Google DeepMind Launches AlphaGenome 2 for Cell-Type-Specific DNA Analysis

Google DeepMind has unveiled AlphaGenome 2, an advanced DNA model capable of predicting variant effects down to specific cell types. Building upon its predecessor, which averaged data across bulk tissues, the new conditional sequence-to-function model can interpret a 1-megabase DNA window alongside a target cell state. According to the release details, the software adapts to unseen single-cell datasets without requiring fine-tuning, marking a significant step forward in computational biology.

The system utilizes a U-Net architecture featuring a convolutional encoder, an 11-block transformer, and a decoder, processing inputs through a 256-dimensional embedding. Trained across four stages—including pre-training on bulk and single-cell tracks, condition refinement, distillation, and post-training—the architecture maps single-cell expression encoders directly into the model's condition space. This allows researchers to pass raw counts from platforms like CELLxGENE to generate accurate condition embeddings.

Performance benchmarks indicate substantial gains over previous iterations and competing models. AlphaGenome 2 achieved a 43% recall rate at 90% sign accuracy on TenK10K single-cell expression quantitative trait loci, outperforming the previous version's 23%. Furthermore, median Pearson correlation for cross-tissue expression changes rose from 0.61 to 0.71. Despite these achievements, access remains limited, as the model is currently available strictly through a hosted API for non-commercial research, with local weights withheld upon initial release.

What to watch next: Watch for expanded academic adoption through the AlphaGenome API and potential updates regarding commercial licensing or local weight releases following final publication.