Google DeepMind announced AlphaGenome 2 on November 11, 2026, a genomics AI model designed to predict how DNA changes affect particular cell types and states. SourceSource
A launch post by @pushmeet says the model predicts what a DNA variant—a change in the DNA sequence—does in a particular cell type or state. The cell isn’t background scenery; it’s part of the prediction. Source
The post says AlphaGenome 2 can predict variant effects across hundreds of human cell types and adapt to previously unseen cell and disease states without retraining. Source
Nearly double, on one immune-cell benchmark
The launch post reports two improvements over the original AlphaGenome on the TenK10K immune-cell dataset:
- Causal-variant identification: 30%, compared with 16% for the original model.
- Direction of effects on genes—whether a DNA change increases or decreases gene activity: 43%, compared with 23%. Source
These are separate benchmark results, not an overall disease-prediction accuracy score. The post doesn’t spell out the evaluation thresholds, denominators or uncertainty estimates behind the percentages, and the figures alone don’t establish diagnostic ability or treatment benefit. Source
Academic access, not an unrestricted invitation
The launch post says AlphaGenome 2 is being made available through the AlphaGenome API for non-commercial academic research. It does not announce broader commercial access. Source
Comments