Single-cell mitophagy signature-based artificial intelligence model enhances prediction of prognosis and immunotherapy response in non-small-cell lung cancer

Abstract

Background

Non-small-cell lung cancer (NSCLC) exhibits pronounced molecular heterogeneity, and current predictive models rarely incorporate mitochondrial quality-control programs such as mitophagy. We hypothesize that an artificial intelligence model based on mitophagy-related genes (MRGs) at single-cell resolution could improve prediction of survival and immunotherapy benefit.

Methods

We analyzed single-cell RNA sequencing data from treatment-naïve NSCLC tumors to evaluate the activity of MRGs and identify genes exhibiting differential expression between cells with high versus low mitophagy levels. These differentially expressed genes were then cross-referenced with mitophagy gene sets to pinpoint candidate prognostic markers. Using LASSO regression combined with multiple machine learning classifiers, we constructed a risk model, which was validated in both internal and external cohorts, including a clinical immunotherapy trial. We further examined the relationship between the risk model, immune cell infiltration, and drug sensitivity in silico. The key MRGs were then experimentally validated in A549 cells using qRT-PCR, Western blotting, immunofluorescence, and functional assays for cell migration and wound healing.

Results

We quantified the mitochondrial autophagy activity of 18,167 single cells. Differential expression yielded 1,668 genes; intersection with the MRG list produced 39 candidates. A six-gene panel (FOS, CANX, EIF4G1, CALCOCO2, HSP90AB1, and PRKAR1A) emerged from LASSO. Gradient boosting machine (GBM) achieved the optimal cross-validated performance (testing set: AUC = 0.80, validation set: AUC = 0.72). SHAP analysis ranked PRKAR1A and CALCOCO2 as the top risk contributors. Patients classified into the high-MRG-score group exhibited consistently shorter overall survival (OS) across all datasets (HR = 3.66, 95% CI 1.72 − 7.81, P < 0.001). Low-MRG tumors displayed elevated immune and ESTIMATE scores with reduced tumor purity, and achieved a significantly higher objective response rate (35% vs 22%, P < 0.05) and prolonged OS (HR = 1.48, 95% CI 1.07 − 2.05, P = 0.017) with immune checkpoint blockade in the clinical-trial setting. Experimental results showed that knockdown of CALCOCO2 or overexpression of PRKAR1A significantly inhibited A549 cell proliferation and reduced mitochondrial membrane potential (P < 0.05), thereby affecting mitophagy.

Conclusion

We developed an MRG-based model that reliably stratifies NSCLC patients according to prognosis and identifies those most likely to respond to immune checkpoint inhibitors, providing a framework for integrating tumor metabolic characteristics into personalized therapeutic decisions.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12931-026-03601-w.

Introduction

Lung cancer has been recognized as a leading cause of cancer-related deaths worldwide, imposing a heavy burden on global public health [1, 2]. Non-small-cell lung cancer (NSCLC) is the most common subtype of lung cancer, accounting for 80–85% of all lung cancer cases. Most patients present with insidious early symptoms, leading to advanced-stage diagnosis in more than 60% of cases. Even among patients receiving first-line treatment, such as epidermal growth factor receptor-tyrosine kinase inhibitors (EGFR-TKIs) for EGFR-mutant NSCLC, 50–70% develop intrinsic or acquired resistance within 1–2 years, resulting in disease progression [3]. Existing prognostic models for NSCLC mainly rely on traditional clinical parameters such as the TNM stage and performance status, or on a limited set of molecular readouts, including circulating tumor DNA, transcriptomic signatures, and serum proteins [4–8]. These approaches treat the tumor as a static entity and overlook the dynamic, cell-autonomous programs. Chief among them is mitophagy, which can directly dictate metastatic propensity and therapy resistance by clearing reactive oxygen species (ROS)-generating organelles. At the same time, mitophagy rewires energy metabolism, dampens intrinsic apoptosis, and shapes the immunopeptidome [9–11]. Exploiting this context-specific sentinel function, we propose a mitophagy-centric prognostic paradigm, offering a novel perspective to dissect NSCLC heterogeneity and to enable early prediction of patient outcomes.

Mitophagy is a conserved cellular process that selectively degrades damaged or dysfunctional mitochondria via autophagic machinery. A growing body of studies has demonstrated the importance of mitophagy in cancer development and progression. The BNIP3/BNIP3L (NIX)-dependent mitophagic pathway has emerged as a key regulator of mitochondrial turnover in hypoxic and nutrient-deprived microenvironments. These conditions are commonly observed in many types of solid tumors [12–14]. In NSCLC, the BNIP3/BNIP3L pathway exhibits abnormal expression patterns associated with disease progression and poorer outcomes [15]. High BNIP3 expression also correlates with increased survival of tumor cells under hypoxic conditions and resistance to platinum-based chemotherapy [16]. This is because mitophagy-mediated clearance of damaged mitochondria reduces the accumulation of ROS and inhibits apoptosis. According to previous reports, enhanced mitophagic activity enables tumor cells to improve their self-repair capacity following chemoradiotherapy [17]. By clearing damaged mitochondria within the cell, tumor cells can maintain survival in the treatment-induced stressful microenvironment. Therefore, inhibiting the mitophagic process may represent a promising therapeutic strategy [18, 19].

To date, although mitophagy plays a critical regulatory role in the metabolic processes, oxidative stress responses, and apoptotic balance of tumor cells, particularly prominent in the hypoxic tumor microenvironment (TME), prognostic investigations on mitophagy-related genes (MRGs) in NSCLC remain relatively limited. Hence, constructing a reliable prognostic model based on MRGs is expected to fill the research gap in this field. Additionally, unlike traditional bulk sequencing that masks low-abundance cell clones (e.g., drug-resistant cells, cancer stem cells), single-cell RNA sequencing (scRNA-seq) resolves tumor heterogeneity cell-by-cell: it pinpoints different functional subpopulations within tumors, uncovers tumor recurrence and drug resistance mechanisms, maps tumor–immune–stroma crosstalk to decipher the TME regulation network, and dynamically tracks gene expression and clonality of tumor cells to guide personalized treatment strategies [20–22].

In this study, we curated NSCLC scRNA-seq data and, by benchmarking state-of-the-art interpretable machine learning algorithms, established a robust prognostic model grounded in MRG expression. This model accurately predicts patient prognosis, quantitatively reflects tumor immune infiltration status, and consequently forecasts the efficacy of immunotherapy, as validated in an independent clinical-trial cohort. The incorporation of interpretable machine learning algorithms further enhanced predictive performance while largely demystifying the “black-box” nature of conventional artificial intelligence (AI) models [23]. Our findings may provide a new perspective for precision management in NSCLC. Our study’s workflow is depicted in Fig. 1.

Fig. 1.

Flowchart of the study

Methods

Database retrieval and data collection

NSCLC patient data were subjected to systematic retrieval via The Cancer Genome Atlas (TCGA) (https://portal.gdc.cancer.gov/) and Gene Expression Omnibus (GEO) databases (http://www.ncbi.nlm.nih.gov/geo/). A total of four NSCLC gene expression profiling cohorts and one melanoma cohort receiving immunotherapy were included in this study, namely TCGA, GSE68465, GSE31210, GSE73403, GSE91061 and GSE135222. We downloaded the matrix files of each GEO cohort and performed normalization for further analysis. For the RNA sequencing data of TCGA, we used the R package TCGAbiolinks to download transcripts per kilobase million (TPM) values of gene expression from the Genomic Data Commons (GDC). Samples lacking survival data or with overall survival (OS) of less than one month were excluded from further analysis. Finally, the remaining 474 samples were used for subsequent analyses. We assessed the differential protein expression between NSCLC and normal tissues using immunohistochemistry (IHC) images obtained from the Human Protein Atlas (HPA) database. (https://www.proteinatlas.org/).

Analysis of single-Cell RNA sequencing data

We obtained the single-cell RNA sequencing (scRNA-seq) datasets GSE117570 and GSE198099 from the GEO database, which included four and two NSCLC samples, respectively, as well as their corresponding adjacent normal tissues. Data quality assessment was performed using the R packages Seurat and SingleR. To maintain the integrity of scRNA-seq data, cells were excluded based on the following criteria: mitochondrial gene content < 15% [24], RNA molecule counts within the range of 2,000–25,000, and genes whose expression was detected in fewer than three cells. These two datasets respectively yielded 18,167 and 8,782 cells. Data integration across multiple samples was performed using the IntegrateData function to correct for batch effects, followed by normalization with ScaleData. To efficiently reduce the dimensionality for downstream clustering and visualization, principal component analysis (PCA) was applied to identify the most informative principal components (PCs). The first 20 PCs were then used with the UMAP algorithm to reveal distinct cell clusters with biological relevance. The FindNeighbors and FindClusters functions were executed with dims = 10 and resolution = 1 to construct the shared nearest neighbor graph and delineate distinct cellular subpopulations. Cell-type annotation was performed based on canonical marker genes curated from CellMarker 2.0 and previously published literature. Finally, the RunUMAP function was used to generate two-dimensional UMAP plots for single-cell visualization.

Mitophagy gene identification and activity evaluation

MRGs were identified from the GeneCards database (https://www.genecards.org/) and genes with a relevance score greater than 1 were included in the subsequent analyses[25–28]. AUCell is a computational tool used to assess the activity of gene sets in scRNA-seq data and to calculate the mitophagy activity score for individual cells. The activity of mitophagy-related gene sets was quantified by computing the area under the curve (AUC) of the cumulative distribution function (CDF).

Cohort division and feature selection

The TCGA dataset was randomly divided into a training cohort (70%, n = 332) and a validation cohort (30%, n = 142). To alleviate multicollinearity in data analysis, LASSO regression analysis was performed using the glmnet R package, with fivefold cross-validation conducted simultaneously. Variables with a P < 0.05 were selected from the univariate analysis. With the penalty parameter λ rising, the predictive variables' coefficients in the model began to diminish steadily toward zero.

Immunotherapy clinical trial data collection

After database screening, the IMvigor210 cohort with complete transcriptomic data and clinical information was selected. The research focus of this clinical trial is to evaluate the therapeutic efficacy of Atezolizumab, a programmed death-ligand 1 (PD-L1) inhibitor [29]. Transcriptomic and clinical data were downloaded from http://research-pub.gene.com/IMvigor210CoreBiologies. The R package DESeq2 was used for data normalization before subsequent analyses. Although the IMvigor210 cohort primarily focuses on urothelial carcinoma, its data share similar immune microenvironment characteristics with NSCLC, which justifies its inclusion in our model validation [30].

Functional enrichment and immune infiltration analysis

Kyoto Encyclopedia of Genes and Genome (KEGG) pathway and Gene Ontology (GO) enrichment analyses were conducted using the clusterProfiler R package. GO terms were analyzed in three categories: biological process (BP), molecular function (MF), and cellular component (CC), to investigate potential biological functions associated with the prognostic MRGs identified in this study [31]. Protein–protein interaction (PPI) analysis was performed using the STRING database (https://string-db.org/), and the interaction network was visualized with Cytoscape software (version 3.8.2; https://www.cytoscape.org). At the statistical level, the Benjamini–Hochberg multiple testing procedure was used to control the false discovery rate (FDR), achieving further adjustment of enrichment P-values to ensure the reliability of the analysis results. Four algorithms, including QuanTIseq [32], CIBERSORT [33], ESTIMATE [34], and TIDE [35], were employed to assess the abundance of immune cells and analyze the relationship between the mRNA expression of MRGs and the tumor immune microenvironment.

Construction and validation of machine learning models

To predict the prognosis of NSCLC, a total of 10 machine learning (ML) models were employed, including support vector machine (SVM), K-nearest neighbors (KNN), categorical boosting (CatBoost), adaptive boosting (AdaBoost), light gradient boosting machine (LightGBM), logistic regression (LR), extreme gradient boosting (XGBoost), neural network (NN), random forest (RF), and gradient boosting machine (GBM). For model performance evaluation, five random resampling iterations were first performed on the training and test sets, followed by basic assessment using the area under the receiver operating characteristic curve (ROC) and decision curve analysis (DCA). Subsequently, on the test and validation sets, indicators such as accuracy, precision, recall, and F1-score were further used to comprehensively verify the efficacy and reliability of the models in predicting the prognosis of NSCLC patients.

SHAP-based interpretable machine learning model

To improve model interpretability, we applied SHAP (SHapley Additive exPlanations), a game-theory-based framework that estimates how each feature contributes to the predictions for individual samples. Derived from the core concept of Shapley values in cooperative game theory, it is particularly suitable for parsing the decision-making mechanisms of complex ML models. Subsequently, the shapviz R package was used to complete the calculation and visualization of SHAP values. In addition, this study incorporated more performance evaluation metrics to comprehensively verify the effectiveness of the model in practical applications. Complete code and parameter settings for the machine learning-based prognostic model have been uploaded to GitHub. For further details, please refer to https://github.com/mikelu1997/machine-learing

Cell lines

The lung adenocarcinoma cell line A549 was obtained from the Cell Bank of the Chinese Academy of Sciences in Shanghai, China. The experiment involved cultivating cells within a DMEM (Gibco) medium, which was enriched with 10% fetal bovine serum (FBS; Gibco), under controlled conditions of 37 °C incubation and 5% CO₂ gas exchange.

Quantitative real-time PCR detection

Total RNA (2 μg) was isolated from A549 cells using TRIzol reagent (#15596026CN, Invitrogen) and checked for quality by the OD260/280 and electrophoresis. cDNA was generated using the EasyScript One-Step gDNA Removal and cDNA Synthesis SuperMix (#AE311, TransGen, China). qRT-PCR was carried out on an Eppendorf Real-Time PCR System using TransScript Green qPCR SuperMix (#AQ211, TransGen, China) on the synthesized cDNA, including technical replicates and NTC/NRC controls. The expression of PRKAR1A, CALCOCO2, EIF4G1, HSP90AB1, FOS, and CANX was normalized to GAPDH and calculated using the 2^–ΔΔCt method, with primer sequences listed in Supplementary Table S1.

Immunofluorescence staining

A549 and BEAS-2B cells were cultured on sterile glass coverslips until 70–80% confluent, fixed with 4% paraformaldehyde for 15 min, and permeabilized with 0.3% Triton X-100 for 10 min. After blocking with 5% BSA for 30 min, cells were incubated overnight at 4 °C with primary antibodies against CALCOCO2 (#12229–1-AP, 1:200, Proteintech) and PRKAR1A (#MA5-24981, 1:100, Invitrogen). After washing, Alexa Fluor 488- or 594-conjugated secondary antibodies (1:500, Invitrogen) were applied for 1 h in the dark, followed by DAPI counterstaining for 10 min. Coverslips were mounted with antifade medium, and fluorescence images were acquired using a confocal microscope (Leica Microsystems, Germany).

Mitochondrial membrane potential assay

Mitochondrial membrane potential (ΔΨm) was measured using the JC-1 dye (Beyotime, China) according to the manufacturer’s instructions. A549 cells were seeded onto glass-bottom dishes and cultured to 70–80% confluence. After treatment or transfection, cells were incubated with JC-1 working solution (5 μg/mL) at 37 °C for 20 min in the dark, then washed twice with JC-1 buffer to remove unbound dye. Fluorescence signals were observed under a confocal microscope (Leica Microsystems, Germany), where JC-1 aggregates emitted red fluorescence indicating intact mitochondria, and JC-1 monomers emitted green fluorescence indicating depolarized mitochondria. The red/green fluorescence intensity ratio was calculated to assess changes in mitochondrial membrane potential.

Cell transfection

A549 lung cancer cells were cultured in DMEM with 10% fetal bovine serum at 37 °C and 5% CO2 to 70%−80% confluence. Transfection was carried out using Lipofectamine™ 3000 reagent (#L3000015) following the manufacturer’s instructions. The siRNA for CALCOCO2 and PRKAR1A, synthesized by GenePharma Biotech (Shanghai, China), was transfected according to Thermo Fisher Scientific's protocol, with sequences listed in Supplementary Table S1.

Western blotting

Total proteins were extracted from cells using RIPA lysis buffer (Yeasen, China), and protein concentrations were measured using the BCA protein assay kit (Yeasen, China) according to the manufacturer’s instructions. Proteins were separated on SDS-PAGE gels and transferred onto nitrocellulose membranes (Millipore, USA). Membranes were blocked with 5% skimmed milk at room temperature for 1.5 h, followed by three washes with 1 × TBST. Primary antibodies against human CALCOCO2 (#12229–1-AP, 1:1000, Proteintech), P62 (#5114, 1:1000, CST), P-AMPK-α (#2531, 1:1000, CST), AMPK (#2532, 1:1000, CST), PRKAR1A (#MA5-24981, 1:2000, Invitrogen), E-cadherin (#14472, 1:1000, CST), Vimentin (#5741, 1:1000, CST), Snail (#3879, 1:1000, CST), P-ERK1/2 (#4370, 1:1000, CST), ERK1/2 (#4695, 1:1000, CST) and GAPDH (#5174 T, 1:10000,CST) were diluted in antibody dilution buffer (#WB500, NCM) and incubated with the membranes overnight at 4 °C. After three washes with 1 × TBST, membranes were incubated with IRDye 800CW infrared fluorescent secondary antibodies (1:10000, #926–32210, #926–32211) at room temperature for 1–1.5 h. Following additional washes, signals were visualized using the Odyssey SA near-infrared imaging system. Experiments were performed in triplicate to ensure reproducibility.

Colony formation assay

For the colony formation assay, 500–1500 cells per well were seeded in 6-well plates and cultured in DMEM with 10% fetal bovine serum for two weeks. Formed colonies were washed with PBS, fixed with 4% paraformaldehyde for 20 min, and stained with 0.2% crystal violet for 25 min. After removing excess dye, colonies containing more than 50 cells were counted.

Transwell migration and invasion assays

Cell migration and invasion were evaluated using Transwell chambers with 8-μm pore membranes. For migration, 2–5 × 10⁴ cells in serum-free medium were seeded into the upper chamber, and the lower chamber was filled with medium containing 10% FBS. After 24–48 h, non-migrated cells were removed, and migrated cells were fixed with 4% paraformaldehyde and stained with 0.1% crystal violet. For invasion, the upper chamber was pre-coated with Matrigel, and the assay was performed following the same procedure.

Wound healing assay protocol

For the wound healing assay, 5 × 10⁴ cells were seeded in 6-well plates and grown to confluence. Linear scratches were created using a 200 µL pipette tip, and detached cells were removed by washing with PBS. Cells were maintained in serum-free DMEM, and images were acquired at 0, 24, and 48 h. Wound closure was quantified using ImageJ software (version 2.1.0).

Statistical analysis

Statistical analyses were conducted using R (version 4.2.1). For continuous variables, differences between two groups were evaluated using the Student’s t test for normally distributed data and the Mann–Whitney U test for non-normal data. Survival curves were generated using the Kaplan–Meier method and compared using Cox proportional hazards regression analysis. All tests were two-sided, and P < 0.05 was considered statistically significant.

Results

Analysis of single-cell sequencing data

Using the scRNA-seq data from GSE117570, we identified 24 distinct cell clusters through a dimensionality reduction algorithm, each characterized by distinct gene expression profiles (Fig. 2A). Based on markers from CellMarker2.0 and previous literature, the clusters were classified into seven different cell types, including T cells, B cells, monocytes/macrophages, fibroblasts, and others (Fig. 2B; Supplementary Table S2). To explore the expression characteristics of MRGs, we evaluated the mitophagy activity of each cell. Cells were categorized into high- and low-expression groups based on the median AUC score of MRGs. Cells with elevated gene expression levels exhibited higher AUC values, a phenomenon predominantly observed in T cells, B cells, monocytes/macrophages, and epithelial cells marked in red (Fig. 2C). To identify the biological mechanisms potentially contributing to the differences in AUC scores, we performed differential expression analysis. By comparing genes between cells with high MRG expression and those with low MRG expression, a total of 1,668 differentially expressed genes (DEGs) were identified, which were subsequently used for candidate gene selection (Fig. 2D). To further evaluate the potential impact of cellular composition on these findings, T cells, B cells, monocytes/macrophages, and epithelial cells were separately extracted, and AUCell-based stratification was conducted within each cell type. Functional enrichment analysis revealed that cells with high mitophagy activity exhibited consistent enrichment patterns in metabolic regulation, immune response, and signaling pathways, while also demonstrating cell-type–specific functional shifts.These findings suggest that mitophagy may contribute to immune modulation and tumor heterogeneity within the tumor microenvironment through cell-type–dependent mechanisms (Supplementary Figure S1A-H). The pseudotime analysis revealed that the mitophagy process was initiated earliest in epithelial cells and monocytes/macrophages, while it occurred latest in T cells and other cell types (Fig. 2E). Annotation of autophagy-related genes revealed that the functions of CALCOCO2 and PINK1 persisted throughout the entire process, with no significant downregulation detected at the late stage (Supplementary Figure S1I-J). This phenomenon is most likely attributed to persistent oxidative stress in the TME, which induces continuous mitochondrial damage.

Fig. 2.

Cell subset annotation and identification of differentially expressed genes (DEGs). A UMAP plot showing the results of dimensionality reduction clustering analysis. B Cells were annotated into seven distinct cell types. C Cells were divided into high- and low-expression groups based on mitophagy-related genes (MRGs). D Differential expression analysis between the high- and low-expression groups. E Pseudotime analysis of the trajectory of cells undergoing mitophagy. F Intercellular communication of the TNF signaling pathway among various cells. G Heatmap of sender and receiver cells of autophagic signals. H Roles of various cells in the TNF signaling pathway

Furthermore, we investigated intercellular communication within the TNF-autophagy pathway and demonstrated that this communication was primarily centered on B cells (Fig. 2F). Specifically, B cells acted as the main senders of autophagic signals, whereas epithelial cells and monocytes/macrophages served as the primary receivers (Fig. 2G). T cells and B cells also functioned as key regulators of autophagy, with both cell types exerting dominant effects on the autophagic process (Fig. 2H). Analysis of the validation dataset (GSE198099) yielded consistent results, attesting to the reliability of our findings (Supplementary Figure S2).

Enrichment analysis and LASSO analysis identify key genes in differential expression profiling

To explore the biological processes associated with MRG-related transcriptional alterations in NSCLC, we performed functional enrichment analyses based on the identified DEGs. GO analysis revealed significant enrichment in oxidative phosphorylation, ATP metabolic processes, and cadherin binding, suggesting alterations in mitochondrial energy metabolism and intercellular interactions within the tumor microenvironment.KEGG pathway analysis demonstrated enrichment in apoptosis, NOD-like receptor signaling, and the PD-1/PD-L1 immune checkpoint pathway (Fig. 3A), indicating potential links between mitophagy-associated transcriptional changes and immune regulation. Gene Set Enrichment Analysis (GSEA) further showed significant enrichment of TNFα signaling via NF-κB (NES = 1.393, adjusted P value (adj. P) = 0.009), a pathway previously reported to be involved in mitochondrial stress responses and mitophagy-related processes. In addition, metabolic pathways such as fatty acid metabolism and cholesterol homeostasis were also enriched (Fig. 3B), reflecting metabolic reprogramming that may accompany altered mitophagy activity. Similar enrichment patterns were observed in the validation dataset (Supplementary Figure S3), supporting the consistency of these findings. Collectively, these enriched pathways reveal the core role of MRGs in immune regulation and tumor progression. We intersected DEGs with MRGs, thereby identifying 39 candidate MRGs (Fig. 3C; Supplementary Table S3). To further clarify the association between these MRGs and NSCLC patient outcomes, we conducted in-depth analyses on these 39 most relevant genes influencing mitophagy activity. Using LASSO regression analysis, under the optimal regularization parameter, six prognosis-related genes were identified: FOS, CANX, EIF4G1, CALCOCO2, HSP90AB1, and PRKAR1A (Fig. 3D). Among these six genes used for model construction, five were risk factors while one was a protective factor (Fig. 3E). Finally, we analyzed the positions of the six MRGs on human chromosomes. The chromosome map showed that two of these MRGs were located on chromosome 17 (CALCOCO2 and PRKAR1A; Fig. 3F).

Fig. 3.

Screening of prognosis-related genes. A GO and KEGG enrichment analyses of DEGs. B GSEA enrichment analysis of DEGs. C Venn diagram showing the common genes at the intersection of DEGs and MRGs. D LASSO Cox regression analysis to screen for prognosis-related genes. E Roles of the six prognosis-related genes. F Chromosomal localization of MRGs

Construction of machine learning models and SHAP interpretability

Through SHAP analysis, we evaluated the importance and impact of MRGs identified by LASSO analysis (FOS, CANX, EIF4G1, CALCOCO2, HSP90AB1, and PRKAR1A) on the predictive performance of machine learning models. Data were split into training and validation sets at a 7:3 ratio. ROC curves compared the prognostic performance of various models, among which GBM achieved the best results. In the validation set, GBM achieved the highest sensitivity (AUC = 0.72, 95% confidence interval [CI] 0.62 − 0.81) and specificity, outperforming other algorithms (Fig. 4A-C; Supplementary Table S4). This highlights the robustness of GBM in making accurate predictions using the identified MRGs. Visualization of SHAP values provided further insights into the contributions of these genes, showing that PRKAR1A and CALCOCO2 had the greatest impact, followed by EIF4G1, HSP90AB1, FOS, and CANX (Fig. 4D-E). SHAP summary plots and waterfall plots further elucidated the direction and magnitude of each gene’s contribution to the model predictions. (Fig. 4F-G). CALCOCO2 showed large positive SHAP values and thus exerted important effects on risk prediction, whereas EIF4G1 and FOS had smaller SHAP values and limited contributions to the model output. By consulting the Human Protein Atlas database, we analyzed IHC data of key molecules with the largest SHAP values and qualitatively observed significant differences in IHC staining performance between normal and NSCLC samples (Supplementary Figure S4).

Fig. 4.

Receiver operating characteristic (ROC) analysis and Shapley Additive Explanations (SHAP) interpretation of different models. A ROC curve of the training set. B ROC curve of the test set. C Decision curve analysis (DCA) of the test set. D SHAP histogram showing the contribution degree of genes. E SHAP beeswarm plot of feature variables for predicting NSCLC prognosis. F SHAP force plot showing each gene’s impact on model prediction. G SHAP waterfall plot illustrating each gene’s contribution direction and magnitude

Prognostic value of MRGs in NSCLC and immune microenvironment analysis

Based on the survival outcome predictions from the GBM model, the optimal cutoff value was selected to categorize patients into low- and high-risk groups. In the training set, Kaplan–Meier survival curves indicated that the high-risk group was significantly associated with shorter OS (hazard ratio [HR] = 3.68, 95% CI: 2.22 − 6.08, P < 0.001; Fig. 5A). We further evaluated the predictive performance of the model using the C-index and time-dependent AUC in both the TCGA training and validation cohorts (Supplementary Figure S5A-F). Similarly, Kaplan–Meier curves in the validation set showed that the high-risk group was also linked to decreased survival (HR = 3.66, 95% CI: 1.72 − 7.81, P < 0.001; Fig. 5B). To evaluate the robustness and generalizability of the model, we applied the MRG-based machine learning prognostic model to three independent external cohorts (GSE68465, GSE31210, and GSE73403) for external validation. We demonstrated that patients with low MRG scores exhibited significantly longer OS, and the model reliably distinguished prognostic differences across cohorts, with ROC curves confirming its predictive accuracy (Fig. 5C–E; Supplementary Figure S5G-I). These results confirm that the MRG-based model demonstrates excellent prognostic performance and potential clinical application.

Fig. 5.

Predictive performance of the MRG-based model and immune landscape differences between high- and low-risk groups. A Survival curve of the training set. B Survival curve of the validation set. C–E External validation of the model in the GSE68465, GSE31210, and GSE73403 datasets. F Evaluation of nine immune cell types using the QuanTIseq algorithm. G Visualization of immune cell fractions derived from the CIBERSORT algorithm. H Relative cell proportions of immune infiltration in high and low MRG expression groups assessed by the CIBERSORT algorithm. I Correlations between six key MRGs and immune cells. J ESTIMATE, immune, and stromal scores for the two risk groups. K Tumor purity and TIDE scores in high- and low-risk groups. (*P < 0.05, **P < 0.01, ***P < 0.001)

Given the prognostic value of the MRG-based model, we explored its potential effects on the tumor immune microenvironment. Cancer initiation and progression are heavily influenced by immune cells within the TME. Thus, we employed four algorithms, including QuanTIseq, CIBERSORT, ESTIMATE, and TIDE, to analyze immune cell infiltration in relation to the two molecular subtypes. In the high-risk group, there was significant infiltration of M1 macrophages, Tregs, and resting CD4⁺ memory T cells, whereas the low-risk group exhibited a higher abundance of follicular helper T cells (P < 0.05, Fig. 5F-G). Comparisons of cell abundance based on CIBERSORT scores showed consistent trends (Fig. 5H). Furthermore, immune cell infiltration analysis revealed that PRKAR1A and CANX were significantly and positively correlated with activated dendritic cells (Fig. 5I). The ESTIMATE algorithm demonstrated that, compared with the low-risk subtype, the high-risk subtype had lower ESTIMATE and immune scores and higher tumor purity (Fig. 5J-K). To assess the potential efficacy of immunotherapy, we calculated the TIDE score, which predicts responses to immune checkpoint inhibitors (ICIs). A higher TIDE score indicates a poorer response to ICIs and suggests immune evasion by the tumor. The low-risk group exhibited significantly lower TIDE scores than the high-risk group, indicating that the low-risk group is more likely to benefit from ICI treatment (Fig. 5K). These results suggest that distinct immune profiles between the two risk groups contribute to their differential responses to immunotherapy. Additionally, multivariate Cox regression analyses adjusting for tumor purity, ESTIMATE stromal and immune scores, and deconvoluted lymphocyte proportions further demonstrated that the risk score remained independently associated with overall survival (Supplementary Figure S6).

Robust performance of MRG-based model across different cohorts

Among patients who received chemoradiotherapy screened from the GSE68465 database, the model still maintained robust performance in predicting OS (HR = 6.64, 95% CI 1.48 − 29.88, P = 0.014; Fig. 6A). In the IMvigor210 cohort, patients receiving anti-PD-L1 immunotherapy with low-risk scores demonstrated significant clinical benefits and markedly prolonged survival (HR = 1.48, 95% CI 1.07 − 2.05, P = 0.017; Fig. 6B). Additionally, patients achieving complete response (CR) or partial response (PR) exhibited significantly lower risk scores compared with those with progressive disease (PD) or stable disease (SD), and patients in the low-risk group showed a significantly higher objective response rate (ORR; CR/PR) (P < 0.05; Fig. 6C-D). The risk score was positively correlated with immune checkpoint molecules, especially PD-L1 and cytotoxic T lymphocyte-associated antigen-4 (CTLA-4) (Fig. 6E). Meanwhile, we validated the predictive value of this prognostic model in the melanoma immunotherapy cohort GSE91061 and the NSCLC immunotherapy cohort GSE135222 (Fig. 6F-G; Supplementary Table S5). Consistently, the low-risk group showed better immunotherapy responses and higher C-index and time-dependent AUC values (Fig. 6H; Supplementary Figure S5N-O). These results indicate that low-risk patients may derive greater benefit from immunotherapy compared with high-risk patients. We further demonstrated a significant association between MRG expression and treatment sensitivity/resistance in NSCLC. The low-risk group exhibited significantly higher predicted sensitivity to multiple anticancer agents (Axitinib, CDK9, Erlotinib, Gefitinib, and Rapamycin), whereas the high-risk group showed increased sensitivity to Selumetinib (Supplementary Figure S7). These results suggest that MRG expression modulates therapeutic sensitivity and resistance, highlighting its potential as a biomarker for personalized treatment strategies to improve patient outcomes.

Fig. 6.

Predictive value of the MRG-based model in immunotherapy response. A Survival analysis of patients receiving chemoradiotherapy in GSE68465. B Survival analysis of the IMvigor210 cohort. C Comparison of risk scores among CR, PR, SD, and PD groups in the IMvigor210 cohort. D Objective response rate (ORR) in high- and low-risk groups of the IMvigor210 cohort. E Correlations between risk scores and immune checkpoint molecules. F Survival analysis of the melanoma cohort (GSE91061). G Survival analysis of the melanoma cohort (GSE135222). (H) ORR in high- and low-risk groups of the GSE135222 cohort. (*P < 0.05, **P < 0.01, ***P < 0.001)

CALCOCO2 knockdown and PRKAR1A overexpression induce mitochondrial depolarization in lung cancer cells

To explore the role of MRGs in NSCLC, we first quantified the expression of six MRGs in A549 cells via qRT-PCR. CALCOCO2, CANX, EIF4G1, FOS, and HSP90AB1 were significantly upregulated, while PRKAR1A was markedly downregulated (Fig. 7A). Based on feature ranking derived from SHAP values, CALCOCO2 and PRKAR1A emerged as the top two MRGs most predictive of mitophagic activity, and were selected for further functional validation. Consistent with the transcript-level findings, Western blot analysis revealed elevated CALCOCO2 protein levels and reduced PRKAR1A levels in A549 cells (Fig. 7B). Immunofluorescence staining further confirmed stronger CALCOCO2 signal in A549 cells compared with BEAS-2B cells, with PRKAR1A showing the opposite pattern (P < 0.05; Fig. 7C–D). Given that loss of ΔΨm is a key trigger of mitophagy [36], we assessed ΔΨm using JC-1 staining. Knockdown of CALCOCO2 significantly decreased the red/green fluorescence ratio, indicating mitochondrial depolarization, whereas overexpression of PRKAR1A similarly induced mitochondrial depolarization (Fig. 7E–H).

Fig. 7.

Expression and mitochondrial functional effects of CALCOCO2 and PRKAR1A in BEAS-2B and A549 cells. A qRT-PCR analysis of MRGs in BEAS-2B and A549 cells. CALCOCO2, CANX, EIF4G1, FOS, and HSP90AB1 were upregulated in A549 cells; PRKAR1A was downregulated. B Western blot of CALCOCO2 and PRKAR1A protein levels in BEAS-2B and A549 cells (n = 3). C-D Immunofluorescence staining and quantification of CALCOCO2 and PRKAR1A expression in BEAS-2B and A549 cells (n = 3). (E–F) JC-1 staining and red/green fluorescence ratio showing that CALCOCO2 knockdown decreases mitochondrial membrane potential (n = 3). (G-H) JC-1 staining showing that PRKAR1A overexpression decreases mitochondrial membrane potential (n = 3). (Data are mean ± SEM, n = 3; *P < 0.05, **P < 0.01, ***P < 0.001)

CALCOCO2 and PRKAR1A regulate lung cancer cell migration through TNF signaling and associated EMT responses

We further performed Transwell and wound healing assays to investigate the roles of CALCOCO2 and PRKAR1A in A549 cell migration and invasion. CALCOCO2 knockdown and PRKAR1A overexpression significantly suppressed the migratory and invasive capacities of A549 cells (Fig. 8A–D), indicating that high CALCOCO2 and low PRKAR1A expression cooperatively enhance metastatic potential in NSCLC cells. Next, we performed Western blot analysis following CALCOCO2 knockdown. This revealed a reduction in P62 and phosphorylated AMPK (p-AMPK) levels, while total AMPK remained unchanged (Fig. 8I), indicating that CALCOCO2 may promote autophagic flux via AMPK activation. Notably, TNF signaling has been implicated in AMPK activation through NF-κB and JNK, suggesting that CALCOCO2 might regulate autophagy in a TNF-dependent manner to maintain mitochondrial homeostasis under stress [37]. Likewise, PRKAR1A overexpression led to decreased phosphorylated ERK1/2, with no change in total ERK1/2 levels (Fig. 8J). As ERK1/2 functions downstream of TNF, this points to a potential role for PRKAR1A in modulating TNF–ERK1/2 signaling, possibly influencing mitochondrial activity and cell migration. PRKAR1A overexpression also resulted in increased E-cadherin and decreased Vimentin and Snail expression (Fig. 8J), suggesting a shift toward an epithelial phenotype. While these markers are commonly associated with the epithelial-mesenchymal transition (EMT), they may represent downstream consequences of TNF pathway modulation rather than an independent regulatory axis. Given the established crosstalk between TNF signaling, mitochondrial function, and EMT programs, these findings likely reflect an integrated stress-adaptive response [38]. Taken together, these results indicate that CALCOCO2 and PRKAR1A mainly regulate the migratory and metastatic behavior of NSCLC cells through TNF and EMT-related signaling.

Fig. 8.

Functional validation and mechanistic analysis of CALCOCO2 and PRKAR1A in NSCLC cell migration. A Representative Transwell migration and invasion images of A549 cells after CALCOCO2 knockdown/overexpression (KD/OE). B Quantification of migrated and invaded A549 cells in Transwell assays after CALCOCO2 KD/OE. C Representative Transwell migration and invasion images of A549 cells after PRKAR1A KD/OE. D Quantification of migrated and invaded A549 cells in Transwell assays after PRKAR1A KD/OE. E Representative wound healing assay images of A549 cells after CALCOCO2 KD/OE. F Quantification of the percentage of wound area in A549 cells after CALCOCO2 KD/OE. G Representative wound healing assay images of A549 cells after PRKAR1A KD/OE. H Quantification of the percentage of wound area in A549 cells after PRKAR1A KD/OE. I Western blot analysis showing expression levels of P62, p-AMPK, and total AMPK in A549 cells after CALCOCO2 knockdown. J Western blot analysis showing expression of p-ERK1/2, E-cadherin, Vimentin, and Snail in A549 cells after PRKAR1A overexpression. (Data are mean ± SEM, n = 3; *P < 0.05, **P < 0.01, ***P < 0.001)

CALCOCO2 and PRKAR1A regulate the expression of tumorigenesis-associated pathway biomarkers

Previous analyses indicated that the expression of PRKAR1A and CALCOCO2 is associated with autophagy, the TNF pathway, and inflammatory responses; thus, we hypothesize that these oncogenic pathways are key mechanisms mediating the progression of NSCLC. To validate this hypothesis, we analyzed the correlations between PRKAR1A/CALCOCO2 and the autophagy-related gene PINK1, the TNF pathway gene NFKBIA, as well as the inflammatory response gene CXCL12. TIMER analysis revealed that the expression levels of both PRKAR1A and CALCOCO2 exhibited a positive correlation with PINK1, NFKBIA, and CXCL12 (Fig. 9A-F). Protein–protein interaction (PPI) analysis based on STRING database further confirmed the interactive relationships between PRKAR1A/CALCOCO2 and tumorigenesis-associated pathway biomarker genes (Fig. 9G). Furthermore, qPCR analysis demonstrated that knockdown of CALCOCO2 markedly reduced the expression of these genes, while overexpression of PRKAR1A significantly increased their expression (Fig. 9H-M), thereby validating our hypothesis. Overall, our results highlight opposing roles of CALCOCO2 and PRKAR1A in NSCLC progression, yet both modulate autophagy, TNF signaling, and inflammatory responses.

Fig. 9.

Regulation of tumorigenic pathway biomarkers by CALCOCO2 and PRKAR1A. A–F PINK1, NFKBIA, and CXCL12 exhibit significant positive correlations (analyzed via TIMER3.0, URL: https://compbio.cn/timer3/). G Protein–protein interaction (PPI) network illustrating the interactions between CALCOCO2, PRKAR1A, and these pathway-related biomarkers. H–M Relative mRNA expression of PINK1, NFKBIA, and CXCL12 in the A549 cell line were quantified by qPCR. (Data are mean ± SEM, n = 3; *P < 0.05, **P < 0.01, ***P < 0.001)

Discussion

NSCLC is characterized by profound molecular heterogeneity and inconsistent responses to currently standard-of-care regimens, from platinum-based chemotherapy to ICIs [39, 40]. Existing prognostic tools, such as TNM staging, liquid biopsy biomarkers or gene expression profiling, often fail to capture the dynamic rewiring of tumor metabolism and its intricate interplay with tumor biology, resulting in imprecise patient stratification and ambiguous treatment decisions [41–45]. By embedding mitophagy biology into an interpretable machine-learning framework, we provide a quantifiable link between mitochondrial quality control and clinical outcomes. At the single-cell resolution, mitophagy activity was unevenly distributed across malignant cells, with enrichment in transcriptionally distinct subclusters expressing immune surveillance and immune escape signatures. This observation aligns with emerging evidence that mitochondrial stress response pathways co-opt developmental programs to drive immune evasion and enhance metastatic fitness [46, 47]. Converting this single-cell insight into a cohort-scale tool, we distilled six core MRGs whose expression pattern effectively predicts survival prognosis and benefits from ICI therapy, thereby addressing a blind spot in current prognostic algorithms.

The foundation of our MRG-based model lies in the systematic identification of clinically relevant MRGs, guided by strategies for mining functional gene sets from high-dimensional data. By quantifying mitophagy activity in individual NSCLC cells, we identified MRGs that were not only differentially expressed across cell clusters but also tightly linked to cell type-specific functions, mirroring how single-cell analysis has been used to map functional gene activity to distinct cellular populations in complex tumors [20]. The identified six-gene MRG signature is biologically relevant and provides a clear mechanism for understanding mitophagy in NSCLC. PRKAR1A, the regulatory subunit of protein kinase A (PKA), is downregulated in NSCLC, and its reduced expression promotes tumor aggressiveness [48]. Previous studies have shown that PRKAR1A inhibits PKA-mediated phosphorylation of DRP1 (Ser637) and FUNDC1 (Ser13), thereby linking PKA signaling to mitophagy regulation, which partly explains the observed mitochondrial depolarization and reduced migratory capacity following PRKAR1A overexpression in A549 cells [49–51]. CALCOCO2 (NDP52) and CANX bridge mitochondrial-derived vesicles to the autophagosome and MHC-I loading compartment, respectively, coupling organelle turnover to antigen presentation [52, 53]. Tumors with high MRG expression exhibit upregulation of HSP90AB1 and EIF4G1, which contribute to EGFR/NRF2 stabilization, apoptosis suppression, and enhanced repair, thereby leading to a worse prognosis [54, 55]. Meanwhile, the NRF2-driven antioxidant program dampens IFN-γ signature and antigenicity, while EGFR signaling recruits M2-polarized tumor-associated macrophages; both processes contribute to an unfavorable response to ICIs, consistent with our findings in this study [56].

The performance of our MRG-based model, constructed using GBM algorithms, underscored the value of machine learning in capturing non-linear relationships between molecular signatures and clinical outcomes. The model achieved an AUC of 0.80 in the test set and 0.72 in the validation set, outperforming other machine learning approaches and demonstrating consistent predictive efficacy across multiple independent cohorts. This cross-cohort validation is critical for ensuring algorithm generalizability, as emphasized in prior prognostic studies that demand evaluation beyond the initial training dataset [57]. We also enhanced the model transparency and interpretability using SHAP analysis, a tool increasingly applied to decode “black-box” machine learning models. SHAP analyses revealed that PRKAR1A and CALCOCO2 were the most influential predictors of NSCLC prognosis, while EIF4G1 and FOS contributed modestly but non-redundantly. This granular breakdown of gene contributions mirrors the use of interpretive tools to link individual biomarkers to model performance, ensuring that the model not only predicts outcomes but also yields mechanistic insights into the biological drivers of prognosis.

Mechanistically, CALCOCO2 and PRKAR1A appear to regulate mitochondrial dynamics through distinct pathways. CALCOCO2 enhance autophagic flux via AMPK activation, whereas PRKAR1A acts mainly through PKA-dependent TNF–ERK signaling. Although they operate through different regulatory axes, both ultimately help preserve mitochondrial polarization and cellular metabolic stability under stress, supporting the adaptive capacity of NSCLC cells [58]. Specifically, CALCOCO2-driven mitochondrial polarization may enhance ATP production and sustain autophagic flux, thereby supporting tumor growth and survival. In contrast, loss of PRKAR1A is associated with sustained mitochondrial polarization and enhanced metabolic activity, contributing to tumor aggressiveness, whereas its overexpression induces mitochondrial depolarization and suppresses cell migration [48]. While we did not directly quantify TNF-α or its canonical downstream effectors (NF-κB, JNK), the PRKAR1A-induced reduction in phospho-ERK1/2, which is an established target of TNF signaling, suggests that TNF-ERK crosstalk may relay metabolic stress cues to the mitochondrial machinery [38]. Further functional studies and in vivo xenograft models are required to explore how the dominance of CALCOCO2 and PRKAR1A may vary in different microenvironments.

A key strength of our model is its ability to bridge MRG expression with clinically relevant phenotypes, including TME composition and treatment response. We found that NSCLC patients with low MRG scores exhibited higher immune and ESTIMATE scores, lower tumor purity, and reduced infiltration of immunosuppressive cells, which were associated with enhanced anti-tumor immunity [59, 60]. This aligns with prior analyses that link functional gene signatures to TME remodeling, highlighting how MRGs may shape the immune landscape of NSCLC [61–63]. In the IMvigor210 trial, patients with low scores exhibited significantly better clinical responses and longer OS under anti-PD-L1 therapy, whereas those with high-score showed greater predicted sensitivity to Selumetinib. These findings extend the prognostic value of MRG score to treatment selection, and may address a critical unmet need in clinical practice: identifying patients most likely to benefit from ICIs versus chemotherapy. This emphasis on therapeutic stratification reflects the growing effort to translate prognostic biomarkers into actionable instruments for clinical decision-making through AI-driven precision oncology [64].

Despite these strengths, our study has several limitations. First, our analysis relied on retrospective datasets, which may carry inherent heterogeneity and potential bias. Nevertheless, we validated the findings in a prospective clinical trial, confirming the robustness of the model for clinical utility. Second, although we have linked these key genes to mitochondrial autophagy and NSCLC prognosis using bioinformatics analyses and in vitro experiments, further functional analyses, such as in situ xenograft models, are needed to establish causality and to elucidate the molecular mechanisms. Last, owing to limited data resources, our model does not integrate clinical variables or other multi-omics data, including proteomics and metabolomics, which warrant further investigations in the future.

In conclusion, our MRG-based model addresses a critical gap in NSCLC management by integrating single-cell-derived mitophagy biology with AI algorithms. The model’s robust performance across cohorts, interpretability via SHAP analysis, and ability to guide treatment decisions highlight its potential for clinical translation. Future work should focus on large-scale real-world validation, functional characterization of core MRGs, and integration of clinical and multi-omics data to refine risk stratification. By prioritizing both biological relevance and clinical utility, our framework not only advances understanding of mitophagy’s role in NSCLC but also provides a practical tool for personalized management and improved clinical outcomes.

Abbreviations

NSCLC

Non-small-cell lung cancer

TCGA

The Cancer Genome Atlas

GEO

Gene Expression Omnibus

MRGs

Mitophagy-related genes

scRNA-seq

Single-cell RNA sequencing

KEGG

Kyoto Encyclopedia of Genes and Genome

GO

Gene Ontology

GSEA

Gene Set Enrichment Analysis

IHC

Immunohistochemistry

EMT

Epithelial-mesenchymal transition

OS

Overall survival

AUC

Area under the curve

Authors’ contributions

M.W.: study conceptualization, formal analysis, visualization, and writing–original draft. Y.W., Y.L. and D.W.: study conceptualization, methodology, and writing–review and editing. M.L. and C.Z: visualization and investigation. Y.L. and N.B.: project administration, validation, and writing–review and editing.

Funding

This work was supported by the National Natural Science Foundation of China (82373216, 82173348, and 12405407), the CAMS Innovation Fund for Medical Sciences (2024-I2M-ZD-004), and the Noncommunicable Chronic Diseases–National Science and Technology Major Project (2023ZD0502100). The funders had no role in the study design, data collection, analysis, interpretation, report writing, or the decision to submit the paper.

Data availability

The TCGA, GSE117570, GSE198099, GSE68465, GSE31210, GSE73403, and GSE91061 datasets analyzed in this study are available from the TCGA (https://portal.gdc.cancer.gov/) and GEO (http://www.ncbi.nlm.nih.gov/geo/). Additionally, the IMvigor210 immunotherapy dataset can be downloaded from (http://research-pub.gene.com/IMvigor210CoreBiologies).

Declarations

Ethics approval and consent to participate

This study did not involve the collection of new human or animal data. All analyses were conducted using publicly available datasets from TCGA, GEO and HPA, which were accessed and used in accordance with the corresponding data access policies and guidelines. As these datasets are de-identified and publicly accessible, additional institutional ethical approval was not required.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.