Google DeepMind has announced AlphaGenome 2, a research model for predicting how genetic variants affect molecular activity in particular cell types or states. The company contrasts the new model with the first AlphaGenome, which produced outputs across broad tissues in parallel. It says the finer distinctions can help researchers prioritize variants that would be harder to see in tissue averages.
What changed from the first model
AlphaGenome 2 conditions a prediction on a specified biological condition, such as a cell type. This changes the question researchers can ask about a variant. A broad average can combine signals from several cell populations. The newer model lets researchers examine a more specific setting and compare its predicted effect with what they observe in experiments.
What DeepMind reports
The team says it trained on bulk and single-cell data, including more than 107 million individual cell profiles. It reports that AlphaGenome 2 identified nearly twice as many variants affecting gene expression in its evaluation as its previous model. The announcement also lists DNA methylation alongside the other molecular measurements the model predicts. The result concerns the reported model evaluation and should be interpreted with its experimental setting.
Who can use the result
DeepMind announced availability for non-commercial research. A geneticist investigating a variant can use a prediction to help decide which cell context warrants closer study. A predicted molecular effect still needs validation, and the announcement addresses research uses. One practical next step is to compare predictions with experimental measurements for the cell types of interest and document which of those cell contexts are supported by the available data.
