Development and validation of an oxidative phosphorylation based prognostic model revealing tumor progression and immune microenvironment in lung adenocarcinoma

Abstract

Background

Lung adenocarcinoma (LUAD) remains a leading cause of cancer mortality worldwide. Although oxidative phosphorylation (OXPHOS) has been linked to tumor progression, its prognostic significance and underlying mechanisms in LUAD remain unclear.

Methods

RNA-seq data from TCGA were used as the discovery cohort, and multiple GEO datasets were used for independent validation. OXPHOS-related genes were obtained from MitoCarta3.0. Prognostic biomarkers were selected by LASSO-Cox regression. Immune cell infiltration was estimated using CIBERSORT and ssGSEA. Single-cell RNA-seq data were analyzed to examine biomarker expression, pathway enrichment, and cell–cell communication.

Results

Thirteen OXPHOS-associated genes were used to construct a prognostic model with robust predictive performance. Model performance was validated across independent cohorts and demonstrated reliable prognostic value. Higher risk scores correlated with poorer overall survival and reduced immune infiltration. UQCC3⁺ and RAB5IF⁺ epithelial cells were enriched in E2F target and G2M checkpoint pathways and showed enhanced interactions via TGF-β and Galectin signaling.

Conclusions

We present an OXPHOS-based prognostic model for LUAD. OXPHOS reprogramming, particularly involving UQCC3 and RAB5IF in epithelial cells, may promote malignant proliferation and immunosuppression. These findings provide novel prognostic biomarkers and potential therapeutic targets, which may support personalized treatment strategies and guide future translational studies.

Supplementary Information

The online version contains supplementary material available at 10.1007/s12672-026-04478-3.

Introduction

Cancer remains a major global health concern. According to estimates from the International Agency for Research on Cancer (IARC), lung cancer was the most frequently diagnosed cancer and the leading cause of cancer-related death in 2022, with approximately 2.5 million new cases and over 1.8 million deaths [1]. Lung adenocarcinoma (LUAD), the predominant histological subtype, accounts for approximately 45% of all lung cancer cases [2]. Despite advances in LUAD treatment, major challenges persist, including limited surgical eligibility, drug resistance, treatment intolerance, and low response rates to therapy, resulting in a five-year overall survival (OS) rate that remains below 28% [3–5]. Traditional indicators such as pathological classification and tumor stage have proven insufficient for accurately distinguishing patients at intermediate risk, underscoring the need for more reliable prognostic tools.

Given these limitations, there is growing interest in molecular biomarkers, particularly those linked to metabolic pathways like oxidative phosphorylation (OXPHOS) [6–8]. OXPHOS is a crucial metabolic pathway that cells utilize to generate ATP [9]. In tumors, the Warburg effect is a metabolic hallmark, characterized by a preference for glycolysis even in the presence of ample oxygen. However, OXPHOS remains indispensable across various cancer types. This is due to the remarkable metabolic flexibility of cancer cells, which allows them to dynamically switch between glycolysis and OXPHOS in response to fluctuations in oxygen and nutrient availability [10]. This flexibility ensures a steady energy supply, thereby enhancing cancer cells’ survival in diverse tumor microenvironments. It also contributes to resistance to multiple therapies and plays a pivotal role in cancer initiation, invasion, and the development of drug resistance [10, 11].

Recent research has revealed associations between OXPHOS status and clinical outcomes across multiple cancer types, including hepatocellular carcinoma, oral squamous cell carcinoma, ovarian cancer and LUAD, prompting exploration of prognostic models based on OXPHOS genes [12–14]. In LUAD, Xu et al. developed a prognostic model based on seven OXPHOS-related genes that can stratify patients into high- and low-risk groups [8]. This model demonstrates promising prognostic potential. However, this study has not elucidated the underlying molecular mechanism and validated the markers in other datasets such as single-cell RNA sequencing data, limiting its ability to better elucidate the heterogeneity of the tumor microenvironment. Additionally, the altered expression of OXPHOS-related genes, such as COA6, HERC5, PDHA1, and COL17A1, has emerged as a potential set of prognostic biomarkers for LUAD [8, 14–17]. Yet, applying these biomarkers in clinical practice remains challenging and requires further investigation into their molecular roles and clinical applicability.

The aim of this study was to investigate the relationship between LUAD and OXPHOS, and to develop a prognostic model for LUAD. By analyzing publicly available RNA-seq data, we identified and validated OXPHOS-related prognostic biomarkers and constructed a prognostic model incorporating these biomarkers along with survival data. Additionally, we investigated the underlying mechanisms of the identified biomarkers in LUAD at a single-cell resolution. Our findings not only provide a prognostic model for the clinical management of LUAD but also offer novel insights into the complex relationship between LUAD and OXPHOS (Fig. 1).

Fig. 1.

Flow chart demonstrates the present study

Materials and methods

Data collection

The set of 169 OXPHOS-related genes (Supplementary Data) were obtained from MitoCarta3.0 (https://www.broadinstitute.org/files/shared/metabolism/mitocarta/human.mitocarta3.0.html). The discovery set consisted of bulk transcriptome data from 568 samples in the TCGA-LUAD cohort (https://portal.gdc.cancer.gov/), and the two validation sets were obtained from GEO: 226 samples from GSE31210 and 398 samples from GSE72094 (https://www.ncbi.nlm.nih.gov/gds/). The single-cell RNA sequencing data, consisting of 25 healthy controls, 23 primary tumor, and 22 metastatic tumor samples, were obtained from GEO (GSE131907 and GSE123902).

Construction and evaluation of OXPHOS-related prognostic biomarkers

Based on the mRNA expression profiles from TCGA-LUAD, along with overall survival time, univariate Cox regression was initially performed to identify 43 OXPHOS-related prognostic biomarkers at p < 0.05. Subsequently, the least absolute shrinkage and selection operator (LASSO) Cox regression was performed to select biomarkers for the risk score model. The risk score was calculated using the following formula:

where coef_i represents the regression coefficient and exp_i denotes the expression level of gene i.

The “survminer” R package (version 0.5.0) was applied to plot Kaplan-Meier survival curves and to determine the optimal cutoff value by surv_cutpoint function. This cutoff value was then applied to stratify the samples into high-risk and low-risk groups. To assess the model’s predictive performance, the “timeROC” package (version 0.4) was utilized to plot the ROC curves for the risk scores at 1, 3, and 5 years, and the areas under the curve (AUC) for each time point were calculated. Finally, the prognostic risk scores were evaluated in two validation sets, and ROC curves were generated.

Construction of a nomogram prognostic model

To quantitatively predict the prognosis of LUAD patients, the risk score and clinical factors, including age, gender, and tumor stage, were integrated to develop a prognostic model. Univariate and multivariate Cox regression was performed to identify potential prognostic factors. “RMS” package (version 6.9-0) was then used to construct a nomogram incorporating variables with p-value < 0.05. Subsequently, the nomogram was applied to predict the 1-, 3-, and 5-year survival probabilities of LUAD patients, and calibration curves were generated to evaluate its accuracy and reliability.

Immune infiltration and immunotherapy-related scoring

To assess the differences in immune cell infiltration between the high- and low-risk groups, enrichment scores for various immune cell types were calculated in tumor samples. Gene sets for 28 immune cells were obtained from previous studies and the single-sample Gene Set Enrichment Analysis (ssGSEA) from the “GSVA” R package (version 1.52.3) were used to calculate immune infiltration scores [18]. Additionally, another literature-reported method was also used to assess immune cell infiltration in tumor tissue [19]. The LM22 signature matrix (LM22.txt), which defines the expression profiles of 22 immune cell types, was downloaded from the CIBERSORTx website (https://cibersortx.stanford.edu/). CIBERSORT was then applied to estimate the relative proportions of 22 immune cell types within each sample.

To further assess the immunological activity of tumors, the Hot score, based on a previously reported 27-gene signature [20], was also calculated using ssGSEA implemented in GSVA. Higher Hot scores indicate a more immunologically active (“hot”) tumor microenvironment. Finally, to evaluate the potential response to immunotherapy, Tumor Immune Dysfunction and Exclusion (TIDE) scores and Immunophenoscores (IPS) were computed for all samples [18, 21], allowing comparison of predicted immunotherapy responsiveness between high- and low-risk groups.

Processing of single-cell transcriptomic data

Single-cell RNA sequencing data were processed using the Seurat R package (version 4.4.0). After filtering out cells with fewer than 200 expressed genes or with mitochondrial gene proportions greater than 25%, a total of 221,792 cells were retained. Data normalization was performed using the logNormalize method in the NormalizeData function. The top 2,000 highly variable genes were then identified. Batch effect correction across different samples was explicitly performed using the Harmony package (version 1.2.1), ensuring that variations due to technical differences were minimized while preserving biological heterogeneity. Principal component analysis (PCA) was conducted, and the top 22 principal components were used for UMAP dimensionality reduction and construction of a shared-nearest neighbor graph. Cell clustering was performed using the FindClusters function with a resolution of 0.2. Finally, cell clusters were annotated based on marker genes from the CellMarker 2.0 database (http://bio-bigdata.hrbmu.edu.cn/CellMarker/).

Calculation of CNV score in epithelial cells

CNV profiles of epithelial cells were inferred using the inferCNV package (version 1.18.1), with 100 B cells and 100 mast cells serving as the reference cell types. Normal diploid regions were assigned a score of 0. A single-copy gain or loss was assigned a score of 1, whereas loss of two copies or gain of more than one copy was assigned a score of 2. The scores for all CNV events within each epithelial cell were summed to generate a CNV score, which served as a quantitative measure of CNV burden.

Gene set enrichment analysis for epithelial cells

Epithelial cells were divided into two groups based on whether the gene of interest was expressed. The difference in gene expression was calculated and converted to log2 fold change (log2FC) using the FindMarkers function. Subsequently, the enrichment of hallmark gene sets from the MsigDB database (https://www.gsea-msigdb.org/gsea/msigdb) was assessed using the “clusterProlifer” package (version 4.12.0).

Cell communication analysis

Cell communication analysis was conducted using the CellChat package (version 1.6.1). Initially, a CellChat object was initialized using the createCellChat function, with CellChatDB.human as the ligand-receptor database. The identifyOverExpressedInteractions function was used to identify overexpressed ligand-receptor interactions, and the projectData function was used to map the expression values of ligand-receptor pairs onto the PPI network. The computeCommunProb function was used to infer the probability of cell interactions, while the computeCommunProbPathway function was applied to infer the cell communication network at the signaling pathway level. Finally, the interaction network was visualized using the netVisual_circle function, and ligand-receptor interactions were displayed as bubble plots using the netVisual_bubble function.

Patient sample collection

Tumor tissues and matched adjacent non-tumorous samples were obtained from lung adenocarcinoma patients undergoing surgical resection, with written informed consent obtained from all participants. Samples were immediately snap-frozen in liquid nitrogen and subsequently stored at − 80 °C until further analysis. The study was approved by the Ethics Committee of Shenzhen Bao’an District People’s Hospital (Approval No. 2021041318093719) and conducted in accordance with the Declaration of Helsinki.

Quantitative real-time PCR

Total RNA was extracted from 12 patient samples using TRIzol reagent (TaKaRa) following the manufacturer’s protocol. RNA concentration and integrity were assessed with a NanoDrop 2000 spectrophotometer and 1.5% agarose gel electrophoresis. cDNA was synthesized using the PrimeScript™ RT reagent Kit with gDNA Eraser (Takara). Quantitative Real-Time PCR (qRT-PCR) was performed on a Bio-Rad CFX96 system with TB Green Premix Ex Taq™ II (Takara). The amplification program consisted of an initial denaturation at 95 °C for 30 s, followed by 45 cycles of 95 °C for 5 s and 60 °C for 30 s, and concluded with a melt curve analysis to confirm specificity. Relative gene expression was calculated using the 2^⁻ΔΔCt method, with all reactions run in triplicate. The primers used were: RAB5IF (F: 5′-CTCCGTCTGGAGTAAGGTGC-3′, R: 5′-TGGATAGGGAGCACTTGGGA-3′), UQCC3 (F: 5′-GATCTCAGTCGCAATGCTGG-3′, R: 5′-CAATAGCTGCTGGGTCCTGG-3′), and β-actin (F: 5′-GCGGGAAATCGTGCGTGACATT-3′, R: 5′-GATGGAGTTGAAGGTAGTTTCGTG-3′).

Statistical analysis

The Wilcoxon rank-sum test was used to compare continuous variables between two groups. The log-rank test was applied to compare the survival outcomes between the two groups. Spearman’s rank correlation was applied to calculate the correlation between two continuous variables. All computations and visualizations were performed using R (version 4.4.0). A p-value of less than 0.05 was considered statistically significant.

Results

Activated OXPHOS was associated with LUAD progression and poor prognosis

To investigate the relationship between OXPHOS and tumor progression, the OXPHOS activity score for each sample from the TCGA-LUAD and two GEO datasets was assessed using ssGSEA. In the TCGA-LUAD cohort, OXPHOS activity was significantly higher in tumor tissues than in adjacent normal tissues (Fig. 2A). Although the difference across tumor stages was not statistically significant, the median of OXPHOS activity scores increased as tumor stages advanced (Fig. 2B). Patients were then divided into high-OXPHOS and low-OXPHOS groups by the optimal cutoff value. Survival analysis in the three cohorts revealed that high-OXPHOS patients exhibited poorer prognosis compared to low-OXPHOS patients (Fig. 2C), suggesting that the expression levels of OXPHOS-related genes may serve as prognostic biomarkers for LUAD.

Fig. 2.

The relationship between OXPHOS activity and tumor progression and prognosis for LUAD. A OXPHOS activity scores between normal and tumor samples in TCGA-LUAD cohort. B OXPHOS activity scores across different stages of tumor samples in TCGA-LUAD cohort. The Wilcoxon rank-sum test was used to compare OXPHOS activity scores between two groups. ****: P < 0.0001; ns: not significant. C Survival analysis based on OXPHOS activity scores in TCGA-LUAD and two GEO (GSE31210 and GSE72094) cohorts. Patients were divided into high- and low-OXPHOS score groups by the optimal cutoff value. The log-rank test was applied to compare the survival outcomes between the two groups

Identification and validation of the OXPHOS-related prognostic model

Based on the TCGA-LUAD cohort, a total of 43 OXPHOS-related genes were identified as being associated with overall survival using univariate Cox regression analysis (P < 0.05) (Supplementary Data). Among them, 13 genes—ATP5F1D, COX6B2, UQCRB, RAB5IF, CYCS, COA6, NDUFA10, TIMMDC1, CMC2, DMAC2, UQCC3, PET100, and FMC1—were selected as biomarkers for constructing a prognostic risk model using LASSO Cox regression in the discovery set (Fig. 3A-C). The risk score for each patient was calculated with the formula:

Fig. 3.

Construction and evaluation of the OXPHOS-related prognostic biomarkers for LUAD. A-C LASSO cox regression identified 13 prognostic biomarkers for LUAD. A, B The curve of the partial likelihood deviation versus log (λ) and the regression coefficient versus log (λ) for OS. C LASSO cox regression coefficients for OXPHOS-related prognostic biomarkers. D Survival curves for risk scores for discovery set (TCGA-LUAD cohort) and validation sets (GSE31210 and GSE72094 cohorts). E ROC curves for 1-, 3-, and 5-year prognostic evaluation of risk scores in discovery and validation sets

To further assess the prognostic value of the risk score, patients in the discovery set (TCGA-LUAD) and two independent validation sets (GSE31210 and GSE72094) were divided into high- and low-risk score groups using the optimal cutoff value in survival analysis. As shown in Fig. 3D, survival analysis across the discovery set and two validation sets consistently demonstrated that patients with high-risk scores had significantly worse OS. In the discovery set, the hazard ratio (HR) for high-risk scores was 1.45 (95% CI 1.06–1.99, p < 0.0001). In the two validation sets, the survival analysis results were consistent with those observed in the discovery set. Specifically, for GSE31210 cohort, the HR was 3.07 (95% CI 1.44–6.55, p = 0.0052); for GSE72094 cohort, the HR was 1.85 (95% CI 1.13–3.04, p = 0.00017). The AUC scores for 1-, 3-, and 5-year survival were as follows: 0.73, 0.70, and 0.66 for the discovery set; 0.69, 0.62, and 0.53 for the GSE31210 cohort; 0.61, 0.64, and 0.51 for the GSE72094 cohort (Fig. 3E). These results suggest that the risk score may serve as a robust predictor for LUAD prognosist, particularly for short-term survival outcomes such as 1- and 3-year survival.

Construction of personalized prognosis prediction model

To develop a more personalized model for LUAD, potential prognostic signatures, including the risk score and clinical information, were integrated for further analysis. Univariate Cox regression analysis showed that the risk score, clinical stage, and pathological T and N stages were significant prognostic factors (p < 0.05). Specifically, the hazard ratios (HR) for these factors were as follows: risk score HR was 4.38 (95% CI 3.01–6.36, p < 10^− 13), stage HR was 1.68 (95% CI 1.46–1.93, p < 10^− 12), pathological T stage HR was 1.52 (95% CI 1.28–1.80, p < 10^− 5), and pathological N stage HR was 1.41 (95% CI 1.23–1.60, p < 10^− 6) (Fig. 4A). Multivariate Cox regression analysis further revealed that the risk score and clinical stage remained independent prognostic factors after adjusting for other clinical variables (Fig. 4B). The adjusted hazard ratios (HR) were as follows: risk score HR was 3.70 (95% CI 2.48–5.53, p < 10^− 9) and stage HR was 1.42 (95% CI 1.16–1.74, p = 0.00056).

Fig. 4.

Construction and evaluation of a nomogram in TCGA-LUAD cohort. A Forest plot of univariate Cox regression analysis for risk score and clinical factors. B Forest plot of multivariate Cox regression analysis for risk score and clinical factors. C A nomogram incorporating independent prognostic factors. D Calibration plots of survival predictions for 1-year, 3-year, and 5-year

By integrating the independent prognostic factors of the risk score and tumor stage, a nomogram was constructed for individualized prediction of overall survival (Fig. 4C). Each variable was assigned a weighted score inversely proportional to its prognostic impact. The calibration curves of the nomogram for 1-, 3-, and 5-year survival probability exhibited strong concordance between the predicted and actual survival outcomes (Fig. 4D).

Evaluation of immune infiltration between high- and low-risk groups

Immune cell infiltration levels in high- and low-risk groups from TCGA-LUAD cohort were assessed by ssGSEA and CIBERSORT, respectively. Most immune cell types (14/17 from ssGSEA and 4/6 from CIBERSORT) had significantly higher enrichment scores in the low-risk group compared to the high-risk group (Fig. 5A). Among them, activated B cell, eosinophil, immature B cell, mast cell, and natural killer cell from ssGSEA, and resting mast cells from CIBERSORT, displayed the most significant differences (p < 0.0001). This suggests a stronger immune activity potential in the low-risk group, which was consistent with the longer OS of these patients (Fig. 3D). However, the proportion of T cells, including activated CD4 + T cell and type 2 T helper cell, was significantly higher in the high-risk group (Fig. 5A). CIBERSORT analysis revealed similar results, with the high-risk group showing notably higher immune score of activated CD4 + memory T cells (p < 0.001) (Fig. 5B). The prognostic role of these T cells in the high-risk group warrants further investigation. Taken together, these results suggest that reprogrammed OXPHOS may affect prognosis by modulating immune cell infiltration within tumor tissues, prompting us to explore whether this would lead to differences in intercellular communication among cells with different OXPHOS biomarkers expression levels.

Fig. 5.

Immune landscape differences between high- and low-risk groups. A Comparison of 28 immune cell infiltration scores assessed by ssGSEA between the high- and low-risk groups. B Comparison of 22 immune cell infiltration scores assessed by CIBERSORT between the two groups. C Comparison of Immunophenoscore (IPS) between high- and low-risk groups. D Comparison of hot-/cold-tumor scores between the two groups. E Comparison of TIDE scores between high- and low-risk groups. *, p < 0.05; **, p < 0.01; ***, p < 0.001; ****, p < 0.0001; ns, not significant

To further evaluate immunotherapy-related characteristics, we analyzed the IPS from the TCIA database and a hot-tumor gene-set–based score. Both IPS and hot-tumor scores were significantly higher in the low-risk group (Fig. 5C, D), suggesting a more active tumor immune microenvironment and potentially favorable immunotherapy response. TIDE analysis showed no statistically significant difference between the two groups (Fig. 5E). Taken together, these findings indicate that reprogrammed OXPHOS may influence prognosis by shaping immune infiltration patterns and immunotherapy-related characteristics within tumor tissues, prompting further investigation into intercellular communication among cells with different OXPHOS biomarker expression levels.

The expression of OXPHOS biomarker in LUAD single cell atlas

The scRNA-seq data analysis was performed on normal, primary tumor, and metastasis tumor tissue from the GSE131907 and GSE123902 datasets. A total of 221,792 cells passed the quality control and were divided into 8 separated cell clusters: ⅰ) T cells highly expressing CD3E, CD3D, and IL7R, ⅱ) B cells highly expressing CD79A, MS4A1, and JCHAIN, ⅲ) myeloid cells highly expressing LYZ, CD68, CD14, and CD4, ⅳ) epithelial cells highly expressing KRT19, EPCAM, and CDH1, ⅴ) mast cells highly expressing TPSAB1 and KIT, ⅵ) endothelial cells highly expressing PECAM1, VWF, CDH5, and CD34, ⅶ) fibroblasts cells highly expressing DCN, LUM, COL1A1, and ACTA2, ⅷ) oligodendrocyte cells highly expressing CLDN11 and MOG (Fig. 6A, C, S1A, B). Though all 8 cell clusters were identified in each sample origin (normal, primary, and metastasis) (Fig. 6B), their distribution was clearly inconsistent. As anticipated, the proportion of T cells gradually decreases from normal tissue to primary tumors and then to metastatic sites, while the proportion of epithelial cells increases progressively (Fig. 6D). The expression levels of genes in prognostic model were compared between bulk RNA(TCGA-LUAD) and scRNA levels (Fig. 6E, F). Among them, 10 genes, including upregulated ATP5F1D, CMC2, COA6, COX6B2, CYCS, NDUFA10, RAB51F, TIMMDC1, and UQCC3 and downregulated DMAC2 in tumor tissue, shared consistent trends in different levels, while the expression of FMC1, PET100, and UQCRB showed inconsistent trends across cohorts and were excluded from further analysis. Furthermore, the expression of the two genes UQCC3 and RAB5IF, which are the focus of subsequent analyses, was validated in clinical samples, showing expression patterns consistent with the transcriptomic data (Fig. 6G, H).

Fig. 6.

Distribution of the OXPHOS-related prognostic biomarkers expression in scRNA-seq data. A UMAP plot of cell clustering and annotation after merging GSE131907 and GSE123902 cohorts. B Origin of samples for each cell. C Dot plot for the expression levels of classic marker genes for each cell type. D Proportional bar chart of each cell type across Normal, Tumor, and Metastasis samples. E Heatmap of mean expression levels of the biomarkers in OXPHOS-related prognostic model for Normal and Tumor samples in TCGA-LUAD cohort. F Heatmap of mean expression levels of the biomarkers in OXPHOS-related prognostic model for Normal, Tumor, and Metastasis samples in scRNA-seq data. G Expression of UQCC3 in clinical samples assessed by qRT-PCR. H Expression of RAB5IF in clinical samples assessed by qRT-PCR. *, p < 0.05

The landscape of biomarker genes in epithelial cells

The expression levels of these 10 OXPHOS biomarker genes were consistent across both the single-cell and bulk datasets and were further investigated. The variable expression of these genes across cell clusters was explored, revealing that most marker genes were highly expressed in epithelial cells (Fig. 7A). Notably, RAB5IF, COA6, UQCC3, CYCS, and NDUFA10 exhibited a progressively increasing trend with disease progression, while ATP5F1D maintained a relatively stable high expression (Fig. 7B). Furthermore, the expression of these six genes was found to have a weak but significant correlation with the CNV score in epithelial cells, with correlation coefficients ranging from 0.1157 to 0.2376 (p < 0.001) (Fig. 7C, S2). Among them, the expression of UQCC3 and RAB5IF displayed the highest correlation coefficient. This prompted us to further explore how the expression of UQCC3 and RAB5IF in epithelial cells mediate poor prognosis. We divided the epithelial cells into two groups by expressed or unexpressed UQCC3 and RAB5IF, respectively. Gene set enrichment analysis (GSEA) revealed that UQCC3-positive and RAB5IF-positive epithelial cells showed the most significantly positive enrichment in pathways related to E2F targets and G2M checkpoint, while UQCC3-positive epithelial cells also had notably downregulation of genes related to allograft rejection (Fig. 7D). These results suggest that the expression of UQCC3 and RAB5IF in epithelial cells may enhance their proliferative and metastatic potential.

Fig. 7.

The landscape of the OXPHOS-related prognostic biomarkers in epithelial cells. A Expression levels of the OXPHOS-related prognostic biomarkers across different cell types. B Violin plots showing the expression levels of the top 6 genes in epithelial cells across Normal, Tumor, and Metastasis groups. C Bar chart showing the Spearman correlation between the top 6 genes and CNV scores in epithelial cells. ***: p < 0.001. D GSEA analysis of epithelial cells comparing UQCC3 + vs. UQCC3- and RAB5IF + vs. RAB5IF- groups

UQCC3 and RAB5IF expression levels contributed to immunosuppression within the tumor microenvironment

To further investigate the impact of UQCC3 and RAB5IF expression in epithelial cells on the immune microenvironment, we applied the CellChat method to analyze the cell-cell interactions between epithelial cells and immune cells, including T cells, B cells, myeloid cells, and mast cells. Our analysis specifically focused on the TGF-β and Galectin pathways, two key signaling pathways involved in tumor immunosuppression [22, 23]. The results revealed that UQCC3 + and RAB5IF + epithelial cells exhibited enhanced communication with immune cells, characterized by an increased number of ligand-receptor pairs (Fig. S3). For instance, the number of ligand secreted by epithelial cells targeting T cells increased from 182 in UQCC3- to 249 in UQCC3+ (Fig. S3). Notably, the probability of communication through immunosuppression-related receptor-ligand pairs, such as LGALS9-CD44, LGALS9-CD45, and TGFB2-TGFBR1 and ACVR1-TGFBR1, was greatly increased (Fig. 8A, B). Similar results were observed in communication between RAB5IF + epithelial cells and T cells (Fig. 8C, D). This may partly explain why the high-risk patient group, despite having more T cell infiltration in tumor tissues, still experiences poorer prognosis (Fig. 5). Meanwhile, the expression of UQCC3 or RAB5IF remodeled the receptor-ligand pairs involved in intercellular communication. On one hand, the ligand-receptor pairs associated with CD45 and CD44 predominated in the communication between epithelial cells and immune cells (Fig. 8), and these pairs have been identified as inhibitory in tumor immune response [24–26]. On the other hand, UQCC3 + or RAB5IF + epithelial cells were more active in establishing connections with immune cells, as demonstrated by the greater number of ligands secreted by epithelial cells for myeloid cells compared to those secreted by myeloid cells for epithelial cells, when compared to UQCC3- or RAB5IF- epithelial cells (Fig. S3). These findings suggest that the expression of UQCC3 and RAB5IF in epithelial cells may potentiate the inhibitory effect of epithelial cells on immune cells.

Fig. 8.

Cell communication between epithelial cells and immune cells. A-B Receptor-ligand pairs that interacted between (A) UQCC3 + and (B) UQCC3- epithelial cells and immune cells in the TGF-β and Galectin pathways. C-D Receptor-ligand pairs that interacted between (C) RAB5IF + and (D) RAB5IF- epithelial cells and immune cells in the TGF-β and Galectin pathways

Discussion

In this study, a total of 43 OXPHOS-related genes were identified as potential biomarkers for LUAD prognosis. After addressing multicollinearity using LASSO-Cox regression analysis, 13 biomarkers were selected for the construction of the nomogram. Further investigation into the potential mechanisms underlying OXPHOS-related regulation of LUAD prognosis was conducted using single-cell RNA-seq data and focused on immune cell infiltration, GSEA and cell-cell interaction.

Aerobic glycolysis is considered a key feature of tumors, enabling adaptation to microenvironment stress and genetic changes, with by-products like lactic acid contributing to tumor progression [27]. However, recent studies have re-emphasized the vital role of OXPHOS in cancer. Upregulated OXPHOS has been reported to promote stemness, metastasis and drug resistance in cancer cells, and therapies targeting OXPHOS, such as gamitrinib, showed a promising therapeutic strategy for various cancers [28–32]. In the tumor microenvironment, OXPHOS enhances resistance to apoptosis, cellular persistence, and antitumor immunity in various immune cells, including interleukin-17-producing CD4 T cells, natural killer cells, and CD8 + T cells [33–35]. Additionally, increasing OXPHOS-based biomarkers associated with patients’ survival have been identified, which regulated tumor prognosis by decreasing immune cell infiltration and upregulating glycolysis [12–14]. Consistent with these findings, our results showed that OXPHOS was significantly activated in LUAD tissues, further confirming that OXPHOS reprogramming is a hallmark of tumor and OXPHOS-based biomarkers have potential for tumor prognostic assessment.

Among the identified OXPHOS-related candidate biomarkers, most were upregulated in tumor tissues, consistent with previous findings that high expression of OXPHOS-related biomarkers correlates with poor prognosis [36]. Several biomarkers, such as COA6 and CYCS, are recognized as potential prognostic markers for lung cancer [37–39]. Elevated COA6 expression is associated with poor response to lung cancer immunotherapy, while CYCS has been associated with lung metastasis in differentiated thyroid carcinoma [38, 39]. More importantly, we identified 11 novel biomarkers associated with LUAD prognosis: DMAC2, FMC1, UQCC3, CMC2, ATP5F1D, COX6B2, RAB5IF, NDUFA10, UQCRB, TIMMDC1, and PET100. The latter eight genes have already been reported as prognostic markers in other cancers. For instance, ATP5F1D has been linked to diagnosis and prognosis in endometrial cancer [40, 41], while elevated COX6B2 expression marks poor long-term prognosis in gastric cancer [42]. Additionally, CMC2, RAB5IF, TIMMDC1, and PET100 have demonstrated prognostic value in glioma, hepatocellular carcinoma, esophageal squamous cell carcinoma, and ovarian cancer, respectively [43–46]. However, the roles of DMAC2, FMC1, and UQCC3 in tumor prognosis remain underexplored. Research has highlighted their roles in cancer development. UQCC3, for instance, regulates angiogenesis in both embryonic and tumor tissues by modulating VEGF expression [47], and drives bioenergetic reprogramming in hepatocellular carcinoma cell for hypoxia adaptation [48]. FMC1 has been shown to be upregulated under acidic pH conditions [49], while DMAC2 has been identified as a biomarker for predicting tumor mutational burden (TMB) in lung cancer [50]. Summarily, our results present novel biomarkers for LUAD prognosis, contributing to a deeper understanding of LUAD prognosis and offering potential for clinical prognostic applications.

Moreover, we noticed that the trend of genes expression changes may differ across various cancer types. In our study, both NDUFA10 and UQCRB were upregulated and recognized as poor prognostic biomarkers for LUAD. However, previous research has shown that NDUFA10 was downregulated in prostate cancer and functioned as an unfavorable prognostic marker [51], while UQCRB exhibited a contrasting trend and prediction performance in colorectal cancer and glioma [46, 52]. These findings suggest that even the same prognostic markers may activate distinct signaling pathways and molecular mechanisms, influencing tumor prognosis differently across various cancer types.

The prognosis of cancer is related to multiple factors. In this study, we investigated the potential mechanisms of identified biomarkers regulating LUAD prognosis, focusing on epithelial cell proliferation, immune infiltration, and suppressive immune microenvironment. Our study revealed the association between OXPHOS expression and immune cell infiltration in tumor tissue. Tumor immunophenotyping studies have revealed that “hot” tumors, characterized by high immune cell infiltration, generally indicate better prognosis. For instance, lung cancer patients with higher levels of naïve B cells, memory B cells, and natural killer cells in tumor samples tended to have longer survival [53, 54]. In our cohort, we observed similar results that high-risk score patients, who exhibited elevated OXPHOS biomarkers expression and shorter survival, showed significant lower immune score for B cells, natural killer cells, and eosinophil cells. This further reinforces the potential of OXPHOS-based biomarkers as prognostic indicators for LUAD. Interestingly, several T cell subtypes, including CD4 T cells and type 2 T helper cell, were more enriched in high-risk patient groups. Subsequent analysis revealed that these T cells were suppressed by OXPHOS-reprogrammed epithelial cells through altered intercellular communication. This suggests that reactivating these infiltrating T cells could potentially improve the prognosis of these high-risk patients.

One more thing in the present is that we applied scRNA-seq data to investigate the effect of identified biomarkers expression on different cell subpopulations. ScRNA-seq technology allows researchers to investigate tumors progression at both the cellular and molecular levels, providing valuable insights for mechanisms investigation and novel precise therapeutic interventions in LUAD [55]. A key finding of our study is that UQCC3 and RAB5IF expressions may promote immunosuppression by enhancing communication between epithelial cells and immune cells through immunosuppression-related receptor ligand pairs. Especially in epithelial cells interacting with T cells, UQCC3 and RAB5IF expression greatly increased the probability of communication through LGALS9-CD44 and LGALS9-CD45 pairs. This may help explain why high-risk patients, despite having more T cell infiltration, have poorer prognosis than low-risk patients. Targeting CD44 and CD45 have been suggested as potential immunotherapies for breast cancer and leukemia [25, 26]. Therefore, further investigation is needed to determine whether targeting CD44 and CD45 can reactivate T cells in the tumor tissues of high-risk patients and improve their prognosis.

Another potential mechanism suggested by our bioinformatic analyses is that OXPHOS-related biomarkers, such as UQCC3 and RAB5IF, may potentially promote both malignant proliferation and immune suppression in LUAD. Their elevated expression was associated with tumor metastasis and increased copy number variation (CNV) scores, indicating more aggressive phenotypes in epithelial cells with high OXPHOS activity. Pathway analysis further suggested enrichment of E2F target and G2M checkpoint pathways, which could contribute to accelerated cell-cycle progression. In parallel, OXPHOS-generated reactive oxygen species (ROS) may activate the NF-κB signaling pathway, potentially driving TGFβ1 transcription and secretion and thereby promoting tumor growth [56]. Enhanced TGF-β signaling may also facilitate epithelial–mesenchymal transition, fibrosis-like remodeling, and suppression of cytotoxic T-cell activity [57, 58]. Additionally, Galectin-mediated interactions could contribute to T-cell exhaustion and impaired antigen presentation [59, 60] Together, these observations provide a plausible hypothesis for how UQCC3 and RAB5IF expression could coordinate proliferative and immunosuppressive programs in LUAD. Positive feedback between UQCC3 and ROS may further support tumor adaptation to stress [48]. Nevertheless, the detailed mechanisms by which these biomarkers influence malignant proliferation in lung epithelial cells remain to be experimentally validated.

Recently, an integrative approach combining single-cell RNA sequencing and machine learning successfully identified an immunogenic-cell-death–related signature for LUAD prognosis and immunotherapy guidance, highlighting the growing value of computational prognostic modeling in LUAD [61]. Building upon this line of research, we identified novel OXPHOS-related biomarkers for LUAD prognosis assessment and developed a prognostic model for long-term clinical management. Our analyses further suggested that reprogrammed OXPHOS may potentially contribute to poor prognosis by promoting immunosuppression in the tumor microenvironment and enhancing epithelial cell proliferation. Nonetheless, several limitations should be acknowledged. Although the prognostic model demonstrated stable performance across datasets, validation in larger, independent cohorts is needed to further support its clinical relevance. Potential dataset heterogeneity and sequencing platform differences may also affect model robustness. While we confirmed the differential expression of key OXPHOS-related biomarkers such as RAB5IF and UQCC3 in patient samples, these findings reflect expression-level alterations only; their mechanistic roles remain inferred from bioinformatics analyses and have not been functionally validated. Comprehensive experimental studies, including gene knockdown, overexpression, and downstream pathway assays in cellular and animal models, will be essential to determine whether these genes directly contribute to LUAD progression and immune dysregulation. Despite these limitations, the present study provides a valuable foundation for subsequent mechanistic and translational investigations.

.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Hongxia Ma: Data curation, Methodology, Formal analysis, Writing – original draft. Shaoshan Zeng: Writing – review & editing, Validation. Changping Xie: Writing – review & editing, Validation. Chengcheng Gao: Visualization, Formal analysis. Siao Jiang: Visualization, Formal analysis. Liuxin Chen: Visualization, Formal analysis.Wenhao Tian: Visualization, Formal analysis. Lizhi Huang: Conceptualization, Methodology, Investigation, Writing – review & editing.

Funding

This rsearch received no external funding.

Data availability

The datasets analyzed in this study are publicly available from the TCGA (https://portal.gdc.cancer.gov/) and GEO (https://www.ncbi.nlm.nih.gov/geo/) databases. Specifically, the TCGA dataset used is TCGA-LUAD, and the GEO datasets include GSE131907, GSE123902, GSE31210, and GSE72094. The reference gene sets and related databases used in this study include MitoCarta3.0 (https://personal.broadinstitute.org/scalvo/MitoCarta3.0/human.mitocarta3.0.html), CellMarker 2.0 (http://117.50.127.228/CellMarker/index.html), MSigDB (https://www.gsea-msigdb.org/gsea/msigdb), LM22/CIBERSORT (https://cibersortx.stanford.edu/), and CellChatDB.human (https://github.com/sqjin/CellChat). The relevant supplementary information files are also accessible. The authors declare that they assume full responsibility for the integrity of the data and the accuracy of the data analysis.

Declarations

Ethics approval and consent to participate

The study was approved by the Ethics Committee of Shenzhen Bao’an District People’s Hospital (Approval No. 2021041318093719) and conducted in accordance with the Declaration of Helsinki. Written informed consent was obtained from all participants prior to sample collection.

Consent for publication

Not applicable.

Competing interests

Authors Chengcheng Gao, Siao Jiang, Liuxin Chen and Wenhao Tian are employees of Hangzhou AstrocyteTechnology Co., Ltd. The other authors affirm that this research was conducted without any commercial or financial affiliations that could be perceived as potential conflicts of interest.