Google DeepMind says its AlphaGenome Atlas gives researchers a score for each of the 9 billion possible single-letter changes in human DNA, to help them decide which variants to study first. DeepMind also released AlphaGenome 2, which predicts effects in individual cell types, and describes both tools as research aids, not clinical ones.
Research scientist Žiga Avsec explained the work in Google’s Ask a Scientist series on November 17, 2026. DeepMind’s own AlphaGenome 2 post, dated November 11, 2026, adds the technical detail.
What is AlphaGenome Atlas?
The human genome has 3 billion base pairs, and each can be swapped for three other letters, which gives 9 billion single-letter changes. Many genetic diseases trace back to one such change, but labs cannot test them all.
DeepMind used AlphaGenome, introduced in June 2025, to predict the outcome of every one of those changes, and published the results as AlphaGenome Atlas in September. Each variant gets one AVI (AlphaGenome Variant Impact) score, which DeepMind says shows the potential impact across the genome and helps researchers pick the most disruptive changes to follow up.
- Input: a partial DNA sequence of 1 million base pairs
- Output: thousands of predictions per sequence
- Scale: predictions for all 9 billion single-letter changes

What does AlphaGenome 2 add?
AlphaGenome was trained mostly on whole tissues or cell lines, and DeepMind names cell type specificity as its key limitation. AlphaGenome 2 takes a 1 million letter sequence plus one condition, such as a tissue or cell type, and predicts effects for that condition. DeepMind says it was trained on more than 100 million individual cell profiles across 68 tissues.
DeepMind reports that it finds nearly twice as many expression-changing variants in cells as the previous state-of-the-art model. On a dataset of 28 immune cell types it reports causal variants at 30% versus 16% for the original model fine-tuned on single-cell data, and directional gene effects at 43% versus 23%. These figures come from DeepMind and its collaborators, and the post reports no independent validation. The paper is a bioRxiv preprint.
What have researchers done with it so far?
According to DeepMind, Gareth Hawkes of the University of Exeter used the Atlas to narrow a list of 1 billion variants. Focusing on the top 1%, he found 19 variant groups linked to body mass index to prioritize.
Laura Covill of the Broad Institute used the AVI score on unsolved rare diseases. The Atlas flagged a critical variant in a gene linked to neurodevelopmental disorders, which gave supporting evidence for a case that was later solved.
Prices
No price is listed. Non-commercial researchers can use the AlphaGenome API and Colab tutorials, according to DeepMind. Commercial research is in a private preview through Google Cloud, with an interest form on the AlphaGenome 2 page. DeepMind also invites use case submissions until November 30, and says selected projects may get extra quota and credits. Check the DeepMind page for current terms.
Who it suits
Genetics researchers. Anyone sifting through millions of variants from a patient sample gets a way to rank them. Avsec gives the example of 4 million variants in one rare disease patient, narrowed to rare ones and then still hundreds or thousands of options.
Drug discovery teams. DeepMind says the tools could help pharmaceutical researchers identify and validate promising therapeutic targets faster.
Agriculture researchers. DeepMind suggests the model design could help with breeding climate-resilient crops, though it gives no results yet.
General readers and patients. The tools do not diagnose anything. DeepMind says AlphaGenome 2 has not been validated or approved for clinical or diagnostic use, and that predictions should be read with other sources of information.
What is still unknown?
DeepMind says it is still working on better ways to measure how confident the model is, and that directional predictions are made only when it is confident. The long-term goal of a full digital simulation of a human cell is a stated aim, not a product.
Next step
Read DeepMind’s AlphaGenome 2 post for the paper link, API access and the interest form, and treat the figures above as the developer’s own results until outside groups report on them.







