Google DeepMind Releases AlphaGenome 2 for Cell-Type-Specific Variant Effects

Executive Briefing
“Google DeepMind has released AlphaGenome 2, a DNA model that processes a 1-megabase genomic...”
Key Takeaways
- 1.AlphaGenome 2 reads a 1 Mb DNA sequence along with a target cell state to predict variant effects in specific cell types.
- 2.The model uses a U-ватися style architecture featuring a convolutional encoder, an 11-block transformer, and a decoder, employing FiLM layers to modulate activations based on a 256-dimensional embedding.
- 3.On TenK10K sc-eQTL benchmarks, AlphaGenome 2 achieved a 30% recall of causal variants at 90% precision, compared to 16% for the fine-tuned predecessor.
- 4.The tool is currently available via an API for non-commercial research, while local weights have not yet been released.
Google DeepMind has released AlphaGenome 2, a DNA model that processes a 1-megabase genomic window alongside a cell state to predict variant effects. The system outperforms its predecessor and competing tools across single-cell gene expression benchmarks.
Key Developments and Technical Context
AlphaGenome 2 reads a 1 Mb DNA sequence along with a target cell state to predict variant effects in specific cell types.
The model uses a U-ватися style architecture featuring a convolutional encoder, an 11-block transformer, and a decoder, employing FiLM layers to modulate activations based on a 256-dimensional embedding.
On TenK10K sc-eQTL benchmarks, AlphaGenome 2 achieved a 30% recall of causal variants at 90% precision, compared to 16% for the fine-tuned predecessor.
The tool is currently available via an API for non-commercial research, while local weights have not yet been released.
Strategic Analysis and Industry Implications
Integrating single-cell states from raw data allows the model to adapt to unseen datasets without requiring costly fine-tuning.
Conditioning the encoder and transformer architectures plays a decisive role in boosting accuracy compared to intervening only at the final decoder layers.
Outlook and Key Factors to Watch
The exact timeline for when Google DeepMind will publish model weights for local deployment remains unspecified.
The system's practical performance outside accessible chromatin regions is not fully established, as direction-of-effect sign prediction scores drop in those areas.







