DeepMind’s first AlphaGenome model, released last year, examined how DNA variants affected molecular activity across broad tissues. Its successor, announced Wednesday, narrows the lens: AlphaGenome 2 is built to predict what the same genetic change may do in a particular cell type or cell state, rather than averaging its effects across a tissue.¹

That distinction matters because heart, immune and other cells carry the same DNA but use it differently. In announcing the model, DeepMind chief executive Demis Hassabis said it predicts variant effects in “different specific cell types,” a capability he framed as crucial to understanding complex disease.Demis Hassabis (demishassabis)²

The company says the system was trained on more than 100 million individual cell profiles spanning 68 tissues. It reports that AlphaGenome 2 identifies nearly twice as many variants that alter gene expression in cells as its earlier leading model, and can make predictions for newly profiled cell states without retraining.¹

The independent framing is more measured, but broadly aligned. The Verge reported that the model could connect variants found in population-scale genetic studies to their effects at the cellular level—potentially exposing disease mechanisms that broader models miss. DeepMind genomics lead Žiga Avsec said such resolution matters because drug development targets “specific proteins and specific cell types.”³

The trade-off emerged just as clearly. Avsec said AlphaGenome 2’s more detailed outputs are “more expensive” to generate, making it impractical for the genome-wide Atlas DeepMind previously produced for every possible single-letter change in human DNA.³

For now, Google has opened the model through the AlphaGenome API for non-commercial research, while commercial access remains in private preview.Sundar Pichai (sundarpichai)⁴ The promise is a more exact map from mutation to disease; the constraint is whether researchers can afford to use that map at scale.