Single-cell transcriptomic analysis and machine learning identify ATAD3A as a key gene that stabilizes mitochondrial-endoplasmic reticulum membranes, promoting bladder cancer progression

Abstract

Background

Mitochondria-associated endoplasmic reticulum membranes (MAM) play a critical regulatory role in cancer, yet their function in bladder cancer (BCa) remains unclear.

Methods

This study conducted a comprehensive analysis of extensive RNA sequencing and single-cell transcriptomic data. A MAM risk feature model was constructed and evaluated using LASSO-Cox regression, Kaplan-Meier survival curves, and receiver operating characteristic (ROC) analysis. Cell communication networks were decoded using CellChat, and tumor-infiltrating immune cells were quantified through CIBERSORT. Machine learning, OncoPredict, Scissor algorithm, and spatial transcriptomics were employed to analyze chemical reactions. The tumor-promoting functions and molecular mechanisms of ATAD3A were validated through cell and animal models, as well as transmission electron microscopy and fluorescence confocal microscopy.

Results

This study systematically delineates the gene expression profile of MAM in BCa and constructs a robust MAM prognostic signature that effectively predicts poor patient outcomes. A high MAM score is significantly associated with an immunosuppressive microenvironment and chemotherapy resistance. Using machine learning, we developed a random forest model that successfully identified ATAD3A as a key gene predicting cisplatin resistance, which is significantly correlated with platinum resistance at both single-cell and spatial transcriptomic levels. Further validation shows that ATAD3A is highly expressed in BCa tissue, and its knockdown significantly suppresses tumor growth. Mechanistically, ATAD3A maintains MAM structural integrity, regulates mitochondrial calcium homeostasis and membrane potential, thereby promoting cellular homeostasis and enhancing chemotherapy resistance.

Conclusion

This study connects MAM to BCa chemotherapy resistance through machine learning and multi-omics analysis, establishing ATAD3A as a prognostic biomarker and potential therapeutic target in BCa.

Graphical Abstract

Supplementary Information

The online version contains supplementary material available at 10.1186/s12967-026-07857-0.

Introduction

Bladder cancer (BCa) is the fourth most common cancer in men and the thirteenth leading cause of cancer-related death globally [1]. In 2020, nearly 600,000 individuals were newly diagnosed with BCa worldwide, a figure projected to double by 2040 [2]. At initial diagnosis, approximately three-quarters of patients present with non-muscle-invasive BCa (NMIBC), while the remaining 25% have muscle-invasive BCa (MIBC), and 5% exhibit metastatic disease [3, 4]. Transurethral resection of bladder tumor (TURBT) is the primary treatment for NMIBC; however, the five-year recurrence rate postsurgery ranges from 31% to 78%, with progression rates ranging from 1% to 45% [3]. For advanced metastatic BCa, cisplatin-based chemotherapy serves as the first-line treatment, although the development of drug resistance significantly compromises the long-term prognosis [5]. Furthermore, BCa is high heterogeneity because of its complex regulatory mechanisms, investigating the underlying mechanisms driving BCa development is essential for predicting long-term patient outcomes.

Mitochondria-associated endoplasmic reticulum membrane (MAM) is membrane contact sites between the endoplasmic reticulum (ER) and mitochondria, first identified in 1990 [6]. The ER and mitochondrial membranes do not fuse but maintain an intermembrane distance ranging from 10 to 25 nm, with varying distances mediating distinct biological functions [7–9]. By serving as communication bridges, MAMs facilitate the frequent exchange of proteins and metabolites—including Ca²⁺, lipids, and reactive oxygen species (ROS)—between these organelles [10, 11]. Emerging evidence highlights the MAM as a critical signaling hub in cancer biology [10, 12]. Substantial evidence indicates that oncogenes and tumor suppressor genes are dynamically recruited to MAMs, where they regulate cancer cell fate by influencing Ca²⁺ exchange, lipid homeostasis, autophagy, ER stress, and mitochondrial dynamics dysregulation [13]. IP3R, the core gene governing Ca²⁺ transfer at MAMs, is upregulated in multiple cancers (e.g., colon, ovarian, liver, and pancreatic cancer) and promotes cancer cell survival [14–17]. Lipid transport-related proteins such as MFN2, ACAT1, and ATAD3A facilitate cancer cell evasion of apoptosis by promoting cholesterol metabolism [18–20]. The Activation or inactivation of the ER stress sensor PERK pathway can exert dual effects, either by promoting tumor growth or mediating cell death [21]. However, despite its established role in other malignancies, the functional landscape of MAMs in BCa progression and therapy response remains largely uncharted. It is unknown whether a coherent MAM-related molecular signature exists in BCa, whether it influences the tumor microenvironment or clinical outcome, and which key regulators might drive its pathogenic functions.

To address these gaps, we performed an integrated transcriptomic analysis to systematically delineate the MAM landscape in BCa. We established a novel MAM-based prognostic signature and dissected its association with the immunosuppressive microenvironment and chemoresistance. Furthermore, we identified and functionally validated ATAD3A as a central oncogene that stabilizes MAM integrity to promote BCa progression. Our study not only establishes MAM dysfunction as a critical driver in BCa but also nominates ATAD3A as a promising prognostic biomarker and therapeutic target.

Results

Evaluation of MAM expression and prognostic value in BCa

A schematic diagram summarizing the study design is provided (Graphical Abstract Image). Transcriptomic analysis across several datasets revealed differential expression of 28 MAM-related genes between BCa and normal tissues. Compared with those in normal samples, the expression of most MAM-related genes in BCa samples was altered. Specifically, genes including ATAD3A, BCAP31, CANX, DGAT2, and VDAC1 were consistently upregulated in BCa tissues, whereas ACAT1 and CAV1 were significantly downregulated (Fig. 1A). Cox regression analysis within the TCGA-BLCA cohort was employed to examine correlations and prognostic value among MAM-related genes. CANX, ATP2A2, and ATAD3A were identified as risk factors and demonstrated strong correlations with other genes, suggesting their potential role as core regulatory genes (Fig. 1B). We integrated BCa single-cell datasets. Following batch correction, dimensionality reduction, and clustering, 13 distinct cell clusters were identified (Fig. 1C). These clusters were annotated into eight major cell populations using established marker genes from the literature (Fig. 1D and E). Notably, the proportion of T cells was greater in the NMIBC group than in the other groups. Differentially expressed genes (DEGs) across cell populations were identified via the FindMarkers function to analyze the distribution of MAM-related genes. BCAP31, CANX, HSPA9, and VDAC1 were widely distributed across various cell types, whereas ATAD3A and DNM1L were predominantly highly expressed in epithelial cells (Fig. 1F). These findings suggest that MAM-related genes may function collaboratively across different cell types to promote BCa development.

Fig. 1.

Expression patterns and prognostic value of MAM-related genes in BCa. (A) Heatmap of MAM-related gene expression across BCa datasets. (B) Prognostic significance (Wilcoxon test) and intergene correlations of MAM-related genes. (C) UMAP visualization of cell clusters after batch correction, dimensionality reduction, and clustering. (D) Dot plot displaying marker gene expression across cell populations. (E) UMAP projection of annotated cell populations. (F) Heatmap of differentially expressed MAM-related genes across cell subpopulations with functional enrichment analysis. ns: not significant; *p < 0.05; **p ≤ 0.01; ***p ≤ 0.001

MAM score influences intercellular communication in BCa

To investigate the regulatory mechanisms of MAMs in BCa and their role in the pathological process, we assessed MAM activity across different groups via three gene set scoring algorithms: PercentageFeatureSet, AddModuleScore, and AUCell. Fibroblasts and endothelial cells presented notably higher MAM scores (Fig. 2A). Comparison of MAM scores between groups using Wilcoxon rank-sum tests revealed significantly higher MAM scores in epithelial cells from the BCa group compared to the normal group (Fig. 2A). Furthermore, epithelial cells in the MIBC group exhibited significantly higher MAM scores than those in the NMIBC group, suggesting a potential role for MAM-related genes in BCa progression (Fig. 2B). Furthermore, fibroblasts in the MIBC group displayed significantly lower MAM scores than those in the NMIBC group did (Fig. 2B). Based on the robustness of the AUCell algorithm in single-cell data analysis and its advantages for biological grouping, we utilized a threshold of 0.5 to classify cells into high and low MAM scoring groups. Subsequently, we employed CellChat to compare the intercellular communication networks between these two groups. Compared with the low-MAM score group, the high-MAM score group presented a substantial increase in both the number and strength of cell‒cell interactions (Fig. 2C). Specifically, within the high MAM score group, communication between epithelial cells and fibroblasts, macrophages, and endothelial cells was significantly enhanced in both quantity and intensity (Fig. 2D). Conversely, interactions among immune cells (including B cells, T cells, macrophages, and NK cells) were attenuated in the high MAM score group relative to the low MAM score group (Fig. 2D). Analysis of signaling pathway strength revealed enhanced activity of multiple pathways—such as SPP1, IGF, and PARs (protease-activated receptors)—in the high MAM score group (Fig. 2E). Further investigation into differential ligand‒receptor interactions revealed several activated pairs, including MDK-SDC1, LGALS9-PTPRC (CD45), LGALS9-CD44, and GDF15-TGFBR2, specifically within the high MAM score group involving epithelial cells and other cell types (Fig. 2F). In contrast, the MDK-NCL interaction was more prominent in communication between epithelial cells and T cells/NK cells within the low MAM score group (Fig. 2F). These results indicate that MAM-related genes modulate intercellular communication patterns within the tumor microenvironment, which may contribute to the development and progression of BCa.

Fig. 2.

Impact of MAM scores on intercellular communication in BCa. A, B. MAM scores per cell population calculated using three gene-set scoring methods, with comparative analysis across groups. C. Bar plot showing differential interaction strength and network size between high/low MAM score groups. D. Circos plot depicting upregulated (red) and downregulated (blue) cellular interactions in high vs. low MAM score groups. E. Pathway activity comparison between MAM score groups. F. Upregulated/downregulated epithelial-ligand pairs in high MAM score group. ns: not significant; *p < 0.05; **p ≤ 0.01; ***p ≤ 0.001

Construction and evaluation of an MAM-based signature for BCa

To investigate the clinical characteristics of MAM molecular subtypes in BCa, samples from the TCGA-BLCA cohort were stratified into two groups (C1 and C2) based on the optimal cluster number (k = 2) (Fig. 3A). KM analysis revealed that patients with the C1 subtype had significantly worse prognosis compared to the C2 subtype (Fig. 3B). This indicates that MAM-related genes can distinguish BCa samples and predict adverse patient outcomes. The LASSO algorithm was applied to construct a prognostic risk model based on the 28 MAM-related genes within the TCGA-BLCA cohort. This analysis identified four key genes (ACSL4, ATAD3A, FUNDC1, and PSEN2) for the MAM signature. Using the TCGA-BLCA dataset as the training set and GSE13507 as the validation set, the distribution of survival status and KM curves were plotted (Fig. 3C and D). KM analysis confirmed the prognostic value of the MAM score, demonstrating that patients with high MAM scores had significantly shorter overall survival (OS) than those with low scores (TCGA-BLCA: p < 0.001; GSE13507: p = 0.008) (Fig. 3E and F). Furthermore, the predictive power of the MAM risk model was assessed using time-dependent receiver ROC curves. The area under the curve (AUC) values in the TCGA-BLCA cohort were 0.725 at 1 year, 0.728 at 3 years, and 0.832 at 5 years, while in the GSE13507 cohort they were 0.680 at 1 year, 0.673 at 3 years, and 0.823 at 5 years (Fig. 3G and H). Univariate and multivariate Cox regression analyses confirmed that a high MAM score was an independent risk factor for shorter OS (p < 0.001; HR = 1.217; 95% CI: 1.159–1.278) (Fig. 3I). Finally, a nomogram was generated to integratively assess the MAM score a longside other clinical features for predicting patient OS at 1, 3, and 5 years (Fig. 3J). In summary, these results demonstrate that the MAM score may serve as a valuable prognostic biomarker in BCa.

Fig. 3.

Construction and validation of the MAM-based prognostic signature in BCa patients. (A) Consensus clustering NMF analysis stratifying TCGA-BLCA patients into two subtypes. (B) Kaplan-Meier survival curves comparing subtypes. C, D. Distribution of MAM scores and survival status in TCGA-BLCA and GSE13507 cohorts. E, F. Kaplan-Meier analysis of high/low MAM score groups. G, H. Time-dependent ROC curves for 1-, 3-, and 5-year survival prediction. I. Univariate/multivariate Cox regression of MAM score and clinical variables. J. Nomogram predicting 1-, 3-, and 5-year OS. *p < 0.05; **p ≤ 0.01; ***p ≤ 0.001

MAM score correlates with immunotherapy and chemotherapy response in BCa

To investigate the relationship between the MAM score and immunotherapy response, patients in the IMvigor210 cohort were stratified into high and low MAM risk groups based on their MAM scores. Patients with high MAM scores exhibited significantly shorter OS than those with low scores (Fig. 4A). The MAM score demonstrated good predictive ability for patient prognosis in the IMvigor210 cohort, with AUC values of 0.715 at 1 year, 0.559 at 3 years, and 0.536 at 5 years (Fig. 4B). Using CIBERSORT analysis in the TCGA-BLCA cohort, we assessed differences in the infiltration levels of 20 tumor-infiltrating immune cell types between the high-risk and low-risk groups. The high-risk group showed significantly reduced infiltration of activated CD4 memory T cells and significantly increased infiltration of regulatory T cells (Tregs) (Fig. 4C). We further analyzed the correlation between the expression of signature genes constituting the risk model and immune cell infiltration levels. ATAD3A expression was significantly positively correlated with Treg infiltration levels (Fig. 4D). Conversely, ACSL4 expression was significantly negatively correlated with Treg infiltration but significantly positively correlated with M2 macrophage infiltration (Fig. 4D). Furthermore, the association between the MAM score and sensitivity to chemotherapeutic agents was evaluated using the R package “oncoPredict”. Cells with high MAM scores exhibited significantly increased half-maximal inhibitory concentration (IC₅₀) values for cisplatin, erlotinib, and paclitaxel, indicating reduced drug sensitivity (Fig. 4E). These results suggest that the MAM score influences the response to both immunotherapy and chemotherapy in BCa.

Fig. 4.

Association of MAM scores with immunotherapy response and chemosensitivity. (A) Kaplan-Meier survival analysis in IMvigor210 cohort. (B) Time-dependent ROC curves for anti-PD-L1 response prediction. (C) Violin plots showing differential immune infiltration between MAM score groups. (D) Dot plot of correlations between signature genes and immune cells. (E) Differential IC₅₀ values for chemotherapeutic agents between groups. *p < 0.05; **p ≤ 0.01; ***p ≤ 0.001

Screening of chemotherapy resistance genes in BCa using machine learning

We further explored the performance of various machine learning models in predicting the response to cisplatin treatment in BCa using MAM. The GSE169455 dataset served as the training set, while GSE5287 and GSE247185 were utilized as validation sets 1 and 2. Results indicated that the Random Forest (RF) model exhibited the best and most stable performance across both independent validation sets, with an AUC of 0.708 for validation set 1 and 0.850 for validation set 2, yielding an average validation AUC of 0.779 and a minimal difference between the two sets (ΔAUC = 0.142) (Fig. 5A). The model combination strategy showed varied results: the glmBoost + RF combination performed optimally in validation set 1 (AUC = 0.784) but declined in validation set 2 (AUC = 0.700) (Fig. 5A). The optimal RF model identified four key gene features. SHapley Additive exPlanations (SHAP) analysis revealed that the RF model constructed on these four key genes exhibited perfect classification performance in the training set (accuracy 100%, sensitivity and specificity both 100%). The SHAP beeswarm plot indicated that ATAD3A and ACSL4 were the two genes contributing most significantly to the model’s predictions, with average absolute SHAP values of 0.105 each (Fig. 5B). Higher expression levels of ATAD3A corresponded to higher SHAP values, indicating a prediction towards positive (cisplatin non-responder) (Fig. 5B). The waterfall plot illustrated the SHAP value decomposition for the first sample in the training set. ATAD3A and FUNDC1 provided positive contributions, while PSEN2 and ACSL4 contributed negatively, ultimately leading to the prediction of this sample as part of the treatment group (Fig. 5C).

Fig. 5.

Screening of chemotherapy resistance genes in bca using machine learning. A. AUC performance of various machine learning models in predicting cisplatin treatment response across two independent validation sets. The RF model demonstrated the best and most stable performance. B-C. SHAP analysis elucidating the decision-making of the RF model. (B) The beeswarm plot indicates that ATAD3A and ACSL4 are the two features with the highest contributions. (C) The waterfall plot illustrates the contribution direction of key genes to the prediction outcome for a single sample in the training set

Investigating the relationship between ATAD3A and chemotherapy resistance in BCa at the single-cell and spatial transcriptomic levels

Based on the significant role of ATAD3A in the model, we further investigated its association with chemotherapy resistance in BCa. Utilizing single-cell sequencing data from two MIBC cohorts (GSE135337 and GSE130001), we performed mitochondrial quality control, dimensionality reduction, clustering, and reannotation into five cell populations (Fig. 6A). Epithelial cells were extracted and analyzed, resulting in the identification of 10 subpopulations after dimensionality reduction and clustering (Fig. 6B). Using GDSC data for cell comparison, the Scissor algorithm was employed to separate cisplatin-resistant (red) and sensitive (blue) cell populations, with the remaining cells marked as undefined (Fig. 6C). ATAD3A expression was significantly higher in the cisplatin-resistant cell population compared to the sensitive group (Fig. 6D). Furthermore, we examined the spatial transcriptomic data to observe the expression distribution of ATAD3A alongside platinum resistance genes. RGB composite visualization showed a significant spatial overlap between regions of high ATAD3A expression (red) and elevated platinum resistance scores (blue), forming a bright purple area, indicating that these spatial locations exhibit both high ATAD3A expression and characteristics of platinum resistance (Fig. 6E).

Fig. 6.

Investigating the relationship between ATAD3A and chemotherapy resistance in BCa at the single-cell and spatial transcriptomic levels. (A) Single-cell transcriptomic atlas of MIBC, revealing five major cell populations. (B) Re-clustering of epithelial cell subpopulations yielded ten distinct subgroups. (C) The Scissor algorithm was applied to identify cisplatin-resistant (red) and sensitive (blue) cell populations within the epithelial cells. (D) ATAD3A expression was significantly higher in the resistant epithelial cell population compared to the sensitive group. (E) Spatial transcriptomic visualization demonstrated a high spatial overlap (bright purple) between regions of high ATAD3A expression (red) and platinum resistance characteristics (blue)

Validation of MAM key gene ATAD3A expression and prognostic value in BCa

Given that ATAD3A, a key gene in the MAM risk model, is predominantly expressed in epithelial cells and is closely associated with immune cell infiltration, we conducted further investigation. IHC staining for ATAD3A was performed on paired BCa and adjacent normal tissues. ATAD3A expression was significantly upregulated in BCa tissues compared to adjacent normal tissues (Fig. 7A and B). ROC curve analysis demonstrated that the IHC score for ATAD3A had high specificity and sensitivity for predicting BCa (AUC = 0.936, 95% CI = 0.890–0.982) (Fig. 7C). Patients were stratified into high and low IHC score groups based on the median IHC score. KM survival analysis revealed that patients with high ATAD3A IHC scores had significantly shorter OS than those with low scores (Fig. 7D). We further analyzed the correlation between ATAD3A expression and sensitivity to chemotherapeutic agents in BCa. The results revealed that ATAD3A expression was significantly negatively correlated with sensitivity to cisplatin, docetaxel, paclitaxel, and vinblastine (Fig. 7E). Collectively, these findings demonstrate that ATAD3A is upregulated in BCa and is closely associated with chemotherapy response.

Fig. 7.

Validation of ATAD3A as a key MAM-related gene in BCa. (A) Representative IHC staining of ATAD3A in normal and BCa tissues. (B) Quantitative comparison of ATAD3A IHC expression. (C) ROC curve assessing diagnostic performance of ATAD3A IHC score. (D) Kaplan-Meier analysis stratified by ATAD3A IHC expression. (E) Correlation between ATAD3A expression and chemosensitivity. “n” denote the number of patient samples

ATAD3A promotes BCa progression

To further investigate the biological function of ATAD3A in BCa, a scratch wound healing assay was employed to assess the impact of ATAD3A on BCa cell migration. ATAD3A knockdown markedly decreased the migration rate of BCa cells at 24 h, whereas ATAD3A overexpression increased their migratory ability (Fig. 8A and B). Silencing ATAD3A significantly attenuated the proliferative capacity of BCa cells, as evidenced by reduced colony formation ability, whereas ATAD3A overexpression enhanced cell proliferation (Fig. 8C and D). Importantly, silencing ATAD3A significantly increased cisplatin sensitivity in bladder cancer (BCa) cells, evidenced by reduced colony formation and decreased cisplatin IC₅₀, whereas overexpression of ATAD3A decreased cisplatin sensitivity (Figs. 8C-F). A subcutaneous xenograft tumor model was established in nude mice using T24 cells stably expressing either ATAD3A knockdown or overexpression constructs (Fig. 8G). During the monitoring period, no significant differences in body weight were observed among the groups (Fig. 8H). Tumor growth was significantly inhibited in the ATAD3A knockdown group, with smaller tumor volumes observed over time compared to other groups (Fig. 8I). Following sacrifice, tumors were excised, rinsed with PBS, and photographed. Gross examination revealed that tumors in the ATAD3A overexpression group were substantially larger, while tumors in the knockdown group were the smallest (Fig. 8J). Measurements of tumor volume and weight further corroborated these findings (Fig. 8K). Collectively, these in vitro and in vivo experimental results demonstrate that ATAD3A promotes BCa progression, highlighting its potential as a promising therapeutic target.

Fig. 8.

ATAD3A promotes BCa progression. A-B. Scratch wound healing assays evaluating migratory ability of BCa cells following ATAD3A knockdown or overexpression. C-D. Colony formation assays assessing proliferative capacity of BCa cells following ATAD3A knockdown or overexpression. E-F. The CCK-8 assay was employed to assess the cisplatin IC50 in BCa cells following ATAD3A knockdown or overexpression. G. Subcutaneous xenograft model schematic. n = 6 mice/group. H. Mouse body weight monitoring during study. I. Tumor growth kinetics. J. Excised tumor photographs. K. Quantitative analysis of tumor volume/weight. Data are derived from at least 3 independent experiments, presented as mean ± standard error of the mean (SEM). * p < 0.05; ** p ≤ 0.01; *** p ≤ 0.001

ATAD3A maintains the integrity of MAM in BCa

We further investigate whether ATAD3A promotes BCa progression by regulating the coupling between mitochondria and ER.Transmission electron microscopy revealed that overexpression of ATAD3A resulted in a reduced distance and increased contact area between mitochondria and the ER (Fig. 9A-C). In contrast, knockdown of ATAD3A led to an increased distance between these organelles (Fig. 9A-C). Immunofluorescence co-localization further confirmed that overexpression of ATAD3A increased the co-localization of mitochondria and ER, while knockdown of ATAD3A reduced this co-localization (Fig. 9D and E). Mitochondria-associated membranes (MAM)-mediated calcium transport between mitochondria and ER is critical for mitochondrial and cellular homeostasis. Knockdown of ATAD3A resulted in calcium overload in mitochondria (Fig. 9F and G). Furthermore, knockdown of ATAD3A led to a decrease in mitochondrial membrane potential, whereas overexpression of ATAD3A maintained the stability of the mitochondrial membrane potential (Fig. 9H and I). Consistently, knockdown of ATAD3A resulted in increased production of ROS (Fig. 9J and K). Knockdown of ATAD3A resulted in an increased proportion of cell apoptosis (Fig. 9L and M). The above results suggest that ATAD3A maintains the integrity of MAM in BCa cells, thereby preserving mitochondrial stability and promoting cellular homeostasis.

Fig. 9.

ATAD3A maintains the integrity of MAM in BCa. A-C. Representative TEM images showing the contact between mitochondria and ER in BCa cells (A). The red line indicates the contact area between mitochondria and ER. The red lines with arrows denote the contact distance between mitochondria and ER. Quantification of the average distance between mitochondria and ER (B) and the percentage of contact surface area between mitochondria and ER (C). D and E. Representative confocal images of BCa cells with ATAD3A knockdown or overexpression, stained with Mito-Tracker (red) and ER-Tracker (green) (D), along with quantification of mitochondrial-associated contact sites (MCC) (E). F and G. Rhod-2 fluorescence staining of calcium ion flux in mitochondria upon ATAD3A knockdown compared to the control group (F), along with representative images and quantification (G). H-I. Representative fluorescence images of mitochondrial membrane potential in ATAD3A knockdown or overexpression cells compared to the control group (H), along with quantification (I). J and K. Representative fluorescence images of ROS detection in ATAD3A knockdown cells compared to the control group (J), along with quantification (K). L and M. Representative fluorescence images of apoptosis detection in ATAD3A knockdown or overexpression cells compared to the control group (L), along with quantification (M). n = 3 independent experiments. * p < 0.05; ** p ≤ 0.01; *** p ≤ 0.001

Discussion

BCa ranks among the most prevalent malignancies associated with mortality [1]. The high tumor heterogeneity of BCa limits the predictive accuracy of single clinical indicators, such as the TNM staging system, for patient overall survival [22]. Consequently, there is an ongoing need to identify novel predictive biomarkers associated with BCa prognosis and treatment strategies. The role of MAM in cancer has garnered increasing attention; however, their function in BCa remains poorly characterized.

Through integration of bulk and single-cell transcriptomic data from public databases, we demonstrated that 28 MAM-related genes have different and stable expression patterns in BCa. A risk signature based on four MAM-related genes robustly predicted poorer overall survival in BCa cohorts. Multivariate Cox regression analysis further confirmed the independent prognostic value of this signature. Therefore, the MAM score serves as an indicator of tumor progression and a valuable tool for predicting BCa patient prognosis. By integrating machine learning methods to systematically identify chemotherapy resistance drivers among MAM-related genes, SHAP interpretability analysis further reveals that ATAD3A is one of the most significant features contributing to the model, with its high expression significantly associated with cisplatin resistance. This finding is supported by single-cell and spatial transcriptomic data. ATAD3A knockdown significantly inhibited BCa cell proliferation and migration in vitro, as well as tumor growth in vivo. Mechanistically, ATAD3A stabilizes the integrity of mitochondria-associated membranes (MAM), thereby regulating calcium ion signaling and maintaining mitochondrial membrane potential, which reduces ROS production and preserves homeostasis in BCa cells.

Cell communication analysis revealed that a high MAM score enhanced epithelial-fibroblast-endothelial cell crosstalk in BCa (e.g., via MDK-SDC1 and LGALS9-CD44 interactions) while suppressing immune cell interactions (e.g., T-cell and NK cell communication through MDK-NCL), suggesting the role of MAM in remodeling the tumor microenvironment (TME) toward an immunosuppressive state. Previous studies indicate that MAM protein interactions facilitate the rapid differentiation of memory T cells into effector T cells [23]. Han et al. demonstrated that elevated cholesterol induces CD8⁺ T cell exhaustion in colon cancer by promoting intermolecular MAM interactions [24]. Modulation of MAM protein interactions has been proposed to influence glycan expression and enhance immune cell recognition, representing a potential therapeutic strategy for glioblastoma [25]. In the IMvigor210 cohort, patients with high MAM scores exhibited poorer survival following anti-PD-L1 therapy, potentially attributable to increased Treg infiltration and reduced activated CD4⁺ memory T cells—a TME phenotype consistent with Chen et al.‘s observations in MAM-dysregulated hepatocellular carcinoma [26]. Xie et al. recently reported that targeting the ATAD3A-PINK1 axis enhances the sensitivity of breast cancer to chemotherapy by promoting the degradation of PD-L1 through mitophagy [27]. The protein interaction network on the MAM may influence the expression or stability of immune regulatory molecules on the surface of tumor cells, such as the localization and degradation of PD-L1. However, the immunomodulatory functions of MAM in BCa require further mechanistic characterization.

Moreover, a high MAM score is associated with chemotherapy resistance in BCa, as evidenced by elevated IC₅₀ values for cisplatin and paclitaxel. The literature strongly supports the critical involvement of MAMs in drug resistance in tumors. Li et al. observed that GRP75 promotes MAM integrity and mediates cisplatin resistance in ovarian cancer patients [28]. Conversely, upregulated IP3R enhances cisplatin cytotoxicity in ovarian cancer cells by increasing Ca²⁺ flux through targeting the BH4 domain of Bcl-2 [29]. PML deficiency reduces MAM-mediated Ca²⁺ transfer, mitochondrial respiration, and ATP production, stimulating autophagy via the AMPK/mTOR/ULK1 pathway and sensitizing cancer cells to 5-fluorouracil [30]. Çoku et al. found that reduced MAM connectivity promotes multidrug resistance (MDR) in neuroblastoma [31]. The role of MAM in BCa and its drug resistance is still poorly understood. Interestingly, ATAD3A expression negatively correlates with sensitivity to chemotherapeutic agents such as cisplatin and docetaxel, positioning it as a predictor of treatment failure. Although additional experimental data is required for confirmation, the molecular network regulated by ATAD3A within the MAM structure represents a highly promising target for overcoming BCa and its resistance. ATAD3A expression is elevated across multiple malignancies including lung adenocarcinoma, prostate cancer, head and neck squamous cell carcinoma, glioma, cervical cancer, and breast cancer [32–35]. In breast cancer, ATAD3A overexpression promotes tumor aggressiveness by increasing plasma membrane cholesterol levels, correlating with poor patient prognosis [36]. Xie et al. demonstrated that targeting the ATAD3A-PINK1 axis promotes PD-L1 recruitment to mitochondria for degradation via mitophagy, thereby sensitizing breast cancer cells to chemotherapy [27]. ATAD3A stabilizes GRP78 to facilitate protein folding and attenuate ER stress, conferring acquired chemotherapy resistance in colorectal cancer [37]. Although Liu et al. observed ATAD3A upregulation in BCa tissues through immunohistochemical staining, associating it with adverse prognosis [38], the biological functions of ATAD3A in BCa remained undefined. We provide the first evidence that ATAD3A promotes BCa progression. As a key molecule on mitochondria-associated membranes (MAM), ATAD3A has been reported to stabilize MAM structure and facilitate calcium transfer between mitochondria and the ER. Mitochondrial calcium overload can lead to irreversible opening of the mitochondrial permeability transition pore, resulting in loss of mitochondrial membrane potential, mitochondrial swelling, and the initiation of downstream irreversible apoptosis. Our observations indicate that ATAD3A stabilizes MAM in BCa cells, preventing mitochondrial calcium overload and reducing ROS production, thereby maintaining mitochondrial stability and cellular homeostasis.

This study acknowledges several limitations. First, the MAM risk model requires validation in larger clinical cohorts. Second, the precise molecular mechanisms through which ATAD3A promotes BCa progression warrant further investigation. The exact pathway linking ATAD3A to immune cell infiltration—particularly Treg recruitment—remains unclear. Future work should investigate whether ATAD3A regulates cytokine secretion (such as TGF - β) or checkpoint ligand expression (such as PD-L1) by modulating MAM, thereby communicating with immune cells such as Tregs.

Conclusion

This study establishes mitochondria-associated membranes (MAM) as a key promoter of BCa progression. We developed a prognostic MAM gene signature and identified ATAD3A as a central oncogene that drives tumor growth and chemoresistance by stabilizing MAM integrity, nominating it as a promising therapeutic target.

Methods

Bulk RNA-seq data collection and processing

Bulk RNA-seq gene expression data and clinical information for BCa samples were obtained from The Cancer Genome Atlas Urothelial Bladder Carcinoma (TCGA-BLCA) cohort and Gene Expression Omnibus (GEO) datasets (accessions: GSE13507, GSE3167, GSE133624, GSE188715, and GSE236932). Transcript expression differences between tumor and normal tissues were analyzed using the limma R package with significance thresholds of |log₂(fold change)| > 1 and adjusted p-value < 0.05 [39]. Univariate Cox regression analysis identified MAM-related genes with prognostic value.

Single-cell transcriptomic data collection and processing

Single-cell transcriptomic datasets (GSE129845, GSE190888, and GSE211388) comprising 3 normal bladder tissues sample, 7 NMIBC sample, and 10 MIBC samples were analyzed. Quality control (QC) and preprocessing were performed using the Seurat R package, wherein cells exhibiting > 10% mitochondrial gene expression, > 1% hemoglobin gene expression, or > 50% ribosomal gene expression were excluded, with additional filtration of low-quality cells based on QC criteria (nFeature_RNA < 200 or nFeature_RNA > 4000) [40]. Data normalization was implemented using the lognormalize method, followed by batch effect correction with the Harmony R package [41]. Subsequent clustering analysis utilized 20 principal components (PCs) for dimensionality reduction, with cell clusters identified via the FindClusters function (resolution = 0.25) and visualized through UMAP. Cell populations were annotated according to established marker genes from literature. Differential gene expression and functional enrichment analyses across cell clusters were visualized using heatmaps generated with the ClusterGVis R package.

Calculation of MAM scores

Twenty-eight MAM-related genes were gathered from the literature, with their associations to MAM functions detailed in Supplementary Table 1. MAM scores were calculated for single-cell data using three distinct algorithms, namely, PercentageFeatureSet, AddModuleScore, and AUCell to assess the activity of MAM-related gene sets. To facilitate visual comparison, the data presented in the figure has been standardized using algorithm-specific Z-scores.

Unsupervised non-negative matrix factorization

Unsupervised non-negative matrix factorization (NMF) analysis was performed on the TCGA-BLCA cohort using the “NMF” R package to determine the prognostic impact of MAM-related genes on BCa patient outcomes [42].

Construction and validation of the MAM signature model

The MAM signature score was calculated using LASSO Cox regression implemented via the “glmnet” R package, incorporating the four identified prognostic genes. The TCGA-BLCA cohort served as the training set, while the GSE13507 dataset functioned as the validation cohort. Prognostic predictive performance was assessed through Kaplan-Meier (KM) survival analysis, time-dependent receiver operating characteristic (ROC) curves, univariate and multivariate Cox regression analyses, and nomogram construction using the “survival”, “survminer”, and “survivalROC” R packages [43].

Cell-cell communication analysis

Intercellular communication networks were decoded from gene expression data using the CellChat R package [44].

Analysis of tumor-infiltrating immune cells

The TCGA-BLCA cohort was stratified into high- and low-MAM score groups based on median MAM score values. The CIBERSORT algorithm was employed to quantify the relative abundance of distinct immune cell subtypes between these groups [45].

Chemotherapy response prediction

The R package “oncoPredict” was employed to evaluate chemotherapeutic drug responses between the high- and low-MAM score groups and to assess the correlation of ATAD3A expression with chemosensitivity [46].

Machine learning and SHAP

A machine learning prediction model was developed using three BCa cisplatin treatment datasets (GSE169455, GSE5287, GSE247185). Gene features were filtered, and Z-score normalization was performed across all datasets based on the mean and standard deviation from the training set GSE169455. Ten core machine learning algorithms, including regularized models, tree-based ensemble methods, and traditional classifiers, were systematically assessed. A two-stage feature selection strategy was employed: initial feature filtering followed by model construction with the selected features. The model was trained on the training set and evaluated on two independent validation sets, with AUC as the primary performance metric. Average validation AUC was calculated to assess model stability, along with corresponding statistical tests and visual analyses. SHAP interpretability analysis was conducted, employing the Random Forest algorithm with 5-fold cross-validation repeated three times. SHAP values for each gene were computed using the kernelshap package to evaluate their contributions to model predictions. The analysis included feature importance ranking, SHAP value distribution visualization, dependency analysis, and interpretation of individual samples. All analyses were conducted in the R environment, primarily utilizing the caret, randomForest, kernelshap, and shapviz packages.

Scissor

The Scissor algorithm was employed to integrate single-cell and bulk transcriptomic data to identify subpopulations of epithelial cells associated with cisplatin resistance in BCa. Epithelial cells were extracted from single-cell data and subjected to dimensionality reduction and clustering. Subsequently, gene matching was performed between GDSC data and single-cell data. A correlation model linking bulk expression to treatment response was constructed using Scissor, allowing the mapping of this correlation back to single cells to identify those significantly associated with resistant or sensitive phenotypes. Results were presented through UMAP visualization and clustering statistics.

Spatial transcriptomics analysis

Spatial transcriptomic data obtained from the 10X Genomics platform (GSE238145) were loaded and normalized. A platinum resistance-related gene set was included, and the AddModuleScore algorithm was used to calculate tumor pathway activity scores and platinum resistance scores for each spatial location. ATAD3A-positive cells were defined as spatial locations co-expressing ATAD3A and the epithelial marker KRT19. A multi-channel visualization strategy was employed, synthesizing ATAD3A expression (red channel) and platinum resistance scores (blue channel) in RGB format to simulate digital multicolor immunofluorescence imaging, enhanced by contrast optimization techniques for improved display.

Cell culture

The human BCa cell lines T24 and 5637 were maintained in DMEM medium (Gibco, Carlsbad, CA, USA) supplemented with 10% fetal bovine serum (Serana, Berlin, Germany) and 1% penicillin/streptomycin at 37 °C in a humidified incubator with 5% CO₂ atmosphere. The cell lines were authenticated by short tandem repeat (STR) profiling upon receipt and used within 6 months of resuscitation.

Lentiviral transfection

T24 and 5637 cells were transfected with control lentivirus or lentiviral constructs expressing ATAD3A shRNA/overexpression vectors (Hanheng Biotechnology, Guangzhou, China), followed by complete medium replacement at 6–8 h post-infection to remove viral particles and ensure cell viability.

Immunohistochemical staining

A BCa tissue microarray (TMA) was obtained from Shanghai Outdo Biotech Co., Ltd. (Shanghai, China), with ethical approval granted by the company’s Institutional Review Board (Ethics Approval No. YBL20250514). All participants provided written informed consent. The TMA contained 70 paired tumor and adjacent normal tissue samples. The detailed patient clinicopathological characteristics are provided in Supplementary Table 2. The staining procedures followed our previously described protocol [47], using a rabbit monoclonal primary antibody against human ATAD3A (PHD9897; Abmart, Shanghai, China; dilution 1:100). IHC scoring was independently performed by two board-certified pathologists blinded to clinical data.

Colony formation and wound healing assays

Colony formation assays were performed as previously described [47], wherein cells (500–1000 cells/well) were seeded in 6-well plates with medium replacement every 3–4 days; after 14–21 days of culture, colonies were fixed and stained with 7.14% crystal violet solution. For wound healing assays, exponentially growing cells were trypsinized into single-cell suspensions and plated in 6-well plates at 6 × 10⁴ cells/well. Upon reaching 100% confluence, uniform scratches were created using pipette tips perpendicular to plate axis. Migration was documented at 0 h and 24 h post-scratching, with migration rate calculated as: (mean initial wound width - mean wound width at t)/mean initial wound width × 100%.

TEM

BCa cells and xenograft tissues were processed for standard TEM analysis. Briefly, samples were fixed in 2.5% glutaraldehyde and 1% osmium tetroxide, then stained with 1% uranyl acetate, dehydrated, and embedded in resin. Ultrathin sections were imaged, and mitochondria-ER contact sites were quantified from high-resolution micrographs using ImageJ. A close association was defined as an intermembrane distance of < 40 nm. The ER-mitochondria distance was measured from three random ER points to the nearest mitochondrial outer membrane, and the percentage of mitochondrial perimeter in close contact with the ER was calculated.

Mitochondria-ER Co-localization

To assess the physical proximity between mitochondria and ER, an indicator of MAM integrity, live BCa cells were co-stained with Mito- Tracker green (Beyotime, C1048) and ER- Tracker red (Beyotime, C1041S) for 30 min at 37 °C. Images were acquired using a confocal microscope with a 100× oil immersion objective. The Mander’s Co-localization Coefficient (MCC), which represents the fraction of mitochondrial pixels overlapping with the ER, was quantified using ImageJ with the JACoP plugin for 60 randomly selected cells per condition. A higher MCC indicates greater organelle contact.

Mitochondrial calcium imaging

To measure mitochondrial calcium levels, live BCa cells were co-loaded with the calcium-sensitive fluorescent probe Rhod-2 AM (2.4 µmol/L) and MitoTracker Green FM (0.3 µmol/L) for 30 min at 37 °C. The co-localization of the Rhod-2 signal (red) with the mitochondrial marker (green) was confirmed via confocal microscopy to ensure specific measurement of mitochondrial calcium.

ROS detection

ROS was assessed using the reactive oxygen species aqssay kit (Jiebeisi Bio, G2706-1). Briefly, cells were loaded with a 1:400 dilution of DHE solution (5mM) for 30 min at 37 °C. The fluorescence intensity was measured by flow cytometry, and the data from 1 × 10⁴ cells per replicate (n = 3) were recorded. The ROS levels are presented as the relative fluorescence intensity normalized to the control group.

Mitochondrial membrane potential (MMP) assay

The MMP was determined using the Enhanced JC-1 Assay Kit (MCE, 3520-43-2). Cells were harvested and incubated with an equal volume of JC-1 staining solution (1:400 dilution) at 37 °C for 30 min, followed by two washes with the provided assay buffer. The fluorescence intensities of JC-1 monomers (green) and aggregates (red) were analyzed by flow cytometry. The MMP was quantified as the ratio of red (aggregates) to green (monomers) fluorescence intensity for each treatment group.

Apoptosis assay

Apoptosis was evaluated using a Propidium Iodide (PI) Staining Kit (Solarbio). Treated cells were stained with PI and imaged by fluorescence microscopy. Apoptotic cells, identified by condensed red nuclei, were counted. The apoptosis rate for each group was quantified as the percentage of PI-positive cells relative to the total cell count from multiple random fields.

Subcutaneous xenograft tumor model

The work has been reported in accordance with the ARRIVE guidelines (Animals in Research: Reporting In Vivo Experiments) [48]. The project has been approved by the Institutional Animal Management and Use (IACUC) Committee of Guangdong Huawei Testing Co., Ltd. (HWT-BG-117b). Twenty-four 4-week-old male BALB/c nude mice were randomized into four experimental groups (n = 6 per group). T24 cells stably expressing ATAD3A knockdown or overexpression constructs were subcutaneously injected into the dorsal flank of each mouse (1 × 10⁶ cells in 100 µL PBS per injection). Tumor dimensions and mouse body weights were recorded every 3 days post-inoculation. At 24 days post-implantation, mice were euthanized and tumors were surgically excised, rinsed with PBS, photographed, weighed, and measured.

Statistical analysis

Statistical analyses were performed using R software (version 4.4.3; R Foundation for Statistical Computing, Vienna, Austria) and GraphPad Prism (version 9.0; GraphPad Software, San Diego, CA, USA). Comparisons between two groups with continuous variables were assessed using independent Student’s t-tests for normally distributed variables and Wilcoxon rank-sum tests for non-normally distributed variables. Statistical significance was defined as p < 0.05.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

HHZ and SSH conducted and conceived the study. HHZ, SSH and ZMS drafted the manuscript. HHZ and ZMS were responsible for the design and execution of the experimental protocols. ZSZ and DJR organized the data analysis and experimental results. BHX, XHJ and FJZ assisted in improving the language quality and completed the final manuscript revision. All authors read and approved the final manuscript.

Funding

This study was supported by the Characteristic Innovation Project of Department of Education of Guangdong Province (2023KTSCX107) (FJZ), Guangzhou Basic and Applied Basic Research Municipal-University(Academy)-Enterprise Joint Funding Project (2023A03J0422) (FJZ), Guangzhou Major Medical Discipline Project (2005–2007), Major technical projects of Guangzhou (2026 C-ZD013), and the National Natural Science Fund of China (82072808) (XHJ).

Data availability

The datasets presented in this study can be found in the TCGA and GEO repositories.

Declarations

Ethics approval and consent to participate

This study follows the declaration of Helsinki. A BCa tissue microarray (TMA) was obtained from Shanghai Outdo Biotech Co., Ltd. (Shanghai, China), with ethical approval granted by the company’s Institutional Review Board (Ethics Approval No. YBL20250514). The subcutaneous xenograft tumor model was approved by the Institutional Animal Care and Use Committee (IACUC) of (Huawei Testing Co., Ltd., Guangdong, China) (Protocol No. HWT-BG-117b).

Consent for publication

All authors are aware of and agree to publish the content of the paper.

Competing interests

The authors declare that no conflicts of interest exist.