GSE136246, GSE68465, GSE3141, GSE31210, GSE37745, GSE50081, and GSE78220 datasets were accessed from the Gene Expression Omnibus (GEO) database (www.ncbi.nlm.nih.gov/geo (accessed on 1 January 2025)). Bulk data for LUAD were acquired from The Cancer Genome Atlas (TCGA) database (https://portal.gdc.cancer.gov (accessed on 1 January 2025)). Additionally, transcriptome data for LUAD patients receiving immunotherapy were obtained from the IMvigor210 cohort. Samples that were duplicated or lacked complete survival and clinicopathological information were excluded from further analysis.

We utilized the Seurat R package to integrate the single-cell RNA sequencing data, applying the following filter criteria: nCount RNA ≥ 1000; nFeature RNA ≥ 200 & nFeature RNA ≤10,000; percent.mt (mitochondrial count) ≤20; percent.ribo (ribosomal count) ≤20. To normalize the data, we employed the LogNormalize method. Marker genes and hypervariable genes for each cell cluster were identified using the FindAllMarkers and FindVariableFeatures functions, respectively. Cell types were delineated according to classical molecule markers, and the UMAP method was utilized for visualization. Using CellChat R package’s programming tools, we measured both the quantity of cell signaling pairs and how actively different cell types interact through these molecular messages.

To uncover potential neutrophil subtypes arising from KRAS/TP53 mutations, we aimed to delineate their developmental pathways and gene expression dynamics through single-cell trajectory analysis. Concurrently, we sought to chart the evolution of epithelial cells within the LUAD microenvironment characterized by these mutations to identify early therapeutic targets. The Monocle2 R package was applied to analyze the evolutionary trajectory of neutrophils and epithelial cells in LUAD patients. Pseudo-time analysis was applied to explore the dynamics of gene expression patterns during the evolution of these cell types.

High-dimensional weighted gene co-expression network analysis (hdWGCNA) enhances the dissection of scRNA-seq data, specializing in the management of high-dimensional datasets. Superseding traditional WGCNA, it effectively navigates vast gene expression data, yielding in-depth insights into cellular heterogeneity and enriching our understanding of gene expression through rigorous gene network construction. It excels at detecting subtle expression differences in single-cell data, demonstrating superior sensitivity and precision. Thus, we looked into the key features of genetic pattern of neutrophils using the hdWGCNA R package, which facilitated the identification of genes highly associated with neutrophils.

Consensus clustering of the TCGA-LUAD cohort based on genes identified through hdWGCNA was analyzed by the ConsensusClusterPlus R package. The vaviable k represented cluster numbers, and the optimum k value was identified by cumulative distribution function (CDF). limma R package, a functional programmed tool for recognization of differentially expressed genes (DEGs), were applied between different clusters. Genes with p < 0.05 and |Log₂FoldChange| > 0.5 were labeled DEGs. All identified DEGs underwent Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) functional enrichment analyses using the clusterProfiler and org.Hs.eg.db R packages.

Based on the TCGA-LUAD cohort, we quantified and compared immune infiltration levels between clusters using single-sample Gene Set Enrichment Analysis (ssGSEA). The StromalScore, ImmuneScore, and ESTIMATEScore were quantified and compared across clusters using the ESTIMATE algorithm. Additionally, the differential expression patterns of immune checkpoints were analyzed. The potential for tumor immune escape was quantified and compared among clusters using the Tumor Immune Dysfunction and Exclusion (TIDE) algorithm.

Least Absolute Shrinkage and Selection Operator (LASSO) regression analysis was utilized to construct the prognostic signature based on DEGs identified in the TCGA-LUAD cohort. A risk score for each patient was calculated, and patients were subsequently divided into high-risk and low-risk groups according to the median risk score. Kaplan-Meier analysis was performed using the survminer R package to compare overall survival (OS) between the high-risk and low-risk groups. Additionally, time-dependent receiver operating characteristic (ROC) analysis was conducted using the timeROC R package, with the area under the curve (AUC) employed to evaluate predictive efficacy. Internal validation cohorts included the GSE37745, GSE3141, GSE50081, GSE68465, and GSE31210 datasets. The efficacy of the prognostic signature in predicting responsiveness to anti-Programmed cell death 1 (PD-1) therapy was externally verified in GSE78220 and IMvigor210 datasets. Atezolizumab was administered in the IMvigor210 dataset, while pembrolizumab and nivolumab were utilized in the GSE78220 dataset.

Risk score across several clinical characteristics were compared. Univariate and multivariate Cox regression analyses were conducted to find prognostic effector. Furthermore, a predictive nomogram was developed using the ggpubr R package. Calibration curves, decision curve analysis, and AUC were employed to systematically assess the performance of the nomogram.

Scoring algorithms (AUCell, UCell, sing score, ssGSEA, and Add) were utilized to quantify the prognostic gene set. The scale function in R was employed to normalize the calculated gene set scores (from zero to one). This normalization involved centering the mean of each feature to zero by subtracting the mean value from every data point, followed by dividing the centered data by its standard deviation to scale the standard deviation of each feature to unity. This technique ensured the comparability of gene set scores derived from different algorithms during subsequent comparative analyses. The sum of corresponding scores was calculated, referred to as the “Scoring”. HighRisk and LowRisk subtypes of epithelial cells were classified according to the median score. Besides, Monocle R package was applied to fit evolutionary trajectory of lung epithelial cells and to indicate expression patterns of the five prognostic genes in pseudo-time.

Two human NSCLC cell lines, A549 (Procell, CL-0016, Wuhan, China) and H1299 (Procell, CL-0165), were selected for further experiments. Both cell lines were cultured in Roswell Park Memorial Institute-1640 (RPMI-1640) medium (Procell, PM150110B) supplemented with 10% fetal bovine serum (FBS) and 1% antibiotic solution. To ensure the integrity of our experimental data, all cell lines used in this study were rigorously tested for mycoplasma contamination. Mycoplasma tests were conducted using the PCR-based Mycoplasma Detection Kit (MedChemExpress, HY-K0552, Shanghai, China) prior to the initiation of experiments and periodically thereafter, following the manufacturer’s instructions. The absence of mycoplasma contamination in all cultured cells was confirmed before conducting any experiments.

Small interfering RNA (siRNA) were transfected to cells. The siRNA sequences targeting Ras homolog family member V (RHOV) are as follows: siRNA1: Sense (SS) sequence: GGACGAUGUCAACGUACUAAU, Antisense (AS) sequence: UAGUACGUUGACAUCGUCCCU; siRNA2: SS sequence: GGCUGGAGAAGAAACUGAAUG, AS sequence: UUCAGUUUCUUCUCCAGCCGG. The sequence of the negative control siRNA (si-NC) is as follows: SS sequence: UUCUCCGAACG, AS sequence: UGUCACGUTT. Transfection was performed using Lipofectamine 3000 (Invitrogen, L3000015, Shanghai, China) according to the manufacturer’s instructions.

Cells were subjected to lysis using a lysis buffer (Beyotime, R0046, Shanghai, China), and protein concentration was assessed utilizing the bicinchoninic acid (BCA) assay. The supernatant containing extracted proteins was separated by denaturing 10% sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE) and subsequently transferred to polyvinylidene fluoride (PVDF) membranes (Beyotime, P0965). Following a blocking step with 5% bovine serum albumin (BSA) (Beyotime, ST2249) for 2 h at room temperature, the membrane was incubated overnight at 4°C with the primary antibodies (RHOV: Abcam, ab54558, 1:200; β-actin: Abcam, ab8227, 1:1000, Shanghai, China). After this incubation, membranes were washed three times and then treated with goat anti-rabbit secondary antibody (Abcam, ab205718, 1:10,000). The images were captured using an Odyssey infrared imaging scanner (LI-COR, Tucson, AZ, USA).

Cell was plated in a 96-well plate with 100 μL of culture medium at a density of 5 × 10⁴ cells/well. Cell proliferative capability was evaluated at 0, 1, 2, 3, 4, and 5 days post-transfection. 10 μL of CCK8 solution (Beyotime, C0042) was added to each well, and the plate was incubated for 2 h at 37°C. Absorbance at 450 nm was then measured using a microplate reader (Molecular Devices, SpectraMax Plus 384, Shanghai, China).

Cells were grown in a 6-well plate (Beyotime, FCP060) until reaching 100% confluence. A linear scratch was created in the cell monolayer using a 200 μL pipette tip. The wells were then rinsed with serum-free medium to eliminate any detached cells and debris, followed by the addition of fresh medium. Images of the scratched region were captured using an inverted microscope (×100) (Fisher Scientific, LMI3PH2, Waltham, MA, USA) at 0, 12, and 24 h.

The transwell chamber (Corning, CLS3412, Shanghai, China) was coated with Matrigel (Corning, 356234). Cells were seeded in the upper chambers with FBS-free medium, while the lower chamber contained medium supplemented with 10% FBS. After a 24-h incubation at 37°C, non-invading cells were removed with a cotton swab. Cells that migrated to the bottom of the chamber were fixed with 4% methanol at 37°C for 10 min and stained with 0.1% crystal violet for 15 min. An inverted microscope (×200) (Fisher Scientific, LMI3PH2) was employed to count the invading cells in three randomly selected fields.

All experiments were conducted with a minimum of three repetitions. Statistical analyses of in vitro experimental data and bioinformatic data were analyzed using R 4.0.3. Survival curves were compared using Cox regression analysis. The Wilcoxon rank sum test was employed to evaluate differences in expression levels between groups. Correlation analysis was conducted using the Pearson correlation coefficient, where |r| > 0.1 was considered relevant and p < 0.05 was considered statistically significant. Throughout the study, * indicates p < 0.05, ** indicates p < 0.01, *** indicates p < 0.001, and **** indicates p < 0.0001.

We obtained 17 distinct clusters of cells through UMAP analysis (Fig. 1A). These 17 clusters of cells were subsequently categorized into nine cell types based on classical gene markers (Fig. 1B,C). Notably, we observed a pronounced abundance of neutrophils in the TME of LUAD with KRAS/TP53 mutations (MUT LUAD) (Fig. 1D).

We identified eight neutrophil subpopulations (Fig. 2A). Compared with the wild-type (WT) group, subpopulations 3, 6, and 7 exhibited a higher proportion in the KRAS/TP53 MUT group (Fig. 2B). Thereafter, we defined subpopulations 3, 6, and 7 as IS MUT neutrophils, while the remaining subpopulations were designated as WT neutrophils. The term “IS MUT” specifically refers to the collective of these three subpopulations (3, 6, and 7) rather than a broader characteristic driven by KRAS/TP53 mutations. Furthermore, the communication among B cells, neutrophils, endothelial cells, macrophages, and epithelial cells was notably robust during carcinogenesis (Fig. 2C,D). Analysis of intercellular signaling revealed that the interaction strength of IS MUT neutrophils was greater than that of WT neutrophils, particularly for signals such as Oncostatin M (OSM), Calcitonin receptor (CALCR), and Interleukin 1 (IL-1) (Fig. 2E). Additionally, three states of IS MUT neutrophils were delineated as pseudo-time progressed (Fig. 3A–C). A heat map illustrated the gene expression patterns during the evolution of neutrophils as pseudo-time changed (Fig. 3D).

A hierarchical hdWGCNA dendrogram was constructed, revealing three gene modules (Fig. 4A). At an optimum soft threshold of 4, the co-expression network was generated (Fig. 4B). Correlations within three identified gene modules were analyzed (Fig. 4C). Hub genes from the three gene modules were also determined (Fig. 4D), with the distribution of these modules observed in neutrophils (Fig. 4E). Notably, the blue and brown modules predominantly encompassed neutrophil subpopulations 3, 6, and 7 (Figs. 2A and 4E). Hence, our focus was directed towards the blue and brown gene modules.

We identified 70 candidate genes from 773 genes in the blue and brown modules via univariate analysis (Fig. 5A). Two clusters, C1 and C2, were divided in the TCGA-LUAD cohort (Fig. 5B,D). Patients in cluster C1 exhibited prolonged survival probabilities compared to counterparts in cluster C2 (Fig. 5E). More abundant infiltration of immune cells were observed in the TME of C1 (Supplementary Fig. S1A). Utilizing the ESTIMATE algorithm, we found that the stromal and immune components of C1 were significantly higher than C2 (Supplementary Fig. S1B). Furthermore, compared to C2, the expression levels of immune checkpoints were significantly elevated in C1 (Supplementary Fig. S1C). TIDE, exclusion, and MDSC scores of C1 were significantly lower than those of C2 (Supplementary Fig. S1D,G,I), indicating that C2 exhibited a stronger immune escape capability. Additionally, dysfunction and IFNG scores of C1 were significantly higher than those of C2, potentially related to T cell exhaustion (Supplementary Fig. S1E,F). In summary, the heterogeneity of the immune microenvironment may serve as a driving factor contributing to the poorer prognosis observed in C2.

494 DEGs were found between cluster C1 and C2 (Supplementary Fig. S2A). The DEGs were significantly enriched in pathways related to immune cell regulation and cell adhesion. Additionally, the DEGs primarily consisted of cell membrane components, such as endoplasmic reticulum and Major histocompatibility complex (MHC) proteins. They were also involved in molecular functions, including chemokine activation and immune receptor-ligand binding (Supplementary Fig. S2B,C). Results from functional enrichment analyses indicated that the identified DEGs related to immune responses and the invasion or metastasis of tumor cells.

From univariate Cox regression analysis, we identified 140 protective genes and 73 risky genes (Fig. 6A). A five-gene prognostic signature was successfully formulated through LASSO regression analysis, with the optimal λ identified as 0.0653 via ten-fold cross-validation (Fig. 6B,D). Risk score = 0.171 ∗ ANLN + 0.092 ∗ FAM83A + 0.094 ∗ RHOV + 0.059 ∗ KRT6A − 0.169 ∗ MS4A1. Worse prognosis of patients with high risk score were observed compared to counterparts with low risk score (p < 0.0001, Fig. 6E). The AUCs for 1-, 3-, and 5-year survival were 0.73, 0.70, and 0.66, respectively (Fig. 6F). Survival probabilities were well stratified across all five internal validation cohorts (p < 0.0047, p = 0.0001, p = 0.018, p = 0.0066, and p = 0.0001, Supplementary Fig. S3A–E). The AUCs for 1-, 3-, and 5-year survival further validated the robustness of the prognostic signature across all five internal validation cohorts (Supplementary Fig. S3A–E).

We observed significant differences in risk scores among patients characterized by various clinicopathological features (Supplementary Fig. S4A–F). Both N stage (p = 0.013, HR = 1.686 [1.116–2.548]) and risk score (p < 0.001, HR = 2.57 [1.822–3.624]) were identified as independent prognostic factors for LUAD through univariate and multivariate Cox regression analyses (Fig. 7A,B). Risk score and other clinicopathological features were further integrated to construct a predictive nomogram for predicting survival outcomes (Fig. 7C). The calibration curve demonstrated that the predicted survival probabilities at 1-, 3-, and 5-year intervals closely aligned with the ideal line (Fig. 7D). Furthermore, the decision curve analysis indicated that the nomogram exhibited optimal clinical benefit across varying threshold probabilities (Fig. 7E). ROC analyses further confirmed the predictive efficacy of the nomogram (Fig. 7F).

Functional characterization revealed that the five prognostic genes were strongly correlated with immune response pathways. Specifically, Membrane spanning 4 domains A1 (MS4A1) and RHOV showed significant positive correlations with immune-related pathways (Supplementary Fig. S5A). The functional annotation was quantified for each TCGA-LUAD sample and represented as a distribution heat map (Supplementary Fig. S5B).

Analysis revealed that in the TME of high-risk group, stromal and immune components were more lower, while tumor component was more higher (Fig. 8A–D). The complex heatmap illustrated the differential infiltration of multiple immune cells across risk groups (Fig. 8E). Nearly all immune cells exhibited differential enrichment in the TME of the risk groups. Additionally, a majority of immune checkpoints were significantly elevated in the low-risk group (Fig. 8F).

We divided patients from the IMvigor210 and GSE78220 cohorts into high- and low-risk groups. Patients in high-risk group exhibited worse prognosis (IMvigor210: p < 0.0001; GSE78220: p = 0.036, Fig. 9A,F). In the IMvigor210 cohort, the overall risk scores of patients with complete response (CR) or partial response (PR) were significantly lower compared to those with progressive disease (PD) and stable disease (SD) (Fig. 9B). The percentage of patients with PD/SD was also higher in the high-risk group compared to the low-risk group (85% vs. 70%, Fig. 9C). Besides, the predictive risk model was more sensitive in forecasting the prognosis of patients with early-stage disease (Fig. 9D,E). In the GSE78220 cohort, no significant difference in risk scores was observed between patients with PD and those with CR/PR (Fig. 9G). The percentage of patients with PD was significantly higher in the high-risk group compared to the low-risk group (57% vs. 38%, Fig. 9H). These results indicate that the prognostic signature effectively predicts the immunotherapy response in patients with LUAD.

To further investigate the five prognostic genes, we scored each gene in MUT LUAD and WT LUAD using five scoring algorithms. The total score (Scoring) for each cell was calculated as the sum of these scoring algorithms (Supplementary Fig. S6A). The results revealed significant differences in Scoring among several cell types between MUT LUAD and WT LUAD (Supplementary Fig. S6B). Notably, the Scoring of epithelial cells was significantly higher in MUT LUAD compared to WT LUAD. Given that LUAD derives from the bronchial mucosa, we focused on epithelial cells for subsequent analyses. HighRisk and LowRisk epithelial cells were classified and distinct intercellular communication pattern was observed between these two subtypes (Fig. 10A). Stronger signaling pattern was determined in LowRisk epithelial cells compared to HighRisk ones (Fig. 10B). Additionally, as the trajectory evolved, RHOV expression was gradually upregulated, while Keratin 6A (KRT6A) expression was gradually downregulated (Supplementary Fig. S6C, Fig. 10C).

Given the progressively elevated expression of RHOV during the evolution trajectory of lung epithelial cells, subsequent functional experiments focused on RHOV. Western blot assay confirmed that both siRNA1 and siRNA2 effectively reduced RHOV protein expression (Fig. 11A). CCK8 assay demonstrated that silencing RHOV impaired the proliferative capabilities of A549 and H1299 cells (Fig. 11B). Consistent results from transwell and wound healing assays indicated decreased capabilities of migration and invasion in A549 and H1299 cells upon downregulation of RHOV (Fig. 11C–F). These findings provided compelling evidence for RHOV as a pivotal gene during LUAD progression.