AlphaGenome 2: Google DeepMind's DNA model now predicts variant effects cell type by cell type

Your own cells, no retraining. Researchers can upload a single-cell RNA-seq snapshot from as few as a few dozen cells, and the model uses it to predict variant effects for that new cell state without retraining. If you give it one measurement, it can also predict unmeasured ones, such as chromatin accessibility.

DeepMind's numbers. On the TenK10K single-cell dataset (28 primary immune cell types), DeepMind reports AlphaGenome 2 identified 30% of causal variants vs. 16% for the original AlphaGenome fine-tuned on single-cell data and other specialized tools, and 43% vs. 23% of directional gene effects. It also says the model reproduced, computationally, a published finding that a distant variant drives inflammatory bowel disease by boosting the ETS2 gene in gut macrophages. These are DeepMind's own results; MODEL has not checked them.

Availability. Available now for non-commercial research via the AlphaGenome API, with interactive Colab tutorials, and in private preview for commercial research. The AlphaGenome docs (opens in a new tab) say the API is free of charge for non-commercial use and that query rates vary with demand. DeepMind says AlphaGenome 2 "Skills" for its Google Antigravity agent platform are coming "soon"; no date given.

Who it's for. Genetics and single-cell researchers who want to rank which DNA variants matter in which cells before running lab experiments. It is not a consumer tool.

The caveats.

• Research use only. DeepMind says AlphaGenome and AlphaGenome 2 have not been validated or approved for clinical or diagnostic purposes.
• Commercial use is a private preview, not open access.
• Benchmarks and the ETS2 example come from DeepMind's own announcement.

Disclosure: Not tested by MODEL. Not sponsored. No affiliate links.

MODEL Editors