The RNA sequencing (RNA-seq) data and clinical information of PRAD samples were obtained from the UCSC Xena database(https://xenabrowser.net/datapages/) (15). To identify PRAD patient samples within the TCGA cohort, we followed the screening methodology outlined in a previous PRAD-related study, excluding samples with incomplete survival data (16). Ultimately, we identified 496 PRAD samples and 52 adjacent normal tissue samples. We also obtained RNA-seq data and clinical characteristics from a cohort of PRAD patients, sourced from the GSE141445 dataset available in the Gene Expression Omnibus (GEO) (https://www.ncbi.nlm.nih.gov/geo/) (17). Single-cell RNA sequencing (scRNA-seq) data from GSE185344 (18) and Cell Reports (PMID: 34936871) (19) verified our results. The TCGA-PRAD, GSE116918 (20), and GSE21034 (21), GSE70769 (22) and DKFZ2018 datasets were utilized for the external validation of the IFRD1 gene. The transcriptomic and clinical data for DKFZ2018 were obtained from the cBioportal database (https://www.cbioportal.org/datasets) (23). These datasets were also employed to assess IFRD1 expression and conduct prognostic analyses. Additionally, we utilized the BEST tool (https://rookieutopia.hiplot.com.cn/app_direct/BEST/) to analyze and download multiple PCa expression and survival datasets, which served as external cohorts for IFRD1 expression and prognostic analysis. Single-cell RNA sequencing (scRNA-seq) reads were log-normalized in Seurat; genes detected in <3 cells, cells with <200 or >7–000 genes, or >10% UPR reads were removed. Batch effects were corrected with Harmony, clusters were identified by canonical markers and visualized with t-SNE. UPR activity was calculated with AddModuleScore; Differentially expressed genes (DEGs) between UPR-high and UPR-low luminal cells (Wilcoxon, adj-P < 0.05) were carried forward.

To quantify the activity of a gene set in each individual tumor, we first derived per-sample UPR scores by applying single-sample gene-set enrichment (ssGSEA) implemented in the GSVA R package to the TCGA-PRAD cohort. Next, we expanded the same ssGSEA framework to the 50 MSigDB hallmark gene sets, generating a pathway-activity matrix. Differential pathway activity between tumors stratified into UPR-high and UPR-low groups was then assessed with limma, and significance was defined after multiple-testing correction.

A scale-free co-expression network (soft power = 6) was built from merged single-cell/bulk DEGs; the module most correlated with UPR (MEturquoise) was retained and its hub genes extracted.

Intersected DEGs were filtered by univariate Cox (P < 0.05), random-forest importance (MDA/MDG) and Least Absolute Shrinkage and Selection Operator (LASSO) (10-fold cross-validation) to yield seven UPR-related prognostic genes. UPRRS = Σ (expression × β) dichotomized patients at the median; performance was assessed with time-dependent Receiver Operating Characteristic (ROC) and log-rank tests.

Seven UPRRGs associated with PFS were identified by log-rank screening (P < 0.05); four remained after univariate Cox (P < 0.05) and random-forest confirmation. LASSO (10-fold cross-validation) converted these into a weighted index (UPRRS = Σ Ei × βi). Patients were split at the median; ROC and Kaplan–Meier (log-rank, P < 0.05) assessed discrimination.

Considering that clinicopathological variables significantly influence survival outcomes, we constructed a nomogram utilizing TCGA clinicopathological parameters alongside UPRRS to enhance the predictive accuracy of UPRRS. Clinicians can leverage the points derived from the nomogram to estimate the patients’ survival probabilities. Calibration curves and ROC curves were generated, and the area under the curve (AUC) was calculated to assess the predictive capability of the nomogram.

To relate the UPRRS to the immune landscape of PCa, we first deconvoluted bulk RNA-seq profiles with Cell-type Identification by Estimating Relative Subsets of RNA Transcripts (CIBERSORT) to obtain relative fractions of 22 infiltrating immune cell types. The robustness of these estimates was confirmed in parallel with ESTIMATE, ssGSEA and XCell. The anti-cancer immune cycle is a crucial component of tumor immunotherapy, which involves several steps to constitute the anti-cancer immune cycle. Finally, immunogenicity was quantified with the Immunophenoscore (IPS) – an integrative, machine-learning metric that incorporates expression signatures of immunomodulators, immunosuppressive cells, MHC molecules and effector cells; high IPS predicts greater benefit from immune-checkpoint blockade. IPS values for TCGA-PRAD tumors were downloaded from The Cancer Immunome Atlas (TCIA).

The drug sensitivity training data were downloaded from the Genomics of Drug Sensitivity in Cancer (https://www.cancerrxgene.org) (24). Subsequently, these data were validated using TCGA data to identify commonly used anti-tumor drugs for our analysis. The results were effectively presented through grouped comparison plots utilizing the “oncoPredict” and “ggpubr” packages. The Wilcoxon test was employed to ascertain the statistical differences between various groups, with a p-value < 0.05 considered statistically significant.

We utilized the ‘TCGAplot’ package to conduct a comprehensive pan-cancer analysis and visualization of multi-omics data related to IFRD1 within TCGA. The source code and pre-built package are available on GitHub (https://github.com/tjhwangxiong/TCGAplot) (25).

Cellular RNA was isolated with the RNAeasy™ Animal RNA Isolation Kit (Beyotime, Shanghai) according to the vendor’s spin-column protocol. Complementary DNA was synthesized and subsequently amplified on a real-time PCR instrument with Taq Pro Universal SYBR qPCR Master Mix (BestEnzymes, Lianyungang). Gene-specific primers for IFRD1 and the housekeeping gene GAPDH were purchased from GENERAL BIOL (Chuzhou). Knockdown experiments were performed with 3 biological replicates, and qPCR was conducted in quadruplicate for each sample (4 technical replicates).

For the internal cohort, Paraffin-embedded prostatectomy specimens (Sept–Dec 2024, Zhongda Hospital) were stained for IFRD1 (ProteinTech) by the hospital pathology unit; expression was scored semi-quantitatively. IRB approval: 2023ZDSYLL056-P01. For the independent external cohort, protein expression levels were assessed using publicly available IHC data retrieved from the Human Protein Atlas (HPA, https://www.proteinatlas.org) database. Representative IHC staining images were selected from PRAD tissue samples, including two high-grade and two low-grade tumors. The staining intensity and subcellular localization patterns were qualitatively evaluated for each sample to compare IFRD1 protein expression between disease aggressiveness categories.

Wound-healing: si-IFRD1 or control PCa cells were grown to confluence in 6-well plates, scratched with a 200 µL tip, washed, and tracked in serum-free medium at 0, 12, 24, 48 h.

Colony formation: 1 × 10³ cells/well were seeded, fed every 3 days, fixed (4% PFA, 30 min) and stained (0.1% crystal violet, 30 min) on day 7–14 before counting.

Analyses were performed in R 4.1.2/3.6.3. Chi-square or Wilcoxon tests compared clinical variables; DEG significance was false discovery rate (FDR)-adjusted. Survival was assessed by Kaplan–Meier/log-rank and Cox models; ROC-AUC used timeROC. Immune correlations were evaluated with Spearman. qRT-PCR and two-group comparisons used Student’s t-test. *P < 0.05 was considered significant.

In the study, we utilized publicly available scRNA-seq data (GSE141445) to clustering analysis (17, 26). Cell clusters were subsequently annotated into seven major cell type - monocytes, mast cells, basal/intermediate cells, fibroblasts, luminal cells, T cells, and endothelial cell - based on canonical marker genes (Figure 1A). Data from seven samples across different batches were integrated to mitigate batch effects, as visualized in Supplementary Figure S1A. Each cell type was characterized by a distinct molecular fingerprint (Supplementary Figure S1B). A heatmap further illustrates the top five marker genes defining each cluster (Supplementary Figure S1C). To quantify UPR activity across distinct cell types, we applied the “AddModuleScore” function in Seurat to calculate a UPR gene set enrichment score for each cell (Figure 1B). Among the seven identified cell types, luminal cells exhibited significantly elevated UPR activity compared to others (Figure 1C). Based on this UPR activity metric, luminal cells were stratified into UPR-high and UPR-low groups. Differential expression analysis between these two groups identified key genes (DEGs) for subsequent investigation (Supplementary Table S2; Supplementary Figures S2A–C). To validate these findings, we analyzed another scRNA-seq dataset (GSE185344) (18). Consistent with our initial results, a subset of epithelial cells displayed UPR-high gene set expression in this cohort (Figures 1D–F). Furthermore, leveraging an additional scRNA-seq dataset (19), we sub-classified luminal epithelial cells into LE-KLK3 and LE-KLK4 subtypes. We observed distinct UPR activity between these subtypes, a finding that corroborates the GSE185344 data. Specifically, UPR activity was significantly higher in the LE-KLK3 subtype compared to the LE-KLK4 subtype (Figures 1G–I). The identified DEGs were upregulated in tumor patients at the bulk transcriptome level, particularly within the LE-KLK3 subtype. Notably, these genes demonstrated significant utility in distinguishing tumor from non-tumor samples in the TCGA-PRAD cohort (Figures 1J–M). To strengthen the clinical relevance of UPR activation, we evaluated its association with standard prognostic indicators. UPR scores showed significant positive correlations with Gleason grade, pathological T stage, and PSA levels, and robustly predicted both PFI and OS in multiple cohorts, underscoring its potential as a clinically actionable pathway (Supplementary Figures S1D, E).

To delineate the molecular programs downstream of UPR activation in PCa, we conducted functional enrichment and trajectory inference analyses comparing UPR-high and UPR-low luminal cells identified from the scRNA-seq dataset (GSE141445) (17). As shown in Figure 2A, Gene Ontology (GO) analysis revealed that the UPR-high luminal cell population was significantly enriched for terms related to the response to unfolded protein, ER stress, and response to topologically incorrect protein. In contrast, the UPR-low population was associated with processes including cytoplasmic translation, ribosome biogenesis, and mitochondrial ATP synthesis coupled electron transport, suggesting a bioenergetic and translational state distinct from ER stress adaptation (27). Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analysis further revealed that DEGs upregulated in UPR-high luminal cells were primarily enriched in pathways such as protein processing in ER, AMPK signaling pathway, and estrogen signaling pathway. Conversely, DEGs from the UPR-low group were enriched in ribosome, amyotrophic lateral sclerosis, and thermogenesis pathways. GSEA results revealed significant positive associations between the UPR DEG signature and hallmark gene sets including the UPR, hypoxia, inflammatory response, and androgen response (Figure 2B; Supplementary Figure S2D). These associations were further validated in the TCGA-PRAD cohort, where Myc targets v1 and v2 were among the most significantly enriched gene sets, implicating the UPR in proliferative and metabolic reprogramming (Supplementary Figures S2F–H) (28). Specifically, the correlations between the UPR DEGs and both the hypoxia and androgen response pathways were positive. Extended functional enrichment analysis of high-UPR luminal cells revealed broader biological consequences beyond protein homeostasis, including activation of MAPK and PI3K-Akt oncogenic pathways, metabolic reprogramming toward glycolysis, and upregulation of immune evasion signatures such as the PD-1/PD-L1 checkpoint axis (Figure 2C). CytoTRACE analysis showed that cellular developmental potential was not correlated with the UPR DEGs expression (Figure 2D; Supplementary Figure S2E), suggesting that UPR activation is not overtly linked to stemness or differentiation trajectories. Finally, we assessed metabolic pathway activities across cell types. Both UPR-high and UPR-low luminal cells were associated with oxidative phosphorylation, glycolysis/gluconeogenesis, and the citrate cycle (TCA cycle), indicating a coordinated metabolic rewiring to sustain ATP production and redox balance under ER stress (Figures 2E, F) (29). To investigate the role of UPR DEGs within the TME at single-cell resolution, we stratified luminal cells into UPR-high and UPR-low groups and systematically analyzed their intercellular communication networks. In the TME, distinct cell populations can function as senders, receivers, mediators, or influencers of specific signaling pathways, thereby shaping directed cellular crosstalk (30). Our analysis revealed that UPR-high luminal cells engage in communication with a broader repertoire of TME cells, particularly endothelial cells (Figures 3A–F), and function as more prominent senders, receivers, mediators, and influencers within key signaling pathways, including TNF, SPP1, IL6, VEGF, NOTCH, and EGF (Figures 3G–L). This prominent communicative role suggests that UPR-high luminal cells may critically regulate innate immune responses, inflammatory processes, tumor cell differentiation, migration, and ultimately, cancer cell survival. To support this notion, further analysis of the global signaling patterns identified UPR-high luminal cells as central hubs within both outgoing and incoming signaling networks (Supplementary Figure S3A). Their elevated connectivity within the TME underscores a role for the UPR that extends beyond cell-autonomous stress adaptation to include paracrine regulation of neighboring immune and stromal cells (31).

In this study, the UPR DEGs derived from single-cell analysis was consistently validated in bulk transcriptomes, confirming elevated UPR activity at the macroscopic tissue level. A volcano plot depicts the DEGs identified between UPR-high and UPR-low luminal cells from the scRNA-seq data, using thresholds of |logFC|>0.5 and p.adj<0.05 (Figure 4A). Then we quantified UPR activity for each sample in bulk transcriptomes using ssGSEA, which calculates the enrichment score of a predefined UPR gene set at the individual sample level. These UPR enrichment scores were subsequently employed as phenotypic traits for WGCNA (32). Following the removal of outlier samples and determination of an optimal soft-thresholding power (β = 6; Supplementary Figure S3B), a scale-free co-expression network was constructed. We then applied WGCNA to the TCGA-PRAD dataset to identify gene modules associated with UPR activity. The set of DEGs (logFC>0.5 and p < 0.05) derived from this single-cell analysis was utilized for co-expression network construction following the removal of outlier samples (Figure 4B). Hierarchical clustering identified five distinct gene modules (Figure 4C). Among these, the MEturquoise module exhibited the strongest positive correlation with UPR activity (cor = 0.45, p = 1e-26; Figure 4D). The functional relevance of the MEturquoise module to UPR was further corroborated by a strong positive correlation between gene significance (GS) for UPR and module membership (MM) within this module (cor = 0.58, p = 2.7e-82; Figure 4E; Supplementary Table S3), indicating that genes highly connected within the module are also central to UPR biology. A Topological overlap matrix (TOM) heatmap visually confirmed the high degree of co-expression connectivity among these UPRRGs within the module (Supplementary Figures S3C, D). Functional enrichment analysis of the UPRRGs revealed their significant involvement in biological processes and pathways directly related to UPR, including the response to ER stress, regulation of protein stability, and protein folding (Figures 4F, G).

To construct a consensus UPRRS, we employed a multi-step machine learning framework to refine the genes within the MEturquoise module (32). An initial subset of 39 genes was selected for further modeling based on two criteria: significance (p < 0.05) in univariate Cox regression, and high importance assessed by both mean decrease accuracy (MDA) and mean decrease gini (MDG) in random forest analysis (Figures 5A–D; Supplementary Table S4). By intersecting the genes, we identified a core set of 18 genes, designated as UPRRGs, which are implicated in the UPR at both bulk-tissue and single-cell resolutions (Figure 5E). These genes were subsequently entered into the LASSO-Cox regression pipeline. The TCGA-PRAD cohort was randomly split into a training set and an internal validation set at a 1:1 ratio. Genes retaining non-zero coefficients from the LASSO regression were further refined using backward stepwise Cox proportional hazards regression, which yielded a final signature comprising eighteen genes. The optimal penalization parameter (λ = 0.023155) for the LASSO model was determined through tenfold cross-validation by minimizing the partial likelihood deviance (Figures 5F–G; Supplementary Table S5). The final UPRRS model was formulated based on the expression levels of seven pivotal genes: IFRD1, REXO2, GPI, TMSB10, EVL, PDCD6, and RUVBL1. In the training cohort, the coefficient for each gene in the signature was positively associated with an increased risk of poor progression-free interval (PFI). Among these seven genes, IFRD1 demonstrated the highest hazard ratio (HR = 1.85, 95% CI [1.33-2.57], p < 0.001; Figure 5E). All seven genes were markedly upregulated in tumor tissues compared to normal counterparts (Figures 5H, I; Supplementary Figure S3E). The prognostic power of each gene was quantitatively assessed using ROC curve analysis. The AUC for predicting poor prognosis were 0.838 (RUVBL1), 0.819 (REXO2), 0.793 (IFRD1), 0.776 (EVL), 0.773 (GPI), 0.747 (PDCD6), and 0.693 (TMSB10), confirming their significant individual predictive value (Figure 5J). To validate the diagnostic robustness of our LASSO-derived signature, we performed Random Forest analysis. The model confirmed IFRD1 as the most influential predictor thereby independently verifying the clinical relevance of the UPRRS (Supplementary Figures S3F–G).

To establish the generalizability and clinical applicability of our model, we performed extensive validation across three independent cohorts, namely TCGA-PRAD, DKFZ2018 (23) and GSE116918 (20). Our results demonstrated that a higher UPRRS was significantly associated with an increased risk of PFI events. Principal component analysis (PCA) confirmed that the UPRRS effectively stratified patients into two distinct groups with significant prognostic separation in the TCGA-PRAD, DKFZ2018, and GSE116918 cohort. Consistent with this, patients classified into the UPRRS-high group exhibited a significantly higher risk score compared to those in the UPRRS-low group across all validation cohorts (Figures 6A–C). To facilitate clinical translation, we developed a PFI-specific nomogram that integrates the UPRRS to provide quantitative predictions of patient survival probabilities at 1, 3, and 5 years (Figure 6D). The robust predictive performance of the UPRRS was further validated by time-dependent ROC analysis. The model demonstrated consistent prognostic accuracy across multiple time points in the TCGA-PRAD cohort [AUC: 0.70 (1-year), 0.72 (3-year), 0.69 (5-year)], the DKFZ2018 cohort [AUC: 0.75 (1-year), 0.78 (3-year), 0.76 (5-year)], and the GSE116918 cohort [AUC: 0.61 (1-year), 0.63 (3-year), 0.75 (5-year)] (Figure 6E). To facilitate individualized survival estimation, we developed a clinically applicable nomogram that integrates the UPRRS with key clinical parameters. Univariate and multivariate Cox regression analyses consistently identified the UPRRS and T stage as the most robust independent predictors of overall survival (OS) in PCa (univariate HR = 3.801, multivariate HR = 2.380, both p < 0.001; Figures 7A, B). In contrast, age and N stage did not retain independent prognostic significance in the multivariate model. Consequently, the final nomogram incorporates only the UPRRS and clinical T (cT) stage to provide a practical tool for predicting survival probabilities in PCa patients (Figure 7C). The integrated nomogram exhibited superior predictive accuracy compared to the UPRRS alone, with time-dependent AUCs of 0.79 (1-year), 0.81 (3-year), and 0.81 (5-year) (Figure 7D). The calibration curves for 1-, 3-, and 5-year survival showed excellent agreement between nomogram predictions and observed outcomes (Figures 7E–G), confirming the model’s high reliability and clinical interpretability.

The UPRRS stratified patients into distinct high- and low-risk groups, effectively capturing the underlying heterogeneity among them. Analysis of immune cell infiltration using Cell-type Identification by Estimating Relative Subsets of RNA Transcripts (CIBERSORT) revealed distinct landscape between the groups. The high-risk group was characterized by a higher abundance of regulatory T cells (Tregs) (Supplementary Figures S4A, B). We next investigated the relationship between UPR and immune activity by comparing the expression profiles of key immune checkpoint molecules between the two groups. Multiple immune checkpoints, including ADORA2A, CTLA4, HAVCR2, IDO1, IL10RB, LAG3, LGALS9, NECTIN2, PDCD1, TGFB1, and TIGIT, were significantly upregulated in the high-risk group, whereas KDR expression was notably lower (p < 0.05; Figure 8A). Furthermore, drug sensitivity analysis revealed that patients in the high-risk group exhibited increased sensitivity to several agents, including docetaxel, tanespimycin, and specific carboplatin-based combinations (e.g., BRD-A02303741-carboplatin, decitabine-carboplatin, tretinoin-carboplatin), but decreased sensitivity to carboplatin alone (Figure 8B). Functional enrichment analysis of the seven genes in UPRRS identified significant involvement in specific biological contexts, notably “cytoskeleton in muscle cells” and the “Neuroactive ligand-receptor interaction” pathway (Figures 8C, D). Collectively, across all three cohorts, high-risk tumor tissues consistently exhibited co-upregulation of oncogenic and metabolic hallmarks coupled with UPR activation. This pattern highlights a critical protein homeostasis-bioenergetics axis that underpins aggressive PCa progression (Figure 8E).

To pinpoint the key regulatory gene within the UPRRS, we analyzed the seven genes and identified IFRD1 as a central hub gene consistently prominent in both single-cell and bulk transcriptomic analyses. Given its significant correlation with poor prognosis, we systematically investigated the biological role of IFRD1 in PCa pathogenesis.

Pan-cancer analysis initially revealed that IFRD1 was significantly upregulated in tumor tissues compared to adjacent normal tissues across multiple cancer types, including PCa (Figure 9A). In the TCGA-PRAD cohort, IFRD1 expression was confirmed to be markedly elevated in primary tumor tissues compared to normal controls (Figure 9B). Interestingly, at the single-cell resolution in the GSE141445 dataset, IFRD1 expression was significantly enriched in UPR-high luminal cells (Figure 9C). To validate the single-cell findings at the bulk tissue level, we examined the correlation between IFRD1 expression and UPR signature scores in the TCGA-PRAD cohort. A significant positive correlation (R = 0.33, p < 0.001) was observed, confirming that the IFRD1-UPR axis observed in individual cells is preserved in whole-tissue transcriptomes (Figure 9D). This expression pattern was consistently validated across two additional independent scRNA-seq datasets (GSE185344 and Cell Reports), where IFRD1 remained highly expressed in UPR-high luminal cells (Figures 9E–G). At the protein level, IHC staining from both internal and external cohort demonstrated substantially stronger IFRD1 expression in PCa tissues compared to matched normal tissues (Figure 9H). To explore the genomic alterations associated with IFRD1-mediated UPR activation, we compared the somatic mutation landscapes between the UPR-high and UPR-low groups. This analysis revealed distinct mutation patterns, with the UPR-high group exhibiting a higher mutational burden in key driver genes such as TP53, SPOP, and FOXA1. This finding suggests a potential mechanistic link between IFRD1-mediated UPR activation and genomic instability in PCa (Figure 9I). Kaplan–Meier survival analysis across multiple independent cohorts - TCGA-PRAD, DKFZ2018, GSE116918, and GSE21034 (21, 22) - consistently demonstrated that elevated IFRD1 expression was significantly associated with a shorter PFI, thereby confirming its robust prognostic value (Figure 9J). Furthermore, IFRD1 expression showed significant positive correlations with key indicators of disease aggressiveness, including advanced Gleason score, and higher T and N stages (Figure 9K). To elucidate the mechanisms underlying IFRD1-driven tumor progression, we performed functional enrichment analysis comparing luminal-IFRD1-positive and luminal-IFRD1-negative subgroups within the GSE141445 dataset. This analysis revealed that high IFRD1 expression was significantly enriched in pathways central to the UPR and ER stress, encompassing protein processing in the ER, ER-nucleus signaling, and estrogen response early pathways. In stark contrast, the IFRD1-low subgroup showed minimal enrichment for these UPR-related pathways, confirming IFRD1’s role as a critical upstream regulator that drives UPR activation, modulates protein homeostasis, and facilitates PCa progression. (Figures 10A, B). GSEA of the DEGs further associated IFRD1 with the activation of hallmark oncogenic pathways, including hypoxia, epithelial-mesenchymal transition (EMT), and androgen response (Figure 10C). These findings were robustly validated in TCGA-PRAD bulk RNA-seq cohort (Figure 10D). Furthermore, to explore IFRD1’s influence on the tumor microenvironment, we performed immune cell deconvolution analysis. High IFRD1 expression was significantly associated with increased infiltration of plasma cells, while rest mast cells and B cells were reduced, indicating that IFRD1 actively remodels the immune landscape toward a pro-tumorigenic state (Figures 10E, F). Integrated analysis revealed that IFRD1 expression correlates with a unique co-expression network, advanced pathology, altered immune infiltration, and poor prognosis as well as immunotherapy non-response across pan-cancer level (Figure 10G; Supplementary Figure S4C–I). This comprehensive profile suggests that IFRD1 functions not merely as a passive marker, but as an active transcriptional regulator and immune modulator, ultimately shaping clinical outcomes. To functionally validate the oncogenic role of IFRD1, we conducted in vitro experiments using the C4–2 PCa cell line. The efficiency of siRNA-mediated IFRD1 knockdown was first confirmed by qRT-PCR (Figure 10H). Functional assays demonstrated that IFRD1 knockdown significantly impaired cellular proliferation and suppressed migratory capacity (Figures 10I, J).

In summary, our multi-faceted investigation establishes IFRD1 as a critical mediator of UPR signaling in PCa, which drives tumor progression by coordinating oncogenic pathway activation and remodeling the tumor immune microenvironment. The compelling prognostic association and functional indispensability of IFRD1 position it as a promising therapeutic target for combating PCa.

Although this study systematically elucidated the prognostic value and therapeutic potential of UPR in prostate cancer through a three-tier integrated strategy spanning single-cell, bulk, and functional dimensions, this study has several limitations that should be acknowledged. First, our analyses are predominantly derived from retrospective analyses of public datasets and thus lack validation in prospective clinical cohorts. Second, although the UPRRS demonstrated robust predictive performance in the cohorts analyzed, its generalizability and clinical utility require further verification in larger, more diverse patient populations. Furthermore, while our in vitro findings support the oncogenic role of IFRD1, its mechanistic functions require further in vivo validation. We have therefore initiated drug discovery efforts to identify clinically applicable drugs that can inhibit IFRD1 or even modulate UPR activity. Finally, given that the UPR is a highly dynamic and context-dependent signaling network, its functional roles and the prognostic value of the UPRRS may vary under different therapeutic pressures, including ADT, radiotherapy, and immunotherapy. Consequently, future investigations should prioritize evaluating the predictive value of the UPRRS within specific treatment contexts.