Metabolic reprogramming signature predicts immunotherapy efficacy in lung adenocarcinoma: Targeting SLC25A1 to overcome immune resistance

Abstract

Objective

Lung adenocarcinoma (LUAD) is the most common subtype of lung cancer. Despite significant advances in immunotherapy, treatment responses vary substantially among individuals. Metabolic reprogramming, as a hallmark of cancer, plays a crucial role in tumor progression and immune evasion. However, the interplay between metabolic features and tumor immune microenvironment in LUAD remains to be systematically elucidated.

Methods

We analyzed data from 1,231 LUAD patients across seven global cohorts and developed an integrated Metabolism-Related Signature (iMRS) using machine learning approaches based on 114 metabolic features. The signature’s ability to predict immunotherapy response was validated using 9 immunotherapy cohorts (n=712, including LUAD, melanoma, and glioma). An in-house LUAD tissue cohort (n=146) confirmed the prognostic significance of SLC25A1, a key gene within the signature, and its spatial relationship with immune cells. In vivo and in vitro experiments investigated SLC25A1’s role in cancer promotion, immune exclusion, and its impact on programmed cell death protein 1 (PD-1) therapy efficacy.

Results

iMRS demonstrated superior prognostic performance in LUAD patients, outperforming 129 published LUAD signatures. In immunotherapy cohorts, responders showed significantly lower iMRS scores. High iMRS was associated with reduced immune activity and “cold” tumor characteristics. SLC25A1 (correlation coefficient=0.54, P<0.05), a key gene in the signature, showed the highest expression in CD8 desert phenotype and correlated with poor prognosis. Multiplexed immunofluorescence revealed exclusion patterns between SLC25A1 and immune cells (CD4+ T cells and CD20+ B cells). SLC25A1 knockdown reduced lung metastasis and enhanced anti-PD-1 efficacy by increasing CD8+ T cell abundance and cytotoxicity [increased interferon-γ (IFN-γ)+/GZMB+ CD8+ T cells].

Conclusions

iMRS provides personalized immunotherapy prediction for LUAD patients. SLC25A1, identified as a novel immune-exclusion related oncogene, represents a promising therapeutic target for LUAD treatment.

Introduction

Lung cancer is the most prevalent and deadliest malignant tumor worldwide (1), with lung adenocarcinoma (LUAD) being the most common subtype of non-small cell lung cancer (NSCLC), accounting for 35%−40% of cases (2). Although surgical intervention remains the preferred treatment for early-stage LUAD (3), the majority of patients present with metastatic disease at initial diagnosis. For patients with advanced metastatic LUAD, chemotherapy and radiotherapy are alternative treatment options. However, cisplatin-based adjuvant chemotherapy shows limited efficacy in improving patient outcomes, with a 5-year survival rate of only 5% (4). The advent of immunotherapy has revolutionized the treatment of advanced LUAD by alleviating the suppression of tumor-infiltrating lymphocytes (TILs) through immune checkpoint inhibition, thereby enhancing TIL activation and promoting the clearance of tumor cells. Despite significant improvements in the prognosis of LUAD patients, the overall survival rate remains low, with a 5-year survival rate of merely 26% (5). Many patients do not respond to immune checkpoint inhibitors (ICIs), partly due to the lack of functional T cell infiltration within the tumor microenvironment (TME) of LUAD, which is characterized as an immunogenic “cold” tumor. This type of tumor exhibits insensitivity to immunotherapy. Previous studies in other tumors suggest that increasing antigenicity, enhancing effector T cell function, and overcoming immune suppressive factors within the TME may hold promise for converting the “cold” tumors of LUAD into “hot” tumors (6).

Moreover, analyzing TILs alone is insufficient to fully characterize the complex tumor immune microenvironment. Additionally, tumors with high levels of TIL infiltration may still exhibit resistance to immunotherapy. Consequently, various biomarkers have been developed to predict responses to immunotherapy, including programmed death ligand 1 (PD-L1) expression levels (7), tumor mutational burden (TMB) (8), microsatellite instability (MSI) (9), and circulating tumor DNA (ctDNA) (10). These biomarkers provide crucial reference points that can assist clinicians in personalizing treatment for patients.

To achieve rapid proliferation, tumor cells undergo metabolic reprogramming, favoring aerobic glycolysis and lactate production, a phenomenon known as the Warburg effect (11). TILs are often impaired in their anti-tumor immune responses due to metabolic dysregulation induced by tumor cell metabolism, including the release of metabolites such as lactate that affect immune molecules (12). Increasing evidence suggests that alterations within the TME significantly influence responses to immunotherapy (13). The TME typically induces metabolic reprogramming, which can modulate the energy consumption of immune cells, leading to immune dysfunction that hinders the clearance of tumor cells (14). In this context, using metabolic modulators to reverse immune suppression and enhance TIL infiltration may serve as a therapeutic strategy (15). In summary, metabolic changes can affect immune function and disrupt the efficacy of immunotherapy (16).

Targeting cancer cell metabolism to overcome immunotherapy resistance and identifying biomarkers to predict therapeutic responses are feasible strategies. Therefore, a systematic understanding of the relationship between immune responses and metabolic regulation is essential. In this study focused on LUAD, we systematically analyzed the expression patterns of 114 metabolism-related genes and their associations with the tumor immune microenvironment. Based on these analyses, we developed an integrated Metabolism-Related Signature (iMRS) to predict patient prognosis and immunotherapy outcomes. Furthermore, we identified and validated key genes within the iMRS and evaluated their potential therapeutic value in LUAD immunotherapy, providing a foundation for precise patient stratification and individualized treatment.

Materials and methods

Dataset source

The model development phase was initiated through the acquisition of LUAD specimens from The Cancer Genome Atlas (TCGA) repository (https://portal.gdc.cancer.gov), which encompassed transcriptomic profiles, histopathological specimens stained with hematoxylin and eosin (H&E), and pertinent clinical documentation. For external validation purposes, six independent transcriptome cohorts were extracted from the Gene Expression Omnibus (GEO) platform (https://www.ncbi.nlm.nih.gov/geo/), with sample sizes ranging from 39 to 227 subjects across different investigations [GSE13213 (17), GSE26939 (18), GSE29016 (19), GSE30219 (20), GSE31210 (21), and GSE42127 (22)]. The immunotherapeutic response prediction capability was evaluated through multiple cohort analyses, including five specialized NSCLC immunotherapy studies [the OAK (23), POPLAR (24), and three GEO datasets: GSE166449 (25), GSE207422 (26), and GSE135222 (27)]. The validation scope was further expanded by incorporating diverse pan-cancer immunotherapy populations, consisting of three independent melanoma research cohorts [identified as GSE91061 (28), GSE78220 (29), and phs000452 (30)] and a dedicated glioblastoma investigation [PRJNA482620 (31)]. The cellular heterogeneity analysis was facilitated through three complementary single-cell sequencing resources: the CRA001963 dataset (32) from the National Genomics Data Center, the GSE123904 collection (33) from GEO, and the PRJNA591860 repository (34) from the National Center for Biotechnology Information Sequence Read Archive. A detailed inventory of all utilized datasets is systematically cataloged in Supplementary Table S1. The metabolism-related genes (MRGs) were identified through previous investigations (35).

Expression matrices from supplementary cohorts were extracted directly from their corresponding data repositories. Statistical normalization was implemented through log2 transformation and z-score standardization. Inter-batch variations were eliminated through the implementation of the “combat” algorithm within the “sva” statistical package (36).

Identification of prognostic MRS

The identification of differentially expressed MRGs (DEMRGs) between neoplastic and non-neoplastic specimens was accomplished through the application of the limma analytical framework. Statistical significance thresholds were established at false discovery rate (FDR) <0.05 and log2 fold change (FC) >1. Prognostic assessment of genetic markers in LUAD patients was conducted via single-variable Cox proportional hazards modeling. The optimization process incorporated a comprehensive evaluation of 101 algorithmic combinations through 10-fold cross-validation methodology. The computational approaches encompassed various statistical frameworks: stepwise Cox regression, Lasso penalization, Ridge regression, partial least squares regression adapted for Cox modeling (plsRcox), CoxBoost enhancement, random survival forest (RSF) methodology, generalized boosted regression modeling (GBM), elastic net (Enet) regularization, supervised principal components (SuperPC) analysis, and survival-oriented support vector machine (survival-SVM) classification. This systematic evaluation aimed to establish an iMRS configuration, with concordance index (C-index) serving as the primary performance metric. The predictive accuracy of the developed iMRS was validated through receiver operating characteristic (ROC) curve analysis and principal component analysis (PCA) methodology. The comparative assessment phase incorporated 129 previously documented prognostic signatures, encompassing both long non-coding RNA and messenger RNA markers, with C-index serving as the comparative performance indicator.

Investigating potential mechanisms and pathways

The functional characterization of DEMRGs was executed through comprehensive pathway analyses. Gene Ontology (GO) annotations and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway mapping were performed utilizing the “clusterProfiler” analytical suite (37). The analytical workflow commenced with the conversion of DEMRGs to their corresponding Entrez identifiers, followed by enrichment analyses with statistical significance determined at an adjusted P value below 0.05. Mechanistic insights were further elucidated through gene set variation analysis (GSVA) methodology. The immunological aspects of cancer progression were examined through detailed analysis of the cancer immunity cycle and immunotherapy response pathways, following previously validated protocols (38,39). The pathway-based investigation was supported by reference gene sets obtained from the MSigDB repository (https://www.gsea-msigdb.org/gsea/msigdb/).

Immune profiling analysis

The evaluation of immunotherapeutic responsiveness in LUAD cases was facilitated through immunophenoscore (IPS) assessment via The Cancer Immunome Atlas platform (https://tcia.at/home). TME characterization was accomplished through the single sample gene set enrichment analysis (ssGSEA) computational approach, which quantified both immune cell infiltration patterns and immune pathway activation states within neoplastic specimens. Comprehensive immune infiltration profiles across TCGA samples were accessed through the TIMER2.0 platform, which synthesized outcomes from multiple computational methods.

Single-cell RNA sequencing (scRNA-seq) data processing

The initial transcriptomic data underwent preprocessing via Seurat R package (40). The analytical inclusion criterion was established at a minimum expression threshold of 10 cells per gene within individual samples. Cell quality filtration was implemented based on predetermined parameters: cellular units were excluded if they expressed more than 5,000 or fewer than 200 genes, or if mitochondrial genome-derived unique molecular identifiers (UMIs) exceeded 10%. Sample integration was accomplished through the harmony R package implementation. The subsequent analytical pipeline incorporated highly variable gene selection for PCA, followed by dimensional reduction utilizing the top 30 significant PCs through Uniform Manifold Approximation and Projection (UMAP) methodology. Subpopulation-specific transcriptional profiles were delineated using the “FindAllMarkers” algorithmic function, with cellular phenotype classification based on established lineage-specific markers from prior investigations (41).

Cell-cell signaling networks

Intercellular communication patterns were decoded through CellChat (42) analytical framework integration with expression profiles. The analysis utilized CellChat’s native ligand-receptor reference database, following standardized protocols. The methodology enabled identification of cell-type specific interaction patterns through detection of preferentially expressed signaling molecules. Communication networks were mapped based on elevated expression patterns of either receptor components or their corresponding ligands within distinct cellular populations.

Patient cohort and sample collection

Clinical specimens included paraffin-embedded tissue sections from the Department of Pathology, Tianjin Medical University Cancer Institute and Hospital. All specimens were pathologically confirmed as LUAD, and only treatment-naive cases before surgical intervention were included. The study was conducted in accordance with the Declaration of Helsinki and was approved by the Ethics Committee of Tianjin Medical University Cancer Institute and Hospital (No: bc2022232), and written informed consent was obtained from all participating subjects. Comprehensive clinical and pathological parameters are shown in Supplementary Table S2.

Immunohistochemistry (IHC) analysis

Tissue processing began with deparaffinization in xylene, followed by rehydration through graded ethanol. After treatment with 3% hydrogen peroxide and blocking with 5% goat serum, sections were incubated overnight at 4 °C with primary antibody against SLC25A1 (1:300, Cat# ab318201, Abcam, UK). Subsequent processing included application of HRP-labeled secondary antibody (30-min incubation), DAB chromogen development, and hematoxylin counterstaining. Assessment was independently performed by two pathologists. SLC25A1 expression was semi-quantitatively evaluated using the immunoreactivity score, calculated as the product of the percentage score (0: no positive cells; 1: ≤10%; 2: 11%−50%; 3: 51%−80%; 4: >80%) and the staining intensity score (0: negative; 1: weak; 2: moderate; 3: strong), with total scores ranging from 0 to 12.

Multiplex IHC analysis protocol

Spatial relationships between SLC25A1 and immune markers were evaluated using multiplex immunofluorescence staining. Tissue processing began with deparaffinization in xylene, followed by rehydration through graded ethanol. After antigen retrieval and blocking with 5% goat serum, sections were sequentially incubated with primary antibodies: SLC25A1 (1:400, Cat# ab318201, Abcam), CD4 (1:600, Cat# ab133616, Abcam), and CD20 (1:100, Cat# ab64088, Abcam), followed by corresponding fluorophore-conjugated secondary antibodies. Nuclear counterstaining was performed with DAPI. T cell infiltration levels were semi-quantitatively assessed by evaluating the percentage of positive cells in the tumor stroma (0: no positive cells; 1: ≤10%; 2: 11%−50%; 3: 51%−80%; 4: >80%). Based on the spatial distribution patterns of T cells, tumors were classified into three phenotypes: inflamed (abundant T cell infiltration within tumor nests), excluded (T cells mainly aggregated in tumor stroma or marginal areas with little or no infiltration into tumor nests), and desert (few or absent T cells in tumor tissue).

Cell culture and RNA interference

LUAD cell lines (A549 and H1299) were obtained from the Institute of Biochemistry and Cell Biology, Chinese Academy of Sciences (Shanghai). Cells were maintained in RPMI 1640 medium supplemented with 10% foetal bovine serum (FBS) and antibiotics (100 U/mL penicillin, 100 mg/mL streptomycin) at 37 °C in a 5% CO₂ incubator. Mouse Lewis lung carcinoma (LLC) cells were cultured under the same conditions. Gene knockdown was achieved through lentiviral transduction, followed by puromycin selection (2 μg/mL) for 48 h. Knockdown efficiency was validated by quantitative real-time polymerase chain reaction (qRT-PCR) analysis.

Cell proliferation assay

Cells were dissociated with 0.25% trypsin, and A549 and H1299 cells were seeded in 96-well plates (2×10³ cells/well, five replicate wells per group) and divided into three groups: blank control (NC), vector control (vector), and SLC25A1 knockdown (sh-SLC25A1). After 24-h adherence, designated as the 0-h time point, cell viability was assessed at 0, 24, 48, 72, and 96 h. At each time point, 10 μL of CCK8 reagent (Dojindo, Kumamoto, Japan) was added to each well and incubated at 37 °C for 2 h in the dark. The optical density (OD) was measured at 450 nm using a microplate reader. The experiment was independently repeated three times.

Murine metastasis model

Female C57BL/6 mice (6−8 weeks, Jiangsu Gempharmatech) were randomly divided into groups (n=6 per group) and received tail vein administration of lentivirus-transduced LLC cells (control or SLC25A1 knockdown groups, 5×10⁵ cells/mouse). The therapeutic protocol comprised intraperitoneal injection of anti-PD-1 antibody (200 μg/mouse) or isotype control on d 6, 9, 12, and 15 post-injection. Terminal analysis occurred on d 21, with humane euthanasia performed by cervical dislocation under isoflurane anesthesia. Pulmonary tissues were collected and either fixed in 4% paraformaldehyde for histological analysis or processed for flow cytometric analysis. All animal experiments were conducted in compliance with the institutional guidelines for the care and use of laboratory animals and approved by the Ethics Committee of the Tianjin Medical University Cancer Institute and Hospital (No. 2023078).

Flow cytometric analysis

Fresh tumor specimens were enzymatically dissociated in RPMI 1640 medium supplemented with collagenase IV (Gibco 17104019, NY, USA) and DNase I (Roche 10104159001, Basel, Switzerland) at 37 °C for 2 h with gentle agitation. Single-cell suspensions were obtained by passing through 70-μm cell strainers (BD Falcon, Franklin Lakes, USA), followed by lymphocyte isolation using Percoll density gradient centrifugation (40%/70%, 800 g for 20 min at room temperature). The isolated cells were stimulated with leukocyte activation cocktail containing PMA/ionomycin and protein transport inhibitor (BD Pharmingen 550583) at 37 °C for 4 h. For immunostaining, cells were first blocked with anti-mouse CD16/CD32 antibody (BD 553141) at 4 °C for 15 min, then incubated with fluorochrome-conjugated antibodies against surface markers: CD45 (BD 553080), CD3e (BD 553064), and CD8α (BD 551162) for 30 min at 4 °C in the dark. Following surface staining, cells were fixed and permeabilized using Cytofix/Cytoperm solution (BD Biosciences 554714) according to the manufacturer’s protocol. Subsequently, intracellular staining was performed with anti-GZMB (Biolegend 372204) and anti-IFN-γ (Biolegend 505808) antibodies at 4 °C for 30 min. Data acquisition was performed on a BD FACS Canto II flow cytometer, and data analysis was conducted using FlowJo software (Version 10.0; BD Biosciences, Franklin Lakes, USA).

Statistical analysis

Statistical analyses were performed using R software (Version 4.2.0; R Foundation for Statistical Computing, Vienna, Austria). For comparisons between two groups, unpaired Student’s t test was applied for normally distributed data, while Mann-Whitney U test was used for non-normally distributed data. For multiple group comparisons, one-way Analysis of Variance (ANOVA) followed by Tukey’s post hoc test was employed for normally distributed data, while Kruskal-Wallis test followed by Dunn’s post hoc test was used for non-normally distributed data. Correlation analysis was performed using Pearson’s correlation coefficient. Data are presented as . Statistical significance was set at P<0.05.

Results

Metabolic reprogramming in LUAD

To explore the role of different metabolic features in LUAD, we established a comprehensive analytical framework as illustrated in Figure 1. We first performed ssGSEA analysis on the TCGA-LUAD dataset to evaluate metabolic differences between LUAD and normal tissues. The results revealed distinct alterations in several metabolic pathways, notably hexosamine biosynthesis, pyrimidine biosynthesis, and biotin metabolism (Figure 2A). Subsequently, we calculated the expression differences of metabolism-related pathway genes between tumor and normal tissues (Supplementary Figure S1) and further illustrated their chromosomal distribution (Figure 2B). After integrating multiple LUAD cohorts and immunotherapy datasets, we applied normalization and batch effect correction to ensure the comparability and robustness of the data (Figure 2D). To elucidate the biological significance of these metabolic changes, we conducted GO and KEGG enrichment analyses on MRGs. The analyses revealed significant enrichment in several key metabolic processes, including carbon metabolism, glycolysis, lyase activity, and ribonucleotide metabolism (Figure 2C,E). We further performed copy number variation (CNV) analysis to explore potential genomic mechanisms underlying these metabolic alterations. The results showed substantial amplifications in several metabolic regulators, particularly ST3GAL5 and SPTLC1 (Figure 2F). Finally, a forest plot demonstrated the prognostic significance of these DEMRGs in the TCGA-LUAD cohort (Figure 2G).

Figure 1.

Schematic overview of the comprehensive analysis workflow integrating metabolic pathways to explore their role in LUAD and guide immunotherapy strategies. The study incorporated multi-cohort analysis, machine learning approaches, and experimental validations to develop and validate the iMRS. The workflow includes clinical cohort analysis, computational modeling, and functional experiments, with a particular focus on SLC25A1’s role in immune regulation and tumor progression. iMRS, integrated Metabolism-Related Signature; LUAD, lung adenocarcinoma; TCGA, The Cancer Genome Atlas; RSF, random survival forest; GBM, generalized boosted regression modeling; NSCLC, non-small cell lung cancer; ROC, receiver operating characteristic; TME, tumor microenvironment.

Figure 2.

Metabolic pathway alterations and differential gene expression analysis in LUAD. (A) ssGSEA analysis revealing differential metabolic pathway activities between LUAD and normal tissues; (B) Chromosomal distribution of DEMRGs, with red dots indicating higher expression in tumor tissues and black dots indicating higher expression in normal tissues. The central PCA plot demonstrates the distribution of samples from seven LUAD cohorts after batch effect removal; (C) KEGG pathway enrichment analysis of DEMRGs; (D) Pie chart showing the distribution of all cohorts used in this study, with segment sizes proportional to sample numbers; (E) Gene Ontology enrichment analysis of DEMRGs, categorized by BP, CC, and MF; (F) Copy number variation frequency of metabolism-related genes across chromosomes, with amplifications shown in orange and deletions in blue; (G) HR analysis of metabolism-related genes in LUAD patients, with orange dots (HR>1) indicating poor prognosis and blue dots (HR<1) indicating better prognosis. LUAD, lung adenocarcinoma; ssGSEA, single sample gene set enrichment analysis; DEMRG, differentially expressed metabolism-related gene; PCA, principal component analysis; KEGG, Kyoto Encyclopedia of Genes and Genomes; HR, hazard ratio; BP, biological processe; CC, cellular component; MF, molecular function.

iMRS development for cancer immunotherapy

Accumulating evidence has demonstrated the crucial role of metabolic reprogramming in modulating tumor immune responses and immunotherapy outcomes. Multiple studies have revealed that metabolic characteristics significantly influence immune cell function and immunotherapy efficacy. However, a comprehensive metabolic signature that can effectively predict immunotherapy response remains to be established. To address this unmet clinical need, we developed an iMRS by systematically analyzing metabolism-related genes and their associations with immunotherapy efficacy. To construct a robust predictive model, we strategically selected 78 DEMRGs that demonstrated significant associations with both metabolic processes and clinical prognosis. We employed multiple machine learning algorithms and systematically evaluated their pairwise combinations to identify the optimal modeling approach. Through comprehensive comparison of various algorithm combinations using mean C-index as the evaluation metric, the Lasso+GBM combination emerged as the best-performing method, showing superior predictive capability across multiple validation datasets (Figure 3A). To establish the reliability and generalizability of iMRS, we conducted a systematic validation process. Initial internal validation using TCGA-LUAD as the training cohort and six independent GEO datasets as validation cohorts demonstrated significant prognostic stratification capability (P<0.05). Consistently across all seven cohorts, high iMRS scores were associated with poor survival outcomes (Figure 3B−H). Given the potential clinical utility in immunotherapy, we extended our validation to diverse immunotherapy cohorts. This external validation encompassed three melanoma studies, one glioblastoma dataset (Figure 3I−L), and five NSCLC cohorts (Figures 3M−Q). Notably, patients with low iMRS scores achieved significantly better responses to immunotherapy, suggesting its potential as a predictive biomarker. To elucidate the underlying biological basis, we performed IPS analysis using TCIA data. This mechanistic investigation demonstrated that the low iMRS group exhibited enhanced responsiveness across various immunotherapy modalities, including PD-1 inhibition, CTLA4 blockade, and combination treatments (Supplementary Figure S2A−D). These findings not only validated the predictive value of iMRS but also provided insights into its biological relevance in immunotherapy response.

Figure 3.

Development and validation of iMRS across multiple cohorts and its predictive value for immunotherapy response. (A) Heatmap showing the performance of different machine learning algorithms in constructing iMRS, with color intensity indicating prediction accuracy; (B−H) Kaplan-Meier survival analysis of seven LUAD cohorts stratified by iMRS scores, consistently showing significantly worse survival in the high-iMRS group; (I−L) Survival analysis in Mel and GBM immunotherapy cohorts demonstrating the broad applicability of iMRS; (M,N) Validation of iMRS in two independent LUAD immunotherapy cohorts (OAK-LUAD and POPLAR-LUAD), showing significant survival differences between high- and low-iMRS groups; (O−Q) Box plot comparing iMRS scores between immunotherapy R and NR in three independent cohorts, with NRs consistently showing higher iMRS scores (Wilcoxon test). iMRS, integrated Metabolism-Related Signature; LUAD, lung adenocarcinoma; Mel, melanoma; GBM, glioblastoma; R, responders; NR, non-responders.

Comparative performance analysis of iMRS

To evaluate the clinical utility of iMRS, we first compared it with conventional clinical parameters [age, gender, stage, epidermal growth factor receptor (EGFR) status] across seven independent datasets. The analysis demonstrated iMRS’s superior predictive capability, consistently achieving higher C-index values (Figure 4A). To investigate the molecular basis of iMRS stratification, we performed PCA on the model gene expression patterns. The results revealed distinct clustering between high- and low-iMRS groups, supporting the biological relevance of our signature (Figure 4B). To assess the temporal stability of prognostic prediction, we conducted time-dependent ROC analysis. The iMRS maintained robust predictive accuracy with area under the curve (AUC) values consistently exceeding 0.55 for one-, three-, and five-year survival predictions across all cohorts (Figure 4C). To establish the superiority of our model, we benchmarked iMRS against 129 previously published LUAD signatures. This comprehensive comparison demonstrated iMRS’s exceptional performance, achieving leading C-index scores among all existing signatures (Figure 4D). Collectively, these comprehensive analyses establish iMRS as a robust and clinically applicable prognostic tool, outperforming existing signatures and conventional clinical parameters in LUAD risk stratification.

Figure 4.

Evaluation of iMRS prognostic performance across multiple LUAD cohorts. (A) Bar plot comparing C-index values of iMRS with various clinical features across different cohorts, demonstrating the robust prognostic capability of iMRS; (B) PCA plot showing distinct clustering of high- and low-iMRS groups (orange and blue dots, respectively) across seven independent LUAD cohorts; (C) Time-dependent ROC curves for 1-, 3-, and 5-year survival predictions across seven cohorts, with corresponding AUC values demonstrating consistent predictive accuracy of iMRS; (D) Comparative analysis of prognostic performance between iMRS and previously published signatures using C-index as the evaluation metric across seven LUAD cohorts. iMRS, integrated Metabolism-Related Signature; LUAD, lung adenocarcinoma; PCA, principal component analysis; ROC, receiver operating characteristic; AUC, area under the curve.

Immune microenvironment analysis

To characterize the immune landscape associated with iMRS stratification, we first examined immune cell infiltration patterns in the TCGA-LUAD cohort. The analysis revealed significantly enhanced immune cell infiltration in the low iMRS group (Figure 5A). To quantitatively assess TME composition, we applied the ESTIMATE algorithm. The results demonstrated significant negative correlations between iMRS scores and stromal, immune, and ESTIMATE scores, while showing positive correlation with tumor purity (all P<0.001) (Figure 5B−E). To further delineate the immune context, we performed ssGSEA profiling, which confirmed differential patterns of immune cell infiltration and pathway activation between iMRS groups (Figure 5F,G). Notably, the low iMRS group exhibited elevated expression of immune checkpoint molecules (TIGIT, BTLA, CTLA4) and costimulatory factors (most P<0.05) (Figure 5H). To validate these findings at the histological level, we reviewed TCGA H&E slides, which confirmed increased lymphocytic infiltration in low iMRS samples (Figure 5I,J). These observations collectively suggest an immunologically active microenvironment in low iMRS tumors that may be more conducive to immunotherapy response (43).

Figure 5.

Comprehensive analysis of immune characteristics between high- and low-iMRS groups. (A) Heatmap showing differential immune cell infiltration patterns between high-and low-iMRS groups based on single-cell analysis; (B−E) Correlation analysis between iMRS and various immune-related scores: stromal score (B), immune score (C), ESTIMATE score (D), and tumor purity (E), with corresponding Pearson correlation coefficients (r) and q-values; (F) Radar plot comparing immune-related pathway activation between high- and low-iMRS groups; (G) Radar plot displaying the relative abundance of different immune cell populations between high- and low-iMRS groups; (H) Expression differences of immune checkpoint-related genes (co-stimulators, co-inhibitors, and MHC molecules) between high- and low-iMRS groups, with statistical significance indicated by asterisks; (I,J) Representative H&E staining images of tumor samples from high (I) and low (J) iMRS groups, with magnified insets showing detailed tissue architecture (×20). iMRS, integrated Metabolism-Related Signature; MHC, major histocompatibility complex; H&E, hematoxylin and eosin. *, P <0.05; **, P <0.01; ***, P <0.001.

Single-cell analysis of iMRS distribution and cellular communication

We integrated three single-cell cohorts and annotated 13 cell types in total (Figure 6A,B). SLC25A1 expression was mainly concentrated in epithelial cells, macrophages, and proliferating cell populations (Figure 6C). The cell-level iMRS score was computed on a per-cell basis by directly applying our existing machine learning model, using its original formula, to the single-cell normalized expression matrix. The violin plot showed that proliferating cell populations exhibited peak iMRS scores (Figure 6D), indicating a direct association between high iMRS and cellular proliferation. Cell type composition analysis demonstrated distinct immune cell distributions, with macrophage enrichment in high-iMRS groups, contrasting with reduced frequencies of B cells, T cells, and natural killer (NK) cells (Figure 6E). Detailed intercellular communication analysis (Figure 6F−I,K−N) revealed enhanced signaling networks in high-iMRS populations, characterized by increased bidirectional signaling intensity (Figure 6K,L). Notably, the high-iMRS group exhibited upregulation of pro-tumorigenic pathways, including EGF, VEGF, and CDH1 signaling cascades (Figure 6J).

Figure 6.

Single-cell analysis of TME and cell-cell interactions in relation to iMRS. (A) UMAP visualization of integrated scRNA-seq data showing major cell types in the TME; (B) UMAP plot colored by data source, showing the integration of three independent scRNA-seq datasets; (C) Expression pattern of SLC25A1 across different cell types; (D) Violin plot showing iMRS score distribution across different cell types (Kruskal-Wallis test, P<2.2e−16); (E) Stacked bar plot showing the relative proportion of cell types in high- and low-iMRS groups; (F,G) Comparison of total number of cell-cell interactions (F) and interaction strength (G) between high- and low-iMRS groups; (H,I) Heatmaps showing differential number of interactions (H) and interaction strength (I) between cell types in the TME; (J) Relative signaling pathway activities between high- and low-iMRS groups, with red indicating enhanced activity in high-iMRS group and blue indicating enhanced activity in low-iMRS group; (K,L) Scatter plot showing incoming vs. outgoing interaction strength for different cell types in low- (K) and high- (L) iMRS groups; (M,N) Network visualization of differential number of interactions (M) and interaction strength (N) between cell types in high- vs. low-iMRS groups. TME, tumor microenvironment; iMRS, integrated Metabolism-Related Signature; UMAP, Uniform Manifold Approximation and Projection; scRNA-seq, single-cell RNA sequencing.

SLC25A1: A novel immunometabolic target in LUAD

Investigation into iMRS-associated biomarkers was conducted, with particular emphasis placed on SLC25A1. This gene was previously established as a significant risk factor, exhibiting elevated hazard ratios (HRs) (HR>1, P<0.05) (Figure 2G). Analysis of Kaplan-Meier plotter database demonstrated diminished survival rates among LUAD cases with heightened SLC25A1 expression (Supplementary Figure S3A). Statistical evaluation revealed substantial positive association between SLC25A1 and iMRS indices (correlation coefficient=0.54, q=0) (Supplementary Figure S3B). Cellular distribution patterns, elucidated through single-cell RNA sequencing, demonstrated preferential SLC25A1 localization within epithelial and proliferative cellular populations (Figure 6C). Given limited published data regarding SLC25A1’s oncogenic mechanisms, comprehensive in vivo and in vitro investigations were undertaken. Inverse correlations between SLC25A1 and CD8A were observed across both TCGA-LUAD and OAK cohorts (r<0, P<0.05) (Figure 7A,B). Stratification analysis based on median SLC25A1 expression levels revealed significantly reduced survival rates among high-expression subjects (P<0.05) (Figure 7C,D). Molecular profiling demonstrated inverse relationships between SLC25A1 and immunological factors, including chemokines, receptors, major histocompatibility complex (MHC) molecules, and immune regulatory elements (Supplementary Figure S3C,D). CD8+ T cell spatial distribution analysis facilitated phenotypic classification into inflammatory, excluded, and desert subtypes. Superior clinical outcomes were documented in inflammatory phenotype cases (Figure 7E), corroborating established findings (42). Differential SLC25A1 expression patterns were observed across phenotypes, with minimal levels in inflammatory, maximal in desert, and intermediate in excluded variants (Figure 7F). Quantitative multiplex IHC revealed enhanced CD20+ B and CD4+ T cell infiltration in specimens with reduced SLC25A1 expression (Figure 7G−J). We collected pathological sections from a clinical cohort and stratified the cases into SLC25A1 high- and low-expression groups according to staining intensity (Figure 7K). Clinical stratification based on SLC25A1 expression levels revealed significantly diminished survival rates in high-expression cohorts (Figure 7L,M). Functional analyses utilizing siRNA-mediated SLC25A1 suppression in pulmonary adenocarcinoma cell lines A549 and H1299 (Supplementary Figure S4A,B) demonstrated reduced cellular proliferation via CCK8 assessment (P<0.05) (Figure 7N,O).

Figure 7.

Correlation between SLC25A1 expression and immune characteristics in LUAD. (A,B) Negative correlation between SLC25A1 expression and CD8A levels in TCGA (A) and OAK (B) cohorts (Pearson correlation); (C,D) Kaplan-Meier survival curves showing overall survival based on SLC25A1 expression levels in TCGA (C) and OAK (D) cohorts; (E) OS survival analysis of patients stratified by CD8+ T cell spatial distribution patterns (Desert, Excluded, and Inflamed); (F) Box plot showing SLC25A1 IHC scores across different CD8+ T cell spatial distribution patterns; (G,H) Quantification of CD20+ B cells (G) and CD4+ T cells (H) in SLC25A1-high vs. SLC25A1-low tumors; (I,J) Representative multiplex immunofluorescence images showing DAPI (blue), CD4 (green), CD20 (purple), and SLC25A1 (red) staining in SLC25A1-high (I) and SLC25A1-low (J) tumors (×40); (K) Representative IHC images showing SLC25A1 expression in SLC25A1-high (upper panel) and SLC25A1-low (lower panel) tumors, with magnified insets (×10); (L,M) Kaplan-Meier curves showing overall survival (L) and disease-free survival (M) based on SLC25A1 expression levels; (N,O) Cell viability assays showing the effects of SLC25A1 knockdown in A549 (N) and H1299 (O) cell lines over time. LUAD, lung adenocarcinoma; IHC, immunohistochemistry; NC, negative control; OD, optical density. **, P<0.01; ***, P<0.001.

SLC25A1 inhibition enhances anti-PD-1 response

The immunomodulatory effects of SLC25A1 were evaluated through an in vivo pulmonary metastasis model. LLC murine LUAD cells were administered via tail vein injection to C57BL/6 mice, followed by treatment protocols utilizing either PD-1 monoclonal antibodies or IgG2a isotype controls (Figure 8A,B). Therapeutic efficacy assessment revealed significantly prolonged survival rates in subjects receiving SLC25A1 knockdown combined with anti-PD-1 intervention (P<0.0001) (Figure 8C). This enhanced therapeutic effect was further evidenced by marked reduction in metastatic burden and pulmonary-to-body mass ratios in the sh-SLC25A1 plus anti-PD-1 combination group vs. sh-NC controls (P<0.05) (Figure 8D−F). Detailed immune profiling demonstrated enhanced TIL populations, particularly CD8+ T cells, in specimens receiving sh-SLC25A1 combined with anti-PD-1 therapy (P<0.05) (Figure 8G,H). While both sh-SLC25A1 and anti-PD-1 independently facilitated CD8+ T cell recruitment and activation, maximal CD8+ T cell enrichment and significantly increased GZMB+ CD8+ T cells and IFN-γ+ CD8+ T cells were observed in TMEs following combined intervention (P<0.05) (Figure 8I,J), indicating that combination therapy not only enhanced CD8+ T cell infiltration but also augmented their cytotoxic function. These observations establish SLC25A1 suppression as a potentiator of CD8+ T cell functionality and anti-PD-1 therapeutic efficacy. The collected evidence suggests SLC25A1-mediated immune evasion occurs through CD8+ T cell exclusion, with its suppression facilitating enhanced immunotherapeutic responses.

Figure 8.

SLC25A1 knockdown enhances anti-PD-1 immunotherapy efficacy in LLC metastasis model. (A) Experimental timeline showing LLC cell intravenous inoculation, anti-PD-1 antibody treatment schedule, and endpoint analysis; (B) Representative images of lung metastatic nodules following tail vein injection of differently treated LLC cells; (C) Kaplan-Meier survival curves of mice from different treatment groups (n=8 per group); (D) Representative H&E staining of lung sections from different treatment groups; (E) Quantification of metastatic foci in lungs across treatment groups; (F) Comparison of lung weights across treatment groups; (G,H) Flow cytometry analysis of CD3+ T cells (G) and CD8+ T cells (H) infiltration in tumor tissues (percentage of CD45+ cells); (I,J) Flow cytometry quantification showing the percentage of IFN-γ+ cells (I) and GZMB+ cells (J) among CD8+ T cells across treatment groups. PD-1, programmed cell death protein 1; LLC, Lewis lung carcinoma; H&E, hematoxylin and eosin; IFN, interferon. ns, not significant. *, P<0.05; **, P<0.01; ***, P<0.001; ****, P<0.0001.

Discussion

Immunotherapy has emerged as a promising treatment modality in LUAD, demonstrating significant tumor-killing effects and improving patients’ quality of life (44-47). However, the effectiveness of immunotherapy is often compromised by immune escape mechanisms and metabolic competition within the TME (48). Due to these factors, many patients remain unresponsive to immunotherapy (49,50). Therefore, a key strategy in converting LUAD from a “cold” to a “hot” tumor involves enhancing effector T cell function while alleviating immunosuppression within the TME. Cancer cells, characterized by their high energy demands, predominantly utilize glycolysis, which generates substantial amounts of lactate (51). The resulting acidic microenvironment, created by lactate accumulation, suppresses immune cell function, leading to suboptimal immunotherapy outcomes. Within the TME, both cancer cells and immune cells undergo metabolic reprogramming in response to environmental changes (52). This intricate relationship between metabolic pathways and the immune microenvironment suggests that targeting specific metabolic pathways or metabolites could enhance immune cell function and consequently improve anti-tumor responses. However, a more comprehensive understanding of the relationship between metabolism and immunotherapy response in LUAD is still needed.

Our study introduces an innovative approach by developing a prognostic model based on key metabolic hub genes related to immunotherapy response. We developed a machine learning-based iMRS to explore the relationship between metabolism and LUAD prognosis. Through the analysis of 114 metabolism-related pathways, we characterized the metabolic landscape of LUAD. To demonstrate the prognostic performance of iMRS, we validated its relationship with immune response across seven different LUAD cohorts, revealing that lower iMRS scores correlate with better immunotherapy responses and improve prognosis. Furthermore, through C-index analysis, we found that iMRS scoring model significantly outperformed other biomarkers, demonstrating robust predictive capability for patient outcomes.

SLC25A1 (also known as CIC) couples mitochondrial citrate export to cytosolic lipid anabolism and is upregulated across cancers including NSCLC, pancreatic, and colorectal malignancies (53-55), where higher expression associates with poorer outcomes. Mechanistic studies in pancreatic cancer show KRAS-driven SLC25A1 upregulation, and pharmacologic/genetic inhibition restores lipid metabolic balance and restrains tumor growth (54), underscoring its druggability. In LUAD, SLC25A1 correlates positively with the immune checkpoint CD276 (56). Consistent with this, we observe that SLC25A1-high tumors display an immunologically “cold” TME with reduced CD4+ T-cell and CD20+ B-cell infiltration and worse prognosis. Functionally, SLC25A1 knockdown suppresses LUAD cell proliferation and, in vivo, cooperates with anti-PD-1 therapy to reduce tumor burden while augmenting IFN-γ+/GZMB+ CD8+ T cells infiltration.These findings have direct translational implications: 1) SLC25A1 may serve as a prognostic and potentially predictive biomarker for PD-1 response; 2) SLC25A1 status could guide selection of PD-1 monotherapy vs. combination regimens; and 3) SLC25A1 inhibitors warrant evaluation in biomarker-enriched trials, particularly in combination with PD-1 blockade. Within our iMRS framework, integrating SLC25A1 with the iMRS score may further refine patient stratification for immunotherapy.

Conclusions

Our research provides new insights into immunotherapy stratification for LUAD patients based on metabolism-related indicators and elucidates SLC25A1 as a potential target for immunotherapy. The current limitations of our study include the selection of immunotherapy cohorts comprising melanoma, glioblastoma, and LUAD patients. In future research, we plan to further validate the iMRS in larger LUAD immunotherapy patient cohorts and explore related mechanisms through in vivo and in vitro experiments.

SUPPLEMENTARY DATA

Supplementary data to this article can be found online.

Supplementary data to this article can be found online.

Funding Statement

None.

Author contributions

Conceptualization: PP Zhang, XQ Liang, JY Liu, CJ Huang, P Luo; Methodology: PP Zhang, XQ Liang; Investigation: PP Zhang, XQ Liang; Data curation: PP Zhang, XQ Liang; Formal analysis: PP Zhang, XQ Liang; Visualization: PP Zhang, XQ Liang; Resources: JY Liu, CJ Huang, P Luo; Project administration: JY Liu, CJ Huang, P Luo; Supervision: JY Liu, CJ Huang, P Luo; Writing – original draft: PP Zhang; Writing – review & editing: BC Ye, ZT Gong, YJ Wang, XF Wang, YM Huang; Validation: PP Zhang, XQ Liang, JY Liu, CJ Huang, P Luo.