Background: Non-small cell lung cancer (NSCLC) remains a leading cause of cancer-related mortality, and although PD-1/PD-L1 immune checkpoint blockade has improved outcomes in some patients, therapeutic responses remain heterogeneous. Tumor-intrinsic heterogeneity within malignant epithelial populations is increasingly recognized as a critical determinant of disease progression and therapy response. Methods: Here, we constructed a comprehensive single-cell atlas of NSCLC by integrating 650,461 cells from 216 tumor and normal samples. Tumor-derived epithelial cells were reclustered to identify transcriptionally distinct subpopulations. Pseudotime analysis, functional experiments, and in vivo validation using a humanized xenograft model were performed to investigate the role of COL3A1. Results: Reclustering of tumor-derived epithelial cells revealed 25 transcriptionally distinct subpopulations. Among these, a high-risk cluster exhibited coordinated activation of epithelial–mesenchymal transition (EMT) and angiogenesis programs and was associated with poor patient survival. Within this aggressive subpopulation, Collagen type III alpha 1 (COL3A1) emerged as a tumor-intrinsic gene associated with extracellular matrix remodeling and angiogenic signaling. Pseudotime analysis indicated that COL3A1⁺ cells represent a late-stage, poorly differentiated malignant state. Functional experiments demonstrated that COL3A1 knockdown impaired NSCLC cell proliferation, migration, and invasion. Virtual knockout further suggested that COL3A1 may be associated with transcriptional programs involved in PD-L1 upstream signaling pathways, indicating a potential indirect link between tumor-intrinsic states and immune regulatory networks. Consistently, in vivo silencing of COL3A1 enhanced the antitumor efficacy of PD-L1 blockade. Conclusions: Collectively, our study identifies COL3A1 as a tumor-intrinsic gene enriched in malignant epithelial cells with mesenchymal features and a potential therapeutic target. These findings provide a rationale for exploring combinatorial strategies integrating tumor-intrinsic pathway inhibition with immune checkpoint blockade in NSCLC.