Google DeepMind Launches AlphaGenome 2 to Map DNA Effects on Single Cells

Google DeepMind has introduced AlphaGenome 2, a biological artificial intelligence model engineered to predict how DNA mutations influence gene regulation across distinct human cell types and cellular states. Google detailed on Wednesday, November 11, 2026, that the architecture serves as the successor to the original AlphaGenome released the prior year. While virtually all human cells share identical genetic sequences, specialized cell lineages such as heart myocytes or immune lymphocytes activate contrasting combinations of genes depending on functional context. By capturing these cell-specific dynamics, AlphaGenome 2 identifies nearly twice as many gene-altering regulatory variants as its predecessor, illuminating the molecular mechanisms that underpin human illness.

Training Dataset Scale and Single-Cell Generalization

DeepMind trained AlphaGenome 2 on more than 107 million individual cellular profiles spanning 68 tissues derived from human and mouse datasets. The model accepts DNA sequences spanning up to one million base pairs alongside a specified tissue, cell lineage, or biological condition. It produces comprehensive molecular readouts including gene expression, chromatin accessibility, histone marks, and DNA methylation patterns for that exact cellular environment. A dedicated cell encoder enables researchers to condition predictions on experimental single-cell data from as few as several dozen cells without requiring model retraining. In benchmark evaluations on the TenK10K atlas across 28 immune cell types, AlphaGenome 2 identified causal genetic variants 30% of the time compared to 16% for leading comparison systems. AlphaGenome 2 reproduced in silico the selective overactivation of inflammatory gene ETS2 in gut mucosal macrophages, according to KIwoche.

Targeted Therapeutics and Virtual Cell Modeling

Google DeepMind research scientist Adam Gayoso told The Verge the big breakthrough is tying human genetics to cell biology. Google DeepMind genomics lead Ziga Avsec explained that developing targeted medicines requires identifying specific proteins and cell types rather than relying on bulk tissue averages. The model effectively converts static cellular atlases into predictive simulation engines, identifying experimental gaps to guide laboratory investigations. Avsec noted that future model iterations may incorporate data across multiple biological species, eventually contributing to virtual cells that digitally model living biological systems.

Computational Demands and API Access Framework

Despite its enhanced resolution, AlphaGenome 2 will operate alongside rather than replace the original AlphaGenome due to substantial computational processing requirements. The heightened compute intensity made it impractical to employ AlphaGenome 2 to construct the complete genome-wide variant atlas published by DeepMind in September. Google has opened AlphaGenome 2 for non-commercial academic research through its AlphaGenome API, while commercial access is restricted to a private preview. DeepMind explicitly notes that the predictive model is not validated for direct clinical care or patient diagnostic applications.