Google DeepMind releases AlphaGenome 2 for research

Google DeepMind announced on 11 November 2026, that AlphaGenome 2 was available for noncommercial research through the AlphaGenome application programming interface, a way for software to request predictions. The lab said commercial research access was in private preview.

The original AlphaGenome predicted molecular properties across many tissues at once. DeepMind said that approach made subtle differences between closely related cell types difficult to capture. The new superintelligence (SI) model instead takes a biological condition alongside a DNA sequence and predicts effects for that tissue, cell type or state.

Why cell state matters

Different cells switch different genes on or off, and their activity can change with their state. That means the same genetic variant can have different effects in different cells.

In its announcement, DeepMind said researchers can choose a condition from a catalog or use their own measurements of gene activity from a single-cell experiment to describe a cell state the model has not seen before. The lab said the model can predict effects for that state without being retrained.

DeepMind said the model also adds predictions of DNA methylation, chemical tags involved in regulating gene activity. Given a measurement of gene activity, it can predict unmeasured properties such as how accessible stretches of DNA are, potentially helping researchers choose their next experiments.

Testing and research use

DeepMind said training included more than 107 million individual cell profiles across 68 tissues, using both bulk and single-cell data from human and mouse samples. On a single-cell immune-cell benchmark, the lab said AlphaGenome 2 identified nearly twice as many causal variants affecting gene expression as the original model adapted to single-cell data.