Non-small-cell lung cancer (NSCLC) is the second most common cancer worldwide, and most deaths are associated with epithelial–mesenchymal transition (EMT). Therefore, this study aimed to explore the role of EMT-related transcriptomic profiles in NSCLC and the effect of EMT-based signatures on clinical diagnosis, prognosis, and treatment responses for patients with NSCLC. After integrating the transcriptomics and clinicopathological data, we first constructed EMT clusters (C1 and C2) using machine learning algorithms, found the significant relationship between EMT clusters and survival outcomes, and then explored the impact of EMT clusters on the tumor heterogeneity, drug efficiency, and immune microenvironment of NSCLC. Prominently, differential-enriched tumor-infiltrated lymphocytes were found between EMT clusters, especially the macrophages and monocyte. Next, we identified the most significantly down-regulated gene SFTA2 in the EMT clusters C2 with poor prognosis. Using RT-qPCR and RNA-seq data from the public database, we found prominently elevated SFTA2 expression in NSCLC tissues compared with normal lung tissues, and the tumor suppressor role of SFTA2 in 82 Chinese patients with NSCLC. After Cox regression and survival analysis, we demonstrated that higher SFTA2 expression in tumor samples significantly predicts favorable prognosis of NSCLC based on multiple independent cohorts. In addition, the prognostic value of SFTA2 expression differs for patients with lung adenocarcinoma and squamous cell carcinoma. In conclusion, this study demonstrated that the EMT process is involved in the malignant progression and the constructed EMT clusters exerted significant predictive drug resistance and prognostic value for NSCLC patients. In addition, we first identified the high tumoral expression of SFTA2 correlated with better prognosis and could serve as a predictive biomarker for outcomes and treatment response of NSCLC patients.