DeepMind's AlphaGenome 2 finds almost twice as many causal DNA changes in immune cells
It predicts what a DNA change does in one cell type, not a whole tissue. The tests are DeepMind's own.
On 11 November 2026 Google DeepMind announced AlphaGenome 2, a model that reads DNA and predicts what a change in it does inside a specific kind of cell. The first AlphaGenome, covered in our earlier entry, predicted averages over whole tissues. The new one is told which cell type or state to predict for, such as a gut macrophage or an activated T cell.
It was trained on more than 107 million individual cells from human and mouse tissue. On TenK10K, a study of 28 immune cell types, it picked out 30% of known causal variants (changes in the DNA that really alter a gene’s activity), against 16% for the original model fine-tuned on single-cell data. It got the direction of the effect, up or down, right for 43% of cases, up from 23%. Both figures are recall at 90% precision, a strict setting where nine in ten of its calls must be right.
DeepMind also reports that it reproduced a published finding that an IBD (inflammatory bowel disease) variant switches on the ETS2 gene in gut macrophages. A few dozen cells of new single-cell data are enough to point it at a cell state it has not seen, without retraining. It adds DNA methylation, a chemical tag on DNA, to its outputs.
It is available for non-commercial research through the AlphaGenome API, with a private preview for commercial research.
What it means
Researchers hunting for the DNA change behind a disease get a shorter list of suspects, and a guess at which cell type is affected. Cambridge’s Sarah Teichmann, a co-founder of the Human Cell Atlas, said it adapts to new single-cell data “without retraining”. It narrows what a lab has to test; it does not replace the testing.
Caveats
- All figures are DeepMind’s own, from a technical report on its own site. We found no peer review or outside replication yet.
- Even at its best setting it recalls 30% of known causal variants, so most are still missed at 90% precision.
- It predicts effects, and a predicted effect still needs a lab experiment to confirm.
- Research use only: DeepMind says neither version is validated for clinical or diagnostic use.