DeepMind's AlphaGenome 2 Predicts DNA Effects Cell by Cell

Google DeepMind has unveiled AlphaGenome 2, the successor to its genome-predicting model AlphaGenome, designed to predict how changes in DNA affect different cells. The Verge reported the launch on 11 November 2026.

The model predicts the effects of genetic variants (single-letter changes or insertions in a DNA sequence) on individual cell types and states, rather than only at the coarser level of whole tissues. It was trained on more than 100 million individual cell profiles across 68 tissues. According to the same report, AlphaGenome 2 identifies nearly twice as many variants that can change gene expression, meaning how strongly a gene is switched on to make its protein, as DeepMind's previous state-of-the-art model.

One capability stands out for practitioners. AlphaGenome 2 can make predictions about new cell types and biological states without being retrained. In practice, the training data does not have to contain a given cell state for the model to score variants within it.

The gains come with a cost. The more detailed predictions are more computationally expensive than those from AlphaGenome, and the report indicates that AlphaGenome 2 is not replacing AlphaGenome imminently. Both models will remain in use for now, with the choice between them likely to depend on whether a project needs cell-type resolution or can accept the lower compute footprint of the earlier model.

The broader context here is the position of gene expression in variant interpretation. Most of the work of connecting a sequence change to a phenotype runs through regulatory effects on expression. A model that scores variants by their predicted effect on expression in specific cells gives researchers a more direct readout than one that operates only on bulk tissue. Whether that readout produces better downstream results will depend on validation work that the announcement does not detail.

In my view, cost per prediction will shape the model's practical reach more than its headline scale. A system that evaluates unseen cell states without retraining is useful for screening hypotheses quickly, but only if the compute required fits the budgets of academic groups and smaller biotechnology teams, not only large industrial programmes.

The optimistic reading is also the straightforward one. If the cell-level predictions hold up in practice, the path from a variant of uncertain significance to a testable mechanistic hypothesis gets shorter. That outcome depends on engineering and validation that will take time, and this announcement marks the start of that process.