A sharper view of gene activity.
A heart muscle cell and an immune cell carry the same DNA blueprint, but they read it differently. Understanding those differences can help scientists trace how a small genetic change contributes to disease.
On November 11, Google DeepMind announced AlphaGenome 2, a model that predicts how DNA sequences and their variations affect specific cell types and states. The advance is a more detailed view of gene regulation: how cells turn genes on or off, and how strongly.
Rather than averaging activity across a broad tissue, researchers can focus on a particular kind of cell. That could help them find disease-related effects that would otherwise disappear into the average.
The cell makes the difference.
The model takes a million-letter DNA sequence together with a biological condition, such as a tissue, cell type or cell state. Google says it trained AlphaGenome 2 on more than 107 million individual cell profiles, alongside bulk data from human and mouse samples.
In one evaluation using TenK10K, a dataset covering 28 primary immune cell types, Google reports that AlphaGenome 2 identified 30% of causal variants affecting gene expression, compared with 16% for the original AlphaGenome fine-tuned on single-cell data. Those results concern a particular research evaluation—not the share of diseases the model can diagnose.
The useful idea is that the cellular setting matters. A genetic variant may change a gene’s activity in one immune cell without having the same effect elsewhere. More specific predictions give researchers a better place to look.
A useful research starting point.
Google illustrates this with a finding about inflammatory bowel disease. A previous experimental study linked a DNA variant in a distant regulatory switch to increased activity of the ETS2 gene in gut immune cells called macrophages.
AlphaGenome 2 reproduced that finding computationally, identifying the variant and its selective effect in those cells. It was not a new clinical discovery or treatment; it was a demonstration that the model could recover an established biological connection at the right cellular level.
Scientists can also supply measurements from their own single-cell experiments to investigate cell states the model has not seen before, without retraining it. Google has precomputed predictions for trait-associated variants across hundreds of human cell and tissue conditions, reducing the need for each research group to repeat large computations.
Opening the toolkit.
AlphaGenome 2 and interactive tutorials are available for non-commercial research through the AlphaGenome application programming interface (API). Commercial research access is in private preview. Google also plans AlphaGenome 2 Skills for Google Antigravity; that integration is not described as already released.
The model is intended for research and has not been validated or approved for clinical or diagnostic purposes. Predictions remain clues to investigate, not medical answers to act on.
Our reading: the constructive step is a better connection between a DNA change and the particular cells where it may matter. For scientists deciding which experiments to run next, that sharper starting point could save effort and open new routes to understanding disease.