Google DeepMind wants to know not just what your DNA says, but which cells are actually reading it. On 11 November 2026 the lab released AlphaGenome 2, a model that predicts how genetic variants affect individual cell types, from heart muscle cells to immune cells fighting an infection. It is the successor to last year’s AlphaGenome, and it marks a shift from broad tissue-level guesses to cell-level predictions that could help researchers find the roots of disease.
Here is what the model does, how it improves on the original, and why the research community cares.
What AlphaGenome 2 Actually Predicts
Nearly every cell in the human body carries the same DNA, but different cell types switch different genes on and off depending on their job and state. That is why the same genetic variant can do very little in one cell and cause trouble in another. Predicting those effects has been an unsolved problem, even as large projects like the Human Cell Atlas map gene expression for millions of cells.
AlphaGenome 2 is built to close that gap. It predicts the effects of genetic variants on individual cell types and states, which means researchers can ask what a specific DNA change does inside, say, a liver cell during inflammation, instead of getting a tissue-wide average that hides the answer.
What Is New Compared With AlphaGenome
The first AlphaGenome, released in 2025, predicted how DNA variants affect molecular activity across broad tissues. DeepMind then expanded it with AlphaGenome Atlas, a platform holding predictions for all 9 billion possible single-letter changes in the human genome. Those tools already help researchers prioritise and interpret variant effects in rare diseases and complex traits.
AlphaGenome 2 zooms in from tissue averages to single cells. According to the DeepMind blog post announcing the model, it was trained on more than 100 million individual cell profiles across 68 tissues. DeepMind says the model effectively turns static cell atlases, the huge data repositories that catalog individual cells, into a predictive engine for human genetics. The architecture was redesigned for the job, and the lab reports a step-change improvement in variant predictive performance.
Two capabilities stand out for working scientists:
- Disease-relevant effects in specific cells that broad-tissue models would miss can now surface in predictions
- The model can make predictions for cell states and disease states it has never seen during training, which covers the rare conditions that small datasets cannot teach directly
Why Cell-Level Predictions Matter for Disease
Most genome-wide association studies end the same way: a list of DNA variants linked to a condition, with no clear answer about mechanism. A variant sits near a gene, but whether it matters in the liver, the brain, or only during an immune response is guesswork. Cell-level prediction turns that list into testable hypotheses.
For drug development, the appeal is speed. If a model can point to the cell type where a variant does its damage, researchers can narrow the search for targets before spending months at the bench. DeepMind positions AlphaGenome 2 as a research tool for exactly that kind of prioritisation, in rare diseases, complex traits and fundamental biology.
What the Model Does Not Do
It is worth being clear about the limits. AlphaGenome 2 does not diagnose anyone, and its predictions are hypotheses, not findings. A predicted effect in a cell type still needs bench validation, and the model’s accuracy will vary by how well a cell type is represented in training data. DeepMind also positions the model as a research tool rather than a clinical product, so any medical use would need its own validation path.
The other limit is interpretation. The model can say a variant likely changes gene activity in a specific cell, but the chain from that change to a disease symptom still runs through biology that researchers must untangle the old way. Think of it as a much better filter for what to test next.
How Researchers Can Use It
- The research paper is published and linked from the DeepMind blog, with the full technical detail on architecture and benchmarks
- AlphaGenome Atlas remains the entry point for single-letter variant lookups across the genome, and the new model feeds the same workflow of prioritising variants for follow-up
- Teams working on rare disease cases can compare variant predictions across cell types to pick which tissue to study next
- Anyone working with Human Cell Atlas data can layer predictions on top of their existing expression maps
The Bigger Picture
AlphaGenome 2 arrives as biology labs absorb what large models can and cannot do. Critics of early genomic AI pointed out that tissue-level predictions were too coarse to act on. The move to single-cell resolution is a direct answer to that complaint. The model is not a diagnosis tool and it does not replace experiments, but it changes which experiments are worth running first.
DeepMind has form in this area. Its protein structure work reshaped structural biology, and the AlphaGenome line is the same playbook applied to the instruction manual of the cell. Whether AlphaGenome 2 delivers on its promise will be decided in the labs that use it, but the direction is clear: prediction at the resolution where disease actually happens. For anyone tracking AI in science, this is one of the releases worth following through its first year of real-world research use.
Frequently Asked Questions
What is AlphaGenome 2?
AlphaGenome 2 is a Google DeepMind AI model released on 11 November 2026 that predicts how genetic variants affect individual human cell types and states. It is the successor to the AlphaGenome model released in 2025.
How is AlphaGenome 2 different from the first AlphaGenome?
The original predicted variant effects across broad tissues. AlphaGenome 2 works at single-cell resolution, trained on more than 100 million cell profiles across 68 tissues, so it can surface effects that tissue-level models hide.
Is AlphaGenome 2 available to researchers?
DeepMind published the research paper alongside the announcement and linked it from its blog. The related AlphaGenome Atlas platform already provides predictions for all 9 billion possible single-letter genome changes.
Can AlphaGenome 2 diagnose diseases?
No. It is a research tool that helps scientists prioritise and interpret genetic variants. Any findings still need to be validated in the lab before they mean anything for patients.
