AlphaGenome 2 Brings Cell Context Into DNA Prediction

Chinese

Google DeepMind announced AlphaGenome 2 on November 11, shifting DNA-effect prediction toward specific cell types and states. Hosted access is available for non-commercial research, with commercial research in private preview. The model supports experimental prioritization; it is not validated for diagnosis.

AI-generated concept illustration: a DNA motif crosses different cellular environments to evoke cell context

A genetic variant can matter in one cellular setting and be difficult to detect in another. That makes context central to a research question: which cells should scientists examine when a stretch of DNA is associated with a trait or disease? A prediction across broad tissue averages may be useful, yet still leave that question unresolved.

Google DeepMind announced AlphaGenome 2 on November 11 to predict DNA effects in specific cell types and states. The model processes a roughly million-letter DNA sequence alongside a selected biological condition. Rather than producing fixed tissue outputs in parallel, it makes a prediction for one condition at a time. [G1]

DeepMind says training incorporated more than 100 million individual cell profiles across 68 tissues. Its claimed advance is more specific interpretation of gene regulation, including conditions represented by new single-cell measurements. The published comparisons are developer results; this report has not independently reproduced them. [G1]

For a laboratory, the potential value is prioritization. A model may help choose which variants, cells or regulatory relationships deserve the next experiment. That can make an investigation more focused, but a ranked prediction remains a hypothesis until it is checked against appropriate evidence. An attractive explanation is not a substitute for a measured biological effect.

Access is available through the AlphaGenome API for non-commercial research, while commercial research remains in private preview. DeepMind says related Skills for Antigravity will be released soon, making that integration a future item rather than an already delivered feature. The announcement explicitly limits the models to research and says they have not been validated or approved for clinical or diagnostic purposes. [G1]

Researchers assessing usefulness should ask what the model contributes beyond existing assays and statistical evidence. Does it distinguish plausible cell contexts, narrow a candidate list, or suggest a test whose outcome could disprove the prediction? A workflow that can reject a model’s suggestion is more informative than one that only collects supporting examples.

The Verge also reported the November 11 release and its focus on how DNA changes affect different cells. That coverage corroborates the event, but does not independently establish model accuracy or downstream treatment benefits. This is a new model announcement, distinct from the earlier AlphaGenome release and the precomputed AlphaGenome Atlas. [G1][G2]

The practical milestone is therefore a more contextual research tool, not a clinical conclusion. Its scientific contribution will become clearer as outside teams document where it changes experimental choices, where it misses relevant biology and how well its predictions survive follow-up work.

Sources:

[G1] Google DeepMind announcement|2026-11-11
[G2] The Verge|2026-11-11