Generative pretraining from large-scale transcriptomes for single-cell deciphering

• DOI: 10.1016/j.isci.2023.106536
• Publication Date: 2023-04-20T10:04:44Z
• Authors: Hongru Shen; Jilei Liu; Jiani Hu; Xilin Shen; Chao Zhang; Dan Wu; Mengyao Feng; Meng Yang; Yang Li; Yichen Yang; Wei Wang; Qiang Zhang; Jilong Yang; Kexin Chen; Xiangchun Li
• Keywords: Proteogenomics; Feature (linguistics); Generative model

Exponential accumulation of single-cell transcriptomes poses great challenge for efficient assimilation. Here, we present an approach entitled generative pretraining from (tGPT) learning feature representation transcriptomes. tGPT is conceptually simple in that it autoregressive models the ranking a gene context its preceding neighbors. We developed with 22.3 million and used four datasets to evalutate performance on analysis tasks. In addition, examine applications bulk tissues. The clusters cell lineage trajectories derived are highly aligned known labels states. patterns tumor tissues learned by associated wide range genomic alteration events, prognosis, treatment outcome immunotherapy. represents new analytical paradigm integrating deciphering massive amounts transcriptome data will facilitate interpretation clinical translation