Overview
- DeepMind publicly launched AlphaGenome 2 and made it available through the AlphaGenome API for non-commercial research while offering a small private preview for commercial users.
- AlphaGenome 2 uses a condition model that takes tissue, cell type, or cell state as input so predictions reflect the specific cellular environment of a DNA sequence or variant.
- The model was trained on over 107 million single-cell profiles from human and mouse across 68 tissues, which DeepMind says improves its ability to distinguish cell-specific variant effects.
- DeepMind reports large performance gains on benchmarks, citing that AlphaGenome 2 identifies roughly 30% of causal expression variants versus 16% for the prior model and raises direction-of-effect accuracy to 43% from 23%.
- DeepMind cautions the tool is for research use only and not clinically validated, so lab experiments remain necessary to confirm findings and users should watch for limits in interpretability and potential misuse.