AlphaGenome 2: Deciphering how cells read DNA to better understand disease

AlphaGenome 2, from Google DeepMind, predicts how a stretch of DNA behaves inside one particular kind of cell. Every cell in the body carries the same DNA, but a heart muscle cell and an activated immune cell switch on different genes, so a single-letter change in the DNA can matter in one cell type and do nothing in another. The model takes up to a million letters of DNA together with a description of the target tissue, cell type or cell state, and predicts measurable activity there: how strongly nearby genes are expressed, how open the DNA is, and chemical marks such as methylation. Comparing the prediction for a variant with the prediction for the usual sequence gives the variant's estimated effect in that cell. In DeepMind's announcement, the team says the first AlphaGenome predicted thousands of tissue-level experiments at once, which spread the model thin and blurred differences between closely related cell types. Feeding the cell condition in at every stage is meant to let the new model learn how the same DNA is read differently from cell to cell.

DeepMind's evidence is that this finds variants the older approach missed. On TenK10K, a large single-cell map of 28 immune cell types, it reports that AlphaGenome 2 identified 30% of causal variants against 16% for the original model fine-tuned on single-cell data and other specialized tools, and the direction of the gene effect in 43% of cases against 23%. It also re-found a published result: a variant in a distant regulatory switch that drives inflammatory bowel disease by overactivating the ETS2 gene in gut immune cells called macrophages. Asked about several gut immune cell types, the model singled out that variant and predicted that it boosts ETS2 in macrophages while leaving the other cell types largely unaffected. That shows it can recover a known answer, not that it has made a new discovery. DeepMind's announcement also does not say how the benchmark's causal variants were established.

The practical change is for researchers trying to turn genetic association studies, which flag broad stretches of DNA linked to a disease, into a specific variant, gene and cell type to study. DeepMind has precomputed AlphaGenome 2's predictions for trait-associated variants across hundreds of cell and tissue conditions, a cell-level counterpart to AlphaGenome Atlas, which stored the first model's predictions for every possible single-letter change in the human genome. The model can also take gene-activity data from a few dozen cells in a state it has never seen and predict there without retraining. In DeepMind's example, a variant linked to severe COVID-19 disrupted gene activity more strongly in vaccine-activated T cells than in resting ones. AlphaGenome 2 is open to non-commercial research through an API and in private preview for commercial use, and DeepMind says it is not validated for clinical or diagnostic use.