Two cells in the same person can contain almost the same DNA and still behave very differently. A heart cell contracts; an immune cell responds to threats. One reason is that each cell reads and regulates the shared genetic instructions differently. That makes a deceptively simple question difficult: what happens when one letter of DNA changes?
Google DeepMind’s AlphaGenome 2 is an AI research model designed to help investigate that question in the context of particular cell types and biological states. Announced on November 11, 2026, it takes a DNA sequence together with information about a chosen cellular context and predicts aspects of gene regulation. Its potential value is to help scientists choose which hypotheses and experiments deserve attention. It is not a medical test or a demonstrated treatment.
In its research announcement, DeepMind describes training on human and mouse data, including more than 107 million individual cell profiles, and reports improvements on selected benchmarks. Those are claims from the model’s developers about their tests; the practical value of a particular prediction still depends on independent evaluation and experimental confirmation.
Why the same genetic change can matter in one cell and not another
Genes are stretches of DNA that provide instructions for making molecules needed by cells. But genes are not simply switched on everywhere at the same level. Nearby and distant regulatory sequences help control when, where and how strongly a gene is used. Cell type, development and environmental signals can all influence that activity.
Imagine the genome as a large reference library shared by many departments. Each department keeps different books open, uses different indexes and responds to different requests. A change to one instruction may be important to a department that relies on it and almost irrelevant to another that rarely consults it. The analogy is imperfect, but it helps explain why the location and context of a DNA change matter.
For researchers, this creates a problem. Studies can identify genetic variants associated with a disease, yet the associated region may contain many differences in DNA. The relevant variant may sit far away from the gene it influences, and its effect may appear in only one kind of cell. Moving from an association to a plausible biological explanation is therefore a separate and difficult step.
What does AlphaGenome 2 take as input?
DeepMind says the model considers a DNA sequence together with a biological condition such as a tissue, cell type or cell state. The sequence can span up to roughly a million DNA letters in the described approach. The condition tells the model which biological setting to consider when generating predictions.
This is called conditioning. It is similar in spirit to asking the same question with an extra piece of context: not merely “what might this sequence do?” but “what might this sequence do in this kind of cell?” The model is not looking directly at a living cell while answering. It is producing computational estimates learned from earlier biological measurements.
According to DeepMind, its training combines bulk measurements, which average signals across groups of cells, with single-cell data. Single-cell methods provide a more detailed view of differences among cells, although the measurements can be noisy and incomplete. Combining those data types may help a model learn patterns that a tissue-wide average can hide.
What does the model predict?
The researchers describe predictions for several kinds of biological readout. Gene expression concerns which genes are active and to what extent. Chromatin accessibility describes how available regions of DNA are for interaction with cellular machinery. Histone modifications and DNA methylation provide additional information about how DNA is packaged or chemically marked.
These measurements are related, but they are not interchangeable. A region that is accessible is not automatically the cause of a disease, and a predicted change in gene expression is not the same as a measured change in a patient. Each output answers a narrower question about a model of the biological system.
DeepMind also says the model can use one type of measurement to estimate another that was not measured in a particular experiment. That may help a laboratory decide which observations to collect next. It does not turn the estimated measurement into an experimental observation: the difference between the two is essential when evaluating evidence.
What did the reported tests show?
DeepMind reports that on a benchmark involving immune-cell data, AlphaGenome 2 identified 30% of a set of likely causal variants compared with 16% for the original AlphaGenome fine-tuned on single-cell data, under the researchers’ stated evaluation. The comparison is encouraging within that setup, but it is not a claim that the model finds every disease-causing variant or performs equally well across all tissues.
The team also describes reproducing a previously studied inflammatory bowel disease example involving a regulatory variant and the ETS2 gene in gut immune cells. Reproducing a known relationship computationally is a useful test of whether a model captures a relevant pattern. It does not mean that the model independently discovered a new treatment for inflammatory bowel disease.
These details matter because benchmark numbers can be tempting to read as universal scores. Results depend on the reference dataset, how candidate variants are defined, which cells are represented and what counts as a correct prediction. A model can improve on a particular test while still failing on unfamiliar data or rare biological situations.
Why cell state adds another layer of complexity
A cell’s identity is not its whole story. An immune cell can be resting or activated; a developing cell can move through stages of maturation. A genetic change might influence the cell differently in those states. DeepMind describes experiments in which the system used single-cell data from previously unseen states without retraining the whole model.
That flexibility could be useful when researchers encounter new experimental conditions. However, the phrase “without retraining” does not mean without data, assumptions or uncertainty. The system still needs a suitable representation of the cellular state, and its predictions can be limited by the kinds of biology represented in its training material.
There is also a practical constraint: DeepMind says the newer model’s cell-specific approach is computationally more demanding than the original. A detailed prediction for selected questions and a faster scan across many regions can serve different purposes. More detail is valuable when it changes a scientific decision, not merely because it produces a more elaborate output.
From prediction to a useful experiment
A plausible workflow might begin with a large genetic study that flags a region associated with a disease. Researchers could compare candidate variants across relevant cell types and use the model to prioritize a possible regulatory mechanism. They could then test whether changing that sequence alters gene activity in an appropriate laboratory system.
Each step narrows uncertainty but introduces its own limitations. Association studies can contain confounding or population-specific effects. Laboratory cells may not fully reproduce a human tissue. A predicted regulatory effect may be correct in direction but too small to be clinically important. Strong evidence comes from the convergence of methods, not from one convincing visualization.
DeepMind says AlphaGenome 2 is available for non-commercial research through its API, with a private preview for commercial research. Availability to researchers is an opportunity for scrutiny and exploration, not evidence of clinical approval. Patients should not use the model’s existence to infer a diagnosis, treatment or individual risk.
What makes the idea interesting?
The deeper idea is that biology depends on context at several levels at once: sequence, cell type, state and environment. A model that brings some of those layers together may help researchers ask more precise questions. Its limitations are equally instructive. A computational prediction is most valuable when it makes an experiment more informative and exposes a hypothesis that can be tested.
For readers following AI and science, AlphaGenome 2 is an example of a system aimed at supporting inquiry rather than replacing it. Whether it becomes a dependable tool will depend on how well it performs in independent studies, across diverse cell types and in the messy conditions of real research.
Explore the book: How to Do Hard Things: The Discomfort Trap
Be First to Comment