Despite significant progress in diagnostic and therapeutic modalities, lung adenocarcinoma (LUAD) still exhibits a high recurrence risk and a low 5-year survival rate. Reliable prognostic signatures are imperative for risk stratification in LUAD patients. This study encompassed 2740 patients from 23 LUAD cohorts, including one single-cell RNA sequencing (scRNA-seq) dataset, five bulk RNA-seq datasets, and 17 microarray datasets. Using scRNA-seq dataset, we defined a group of epithelial-specific transcription factors significantly over-represented in the epithelial-to-mesenchymal transition (EMT) gene set (enrichment ratio [ER] = 5.80, Fisher's exact test p <
.001), and the corresponding target genes were significantly enriched in the cancer driver gene set (ER = 2.74, p <
.001), indicating of their crucial roles in the EMT process and tumor progression. We constructed a single-cell gene pairs (scGPS) signature, composed of 3521 gene pairs derived from the epithelial cell-specific transcription factor regulatory network, to predict overall survival (OS) of LUAD. High-risk patients identified by scGPS in the discovery cohort exhibited significantly worse OS compared to low-risk patients (Hazard ratio [HR] = 1.78, 95% CI: 1.29-2.46, log-rank p = 1.80 × 10