AlphaGenome 2: Deciphering how cells read DNA to better understand disease
AlphaGenome 2 is a conditional DNA‑sequence model that predicts the impact of genetic variants on gene expression in specific human cell types and states, improving variant detection nearly two‑fold over its predecessor. Trained on more than 100 million single‑cell profiles spanning 68 tissues, it ingests a 1‑million‑base DNA segment together with a cell‑type condition at every architectural stage. The model is accessible via the AlphaGenome API for non‑commercial research and in private preview for commercial use, enabling engineers to embed cell‑type‑specific variant predictions into genomics pipelines.
⚡ Key Takeaways
- Identifies nearly twice as many variants that change gene expression in cells compared to AlphaGenome 1.
- Uses a conditional architecture that incorporates the target biological condition (tissue, cell type, or state) alongside the DNA sequence at each layer.
- Trained on >100 million individual cell profiles across 68 tissues, providing fine‑grained predictive power.
- Accessible through the AlphaGenome API for non‑commercial research (private preview for commercial).
- Requires a 1‑million‑letter DNA sequence input and a defined cell‑type condition; predictions for unseen cell states rely on the model’s encoder.
Engineers can now integrate precise, cell‑type‑specific variant effect predictions into disease‑association studies and drug‑target pipelines, reducing reliance on coarse bulk‑tissue models and accelerating discovery of cellular disease mechanisms.
✅ Practical Steps
- Call the AlphaGenome API with a 1‑million‑base DNA sequence and a cell‑type identifier from the reference catalog to obtain variant effect predictions.
- Encode custom single‑cell experiment data using the provided cell encoder to generate on‑the‑fly condition inputs for novel cell states.
- Deploy the API calls within your variant‑prioritization workflow to replace bulk‑tissue predictions with cell‑type‑specific scores.
Want the full story? Read the original article.
Read on Google DeepMind ↗