Google DeepMind on November 11 unveiled AlphaGenome 2, a new artificial intelligence model that predicts how changes in DNA affect gene activity inside individual human cell types, a step the company says could help scientists trace the cellular roots of disease. Trained on more than 100 million individual cell profiles, the model is available now to researchers for non-commercial work through an API, though DeepMind stresses it is a research tool only and has not been validated for clinical or diagnostic use.
 |
| An illustration of a DNA double helix. AlphaGenome 2 predicts how changes in DNA affect gene activity in specific human cell types. (Illustration) |
Why Reading DNA Cell by Cell Matters
Every cell in the human body carries the same DNA, yet a skin cell, a brain cell, and an immune cell behave in profoundly different ways. The difference lies in which genes each cell switches on and off, and a single-letter change in DNA can matter enormously in one cell type while doing nothing at all in another. That basic fact has long been a blind spot for genetic research: most tools that predict the effect of a DNA variant average their answers across whole tissues, diluting or even hiding signals that exist only in a rare or specialized cell.
The stakes are high because the majority of DNA variants linked to disease do not sit inside the genes that build proteins. Instead, they fall in the stretches of DNA that control when and where genes are used, and their effects often show up only in a particular cell type or in a cell that has been switched on by disease. Finding those effects has traditionally required slow, expensive laboratory experiments, one cell type at a time.
That is the gap AlphaGenome 2 is built to close. According to DeepMind, the model takes a stretch of DNA of up to one million letters, together with a specific cell type or cell state, and predicts how that DNA is read in that cell: which genes are active, how the DNA is packaged and accessed, and how a single-letter change would alter any of it. The prediction happens in software, without anyone running a lab experiment first.
How AlphaGenome 2 Works
The first AlphaGenome, published in the journal Nature in January 2026, read DNA across thousands of tissue samples at once, producing fixed predictions for bulk tissues and cell lines simultaneously. AlphaGenome 2 changes the design. It is what DeepMind calls a conditional model: it evaluates one condition at a time, and supplies that condition throughout the network, so the same stretch of DNA can be read differently depending on the cell it is placed in.
Researchers can choose the condition from a reference catalog of cell types, or generate it on the fly from their own single-cell experiment using a part of the system called a cell encoder. That flexibility matters because real biology rarely matches a catalog: a patient's cells, or cells in a particular state of disease, may not appear in any standard reference.
The scale of the training data is large by any measure. DeepMind says the model was trained on more than 100 million individual cell profiles across 68 tissues, drawn from human and mouse samples, combining bulk data with single-cell data. In September, DeepMind also released AlphaGenome Atlas, a precomputed resource covering predictions for all 9 billion possible single-letter changes in the human genome. AlphaGenome 2 is a separate model release, not an extension of that atlas, and it is designed for researchers who want to ask their own questions rather than look up a precomputed answer.
 |
| The Google DeepMind logo. The AI lab announced AlphaGenome 2 on November 11, 2026. |
What the Tests Showed
DeepMind's headline numbers come from its own testing, and the company reports a clear improvement over the first model. AlphaGenome 2 identifies nearly twice as many variants that change gene expression in cells, compared with the previous state-of-the-art model, according to the announcement.
The flagship test used a dataset called TenK10K, a single-cell map of 28 primary immune cell types. On that test, DeepMind reports that AlphaGenome 2 recovered 30 percent of the variants scientists believe are causal for changes in gene expression, at 90 percent precision, compared with 16 percent for the first AlphaGenome model fine-tuned on single-cell data. It also called the direction of gene effects correctly in 43 percent of cases, against 23 percent for the earlier model.
Those figures deserve a careful reading. Even at high precision, the model still recovers only a minority of the variants that statistical analyses nominate as likely causal, a limitation DeepMind itself acknowledges in its technical paper. The paper is a preprint that had not been peer-reviewed at launch, and its headline test was an immune-cell dataset, so performance in other parts of the body is less established. The numbers, in other words, show real progress on a hard problem, not a solved one.
Independent coverage has struck a similar note. Nature reported on the release the same day, and analysts who read the announcement in full have emphasized that all performance figures are company-reported and that the model is for research use only. The Wellcome Sanger Institute is named as an ongoing collaborator, and Genomics Limited chief executive Peter Donnelly said he looked forward to incorporating AlphaGenome 2 into the company's Mystra platform.
What This Means
For researchers, the practical change is speed and focus. A scientist studying a disease can now ask, in software, which cell types a suspect variant is likely to act in, and design laboratory experiments around the most promising leads instead of testing blindly. That could matter most for diseases where the relevant cell type is rare or hard to grow in a dish, and for the long tail of variants that genome-wide studies have linked to conditions without explaining how they work.
The model also fits a wider shift in biology toward single-cell resolution. Medicine has spent decades reasoning about tissues as averages; the new generation of tools treats each cell as its own context. AlphaGenome 2 sits squarely in that shift, turning vast single-cell atlases into what DeepMind calls a predictive engine for human genetics.
There are limits that should not be glossed over. The model cannot diagnose disease, and DeepMind states plainly that neither AlphaGenome model has been validated or approved for clinical or diagnostic purposes. It cannot be downloaded and run locally; access runs through DeepMind's API. And because the evidence so far rests on a preprint and company-reported benchmarks, independent replication will be the real test of how far the advance reaches.
 |
| A researcher reviews DNA sequencing data on a monitor in a laboratory. AlphaGenome 2 is available to researchers through an API for non-commercial research. (Getty Images) |
What Happens Next
AlphaGenome 2 is available now for non-commercial research through the AlphaGenome API, with commercial research in private preview, according to DeepMind. Interactive tutorials for expression conditioning are available alongside it, and DeepMind says it will soon release AlphaGenome 2 Skills in Google Antigravity, its agentic development platform.
The next milestones to watch are outside DeepMind's control. Peer review of the technical paper will test the headline claims, and independent research groups will now put the model through its paces on their own data, in cell types and diseases far from the immune system. If the gains hold up beyond the benchmark, AlphaGenome 2 could become a standard first step in the long journey from a genetic clue to a biological explanation, and eventually, to a treatment.
For now, the announcement marks another step in a fast-moving year for AI in science. DeepMind's first AlphaGenome arrived in January; the Atlas followed in September; and the second model, built to read DNA one cell at a time, arrived on November 11. The genome has not changed. What is changing is how precisely scientists can read it.
Post a Comment