Clarity for AI buildersSponsored by AI Routing
adoption

AlphaGenome 2 targets cell-specific genetic effects, with improved results reported on immune-cell benchmark

AlphaGenome 2 targets cell-specific genetic effects, with improved results reported on immune-cell benchmark
AlphaGenome 2 predicts genetic effects for specific cell states, The Decoder reports. Google DeepMind has released the genomics model for noncommercial research through an API, with commercial access offered through a closed preview, according to The Decoder. The model is designed to predict how DNA variations affect particular human cell types rather than produce broad predictions across many tissues at once. It is not approved for clinical diagnostics. The central change is how researchers specify the biological context. The original AlphaGenome processed a DNA sequence one million letters long and predicted properties across thousands of experiments spanning multiple tissues. AlphaGenome 2 instead takes a biological state as an additional input and calculates effects for that condition. The Decoder attributes to DeepMind the explanation that the earlier approach spread the model’s capacity too thin to capture subtle differences between related cell types. That distinction matters because a shared DNA sequence does not imply a shared response to a genetic variant. Different cell types activate different genes. Conditioning predictions on a cell state is therefore intended to help researchers distinguish where a variant has an effect, rather than treating its consequences as uniform across tissues. The model was trained on more than 107 million individual cell profiles from 68 tissues in humans and mice, the report says. Its predictions cover gene expression, genome accessibility and DNA methylation. It can also use one measured property, such as gene activity, to infer other properties for which measurements are missing. These inferred properties remain predictions, not additional laboratory observations. Their practical role, as described in the report, is to help researchers prioritize experiments. The distinction is important when interpreting the model’s ability to fill gaps in a dataset: it can suggest what to measure next without replacing the missing measurement. On the TenK10K benchmark covering 28 primary immune cell types, AlphaGenome 2 identified about 30% of causal variants, compared with 16% for a fine-tuned version of AlphaGenome 1. It correctly predicted whether a variant increased or decreased gene activity in 43% of cases, versus 23% for the predecessor, according to The Decoder. Those results support an improvement on the reported benchmark, not a claim that the model resolves most genetic effects. Identifying a causal variant and determining the direction of its effect are also distinct tasks; the two reported percentages should not be read as interchangeable measures of accuracy. The report describes a computational reproduction of a previously studied inflammatory bowel disease mechanism involving selective overactivation of ETS2 in gut macrophages. In a separate application using COVID-19 mRNA vaccine study data, the model distinguished activated T cells from resting cells and predicted that a variant linked to severe COVID-19 disrupted activity mainly in the activated state. Together, these examples illustrate why cell state can matter alongside cell type: an effect visible after activation may be missed in resting cells. They do not establish a new clinical finding or demonstrate diagnostic readiness. DeepMind has also precomputed predictions for hundreds of cell and tissue conditions, according to the report. These are intended to help interpret genome-wide association studies by narrowing candidate causal variants and identifying affected cell types. The supplied evidence is The Decoder’s account; it does not include the underlying study or evaluation materials for independent verification.