Machine learning identifies TIME subtypes linking EGFR mutations and immune states in lung adenocarcinoma
Zetian Gong, Mingjun Du, Ying Li, Bicheng Ye, Yuming Huang, Hui Gong, Wei Wang, Liang Chen, Zongli Ding, Pengpeng Zhang
Abstract
Epidermal growth factor receptor (EGFR) mutation is a key oncogenic driver in lung adenocarcinoma (LUAD), but its impact on the tumor immune microenvironment (TIME) remains unclear. By integrating single-cell transcriptomes from 153 LUAD samples using machine learning, we generated an atlas of over one million cells that delineates immune heterogeneity. EGFR-mutant tumors exhibited enrichment of TIGIT + regulatory T cells, neutrophils, and macrophages, whereas wild-type tumors contained abundant ZNF683 + CD8 + tissue-resident memory T cells, diverse memory B cells, and FGFBP2 + CD16 high natural killer cells, reflecting an immune-active TIME. Non-negative matrix factorization defined five TIME subtypes, with EGFR-mutant patients clustering into immunosuppressive profiles linked to poor prognosis. Flow cytometry and mouse models confirmed the cytotoxic and PD-1 blockade-enhancing functions of FGFBP2 + NK cells. These findings reveal distinct TIME landscapes in EGFR-mutant LUAD and illustrate the potential of machine learning-based immunogenomic analysis to inform precision immunotherapy.