Google DeepMind Releases AlphaGenome 2 for Cell-Level DNA Predictions

Google DeepMind on Wednesday released AlphaGenome 2, a DNA model that predicts how a genetic variant changes gene activity in a specific human cell type. It identified 30% of causal variants on the TenK10K immune-cell dataset against 16% for the prior approach, and is available today for non-commercial research.

The announcement came from DeepMind's research team, led by Adam Gayoso, Natasha Latysheva, Pushmeet Kohli and Ziga Avsec. According to the company's blog post, every cell in the body carries the same DNA, yet each cell type switches different genes on or off. Earlier tools such as AlphaGenome, launched last year, predicted how variants affect molecular activity across broad tissues, which can dilute signals that matter only in specialized cells.

AlphaGenome 2 is a conditional model: alongside a DNA sequence of about one million letters, it takes a specific biological condition, such as a tissue, cell type or state, as input at every stage of its architecture. DeepMind said this lets it learn how the same stretch of DNA is read differently across cell types, and that it can reproduce findings such as an IBD-linked variant overactivating the ETS2 gene in gut macrophages while leaving other cells largely unaffected.

The model was trained on bulk and single-cell data from human and mouse samples, including over 107 million individual cell profiles across 68 tissues. It predicts the same core readouts as AlphaGenome, including gene expression, histone modifications and chromatin accessibility, plus DNA methylation, and can generalize to new cell states from a single-cell RNA-seq snapshot of just a few dozen cells without retraining. SiliconANGLE, which reported the release, noted that the predictions guide research rather than replace experimental validation and that the model is not approved for clinical or diagnostic use.

Access is split: non-commercial researchers can use it through the AlphaGenome API, with commercial research access in a private preview. DeepMind also precomputed predictions for trait-associated variants across hundreds of conditions and said AlphaGenome 2 Skills are coming to Google Antigravity. Ziga Avsec told The Verge, as cited by SiliconANGLE, that the finer-grained predictions are more computationally expensive, so AlphaGenome 2 does not replace the original model or the genome-wide Atlas released in September.

Alphabet (GOOGL) shares were up 2.1% in delayed trading on Thursday. The next catalyst for the company's AI research story is not dated in the sources.