Non-small cell lung cancer (NSCLC) accounts for more than 80% of all lung cancers and is characterized by poor prognosis, particularly when diagnosed at advanced stages. A fundamental but understudied aspect of NSCLC is its cellular origin: type II alveolar epithelial cells (type II pneumocytes) are one of the major sources from which NSCLC cancer cells arise, making them a critical entry point for understanding the disease.
Type II pneumocytes are multifunctional lung cells that normally produce surfactant (the substance that keeps air sacs open), participate in lung injury repair, and interact extensively with the immune system. In NSCLC, these cells cross-talk with immune cells in the tumor microenvironment through secreted cytokines - for example, by releasing TGF-beta or GM-CSF to polarize macrophages toward the tumor-promoting M2 phenotype that suppresses immune responses and facilitates cancer growth.
Despite their biological importance, no reliable biomarkers derived from type II pneumocytes had been systematically identified and validated for clinical use in NSCLC. Such biomarkers could reveal both the molecular mechanisms of NSCLC progression and predict which patients will respond to immunotherapy, where response rates of only 27-45% indicate that better patient selection tools are urgently needed.
This study used a multi-layered computational approach to fill this gap: single-cell RNA sequencing to map the tumor microenvironment at cellular resolution, high-dimensional weighted gene co-expression network analysis (hdWGCNA) to identify gene networks specific to type II pneumocytes, and two machine learning algorithms to select the most informative biomarker genes from those networks.
Single-cell RNA sequencing (scRNA-seq) data from four untreated early-stage NSCLC patients (dataset GSE117570) was processed using the Seurat bioinformatics package. After quality filtering to remove low-quality cells and those with high mitochondrial gene content (indicating dying cells), 3,883 cells were retained for analysis. Dimensionality reduction and clustering identified seven distinct cell populations within the tumors.
CellChat analysis was used to map communication between cell subpopulations by identifying ligand-receptor pairs - the molecular signals one cell type sends and another receives. This approach revealed which cell types communicate with type II pneumocytes and through which specific molecular interactions, providing insight into how these cells influence the tumor immune environment.
High-dimensional weighted gene co-expression network analysis (hdWGCNA) was applied to the type II pneumocyte population specifically. This technique groups genes that are co-expressed together (turned on and off in coordinated patterns) into modules, then identifies which genes within each module have the most central, hub-like connections. Hub genes are the most likely candidates for driving the biological function of that module.
The M6 co-expression module showed the highest expression ratio and expression level specifically within type II pneumocytes compared to all other cell subpopulations, making it the target module for downstream biomarker identification. Its hub genes were then intersected with differentially expressed genes between NSCLC and normal tissue to narrow to candidates most relevant to cancer progression.
Two machine learning algorithms were applied in parallel to identify the most informative genes from the candidate set of five type II pneumocyte-associated hub genes that were also differentially expressed in NSCLC: LASSO (Least Absolute Shrinkage and Selection Operator) and SVM-RFE (Support Vector Machine Recursive Feature Elimination).
LASSO regression works by fitting a predictive model while simultaneously penalizing the model for including too many genes - genes that contribute less predictive information are progressively shrunk to zero and removed. Using 3-fold cross-validation to select the optimal penalty, LASSO identified three genes from the five candidates.
SVM-RFE works differently: it trains a classifier, ranks genes by their importance to the classifier's decisions, removes the least important gene, and repeats - iteratively eliminating genes while measuring which subset achieves the best classification performance. Applied to the same five genes, SVM-RFE identified a set of feature genes with the lowest cross-validated error.
Taking the intersection of both algorithms' results yields genes identified as informative by two independent computational methods - a stronger validation than either approach alone. This intersection produced the three final biomarkers: HN1 (Hematological and Neurological Expressed 1), OCIAD2 (Ovarian Carcinoma Immunoreactive Antigen Domain 2), and SFTA2 (Surfactant-Associated Protein 2).
Seven major cell subpopulations were identified in NSCLC tumor samples: macrophages, type II pneumocytes, T cells, ciliated cells, B cells, endothelial cells, and plasma cells. Each population was identified by distinct marker genes - for example, type II pneumocytes were marked by SFTPB, NAPSA, and SLC34A2 (all involved in surfactant production or lung epithelial function).
Type II pneumocyte-specific genes were enriched in pathways involved in epithelial cell proliferation and the ERBB signaling pathway - a network of growth factor receptors including HER2 that is frequently amplified or mutated in cancer and that drives rapid cell division when overactivated.
CellChat identified extensive two-way communication between type II pneumocytes and macrophages. Three key ligand-receptor interaction pairs mediate this crosstalk: ITGB2-ICAM1 (an adhesion-mediated contact signal), HLA-DPB1-CD4, and HLA-DQB1-CD4. The HLA-CD4 interactions are particularly significant - they represent antigen presentation signals connecting type II pneumocytes with CD4+ helper T cells that can coordinate immune responses to cancer.
The extensive type II pneumocyte-macrophage communication is biologically meaningful: macrophages are the most abundant immune cell in the NSCLC tumor microenvironment, and their polarization toward either the anti-tumor M1 or pro-tumor M2 phenotype fundamentally determines whether the immune environment supports or suppresses cancer growth. The signals from type II pneumocytes could be actively driving macrophage behavior in NSCLC.
CIBERSORT immune cell deconvolution analysis revealed markedly different immune landscapes between NSCLC tumors and normal lung tissue: tumor samples showed higher plasma cell and macrophage infiltration, while normal tissue had more CD8+ T cells, activated mast cells, and neutrophils. The reduced CD8+ T cell presence in tumors is consistent with immune evasion.
HN1 and OCIAD2 showed opposing immune correlations to SFTA2. HN1 and OCIAD2 expression were significantly negatively correlated with immune score, stromal score, and ESTIMATE score (which together measure immune cell and stromal cell infiltration levels). This means that higher HN1 and OCIAD2 expression correlates with a less immune-infiltrated, more immunologically cold tumor environment - a hallmark of tumors that are harder to treat with immunotherapy.
SFTA2, in contrast, was significantly positively correlated with all three immune infiltration scores, and also positively correlated with infiltration of granulocytes, monocytes, dendritic cells, eosinophils, and neutrophils. High SFTA2 expression indicates an immune-hot tumor microenvironment with greater immune surveillance. This pattern is consistent with prior data linking higher SFTA2 expression to favorable prognosis in NSCLC.
The divergent correlations of HN1/OCIAD2 versus SFTA2 suggest these genes may serve as complementary markers: HN1 and OCIAD2 as markers of aggressive, immune-cold disease, while SFTA2 marks immune-active, potentially more treatment-responsive tumors. The ability to distinguish these immune phenotypes from gene expression profiling could inform immunotherapy patient selection.
GSEA (Gene Set Enrichment Analysis) was used to understand the biological functions associated with high versus low expression of each biomarker. Patients were divided into high and low expression groups by median biomarker level, and enrichment of HALLMARK gene sets was assessed - identifying which established biological pathways are most active in each group.
High HN1 expression enriched E2F TARGETS and MYC TARGETS (V1 and V2), both of which are core cancer cell cycle and proliferation programs. E2F transcription factors drive cells through the cell cycle checkpoints, and MYC is a master regulator of cell growth and metabolism that is overactive in many cancers. This enrichment pattern explains why high HN1 expression correlates with aggressive tumor behavior.
Low HN1 expression showed enrichment of inflammatory response and apoptosis pathways, which are typically associated with cancer resistance mechanisms - that is, the tumor microenvironment when HN1 is low may be responding to immune attack and activating survival pathways. Notably, SFTA2 pathways were opposite to HN1, consistent with the inverse relationship between these two genes in tumor biology.
High OCIAD2 expression enriched angiogenesis, coagulation, IL6-JAK-STAT signaling, KRAS signaling, and TNF-alpha/NF-kB signaling pathways. These pathways are all established cancer-promoting programs: angiogenesis feeds tumor blood supply, IL6-JAK-STAT drives immune evasion and tumor growth, and KRAS/NF-kB signaling promotes cancer cell survival and invasion. OCIAD2 thus appears to be a regulator of multiple interconnected tumor-promoting programs simultaneously.
Laboratory validation experiments confirmed that HN1 and OCIAD2 are significantly overexpressed in two NSCLC cell lines (NCI-H838 and A549) compared to normal bronchial epithelial cells (BEAS-2B), while SFTA2 is downregulated in these cancer cell lines. This expression pattern in cell culture mirrors what was found in patient tissue data, supporting the clinical relevance of the computational findings.
HN1 was selected for functional validation by designing two independent siRNA sequences (si-HN1#1 and si-HN1#2) to knock down its expression in both NSCLC cell lines. siRNA (small interfering RNA) is a tool that specifically silences one gene while leaving all others unaffected, allowing researchers to determine what happens to cancer cells when HN1 function is removed.
HN1 knockdown reduced cancer cell proliferation, as measured by the CCK-8 cell viability assay. It also significantly inhibited migration (measured by wound healing assay, where cells fill a gap scratched in a cell monolayer) and invasion (measured by Transwell assay, where cells penetrate through a membrane barrier). These three behaviors - proliferation, migration, and invasion - are the fundamental capabilities that allow cancer to grow and spread.
These functional results provide causal evidence that HN1 is not merely a bystander marker but actively promotes cancer cell aggressiveness. The inhibition of all three tumor behaviors upon HN1 silencing suggests it could be a therapeutic target - drugs that block HN1 function might limit NSCLC growth and metastatic potential.
This study successfully identified three NSCLC biomarkers - HN1, OCIAD2, and SFTA2 - by combining single-cell RNA sequencing, hdWGCNA co-expression network analysis, and two independent machine learning algorithms (LASSO and SVM-RFE). These biomarkers are derived from type II pneumocyte biology, functionally connected to the immune microenvironment, and validated in cancer cell lines.
The three biomarkers have distinct clinical implications: HN1 and OCIAD2 mark immune-cold, aggressively growing tumors (high proliferation pathway activity, low immune infiltration), while SFTA2 marks immune-active, potentially better-prognosis disease. This dichotomy could guide immunotherapy patient selection and prognostic risk stratification beyond current methods.
Important limitations must be acknowledged. The single-cell data came from only four patients, and the bulk transcriptomic dataset was a single public cohort without independent external validation. No survival outcome analyses were performed, and causal functional validation was only completed for HN1 - OCIAD2 and SFTA2 lack systematic gain-of-function/loss-of-function experiments. The CIBERSORT deconvolution has methodological uncertainties.
Future research priorities include independent multicenter cohort validation with clinical outcome data, in vivo mouse model experiments for all three biomarkers (including conditional type II pneumocyte lineage models), spatial transcriptomics or multiplex immunohistochemistry to validate spatial colocalization, and construction of clinical predictive models with calibration curve and decision curve analysis to define threshold values for clinical application.