Google DeepMind on Wednesday announced AlphaGenome 2, an AI model it says predicts how DNA sequences and genetic variants affect specific human cell types and states, and said the model is available for non-commercial research through the AlphaGenome API^1,2.

The company described the system as a successor to AlphaGenome, released last year to predict variant effects on molecular activity across broader tissues, later expanded with AlphaGenome Atlas predictions for single-letter genome changes^1,2. DeepMind said nearly every cell carries the same DNA, but cell types switch different genes on and off, so the same variant can have different effects depending on cell type and state^1,2.

According to DeepMind, AlphaGenome 2 was redesigned as a conditional model that predicts DNA effects for one condition — a tissue, cell type, or state — at a time, rather than predicting many tissues in parallel². It said the model was trained on bulk and single-cell data from human and mouse samples, including more than 107 million individual cell profiles; The Verge reported training on more than 100 million profiles across 68 tissues^1,2. DeepMind said the model predicts gene expression, histone modifications, chromatin accessibility, and DNA methylation².

DeepMind said AlphaGenome 2 identifies nearly twice as many variants that change gene expression in cells as its previous state-of-the-art model^1,2. On the TenK10K single-cell map of 28 primary immune cell types, it said the model identified nearly twice as many causal variants (30% versus 16%) and directional gene effects (43% versus 23%) as the original AlphaGenome fine-tuned on single-cell data and other specialized tools². Those benchmark figures come from DeepMind and were not independently assessed in the reporting reviewed here².

The company said the model can generalize to new single-cell experiments without retraining, including predicting missing experimental readouts such as chromatin accessibility from a gene-activity measurement^1,2. DeepMind cited examples including computationally reproducing a published finding that a distant regulatory variant drives IBD by overactivating ETS2 in gut macrophages, and analyzing unseen COVID-19 mRNA vaccine single-cell data².

Adam Gayoso, a Google DeepMind research scientist on the project, told The Verge the advance is “tying human genetics to cell biology”¹. Genomics lead Žiga Avsec told The Verge that drug development targets specific proteins and cell types, and that the finer predictions are more computationally expensive than the first model, which he said made AlphaGenome 2 impractical for the genome-wide Atlas DeepMind released in September¹.

DeepMind said it precomputed variant scores across hundreds of cell and tissue conditions to help map enhancer–gene links and interpret genome-wide association study signals for traits including height and type 2 diabetes². It said it is working with collaborators including the Wellcome Sanger Institute to profile more cell types and states, and plans to release AlphaGenome 2 skills in Google Antigravity². Google said a limited private preview is available for commercial research^1,2. DeepMind stated that AlphaGenome and AlphaGenome 2 are for research use only and have not been validated or approved for clinical or diagnostic purposes².