’s maker, not by independent researchers.
- DeepMind says it isn’t validated for personal genome or clinical use, and that switches sitting far from the gene they control (more than 100,000 letters away) are still hard to predict.
- Per DeepMind’s page the trained model itself isn’t released, only access to it; one secondary source claims otherwise.
- Some write-ups say “25 of 26”; DeepMind’s blog says 24 of 26.