A multi-programmed cell death-related LncRNA signature for prognosis and immune microenvironment evaluation in kidney renal clear cell carcinoma

Abstract

Kidney renal clear cell carcinoma (KIRC) shows substantial prognostic heterogeneity, motivating robust transcriptome-based risk stratification. Ferroptosis, cuproptosis and disulfidptosis are distinct stress-related cell-death programs whose gene-expression patterns may capture diverse tumour states. Using TCGA-KIRC RNA-seq data (n=533), tumour samples were randomly split 1:1 into a training set (n=267) and an independent test set (n=266); In the training cohort, lncRNAs associated with curated genes for each programmed cell death (PCD) modality were screened using EPIC-adjusted partial Spearman correlations (—r—>0.4, FDR<0.05) and differential expression versus adjacent normal tissues (p<0.05). The intersection across modalities yielded 176 multi-PCD-related lncRNAs. Univariate Cox analysis, LASSO-Cox regression and multivariate Cox modelling identified a six-lncRNA signature (AC026401.3, GAS5-AS1, SNHG8, AC084024.4, LINC02027 and AC115522.1) to compute a risk score. High-risk patients had significantly poorer overall and progression-free survival in both training and test cohorts, with stable time-dependent discrimination (AUCs 0.765/0.752/0.751 at 1/3/5 years in the full cohort). Computational analyses suggested risk-associated differences in immune contexture and predicted immune escape, and mutation profiling indicated higher VHL and PBRM1 alteration frequencies in the high-risk group. This signature supports prognostic stratification in KIRC and provides an association-based framework for further validation.

Introduction

Renal cell carcinoma (RCC) is one of the most common malignant tumors of the urinary system, with a continuously rising global incidence¹. Kidney renal clear cell carcinoma (KIRC) is the predominant histological subtype of RCC, accounting for approximately 70–80% of all cases². KIRC is characterized by pronounced heterogeneity and aggressiveness and shows limited sensitivity to conventional radiotherapy and chemotherapy. Although vascular endothelial growth factor receptor tyrosine kinase inhibitors (VEGFR-TKIs) such as sunitinib and immune checkpoint blockade targeting programmed death-1 (PD-1) or its ligand (PD-L1) have improved outcomes in advanced KIRC, treatment responses and survival still vary substantially across individuals^3,4. A considerable proportion of patients develop primary or acquired resistance, ultimately leading to treatment failure and disease progression⁵. Therefore, improving prognostic risk stratification and identifying robust, clinically actionable biomarkers remain urgent needs in KIRC.

Programmed cell death (PCD) contributes to tissue homeostasis through the elimination of abnormal cells and may influence tumor development and treatment response. Beyond classical apoptosis, multiple non-apoptotic PCD modalities have been described in recent years, including ferroptosis, cuproptosis, and the recently reported disulfidptosis^6–8. These processes are linked to metabolic reprogramming and oxidative/disulfide stress and have been associated with tumor progression, immune contexture, and therapy response in several cancer types. Ferroptosis is an iron-dependent form of cell death driven by lipid peroxidation; in KIRC, prior studies have reported associations between ferroptosis-related pathways and the hypoxia-inducible factor (HIF) axis, tumor immune features, and targeted-therapy response patterns^9,10. Cuproptosis can be triggered by intracellular copper overload, whereby copper binds lipoylated components of the tricarboxylic acid (TCA) cycle, leading to protein aggregation and proteotoxic stress^7,11. Disulfidptosis has been linked to disulfide stress and actin cytoskeleton collapse, and its potential roles in cancer biology are increasingly being recognized⁸. Importantly, these modalities may not operate in isolation; rather, they can reflect partially overlapping stress-response landscapes shaped by redox imbalance, mitochondrial metabolism, and cytoskeletal perturbation, which may correspond to heterogeneous malignant phenotypes and treatment-response tendencies. However, most existing studies model individual PCD modalities separately, and few have systematically tested whether integrating multiple PCD-related transcriptional programs yields incremental prognostic value within the same cohort.

Long non-coding RNA (LncRNA) refers to non-coding RNA transcripts longer than 200 nucleotides. Accumulating evidence suggests that lncRNAs are implicated in diverse biological processes—including stress-response programs relevant to PCD—through interactions with chromatin regulation, transcriptional control, and post-transcriptional events^12,13. For example, lncRNA GAS5 has been reported to be associated with tumor drug-response phenotypes, partly through links to autophagy- and ferroptosis-related gene expression¹⁴; lncRNA PANDAR has been proposed to relate to cuproptosis-relevant pathways by influencing copper metabolism-associated proteins¹⁵. Owing to their tissue specificity and disease relevance, lncRNAs are regarded as promising diagnostic and prognostic biomarkers¹⁶. Indeed, multi-lncRNA prognostic models have shown encouraging performance in KIRC^17,18.

Against this backdrop, we integrated curated gene sets for disulfidptosis, cuproptosis, and ferroptosis to identify lncRNAs that show statistically significant transcriptomic associations with these modalities in the TCGA-KIRC cohort. To mitigate confounding from tumor purity and cellular composition, we incorporated EPIC-based tumor purity adjustment during the association screening step. In addition to gene-level associations, we evaluated associations between candidate lncRNAs and pathway activity scores derived from PCD gene sets (ssGSEA), and examined their integration with canonical regulators (e.g., GPX4, SLC7A11, FDX1, ACTB) to provide convergent, hypothesis-generating evidence. Using machine-learning and Cox regression frameworks, we constructed and validated an integrated prognostic risk model based on multi-PCD-associated lncRNAs. We further characterized differences between risk groups in functional enrichment patterns, tumor immune-related features inferred from established computational tools, genomic alterations, and drug-response tendencies estimated from pharmacogenomic reference data, emphasizing that these results reflect predicted tendencies rather than direct immunological or pharmacological phenotypes. Collectively, our study aims to provide an association-based and clinically relevant biomarker framework to support risk stratification in KIRC while motivating future functional and clinical validation.

Methods

Data collection

This study did not generate original sequencing data. The TCGA-KIRC RNA-seq expression profiles, somatic mutation (SNP/MAF) files, and corresponding clinical information were downloaded from the Genomic Data Commons (GDC) portal (https://portal.gdc.cancer.gov). Expression values were harmonized at the gene-symbol level as provided by TCGA. Patients with missing overall survival information or incomplete key clinical variables were excluded from analyses when required. All other relevant materials and parameters are provided in the Supplementary Materials.

Construction of a prognostic risk model based on LncRNAs associated with multi-programmed cell death

We retrieved ferroptosis-related genes from FerrDb and collected disulfidptosis- and cuproptosis-related genes from published literature^19–21.The gene list is provided in Supplementary Material 1. TCGA-KIRC patients were first randomly split into training and testing sets (1:1), with paired adjacent normal tissues assigned to the same cohort at the patient level. and the testing set was used only for independent evaluation. In training tumors, tumor purity was estimated using EPIC, and lncRNA–PCD gene associations were quantified by partial Spearman correlation with purity as a covariate; multiple testing was controlled by Benjamini–Hochberg, retaining lncRNAs with |r|>0.4 and FDR<0.05. Differential expression between training tumors and adjacent normals was performed, and lncRNAs meeting both criteria (p<0.05) were retained, yielding DiRLncRNAs (Supplementary Material 2), CuRLncRNAs (Supplementary Material 3), and FeRLncRNAs (Supplementary Material 4). Their intersection was defined as MPCDLncRNAs (Supplementary Material 5).

Prognosis-associated MPCDLncRNAs were identified by univariate Cox regression in the training set (p<0.05), followed by LASSO-Cox regression with cross-validated penalty selection and multivariate Cox modeling to construct:

Patients were stratified by the training-set median RiskScore; coefficients and cutoff were then fixed and directly applied to the testing set without refitting or re-optimization. Performance was assessed using Kaplan–Meier analysis and time-dependent ROC curves (1-, 3-, and 5-year AUC).

Validation and evaluation of the prognostic risk model

Model performance was evaluated in the training set and independently assessed in the hold-out testing set using Kaplan–Meier survival analysis, time-dependent ROC curves (1-, 3-, and 5-year AUC), and Harrell’s C-index, with the training-derived cutoff applied unchanged to the testing set. Univariate and multivariate Cox regression analyses were conducted to test whether RiskScore was an independent prognostic factor after adjusting for clinical variables (e.g., age, gender, and stage). A nomogram integrating RiskScore and clinicopathological variables was constructed in the training set, internally validated by bootstrap resampling, and assessed by calibration curves.To assess stability, we further performed resampling-based internal validation in the training set, including repeated bootstrap resampling for hazard ratio stability and repeated k-fold cross-validation to estimate C-index variability and optimism (Supplementary Material 6). Finally, to quantify the incremental value of integrating ferroptosis, cuproptosis, and disulfidptosis, we compared the proposed model with representative published single-PCD lncRNA signatures (Lai_signature, Xu_signature, Wang_signature) implemented using the original coefficients (Supplementary Material 7) and evaluated under identical endpoints and metrics (C-index and time-dependent AUC at 1, 3, and 5 years)^22–24.

Differentially expressed genes and functional enrichment analysis between high- and low-risk groups

To explore biological differences between the high- and low-risk groups, differentially expressed genes (DEGs) were identified using the Wilcoxon test with the following criteria: and FDR <0.05. Gene Ontology (GO) enrichment and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analyses were performed using the clusterProfiler package²⁵. Enrichment results were summarized across Biological Process (BP), Cellular Component (CC), and Molecular Function (MF) categories, as well as KEGG pathways. Additionally, gene set enrichment analysis (GSEA) based on KEGG gene sets was conducted, and the top 5 most significant pathways for each risk group were visualized.

PCD pathway activity scoring and integration with canonical regulators

To provide independent pathway-level evidence beyond gene-wise co-expression, we quantified the activity of curated PCD gene sets using single-sample gene set enrichment analysis (ssGSEA). Briefly, ssGSEA scores for ferroptosis, cuproptosis, and disulfidptosis gene sets were computed on tumor expression profiles, and associations between candidate lncRNAs and pathway activity scores were evaluated using Spearman correlation and purity-adjusted partial Spearman correlation. In addition, to strengthen biological plausibility, we examined purity-adjusted associations between candidate lncRNAs and canonical regulators of each modality, including GPX4 and SLC7A11 (ferroptosis), FDX1 (cuproptosis), and ACTB (disulfidptosis), using the same partial-correlation framework. These integrative analyses were performed to support a coherent, association-based interpretation and do not establish causality.

Comprehensive analysis of tumor immune microenvironment, somatic mutations, and therapeutic potential

To characterize tumor microenvironment features, the ESTIMATE algorithm was applied to calculate stromal score, immune score, and ESTIMATE-inferred tumor purity for each sample²⁶. EPIC-inferred tumor purity was used specifically for confounder control during the PCD–lncRNA association screening step described above. The CIBERSORT method was used to estimate the relative fractions of 22 immune cell subtypes²⁷. Single-sample gene set enrichment analysis (ssGSEA) was performed to quantify the activity of 13 immune-related functional pathways²⁸, and the expression patterns of key immune checkpoint genes were compared between risk groups. These immune analyses represent computational inferences of relative abundance or activity derived from bulk transcriptomic data, rather than direct immunophenotypic assessments.Somatic mutation profiles were analyzed using the maftools package to visualize the 15 most frequently mutated genes and to calculate tumor mutational burden (TMB). The optimal cutoff for TMB stratification was determined using surv_cutpoint, and combined analyses of TMB and the risk score were conducted for survival evaluation. In addition, the Tumor Immune Dysfunction and Exclusion (TIDE) framework was used to infer immune-escape potential²⁹, and indicators including TIDE score and interferon-γ response were compared between groups. Group comparisons were performed using the Wilcoxon signed-rank test, with multiple testing adjusted by FDR and significance defined as adjusted p<0.05.

Prediction of drug sensitivity using model

Drug sensitivity was estimated using the oncoPredict framework with data from the Genomics of Drug Sensitivity in Cancer (GDSC) database (https://www.cancerrxgene.org/). Batch effects were corrected using empirical Bayesian methods, and low-variability genes were filtered as recommended. Expression profiles were used to impute IC50 values for each compound. After log-transformation of IC50 values, differences between the high- and low-risk groups were evaluated using the Wilcoxon signed-rank test.

Cell culture and establishment of sunitinib-resistant cell lines

The KIRC cell line ACHN was obtained from Shanghai Procell Biotechnology (Procell, Cat. No.: CL-0006). Cells were cultured in RPMI-1640 medium supplemented with 10% fetal bovine serum (FBS; ExCell Bio, Cat. No.: FSP500) and 1% penicillin–streptomycin (Beyotime, Cat. No.: C0222) at 37C with 5% CO₂. To establish sunitinib-resistant cells, a stepwise induction strategy was used. Briefly, logarithmically growing cells were treated with an initial concentration of 0.5 μM sunitinib (MCE, Cat. No.: HY-10255A) for 72 h. The medium was replaced with fresh drug-containing medium at each passage, and the drug concentration was gradually increased every 2–3 passages (0.5 μM → 1 μM → 2 μM → 3 μM → 4 μM → 5 μM) according to cell viability until stable proliferation at 5 μM was achieved. The induction lasted for more than 6 months, and resistance was confirmed by CCK-8 assays. The resulting resistant line was termed R-ACHN.

CCK-8 assay

Cell viability and drug response were assessed using the CCK-8 kit (Beyotime, Cat. No.: C0038). Cells in the logarithmic growth phase were seeded into 96-well plates and cultured for 24 h, followed by treatment with a range of drug concentrations. After 48 h, 10 μL of CCK-8 solution was added to each well and incubated at 37C for 2 h. Absorbance at 450 nm was measured using a microplate reader (BioTek, Synergy H1), and cell viability was calculated. Each condition included three technical replicates, and experiments were independently repeated three times.

Processing of clinical tissue samples and quantitative reverse transcription PCR

Paired KIRC tumor and adjacent normal tissues (n = 30 pairs) were collected from the Second Affiliated Hospital of Nanchang University and confirmed by pathological examination. Samples were cryopreserved in liquid nitrogen immediately after collection. Total RNA was extracted from cells or tissues using RNAiso Plus (TaKaRa, Cat. No.: 9109), and cDNA was synthesized using the HiScript III RT SuperMix (Vazyme, R323). Quantitative real-time PCR was performed using ChamQ Universal SYBR qPCR Master Mix (Vazyme, Q711) with GAPDH as the internal control. Relative expression levels were calculated using the method. Primer sequences are provided in Supplementary Material 8. None of the enrolled patients received any antitumor therapy prior to surgery. This study complied with the Declaration of Helsinki and was approved by the Ethics Committee of the Second Affiliated Hospital of Nanchang University. Written informed consent was obtained from all participants prior to surgery.

Statistical analysis

All analyses were performed using R (version 4.5.0) and GraphPad Prism (version 9). Continuous variables are presented as mean ± standard deviation. Between-group comparisons were conducted using Student’s t-test or the Mann–Whitney U test as appropriate, while categorical variables were compared using the test. Survival differences were evaluated using Kaplan–Meier curves and the log-rank test. Cox regression models were used to estimate hazard ratios (HR) with 95% confidence intervals. Correlations were assessed using Spearman tests for association screening and partial Spearman correlation for tumor-purity adjustment when indicated. Multiple testing was controlled by FDR correction, and adjusted p<0.05 was considered statistically significant.

Results

Screening of MPCDLncRNAs and development of a KIRC prognostic risk model

To develop a KIRC prognostic model based on lncRNAs linked to multiple programmed cell death (PCD) modalities (disulfidptosis, cuproptosis, and ferroptosis), we curated 23 disulfidptosis- and 10 cuproptosis-related genes from the literature and obtained 55 ferroptosis-related genes from FerrDb (http://www.zhounan.org/ferrdb/current/). TCGA-KIRC tumors were randomly divided into a training cohort (n=267) and an independent test cohort (n=266); the test cohort was not used for feature selection and served only for model evaluation. All screening steps were conducted in the training cohort. Using training-set transcriptomes, lncRNA–PCD gene associations were assessed in tumors by EPIC-adjusted partial Spearman correlation, with Benjamini–Hochberg correction; lncRNAs meeting |r|>0.4 and FDR<0.05 were retained. In parallel, differential expression between training tumors and adjacent normal tissues was performed, and lncRNAs with p<0.05 were kept. lncRNAs satisfying both criteria defined three sets: DiRLncRNAs, CuRLncRNAs, and FeRLncRNAs, with corresponding co-expression networks shown in Fig. 1A–C. Their intersection yielded 176 multi-programmed cell death-related lncRNAs (MPCDLncRNAs) (Fig. 1D; Supplementary Material 5). Prognosis-associated MPCDLncRNAs were screened by univariate Cox regression (p<0.05) in the training cohort, followed by LASSO-Cox regression (Fig. 1E,F) and multivariate Cox regression to construct the final model (Fig. 1G), with resampling-based internal validation for stability. The resulting six-lncRNA signature comprised AC026401.3, GAS5-AS1, SNHG8, AC084024.4, LINC02027, and AC115522.1, and the risk score was calculated as:

Baseline characteristics are summarized in Table 1. Correlations between PCD genes and the six MPCDLncRNAs are shown in Fig. 1H–J. Expression heatmaps and risk distributions indicated higher levels of GAS5-AS1, SNHG8, LINC02027, and AC115522.1 in the low-risk group (protective coefficients), whereas AC026401.3 and AC084024.4 were elevated in the high-risk group (risk coefficients) (Fig. 1K,L). Kaplan–Meier analyses demonstrated significantly poorer OS and PFS in the high-risk group (Fig. 1M,N). To support the “multi-PCD-related” designation beyond gene-wise co-expression, we additionally examined pathway- and regulator-level consistency: after purity adjustment, the six lncRNAs showed concordant associations with ssGSEA-derived multi-PCD pathway activity scores and canonical regulators (e.g., GPX4/SLC7A11, FDX1, and ACTB) (Supplementary Material 9 and Supplementary Material 10). These findings are association-based and do not imply causality.

Fig. 1.

Construction of a prognostic risk model based on MPCDLncRNAs. (A–C) Co-expression networks of disulfidptosis-related genes (A), cuproptosis-related genes (B), and ferroptosis-related genes (C) with their significantly correlated lncRNAs (purity-adjusted in tumor samples). (D) Venn diagram showing the intersection of DiRLncRNAs, CuRLncRNAs, and FeRLncRNAs. (E, F) LASSO regression analysis of prognosis-associated MPCDLncRNAs in the training set. (G) Forest plot of multivariate Cox regression for six key MPCDLncRNAs. (H–J) Heatmaps showing correlations between six MPCDLncRNAs and cuproptosis-related genes (H), ferroptosis-related genes (I), and disulfidptosis-related genes (J). (K, L) Expression heatmap of six MPCDLncRNAs with risk score and survival status distribution in the training set. (M, N) Kaplan–Meier curves of PFS (M) and OS (N) between high- and low-risk groups in the training set. *P<0.05; **P<0.01; ***P<0.001.

Table 1.

Baseline clinical characteristics of KIRC patients in the TCGA cohort.

Covariates	Type	Total (n=533)	Test (n=266)	Train (n=267)	P-value
Age	≤65	349 (65.48%)	182 (68.42%)	167 (62.55%)	0.182
	>65	184 (34.52%)	84 (31.58%)	100 (37.45%)
Gender	Female	188 (35.27%)	95 (35.71%)	93 (34.83%)	0.902
	Male	345 (64.73%)	171 (64.29%)	174 (65.17%)
Grade	G1	14 (2.63%)	7 (2.63%)	7 (2.62%)	0.996
	G2	229 (42.96%)	114 (42.86%)	115 (43.07%)
	G3	206 (38.65%)	103 (38.72%)	103 (38.58%)
	G4	76 (14.26%)	38 (14.29%)	38 (14.23%)
	Unknown	8 (1.50%)	4 (1.50%)	4 (1.50%)
Stage	Stage I	267 (50.09%)	144 (54.14%)	123 (46.07%)	0.033
	Stage II	57 (10.69%)	30 (11.28%)	27 (10.11%)
	Stage III	123 (23.08%)	47 (17.67%)	76 (28.46%)
	Stage IV	83 (15.57%)	43 (16.17%)	40 (14.98%)
	Unknown	3 (0.56%)	2 (0.75%)	1 (0.37%)
T stage	T1	272 (51.03%)	137 (51.50%)	135 (50.56%)	0.865
	T2	69 (12.95%)	37 (13.91%)	32 (11.99%)
	T3	181 (33.96%)	87 (32.71%)	94 (35.21%)
	T4	11 (2.06%)	5 (1.88%)	6 (2.25%)
M stage	M0	422 (79.17%)	210 (78.95%)	212 (79.40%)	0.398
	M1	79 (14.82%)	44 (16.54%)	35 (13.11%)
	Unknown	32 (6.00%)	12 (4.51%)	20 (7.49%)
N stage	N0	240 (45.03%)	121 (45.49%)	119 (44.57%)	1.000
	N1	16 (3.00%)	8 (3.01%)	8 (3.00%)
	Unknown	277 (51.97%)	137 (51.50%)	140 (52.43%)

Validation and evaluation of the 6-MPCDLncRNA prognostic risk model on the test set

To evaluate robustness and generalization, we performed multi-dimensional validation in the test set (n=266). Risk score and survival status distribution plots showed denser mortality events and shorter survival times in the high-risk group (Fig. 2A). Heatmaps confirmed consistent expression patterns of the six MPCDLncRNAs with the training set: GAS5-AS1, SNHG8, LINC02027, and AC115522.1 were higher in the low-risk group, while AC026401.3 and AC084024.4 were elevated in the high-risk group (Fig. 2B). Kaplan–Meier analysis further demonstrated significantly poorer OS and PFS in the high-risk group (Fig. 2C,D).

Fig. 2.

Validation of the 6-MPCDLncRNA prognostic model in the test set. (A, B) Heatmap of six MPCDLncRNAs with risk score and survival status distribution in the test set. (C, D) Kaplan–Meier curves for OS (C) and PFS (D) between high- and low-risk groups. (E) Time-dependent ROC curves for predicting 1-, 3-, and 5-year OS in the test set. (F) ROC comparison between the risk score and conventional clinical variables in the test set (at a representative time point). (G, H) PCA based on 298 MPCDLncRNAs (G) and the six-lncRNA model (H).

Time-dependent ROC analysis yielded AUCs of 0.780, 0.775, and 0.751 for 1-, 3-, and 5-year OS prediction in the test set (Fig. 2E), indicating sustained predictive ability. For clinical comparison, the risk score showed higher discrimination than age and sex and was comparable to grade and stage (Fig. 2F). PCA based on 176 MPCDLncRNAs (Fig. 2G) and on the six-lncRNA model (Fig. 2H) showed clear separation between risk groups in three-dimensional space, supporting the model’s capability to stratify KIRC patients.

Risk scoring as an independent prognostic factor and its clinical translation value

Univariate and multivariate Cox regression analyses in the entire cohort demonstrated that the risk score remained an independent prognostic factor after adjustment for clinical variables, together with age, grade, and stage (Fig. 3A,B). Across the entire cohort, the model achieved AUCs of 0.765, 0.752, and 0.751 for predicting 1-, 3-, and 5-year OS (Fig. 3D). The risk score showed superior discrimination compared with age and sex and was comparable to grade and stage (Fig. 3C), supporting its stable predictive capability.

Fig. 3.

Risk score as an independent prognostic factor and assessment of clinical utility. (A, B) Forest plots of univariate (A) and multivariate (B) Cox regression analyses. (C) ROC comparison between the risk score and conventional clinical variables in the entire cohort. (D) Time-dependent ROC curves of the risk model for 1-, 3-, and 5-year OS in the entire cohort. (E) Nomogram integrating the risk score with clinical variables. (F) Calibration curve and C-index for the nomogram.

To facilitate clinical translation, we built a nomogram integrating the risk score with age, sex, grade, and stage (Fig. 3E). Calibration curves showed good agreement between predicted and observed survival probabilities, and the nomogram achieved a C-index of 0.790 (Fig. 3F). Subgroup analyses further demonstrated consistent prognostic stratification by the risk score across age, grade, stage, T stage, and metastasis status (all p<0.05; Fig. 4A–K), indicating robustness across clinical subpopulations. To assess whether integrating ferroptosis, cuproptosis, and disulfidptosis provides incremental prognostic value beyond modeling a single PCD modality, we conducted head-to-head comparisons in the same TCGA-KIRC cohort. We benchmarked our MPCD_signature against representative published lncRNA signatures, including Lai_signature (ferroptosis-associated), Xu_signature (cuproptosis-associated), and Wang_signature (disulfidptosis-related). The MPCD_signature achieved the highest C-index (0.790), outperforming Lai_signature (0.714), Wang_signature (0.721), and Xu_signature (0.697) (Fig. S1A). Time-dependent ROC analyses showed that the MPCD_signature yielded AUCs of 0.765, 0.752, and 0.751 at 1, 3, and 5 years, respectively. In comparison, Lai_signature achieved AUCs of 0.754/0.732/0.755, Xu_signature achieved 0.747/0.718/0.738, and Wang_signature achieved 0.748/0.751/0.778 at 1/3/5 years (Fig. S1B–E). Collectively, these findings indicate that multi-PCD integration provides incremental prognostic information in this cohort, particularly reflected by improved concordance and competitive time-dependent discrimination across clinically relevant time horizons.

Fig. 4.

Subgroup validation of the prognostic performance of the risk model in TCGA-KIRC patients. (A, B) Age. (C, D) Grade. (E, F) Stage. (G, H) T stage. (I, J) M stage. (K) N stage.

KEGG, GO, and GSEA enrichment analysis

Differential expression analysis between the high- and low-risk groups identified 774 DEGs (Supplementary Material 11), which were subjected to GO, KEGG, and GSEA analyses. GO enrichment indicated that these DEGs were mainly involved in immune-related processes and transport/metabolic functions (Fig. 5A). At the biological process (BP) level, DEGs were enriched in humoral immune-related processes such as “antimicrobial humoral response,” “humoral immune response,” and “acute-phase response.” Cellular component (CC) terms included “immunoglobulin complex,” “blood microparticle,” and “collagen-containing extracellular matrix,” while molecular function (MF) terms were enriched in “antigen binding” and transporter-related activities. These results suggest pronounced immune dysregulation and microenvironment remodeling associated with risk stratification.

Fig. 5.

Functional enrichment analyses. (A) GO enrichment. (B) KEGG enrichment. (C) GSEA results for the high-risk group. (D) GSEA results for the low-risk group.

KEGG pathway analysis revealed enrichment in “Complement and coagulation cascades,” “Cytokine–cytokine receptor interaction,” “PPAR signaling pathway,” “TGF-β signaling pathway,” and “Neuroactive ligand–receptor interaction” (Fig. 5B). Notably, pathways related to renal physiology, including “Proximal tubule bicarbonate reclamation” and “Mineral absorption,” were also enriched, indicating that the risk model may reflect tissue-specific metabolic alterations in KIRC. GSEA further showed that the high-risk group was enriched in pathways such as retinol metabolism and complement/coagulation cascades (Fig. 5C), whereas the low-risk group was enriched in pathways related to metabolic homeostasis and cellular maintenance, including valine, leucine and isoleucine degradation, neurotrophin signaling, and endocytosis (Fig. 5D). Together, these analyses highlight distinct biological states between risk groups and provide functional context for the prognostic model.

Analysis of tumor immune microenvironment, immune checkpoint expression, and therapeutic response based on the 6-MPCDLncRNA risk model

We profiled tumor immune microenvironment (TIME) features using transcriptome-based computational methods, including ESTIMATE, CIBERSORT, ssGSEA, and TIDE, to compare immune-related patterns between risk groups. ESTIMATE analysis showed significantly higher stromal and ESTIMATE scores in the high-risk group (Fig. 6A), suggesting differences in non-tumor components inferred from bulk expression. CIBERSORT deconvolution indicated distinct estimated immune-cell composition: the high-risk group showed higher inferred fractions of plasma cells, activated CD4 memory T cells, follicular helper T cells, regulatory T cells (Tregs), and M0 macrophages, whereas the low-risk group showed relatively higher inferred fractions of resting CD4 memory T cells, monocytes, M1/M2 macrophages, resting dendritic cells, resting mast cells, and neutrophils (Fig. 6B,C). In addition, ssGSEA suggested that the low-risk group had higher inferred activity scores for several immune functions, including APC co-stimulation, CD8 T-cell-related signatures, cytolytic activity, T-cell co-stimulation, and type II interferon response (Fig. 6D).

Fig. 6.

TIME and immunotherapy potential based on the 6-MPCDLncRNA risk model. (A) ESTIMATE-derived ImmuneScore, StromalScore, ESTIMATEScore, and tumor purity. (B, C) CIBERSORT-estimated infiltration proportions of 22 immune cell subtypes. (D) ssGSEA of immune-related functions. (E) Differential expression of immune checkpoint genes. (F) TIDE score comparison. *P<0.05; **P<0.01; ***P<0.001.

Immune checkpoint analysis revealed higher expression of inhibitory checkpoint genes (e.g., LAG3 and PD-1) in the high-risk group (Fig. 6E), a pattern that is consistent with an exhausted or immunoregulatory context in transcriptome-level inference. TIDE analysis showed higher TIDE scores in the high-risk group (Fig. 6F), indicating increased predicted immune escape potential and a potentially less favorable immunotherapy-related signature. Collectively, these analyses suggest that the 6-MPCDLncRNA risk model is associated with distinct computationally inferred TIME patterns;

Genomic variation characteristics and prognostic value of tumor mutational burden in high- and low-risk groups

We analyzed somatic mutation data to explore genomic differences between risk groups. As shown in Fig. 7A,B, both groups exhibited high mutation prevalence (80.59% in the high-risk group and 77.23% in the low-risk group). Recurrently mutated genes such as VHL, PBRM1, TTN, SETD2, and BAP1 were observed in both groups, consistent with known KIRC driver events. Notably, VHL (41% vs. 34%) and PBRM1 (41% vs. 16%) mutation frequencies were higher in the high-risk group, indicating potential associations with unfavorable outcomes.

Fig. 7.

Somatic mutations and TMB analyses. (A, B) Mutation landscapes in high- and low-risk groups. (C) TMB comparison between groups. (D) Kaplan–Meier OS by TMB status. (E) Kaplan–Meier OS stratified by combined TMB and risk score.

TMB levels did not differ significantly between risk groups (p=0.081; Fig. 7C). Nevertheless, survival analysis in the overall cohort showed superior OS in patients with high TMB compared with low TMB (Fig. 7D). Combined stratification using TMB and the risk model further separated survival outcomes (Fig. 7E), suggesting that TMB may provide complementary prognostic information when interpreted alongside the risk score.

Drug prediction analysis

Using the oncoPredict framework, we compared model-predicted drug response profiles between risk groups by estimating IC50 values from bulk transcriptomic data with GDSC-trained models. In this analysis, the low-risk group exhibited significantly lower predicted IC50 values for multiple VEGFR inhibitors (Fig. 8A–D), including sunitinib, pazopanib, axitinib, and cabozantinib, whereas the high-risk group showed lower predicted IC50 values for selumetinib, fludarabine, dabrafenib, and lapatinib (Fig. 8E–H). These results indicate differential inferred drug susceptibility patterns associated with the risk score and may serve as a hypothesis-generating resource for prioritizing candidate agents for further validation.

Fig. 8.

Drug sensitivity prediction by risk group. (A) Sunitinib. (B) Cabozantinib. (C) Pazopanib. (D) Axitinib. (E) Selumetinib. (F) Lapatinib. (G) Fludarabine. (H) Dabrafenib.

Expression patterns of six MPCDLncRNAs in KIRC

To validate expression patterns of the six MPCDLncRNAs in KIRC, we compared tumor and adjacent normal tissues in TCGA-KIRC. AC026401.3 and AC084024.4 were significantly upregulated in tumors (Fig. 9A,B). AC115522.1 and SNHG8 were also higher in tumors (Fig. 9C,D), whereas LINC02027 and GAS5-AS1 were downregulated (Fig. 9E,F). Although AC115522.1 and SNHG8 were upregulated in tumor tissues, they were associated with favorable prognosis in the multivariate Cox model, suggesting context-dependent effects. These expression trends were further validated by qRT-PCR in independent paired clinical samples (Fig. 9G–L).

Fig. 9.

Expression validation of six MPCDLncRNAs. (A–F) TCGA-based expression comparison between tumor and adjacent normal tissues: (A) AC084024.4, (B) AC026401.3, (C) AC115522.1, (D) SNHG8, (E) LINC02027, (F) GAS5-AS1. G–L. qRT-PCR validation in independent paired samples. *P<0.05; **P<0.01; ***P<0.001.

Prognosis and drug resistance analysis of key risk factors in KIRC

We further examined two risk-associated lncRNAs (AC026401.3 and AC084024.4) with respect to prognosis and sunitinib resistance. In the TCGA-KIRC cohort, higher expression of AC026401.3 and AC084024.4 was associated with poorer OS (Fig. 10A,B), consistent with their positive coefficients in the risk model.

Fig. 10.

Prognostic and resistance-associated expression patterns of AC026401.3 and AC084024.4. (A, B) Kaplan–Meier OS stratified by high vs. low expression of (A) AC026401.3 and (B) AC084024.4 in TCGA-KIRC. (C) Reduced sunitinib responsiveness in sunitinib-resistant ACHN cells assessed by CCK-8 assay. (D, E) qRT-PCR showing increased expression of (D) AC026401.3 and (E) AC084024.4 in resistant versus parental ACHN cells. *P<0.05; **P<0.01; ***P<0.001.

To assess whether these lncRNAs are associated with a sunitinib-resistant phenotype at the transcriptomic level, we established sunitinib-resistant ACHN cells and confirmed reduced drug responsiveness using CCK-8 assays (Fig. 10C). qRT-PCR showed that AC026401.3 and AC084024.4 were upregulated in resistant cells compared with parental cells (Fig. 10D,E). Collectively, these results support a correlation between lncRNA upregulation and the resistant phenotype in this in vitro model. Accordingly, AC026401.3 and AC084024.4 may serve as candidate resistance-associated biomarkers; however, these data do not establish causality, and mechanistic involvement should be tested in future studies using functional perturbation and rescue experiments.

Discussion

Kidney renal clear cell carcinoma (KIRC) remains therapeutically challenging due to pronounced inter- and intra-tumoral heterogeneity and limited sensitivity to conventional radiotherapy and chemotherapy¹. Although immune checkpoint blockade and VEGFR-targeted agents have improved outcomes in advanced disease, responses are variable and durable benefit is achieved only in a subset of patients³⁰. Robust biomarkers for prognosis prediction and treatment stratification are therefore of clinical importance.

Programmed cell death (PCD) is increasingly recognized as part of tumor stress adaptation and therapy response. Beyond apoptosis, ferroptosis, cuproptosis, and disulfidptosis have been linked to redox homeostasis, mitochondrial metabolism, and cytoskeletal stress in cancer^31–33. While these modalities are conceptually distinct, they may reflect partially overlapping stress-response landscapes in tumors. Long non-coding RNAs (lncRNAs) can participate in stress-related transcriptional regulation and have been implicated in multiple PCD-associated contexts¹².

In this study, we developed a multi-PCD prognostic signature consisting of six lncRNAs (6-MPCDLncRNA) and observed stable discriminatory ability in the training set, testing set, and the entire TCGA-KIRC cohort. Time-dependent ROC curves showed AUC values above 0.75, and multivariate Cox regression confirmed that the risk score remained an independent prognostic factor after adjusting for clinicopathological variables. A nomogram integrating the risk score with clinical parameters achieved a C-index of 0.790 with good calibration. In the same TCGA-KIRC cohort, we performed a head-to-head comparison with representative published lncRNA signatures associated with ferroptosis, cuproptosis, and disulfidptosis. The integrated MPCD_signature achieved the highest concordance index (C-index = 0.790), outperforming the Lai_signature (0.714), Xu_signature (0.697), and Wang_signature (0.721), supporting its potential utility for risk stratification.To place the signature in a biological context, functional enrichment analyses indicated that genes differentially expressed between risk groups were enriched for immune- and inflammation-related processes (e.g., complement/coagulation cascades and cytokine–cytokine receptor interaction). These findings are consistent with KIRC being an immune-infiltrated tumor in which inflammatory programs can accompany angiogenesis and tumor progression^34,35. Nevertheless, enrichment of immune-related terms does not necessarily imply effective antitumor immunity; rather, it may reflect immune activation, dysregulation, or stromal remodeling depending on the underlying cellular composition. Accordingly, these pathway-level results should be interpreted as association-based context for the observed prognostic separation.

Computational analyses of the tumor immune microenvironment suggested distinct immune-related patterns between risk groups. The high-risk group showed higher ESTIMATE-derived immune/stromal scores, while CIBERSORT-based deconvolution suggested increased fractions of immunoregulatory populations (e.g., Tregs and macrophage-related signals) and immune checkpoint expression differences (e.g., LAG3 and PD-1). In contrast, immune functional scores derived from ssGSEA indicated relatively higher effector-related activity in the low-risk group (e.g., cytolytic activity and type II interferon response). Collectively, these results are consistent with a scenario in which the high-risk group may exhibit “higher inferred infiltration but lower inferred effector function.” However, these observations are inference-based and, in the absence of orthogonal validation, should not be interpreted as direct immunophenotypic evidence. Likewise, the TIDE-based results suggest a higher immune evasion potential in the high-risk group, but such computational estimates cannot substitute for actual clinical response data.Genomic profiling further linked the risk model to established molecular features of KIRC. VHL and PBRM1 mutation frequencies were higher in the high-risk group. VHL loss is a hallmark of KIRC and promotes angiogenesis and metabolic adaptation via HIF stabilization³⁶, whereas PBRM1 alteration has been implicated in tumor evolution and immune-related phenotypes³⁷. Although tumor mutational burden did not differ significantly between risk groups, combined stratification using TMB and the risk score provided complementary prognostic separation, suggesting that genomic alterations and microenvironment-associated signals capture nonredundant aspects of outcome.

Beyond prognosis, our analyses also explored treatment-related associations. Drug response estimation using oncoPredict suggested that the low-risk group may have lower predicted IC50 values for several VEGFR inhibitors (including sunitinib and pazopanib), consistent with the central role of the VHL-HIF-VEGF axis in KIRC³⁸. We also observed elevated expression of AC026401.3 and AC084024.4 in a sunitinib-resistant ACHN model. However, these findings do not establish a causal role in resistance; they merely indicate an association between lncRNA expression and the drug-resistant state. Therefore, these lncRNAs should be considered candidate resistance-associated biomarkers warranting future functional validation.

This study has several limitations. First, our risk model lacks validation in an independent external cohort; the generalizability of the fixed risk formula and its robustness to cross-platform batch effects remain to be confirmed in independent multicenter cohorts and prospective studies. Second, the immune landscape and drug responses were inferred computationally and require validation using orthogonal deconvolution approaches and phenotype-level assays. Therefore, our findings should be interpreted as hypothesis-generating evidence rather than definitive proof of drug sensitivity or resistance. Third, although tumor purity adjustment was implemented, residual confounding from cellular composition, unmodeled batch effects, and broader transcriptional gradients cannot be fully excluded. Finally, while single-sample gene set enrichment analysis (ssGSEA) and regulator association analyses provide additional transcriptome-level support, establishing mechanistic causality will require targeted perturbations (gain- and loss-of-function), pathway-specific readouts, and in vivo validation.

Conclusion

We developed a six-lncRNA prognostic signature integrating ferroptosis, cuproptosis, and disulfidptosis pathways, which robustly stratified KIRC patients by survival in training and test cohorts. The model outperformed single-PCD signatures and was independently prognostic. Multi-omics analyses linked the risk score to distinct immune microenvironments, mutation patterns (VHL/PBRM1), and predicted drug responses. This multi-PCD framework provides a robust stratification tool and generates hypotheses for functional validation.

Abbreviations

DEGs

Differentially expressed genes

GO

Gene ontology

GSEA

Gene set enrichment analysis

KEGG

Kyoto encyclopedia of genes and genomes

KIRC

Kidney renal clear cell carcinoma

LncRNA

Long non-coding RNA

MPCDLncRNA

Multi-programmed cell death-related long non-coding RNA

OS

Overall survival

PCA

Principal component analysis

PCD

Programmed cell death

PFS

Progression-free survival

RCC

Renal cell carcinoma

TCGA

The cancer genome atlas

TMB

Tumor mutational burden

Author contributions

Yi Yu designed the project and proofread the manuscript. Shaochao Zhan and Yuchen Li conducted the experiments and finished the manuscript (Yuchen Li and Shaochao Zhan contributed equally to this work as co-first authors). Zhaohui Guo and Ting Li analyzed the data, Chengjian Xie and Tao Yang, Zitong Luo participated in the experimental work and collected clinical data. All authors checked and approved the final manuscript.

Funding

This study was supported by the National Natural Science Foundation of China (Grant No. 82060466) and the Natural Science Foundation of Jiangxi Province (Grant No. 20224BAB206081).

Data availability

This study did not generate any original sequencing data. The datasets used in this study (TCGA-KIRC RNA-seq data, SNP, and clinical data) can be found in the following public data repositories (https://portal.gdc.cancer.gov). Other data can be found in the supplementary files.The R code used in this study is available at: (https://github.com/zsc171/MPCDLcnRNA)

Declarations

Competing interests

The authors declare no competing interests.

Ethics approval and consent to participate

The study protocol was approved by the Institutional Review Board: all KIRC tissue samples were obtained from the Second Affiliated Hospital of Nanchang University, Nanchang, China. Ethical approval for all procedures was obtained from the Medical Research Ethics Committee of the Second Affiliated Hospital of Nanchang University.

Consent for publication

Consent to publish has been obtained from all authors.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1038/s41598-026-58577-w.