1059-TranscriptFormer: Atlas Across Evolutionary History

The researchers introduced TranscriptFormer, a family of generative foundation models designed to analyze single-cell transcriptomic data across 1.53 billion years of evolution. By training on up to 112 million cells from 12 diverse species, the model learns to represent cellular states and gene relationships in a species-agnostic embedding space. It achieves state-of-the-art performance in classifying cell types, predicting disease states, and identifying drug-induced changes, even for organisms not seen during training. Beyond simple classification, the model functions as a "virtual instrument," allowing scientists to simulate biological inquiries like gene regulatory network interactions through generative prompting. This work demonstrates that universal principles of cellular organization can be computationally captured and predicted across the entire tree of life.

References:

• Pearce J D, Simmonds S E, Mahmoudabadi G, et al. TranscriptFormer: A generative cell atlas across 1.5 billion years of evolution[J]. Science, 2026: eaec8514.