What happened
On November 11, Google DeepMind introduced AlphaGenome 2, a research model designed to predict how changes in DNA affect gene activity in particular cell types and biological states. Unlike its predecessor, which predicted molecular properties across broader tissues, the new system conditions predictions on a specified cell type or state. DeepMind says it trained the model on more than 100 million individual cell profiles spanning 68 tissues. In its reported evaluation using immune-cell data, the model identified about 30% of causal variants, versus 16% for the earlier approach. The Verge independently reported the launch and explained that the model can analyze unfamiliar cellular states without retraining. Noncommercial researchers can access it through an API, while commercial use is in limited preview.
Why it matters
Many disease-associated genetic changes do not alter a protein directly. Instead, they affect regulatory instructions that determine which genes are active, and their consequences can differ across cells. A tool that connects a variant to a specific immune or other cell population could help researchers prioritize laboratory experiments and investigate disease mechanisms more precisely. DeepMind described an example involving an inflammatory bowel disease-associated regulatory change and the ETS2 gene in macrophages. That example is a computational reproduction of an existing biological finding, not a newly proven therapy. More detailed predictions also demand more computing resources than the earlier model, limiting the practicality of using this version for every genome-wide screening task.
What to watch next
Independent researchers will need to test how well AlphaGenome 2 generalizes to underrepresented tissues, disease states and populations, especially when experimental data are sparse. The next important evidence will come from prospective laboratory validation: whether predictions consistently identify functional variants that conventional methods miss, and whether those findings improve experimental decisions. Broader access, transparent benchmark comparisons and reproducible studies will matter more than headline model rankings. Researchers should also track the computational cost of applying the system at scale. Its immediate role is to generate and prioritize biological hypotheses; clinical utility would require substantially more validation.