Under the leadership of Nobel laureate Hassabis, DeepMind has launched AlphaGenome 2, its next-generation DNA foundation model. The new model focuses on the genomic "dark matter" in human DNA—the non-coding regions that make up 98% of the genome and harbor over 90% of pathogenic mutations. It achieves four core breakthroughs. First, its prediction accuracy has been refined to the single-cell level by integrating over 107 million single-cell data points from 68 human tissues, and its ability to identify pathogenic mutations is nearly double that of the previous generation. Second, without any biological experiments, it has computationally reproduced findings on inflammatory bowel disease-related pathogenic mutations that took Nature years to complete, and has accurately identified the targets of an already-marketed CRISPR gene therapy for sickle cell disease. Third, when faced with previously unseen lesions, it requires no retraining—researchers simply input dozens of single-cell sequencing data points to rapidly predict mutation effects, far outperforming similar models in tests on Parkinson's disease and COVID-19-related samples. Fourth, it can infer multiple high-cost detection metrics from routine single-cell gene expression data. The official resources, including an impact score library covering nearly 7 million gene variants, are freely open. The model will subsequently be packaged as Skills and integrated into relevant platforms for Gemini 4 to invoke, potentially greatly shortening research cycles for complex diseases and new drug development.