Google DeepMind released AlphaGenome 2, a redesigned model that predicts how DNA variants affect gene activity in specific human cell types and states rather than in broad tissue averages, the company announced Wednesday.
The new model “identifies nearly twice as many variants that change gene expression in cells, compared to our previous state-of-the-art model,” DeepMind wrote. It was trained on more than 100 million individual cell profiles across 68 tissues — the company later specified more than 107 million profiles from human and mouse samples, including both bulk and single-cell data. AlphaGenome 2 is available now for non-commercial research through the AlphaGenome API and in private preview for commercial research, the post said.
Last year’s original AlphaGenome took a million-letter DNA sequence and predicted molecular activity across many tissues at once. DeepMind said that approach spread the model’s capacity thin. AlphaGenome 2 is a conditional model: it takes the target biological condition — a tissue, cell type or state — as an input at every stage and predicts effects for one condition at a time. That design, the company said, also allows it to impute missing experimental data, for example predicting chromatin accessibility from a single gene-activity measurement, and to make predictions for cell and disease states it has not seen before. In addition to gene expression, histone modifications and chromatin accessibility, the new model predicts DNA methylation, which DeepMind called a widely requested readout.
The company said the work builds on AlphaGenome Atlas, which contains predictions for all 9 billion possible single-letter changes in the human genome, and is meant to help researchers prioritize variants in rare disease, complex traits and basic biology. The performance claims and the “nearly twice as many” figure come from DeepMind’s own announcement and have not been independently audited in the reporting reviewed here. The practical test will be whether outside labs can use the API to find disease-relevant variants that tissue-level models missed.