AI
Google DeepMind's AlphaGenome 2 predicts how DNA changes ripple through individual cells
A new AI model reveals how genetic variants affect specific cell types and states, potentially uncovering cellular roots of disease.
1/2
The Verge
Google DeepMind has introduced AlphaGenome 2, an AI system that offers researchers an unprecedented lens into how genetic mutations influence cells at granular detail. By modeling the effects of DNA variations on specific cell types and their conditions, the tool could unlock insights into disease mechanisms and inform new treatment strategies.
How it works differently#
While humans share the same DNA across nearly all cells, different cell types—muscle, skin, immune cells—activate different genes depending on their function and environment. The same genetic change can therefore produce radically different outcomes across cell types. AlphaGenome 2 is engineered to capture this variation by predicting how genetic variants alter gene expression in individual cell types and states.
The system was trained on more than 100 million cell profiles across 68 tissues. This training converts static cell atlases—comprehensive catalogs that detail individual cells—into a tool that forecasts genetic effects. The result: AlphaGenome 2 identifies nearly twice as many gene-altering variants as prior leading models, according to DeepMind.
Why cellular-level precision matters#
For drug developers, this specificity is crucial. As Žiga Avsec, DeepMind's genomics lead, explained to The Verge, therapeutics target particular proteins in particular cell types, making this kind of resolution essential for translation into treatments. The model also works without retraining on new cell types and can flag gaps in experimental data, helping scientists prioritize which questions to investigate next in the lab.
Research scientist Adam Gayoso characterized the key breakthrough as "tying human genetics to cell biology"—enabling researchers to zoom from population-level genetic findings down to cellular-level effects that might drive disease.
Current limitations and future directions#
AlphaGenome 2 comes with computational trade-offs: its detailed predictions are more expensive to generate than those from its predecessor, AlphaGenome. That cost means it won't replace the original model for genome-wide atlases covering every possible DNA change, which DeepMind released in September.
Going forward, Avsec said the team plans to expand training data, particularly for under-represented cell states, and eventually incorporate information across multiple species to broaden biological insight. Longer-term applications could include training models to power ""—digital simulations that mimic living biological systems.

