1
0The research article introduces TranscriptFormer, a sophisticated generative foundation model designed to analyze single-cell transcriptomics across a vast evolutionary timeline of 1.5 billion years. By utilizing protein sequence embeddings and an autoregressive architecture, the model maps gene expression data into a shared embedding space that allows for direct comparisons between highly diverse species without requiring common orthologous genes. The study demonstrates that TranscriptFormer excels at cell type classification and zero-shot disease detection, significantly outperforming existing models like UCE and Geneformer in both accuracy and evolutionary reach. Beyond classification, the model serves as a "virtual instrument" capable of simulating cellular responses and predicting gene regulatory networks through interactive prompting. Ultimately, the research proves that universal biological principles are learnable by AI, establishing a powerful new framework for comparative cellular biology and the study of evolutionary conservation.
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, 393(6806): aec8514.
