AlphaGenome 2 Tries to Find the Cell Where a DNA Variant Matters

A genetic variant can look unremarkable when averaged across an entire tissue and still matter in one small population of cells. Google DeepMind’s AlphaGenome 2 is designed to ask a more precise question: given this DNA sequence, what changes in this particular cell type or state?

The model announcement, published November 11, describes a shift from broad tissue averages toward cell-conditioned predictions. That matters because disease biology often depends on where a regulatory switch is active, not just whether a gene changes somewhere in a sample. It is a research tool for narrowing hypotheses, not a clinical test.

From one sequence to one biological condition

The first AlphaGenome model predicted many experimental readouts across tissues in parallel from a long DNA input. DeepMind says AlphaGenome 2 instead receives the target biological condition—such as a cell type or state—alongside a DNA sequence of about one million letters. That condition enters the model throughout its architecture, allowing it to represent that the same regulatory sequence may behave differently across cell contexts.

The design also supports a form of missing-data prediction. If researchers provide one measurement from a single-cell experiment, such as gene activity, the model can estimate other properties, such as chromatin accessibility. That may help decide which laboratory experiments to run next. A predicted measurement is not an observed one; it should be treated as a guide for experiment design, then checked in the lab.

DeepMind says the model was trained on more than 107 million individual cell profiles from human and mouse samples and predicts gene expression, chromatin accessibility, histone modifications and DNA methylation. More training data does not automatically mean that predictions are reliable in every tissue or population, so the boundaries of the evaluation matter.

What the headline result measures

On the TenK10K dataset, which contains a single-cell map of 28 primary immune cell types, DeepMind reports that AlphaGenome 2 identified 30% of causal variants, compared with 16% for the previous model fine-tuned on single-cell data and other specialized tools. It reports directional gene effects for 43% versus 23% on the same comparison.

Those are company-reported benchmark figures, not evidence that the model has found all causal variants or can predict an individual patient’s disease. They describe performance on a particular dataset and task. The useful question is whether a model can prioritize candidate variants and identify a plausible cell type well enough to direct scarce experimental time.

DeepMind offers an example involving inflammatory bowel disease: a previously studied variant in a distant regulatory region was linked to increased ETS2 activity in gut macrophages. The company says AlphaGenome 2 reproduced that cell-specific relationship computationally. This is a retrospective test against a known finding, which supports the model’s ability to recover a mechanism; it does not establish that every new prediction is correct.

Generalizing to cell states researchers have just measured

A notable feature is conditioning on a researcher’s own single-cell data. DeepMind says a small snapshot of dozens of cells can be compressed into a cell-state representation so the model can estimate variant effects for an unseen state without retraining. In an example using immune cells after COVID-19 vaccination, it distinguished activated T cells from resting cells and predicted a stronger effect from a severe-COVID-associated variant in the activated state.

This could make variant interpretation more context-aware. But new cell states also create a validation challenge: the model may be extrapolating beyond the conditions represented in training, and benchmark results may not transfer to a different assay, ancestry group or disease cohort.

AlphaGenome 2 is available through an API for noncommercial research and in private preview for commercial research, according to DeepMind. The company explicitly says neither AlphaGenome model is validated or approved for clinical or diagnostic use. The practical next step for scientists is to test whether its rankings survive independent datasets and targeted experiments. See also our prior reporting on the AlphaGenome Atlas and variant prediction evidence. The model can help choose where to look; it cannot replace the evidence that establishes what a variant does.

Predictions can focus experiments, but they do not settle causality

A cell-aware model can improve the first step in a long chain: choosing which variant, gene and cell type to investigate. Researchers could use those rankings to prioritize CRISPR perturbations, expression assays or follow-up samples. If the model can transfer a cell-state representation from a small experiment, it may be particularly useful where generating a full labeled dataset would be expensive.

The second step is to determine whether a predicted effect survives controlled intervention. A variant may correlate with a signal because of linkage to another variant, a dataset artifact or a biological context the model does not represent. Predictions may also be less reliable in rare cell types or underrepresented populations. A promising score should therefore be compared with independent cohorts and experiments that manipulate the candidate sequence or pathway.

The benchmark numbers need careful reading. “Identified 30%” means that a specified benchmark judged a portion of causal variants to be recovered under its evaluation setup. It does not mean that 30% of patients will receive a correct diagnosis or that the model is 30% accurate in general. The comparison with other methods is only as informative as the dataset, split, definitions and baselines behind it. DeepMind describes the test set and comparison in its announcement, but external replication will be important for researchers deciding whether to build workflows around it.

Availability also shapes reproducibility. DeepMind says noncommercial researchers can access the API and commercial access is in private preview. Scientists should record model versions, inputs and output settings, and consider whether their data can be sent to an external service under their institutional rules. These practical constraints are part of research quality, not an afterthought.