DeepMind released AlphaGenome 2, which found nearly twice as many gene-expression variants as its predecessor
On a map of 28 immune cell types, AlphaGenome 2 identified 30% of the variants that change gene expression, against 16% for the model it replaces. DeepMind calls that nearly twice as many.
Google DeepMind released AlphaGenome 2 on 11 November. It is a DNA sequence model that takes a million letters of DNA and a biological condition, a tissue, a cell type or a cell state, and predicts how that sequence is read. It was trained on over 107 million individual cell profiles from human and mouse samples, across 68 tissues. The same API is open today for non-commercial research, and in a private preview for commercial research. DeepMind says the model is for research use only and has not been validated for clinical or diagnostic purposes.
Why this one is different
The first AlphaGenome predicted effects across many tissues at once, which DeepMind says spread the model's capacity thin. AlphaGenome 2 is a conditional model: it predicts for one condition at a time, and it can take a scientist's own single-cell experiment as that condition without being retrained. It also predicts DNA methylation, a readout the first model did not. On TenK10K it also doubled directional gene effects, 43% against 23%.
30% against 16% is DeepMind's own comparison, on one immune-cell map.
How we got here
- 2025DeepMind launches AlphaGenome, which predicts how DNA variants affect molecular activity across broad tissues.
- Sep 2026DeepMind publishes AlphaGenome Atlas, predictions for all 9 billion possible single-letter changes in the human genome.
- 11 Nov 2026AlphaGenome 2 moves from tissue averages to individual cell types, and is available through the AlphaGenome API.
What it does and does not mean
The 30% and 16% are DeepMind's figures on one dataset, and they do not make the model ready for the clinic. DeepMind says so itself: the model is research use only. Peter Donnelly, the CEO of Genomics Limited, wrote that even with the gains the model makes confident predictions for only a small subset of the variants that can be queried. What the release does show is a clear step in cell-type resolution, and that a finding about a variant that drives IBD in gut immune cells, first made in the lab, can now be reproduced in the model. What researchers do with that is not yet known.