Google DeepMind Releases AlphaGenome 2 for Cell-Type-Specific Variant Effects
Summary
- Google DeepMind has released AlphaGenome 2, a conditional sequence-to-function DNA model that predicts variant effects in specific cell types.
- The model reads 1 Mb of DNA and a target cell state, utilizing a 256-dimensional embedding to modulate activations via FiLM layers.
- Training involved four stages, including pre-training on bulk and single-cell tracks, condition refinement, distillation, and post-training with an expression encoder.
- AlphaGenome 2 outperforms its predecessor and other models like Borzoi and Scooby on single-cell eQTL and enhancer-gene benchmarks.
- The model is currently accessible via the AlphaGenome API for non-commercial research, with weights planned for release upon final publication.
Why it matters
AlphaGenome 2 advances genomics research by providing high-resolution, cell-type-specific variant effect predictions without requiring model fine-tuning, offering improved tools for biological discovery.