AlphaGenome 2: Deciphering Cellular DNA Interpretation for Disease Research
A new AI model predicts how genetic variants impact specific human cell types.
1Advancing Genomic Predictive Capabilities
AlphaGenome 2 is a next-generation model designed to predict how DNA sequences and variations impact specific human cell types. By bridging single-cell biology with human genetics, it provides scientists a powerful tool to identify cellular drivers of disease. This model represents a significant evolution from its predecessor, AlphaGenome, launched last year.
2Redesigning Model Architecture
Unlike the original model that predicted properties across broad tissues in parallel, AlphaGenome 2 uses a conditional architecture. It takes a specific biological condition, such as a cell type or state, alongside a million-letter DNA sequence. This allows the model to learn how identical DNA stretches function differently across various cell types.
3Training on Massive Datasets
The model was trained on over 107 million individual cell profiles sourced from both human and mouse samples across 68 distinct tissues. This extensive training data enables the system to resolve activity within specialized cells, such as heart muscle or immune cells, which were previously obscured by broad tissue-level averages.
4Generalizing to New States
AlphaGenome 2 can generalize to entirely new single-cell experiments without requiring retraining. By uploading a small snapshot of gene activity, scientists can condition the model on unseen states. For example, it successfully distinguished vaccine-activated immune T cells from resting ones, revealing variants that specifically disrupt gene activity in activated states.
5Expanding Predictive Readouts
Beyond standard readouts like gene expression and chromatin accessibility, the model now predicts DNA methylation. This addition was widely requested by the research community to better understand epigenetic regulation. The model's ability to perform cross-modality imputation allows researchers to predict unmeasured properties from a single measurement, helping prioritize future experimental efforts.
6Bridging Genetics and Drug Discovery
Connecting trait-associated variants from genome-wide association studies to single-cell maps remains a major challenge in drug discovery. AlphaGenome 2 precomputes predictions for these variants across hundreds of conditions. This helps researchers bridge the gap between association signals and actionable insights, potentially accelerating the development of new therapeutic programs for complex diseases.
7Availability and Future Integration
The model is currently available for non-commercial research via the AlphaGenome API, with a private preview for commercial entities. Future plans include releasing AlphaGenome 2 Skills within the Google Antigravity platform. This integration aims to combine specialized biological models with agentic science tools to further accelerate discoveries in the research community.
8Research-Only Clinical Constraints
While AlphaGenome 2 provides advanced insights into basic biology and disease mechanisms, it is strictly intended for research use only. The developers emphasize that the model has not been validated or approved for clinical or diagnostic purposes. Users are encouraged to provide feedback to help shape the future roadmap of the platform.