Google DeepMind releases AlphaGenome 2, a model predicting genetic variant effects at individual cell-type and cell-state level, via API for non-commercial research

Google DeepMind released AlphaGenome 2, a model that predicts how genetic variants change gene expression in specific cell types and cell states. It was trained on over 100 million single-cell profiles across 68 tissues. DeepMind says it finds nearly twice as many expression-altering variants as its previous best model and can handle new cell types and states without retraining. It is available now through the AlphaGenome API for non-commercial research, with a limited private preview for commercial research. It does not replace the original AlphaGenome because it costs more to run.

The main users are academic and non-profit genomics researchers, who can now test variant effects at cell-type level through an API without building their own models. Pharmaceutical and biotech teams get only a limited preview, so commercial adoption will lag. The practical decision for labs is whether the extra compute cost is worth it over the original AlphaGenome for their questions. The real impact depends on whether its predictions hold up against lab experiments, and nothing in the coverage shows that yet.

The "nearly twice as many variants" and "generalises to new cell types" claims come from DeepMind alone, with no independent benchmark or peer-reviewed comparison cited. It is also unclear what counts as a correctly identified variant, or how the comparison model was chosen. The talk of revealing disease roots, guiding drug targets and feeding "virtual cell" efforts is aspiration, not something the release demonstrates.

1. 01Independent benchmarks or preprints from outside labs that test AlphaGenome 2 against experimental data, especially for the 'new cell types without retraining' claim.
2. 02Usage and cost details for the API: rate limits, pricing for the commercial preview, and how many research groups publish results using it in the next 6 to 12 months.
3. 03Whether the commercial preview widens into general availability, and whether any drug developer reports a named target or variant finding that came from the model.