Google DeepMind unveils AlphaGenome 2, predicting DNA variant effects cell by cell

Google DeepMind on November 11 unveiled AlphaGenome 2, a next-generation DNA sequence model that predicts how DNA variants affect gene activity in specific human cell types, helping scientists uncover the cellular drivers of disease. Unlike its predecessor, which predicted tissue averages, the new model can tell how the same stretch of DNA is read differently across cell types, nearly doubling performance on a key benchmark.

From tissue averages to individual cell types

Every cell in the body carries the same DNA, yet heart muscle cells, immune cells and other cell types switch different genes on and off. The first AlphaGenome predicted thousands of tissue experiments in parallel, which spread its capacity thin and made it hard to capture subtle differences between closely related cell types. AlphaGenome 2 is a conditional model: alongside a DNA sequence of up to one million letters, it takes a target biological condition — a tissue, cell type or state — as input at every stage, predicting the DNA effect for that condition one at a time. It was trained on more than 107 million individual cell profiles across 68 tissues, and adds DNA methylation, a widely requested readout.

Nearly double the detection rate, reproducing a key bowel-disease finding

On the TenK10K dataset, a large single-cell map of 28 primary immune cell types, AlphaGenome 2 reliably identified 30% of causal variants, against 16% for the original model, and judged the direction of gene effects correctly 43% of the time, against 23%. In a landmark study, scientists found that a variant in a distant regulatory switch drives inflammatory bowel disease by overactivating the ETS2 gene. AlphaGenome 2 reproduced that finding computationally, singling out the causal variant and showing that it selectively boosts the ETS2 gene in gut macrophages, leaving other cell types and genes largely unaffected.

New cell states without retraining

AlphaGenome 2 can also generalise to entirely new single-cell experiments without retraining: scientists upload a single-cell RNA-sequencing snapshot from just a few dozen cells, and the model infers unseen cell states. Given unseen data from a COVID-19 mRNA vaccine study, for example, it distinguished vaccine-activated immune T cells from resting ones, and revealed that a variant linked to severe COVID-19 disrupts gene activity more strongly in the activated state — an effect that would be missed in unactivated cells. The model can also impute unmeasured properties from a single measurement, such as predicting chromatin accessibility from gene activity, helping researchers prioritise experiments.

Access and limits

AlphaGenome 2 is available from today for non-commercial research via the AlphaGenome API, and in private preview for commercial research. DeepMind has also precomputed predictions for trait-associated variants across hundreds of human cell and tissue conditions, helping researchers connect variants found in genome-wide association studies to specific genes and cells, across complex traits from height to type 2 diabetes. DeepMind stresses that AlphaGenome and AlphaGenome 2 are intended for research use only and have not been validated or approved for clinical or diagnostic purposes; the technical paper is a preprint and has not yet been peer-reviewed.

(Sources: Google DeepMind; Nature news.)