What happened?
AlphaGenome is an AI model that reads up to one million letters of DNA and predicts thousands of measurements about how genes are controlled DeepMind blog ↗. Its main job is to estimate how a single change in DNA, called a variant, might alter those measurements DeepMind blog ↗. DeepMind previewed it in June 2025 and published it in Nature in January 2026 DeepMind blog ↗ Nature paper summary ↗.
Most human DNA does not hold recipes for proteins. Much of it acts like switches and dials that decide when, where and how much each gene is used, and about 98% of genetic changes fall in these non coding regions Nature paper summary ↗. Spotting which changes matter is hard, so scientists use computer models to decide what to test first.
AlphaGenome predicts things like where genes start and end, how RNA is cut and joined (called splicing), how much RNA is made, and which DNA is open for proteins to bind, across hundreds of human and mouse cell types DeepMind blog ↗. The Nature paper reports 5,930 human and 1,128 mouse signals predicted at once, and says it matched or beat leading models on 25 of 26 variant effect tests Nature paper summary ↗. The earlier preview reported 24 of 26 DeepMind blog ↗.
One model that looks at long stretches of DNA and many kinds of signal at once could help scientists sort a huge list of DNA changes down to the few worth testing.
Leapscope interpretation of the reported result.How did AI help?
The AI model is the result. DeepMind researchers trained it on large public datasets of lab measurements from human and mouse cells Nature paper summary ↗. In one example, the team checked whether it could explain known cancer causing mutations near a gene called TAL1 in a type of leukaemia, and its prediction matched the known mechanism DeepMind blog ↗. Scientists at the Wellcome Sanger Institute said they tested it against over half a million new experiments and found it performed very well Expert reaction ↗.
Figures from the Nature paper summary Nature paper summary ↗; the input length also appears in DeepMind's announcement DeepMind blog ↗.
Outside experts also named real limits Expert reaction ↗. One geneticist said it performs as well as the best existing tools but not better, and that current AI models are not reliable enough for patient care. DeepMind itself says it struggles with DNA switches more than 100,000 letters away from a gene, was not designed for predicting an individual's genome, and is not validated for clinical use DeepMind blog ↗.
Which fields could this affect?
Some uses are relevant now and others are possible future value; the connections below are our assessment.
Genomics research
Researchers can query the model through a free API to rank DNA changes for follow up GitHub ↗. This can save lab time by narrowing what to test first.
Explore scienceSplicing studies
One outside expert praised its splicing predictions and said it could soon replace older standard tools there Expert reaction ↗. This is useful for studying genetic diseases caused by faulty RNA cutting.
Explore scienceRare disease diagnosis
Better ranking of variants could help doctors find the cause of rare conditions. Experts warn the model is not yet reliable enough for patient care Expert reaction ↗.
Explore healthcarePersonal genome reading
Predicting traits or disease risk from one person's DNA is not what the model was built or checked for DeepMind blog ↗. This report demonstrates no personal medical use.
What has been checked?
The evidence is a peer reviewed Nature paper with benchmark tests, independent expert reactions, and a public research API. Leapscope reviewed these sources; we did not repeat the experiments.
Shown so far
- The Nature paper reports it matched or beat leading models on 25 of 26 variant effect tests Nature paper summary ↗.
- The Sanger Institute tested it on over half a million new experiments and reported strong performance Expert reaction ↗.
- It reproduced a known disease mechanism near the TAL1 gene in leukaemia DeepMind blog ↗.
Still unknown
- How well it captures differences between specific cell types, which experts call an important challenge Expert reaction ↗.
- How accurate it is for very rare variants with little training data Expert reaction ↗.
- Whether its rankings improve real diagnoses once tested in clinical studies.
Evidence status: Research preview. Stage: Usable. Available to researchers for non commercial use.
From prediction to diagnosis
This is our suggested way to follow it, not a promised timetable.
Can I use it today?
Researchers can get a free API key for non commercial use and run small to medium analyses through Python GitHub ↗. It is not a consumer product and is not validated for medical decisions DeepMind blog ↗.
A few things you might be wondering
Can AlphaGenome read my DNA and tell me my disease risk?
No. DeepMind says it was not designed or validated for personal genome prediction or clinical use DeepMind blog ↗.
Did it discover new biology?
One outside expert said it reveals no new biology and performs about as well as the best existing tools Expert reaction ↗. Its value is doing many predictions together across long DNA stretches Nature paper summary ↗.
Is it better than older tools?
The Nature paper reports it matched or beat them on 25 of 26 variant tests Nature paper summary ↗. Experts still call it far from perfect, mainly because training data is limited Expert reaction ↗.
Go straight to the sources
Checked Oct 8, 2026. The first source is the original announcement or research. Later sources add independent context; background pages do not validate the result on their own.
01The preview announcement with capabilities, benchmark results, examples and stated limits.
Summary of the peer reviewed Nature paper with key figures.
Comments from seven outside scientists on strengths and limits of the model.
The Python client, access terms and usage limits for the research API.