Google DeepMind has released AlphaGenome 2, a new AI model that predicts what a change in someone’s DNA does, and, for the first time in its AlphaGenome line, in which specific kinds of cells it does it. That matters because most DNA variants linked to disease sit outside genes, in stretches that act as on-off switches, and the same switch can behave very differently in an immune cell than in a heart or brain cell.
DeepMind announced the model on Wednesday 11 November UK time (early Thursday in eastern Australia), with a technical paper on its website. Scientists outside the company have called it a big step. The model’s headline test also leans on an Australian dataset: blood samples from nearly 2,000 Australians, collected in Sydney and Hobart. Here’s what AlphaGenome 2 does, how strong the claims are, and who can actually use it.
What AlphaGenome 2 does
Like the original AlphaGenome, released last year, the new model reads a stretch of DNA about a million letters long and predicts what that sequence does inside a cell. That includes which genes are switched on and how strongly, how open or tightly packed the DNA is, and chemical marks such as DNA methylation, which is new in this version. To estimate what a variant does, researchers run the prediction twice, once with the usual DNA letter and once with the changed one, and compare the results.
The big change is how cell types are handled. AlphaGenome 1 predicted a fixed list of lab measurements, mostly from “bulk” tissue samples that blend many kinds of cells together. AlphaGenome 2 instead takes the target cell type or state as an input, alongside the DNA, and makes its prediction for that one condition. According to DeepMind’s announcement, it was trained on more than 100 million individual cell profiles across 68 tissues, and the company says this “turns static cell atlases into a predictive engine for human genetics”.
It can also work with cells it never saw in training. A researcher can upload single-cell gene-activity data from a few dozen cells, for example from a disease or after vaccination, and the model adapts its predictions without being retrained. Nature reports the model was trained on 637 human and 311 mouse cell types. Professor Sarah Teichmann, a co-founder of the Human Cell Atlas consortium, is quoted in DeepMind’s post: “What makes AlphaGenome 2 exciting is how it adapts directly to new single-cell data without retraining.”
How strong are the claims?
The headline numbers come from DeepMind’s own testing, described in a paper the company published itself; it has not yet been through journal peer review. On a large map of genetic effects across 28 types of immune cell, AlphaGenome 2 picked out 30% of the likely causal variants at a strict 90% precision threshold, against 16% for the original model fine-tuned on the same kind of single-cell data. For predicting whether a variant turns a gene up or down, the figure rose from 23% to 43%. That’s the basis for DeepMind’s claim that the new model finds “nearly twice as many” variants that change gene activity in cells.
The most striking example is a case from inflammatory bowel disease. In 2024, a team at London’s Francis Crick Institute showed that a variant in a stretch of DNA far from any gene drives the condition by overactivating a gene called ETS2 in gut immune cells. DeepMind says AlphaGenome 2 reproduced that finding computationally, singling out the variant and the cell type. “They’ve managed to figure out the answer in 30 minutes, and it took us a year and a half of experiments,” the 2024 study’s leader, James Lee, told Nature.
Independent experts are positive but measured. Professor Sir Peter Donnelly, chief executive of UK drug-discovery company Genomics, writes that the model is a “step change” and that its predictions of whether a variant turns a gene up or down are “markedly better” than its predecessor’s. But he adds: “Even with these improvements, AlphaGenome 2 makes confident predictions for only a small subset of the variants that can be queried.” By his reading of the paper, it makes confident calls for about 2% of the roughly 6.8 million variants scored, though far more often among variants already flagged as likely causes. When the model stays quiet, he says, that “should be read as ‘no information’, not as ‘no effect'”. Genomics says it will now test the model across its own disease-genetics data.
DeepMind’s paper is also upfront about the gaps. “At high precision, AlphaGenome 2 recovers only a minority of fine-mapped eQTLs and putatively causal GWAS variants,” it says, noting the model can underestimate effects outside accessible regions of DNA and still struggles with fine-grained states such as subtypes of brain cells. Google DeepMind’s genomics lead, Žiga Avsec, told The Verge the new model’s predictions are “more expensive” to produce than the original’s, which is why it won’t replace AlphaGenome 1 for genome-wide jobs any time soon.
The Australian data in the headline test
The immune-cell benchmark behind DeepMind’s biggest claim is TenK10K, a project led from Sydney’s Garvan Institute of Medical Research. Its first results were published in Nature on 28 October. “In the first phase of our TenK10K project, we’ve sequenced the genomes of nearly 2,000 Australians and used the latest single-cell technologies to measure gene activity in more than 9 million of their immune cells,” Garvan’s Professor Joseph Powell said at the time.
According to the TenK10K project site, phase one covers 1,925 donors from two cohorts: 950 adults recruited in Hobart through the Tasmanian Ophthalmic Biobank, and 975 adults attending for cardiac imaging or treatment in Sydney through the BioHEART-CT study. DeepMind’s paper describes TenK10K as an example of the larger cohorts needed to map genetic effects in single cells, and lists it among the datasets used in its work.
One caveat applies to anyone reading the results. TenK10K’s own site notes that “Donors are of genetically inferred European ancestry”, and because BioHEART recruits through cardiac care, it isn’t a random sample of the population. So a strong score on this benchmark says less about how well the model works for people of other ancestries. Garvan says the next phase will increase the genetic diversity of participants.
Who can use it, and what it costs
- Researchers (non-commercial): AlphaGenome 2 is available now through the AlphaGenome API, which DeepMind’s GitHub page says is “offered free of charge for non-commercial use”, with Colab tutorials. Query rates vary with demand, and live model predictions suit small analyses rather than genome-wide runs.
- Companies: commercial research access is in private preview only. DeepMind’s GitHub page also says an AlphaGenome model is offered for commercial use through Google Cloud.
- Downloads: DeepMind has published precomputed predictions, including scores for about 7 million common and trait-linked variants. The paper says the model’s code and weights “will be provided upon final publication”.
- Coming soon: DeepMind says AlphaGenome 2 “Skills” are coming to Google Antigravity, its agentic development platform, alongside Gemini 4.
The terms of use say the services are “only available for use by individuals and non-commercial organizations (universities, non-profit organizations and research institutes, educational and government bodies), or for journalism”. They bar commercial work, including research on behalf of companies, and clinical use. The AlphaGenome-specific terms don’t list any country restrictions, so Australian universities and institutes appear able to sign up under the same rules as everyone else; users outside the European Economic Area and Switzerland contract with Google LLC.
It isn’t a consumer health tool. DeepMind says AlphaGenome and AlphaGenome 2 “are intended for research use only and have not been validated or approved for clinical or diagnostic purposes”. If you have results from a consumer DNA test, this won’t tell you your disease risk, and the terms say outputs shouldn’t be relied on for medical advice.
Our read: the real advance here is the cell-level detail. As Avsec put it to The Verge, “When you develop drugs, you’re targeting specific proteins and specific cell types.” A model that can suggest which variant matters, whether it turns a gene up or down, and in which cell, could save labs months of work choosing experiments. But the strongest results so far are DeepMind marking its own homework, on a benchmark built largely from people of European ancestry, and the model is silent on most variants. Independent checks like the one Genomics has planned will show how much it adds in practice. For Australian researchers, the free non-commercial access, and the fact that home-grown data like TenK10K is now part of how these models are measured, are the parts worth watching.
It’s the latest health-related result from Google’s AI research, after its Nature study on spotting high blood pressure with a smartwatch. The Gemini 4 model DeepMind mentions for Antigravity is covered in our Gemini 4 Carbon explainer.
Details in this article come from Google DeepMind’s announcement and technical paper, the AlphaGenome API documentation and terms, the Garvan Institute and TenK10K, and reporting by Nature, The Verge and Genomics, all published between 28 October and 11 November 2026. Performance figures are DeepMind’s own and haven’t yet been peer reviewed.
Featured image: the entrance of 6 Pancras Square in London, the building that houses Google and Google DeepMind, December 2017. Photo: Gciriani via Wikimedia Commons, licensed under CC BY-SA 4.0; cropped to 3:2, resized to 1200×800 and converted to PNG. The adapted image is shared under the same licence. Google trademarks shown are for editorial purposes and do not imply endorsement.
