AI
DeepMind's AlphaGenome 2 spots disease variants in individual cell types with twice the accuracy
A new AI model trained on over 100 million single-cell profiles can now predict how genetic mutations affect specific human cells, advancing the hunt for disease-causing variants.
1/3
Google DeepMind Blog
Google DeepMind has unveiled AlphaGenome 2, a DNA analysis tool that represents a significant leap forward in understanding how genetic variants influence human disease at the cellular level. The model advances beyond its predecessor by examining gene activity in individual cell types rather than averaging predictions across entire tissues, enabling researchers to identify disease-driving mutations that would otherwise remain hidden.
Redesigning for cellular precision#
The original AlphaGenome processed DNA sequences while simultaneously predicting effects across thousands of tissue samples. This parallel approach sacrificed precision—it couldn't capture how the same genetic mutation behaves differently in neighboring cell types.
AlphaGenome 2 inverts this strategy. It works as a conditional model, taking both a DNA sequence and a specific biological condition (a cell type, tissue, or cellular state) as inputs. This architecture lets the system learn exactly how the same DNA stretch gets "read" by different cells. Trained on more than 100 million individual cell profiles spanning 68 tissues, the new model can resolve activity inside specialized cells—such as distinct heart muscle cells or blood immune cells—to reveal DNA variants that act only in specific cell types.
The redesign also allows the model to infer missing experimental measurements. If researchers measure just gene activity in a cell, AlphaGenome 2 can predict other properties like chromatin accessibility, helping scientists decide which experiments to prioritize next.
Doubling down on variant discovery#
AlphaGenome 2 identifies nearly twice as many variants that alter gene expression compared to its predecessor. On the TenK10K dataset—a large map of 28 primary immune cell types—the model reliably flagged 30% of causal variants versus 16% for the original system, and correctly predicted the direction of genetic effects 43% of the time versus 23%.
One concrete success involved a recent disease finding: researchers had identified a variant in a distant regulatory switch that drives inflammatory bowel disease by overactivating the ETS2 gene specifically in gut immune cells (macrophages). "AlphaGenome 2 reproduced this finding computationally," singling out the causal variant and showing it selectively boosts ETS2 in gut macrophages while leaving other cell types and genes largely unaffected.

