Google’s AlphaGenome 2 zooms in on the cellular roots of disease

The same genome can tell different stories in different cells. A variant that appears harmless in a tissue-wide average may alter gene activity in an activated immune cell or a gut macrophage, precisely the kind of distinction that can matter when researchers are looking for disease mechanisms or drug targets. Google DeepMind’s AlphaGenome 2 is designed to make that distinction computationally.

The model, announced November 11, was trained on more than 100 million individual cell profiles spanning 68 tissues. Its central change is not simply a larger training set. AlphaGenome 2 evaluates one selected tissue, cell type or cell state at a time, carrying that condition through the network so the same DNA segment can be interpreted differently depending on the cellular context. It predicts gene expression, RNA splicing and chromatin accessibility, and adds CpG DNA methylation to the original model’s outputs. Researchers can also supply single-cell RNA-seq data from their own experiment; a cell encoder converts that snapshot into a condition without further fine-tuning, allowing predictions for cell states the model has not previously seen.

That is a meaningful shift from the first AlphaGenome, which produced fixed outputs for bulk tissues and cell lines from a one-million-letter DNA sequence. It is also distinct from AlphaGenome Atlas, the precomputed set of predictions for all nine billion possible single-letter changes in the human genome. AlphaGenome 2 is a separate model release, not an expanded atlas. The practical distinction matters: instead of asking which variants might matter somewhere in a tissue, a researcher can ask what a variant is predicted to do in a particular disease-activated cell.

The strongest reported benchmark comes from TenK10K, a single-cell immune dataset covering 28 cell types. On that evaluation, AlphaGenome 2 recovered 30% of putatively causal expression-linked variants at 90% precision, compared with 16% for a fine-tuned AlphaGenome 1. It recovered 43% of variant effect directions at 90% accuracy, against 23%. In genomic regions excluded from training, it explained a median 34% of within-tissue, cell-type-specific expression variance, versus 23% for the predecessor. These are gains on regulatory differences between related cells, the exact problem the conditional design was built to address, rather than a general claim that the model has solved variant interpretation.

Two biological examples show how the specificity could become useful. That difference in evidentiary status is easy to blur, and it should not be.

The model also offers a way to fill gaps in experiments. Given gene-expression data alone, it can predict unmeasured properties such as chromatin accessibility, helping scientists decide which additional assays to run. The benefit is prioritization rather than substitution: a prediction can direct attention, but it does not replace the measurement that would confirm it. DeepMind’s own researchers also acknowledge a limitation in regions where chromatin is not readily accessible, where the model tends to underestimate effects.

Access is broad for research but not yet unrestricted for commercial use. AlphaGenome 2 is available for non-commercial research through the AlphaGenome API, with a limited private preview for commercial organizations. The computational price is also higher than the predecessor’s: because each condition requires its own prediction, the model is more expensive to run than fixed-track multitask systems. That trade-off may be acceptable for a focused drug-discovery question and less attractive for broad screening.

The decision-making power shifts toward researchers who can supply well-characterized single-cell data and ask cell-specific questions. The risk falls on those who treat a predicted regulatory effect as a causal finding, particularly when the relevant cell state is rare, poorly represented or absent from training. The next evidence that would matter is experimental validation of predictions in unseen states, followed by evidence of whether the model changes target selection or trial design in practice. Until then, AlphaGenome 2’s most credible contribution is narrower than “reading disease in cells”: it is making the cellular context of a genetic variant explicit enough to test.