AlphaGenome 2 maps DNA effects across individual cell types
DeepMind’s new model uses more than 100 million cell profiles to predict how genetic variants behave differently across tissues and cell states.
A more conditional view of genetics
Google DeepMind introduced AlphaGenome 2, a DNA sequence model designed to predict how genetic variants affect specific human cell types. The system can condition its predictions on a reference cell type or on a cell state generated from a scientist’s own single-cell experiment.
DeepMind says the model was trained on more than 100 million individual cell profiles spanning 68 tissues. Compared with the previous system, it identifies nearly twice as many variants that alter gene expression in cells, according to the company’s evaluation.
From tissue averages to cellular detail
The first AlphaGenome model produced many predictions in parallel across tissues and cell lines. AlphaGenome 2 instead focuses on one biological condition at a time, allowing the architecture to model how the same DNA sequence may be interpreted differently by heart muscle, immune or other specialized cells.
That shift addresses a central problem in human genetics. A disease-associated variant may have little visible effect in an average tissue measurement while strongly changing gene activity in a rare or highly specific cell population. A model that can make those distinctions could help researchers prioritize variants for laboratory testing and connect genetic changes to disease mechanisms.
The conditional design also supports cross-modality imputation. Given one measurement from a single-cell experiment, the system can estimate other unmeasured properties, such as chromatin accessibility. DeepMind says it can adapt to new laboratory data without retraining the entire model.
Access and limits
AlphaGenome 2 is available today through an API for non-commercial research, with commercial research in private preview. The release remains a prediction system, not a clinical diagnostic tool. Its usefulness will depend on experimental validation, data quality and whether researchers can translate correlations into causal biological findings.
The important development is not simply a larger genomic model. It is the move toward cell-aware predictions, where the relevant unit of genetic risk is the particular cell state in which a variant acts.