TCGA (https://www.cancer.gov/ccg/research/genome-sequencing/tcga) and Gene Expression Omnibus (https://www.ncbi.nlm.nih.gov/GEO/) and BCa RNA sequencing data (HRA009938, https://ngdc.cncb.ac.cn/gsa-human) from Shanghai Tenth People’s Hospital (STPH) were used in this study. The total RNA extraction, bilateral library generation, RNA sequencing protocol, patient inclusion and exclusion criteria included in STPH’s BCa sequencing library establishment process have been described in our published article [23]. The study was conducted according to the guidelines of the Declaration of Helsinki, and approved by the Shanghai Tenth People’s Hospital Ethics Committee (Approval No. 2021KN108 and No. 24KT68) and Guangdong Second Provincial General Hospital Ethics Committee (Approval No. 2024-KY-KZ-128-01), and informed consent was collected from all participants. The animal experiments were approved by the Shanghai Tenth People’s Hospital Ethics Committee (Approval No. SHDSYY-2024-1533). At the same times, the best databases [24] were also used in this study.

T24 (CBP60649), 5637 (CBP60309), UMUC3 (CBP60753), J82 (CBP60311), and RT4 (CBP60313) (All cell lines were sourced from the Chinese Academy of Sciences, Shanghai, China) were used as human BCa cell lines and MB49 (CBP61301, Chinese Academy of Sciences) was used as mouse BCa cell line in this study. Immortalized human normal bladder epithelial cell line SV-HUC-1 (CBP61685, Chinese Academy of Sciences) was also used in this study. MB49, T24, 5637, and UMUC3 cell lines were cultured in RPMI-1640 medium (Thermo Fisher Science, 11875085, Waltham, MA, USA). J82 was cultured in DMEM medium (Thermo Fisher Science, 11965092). RT4 cell lines were cultured in McCoy’s 5A medium (Thermo Fisher Science, 16600082). SV-HUC-1 was cultured in F12k medium (Sigma Aldrich, N3520, St. Louis, MO, USA). The media used in the culture process were all supplemented with a final concentration of 1% penicillin/streptococcus (Hyclone, SV30082.01, Logan, UT, USA) and 10% fetal bovine serum (FBS; Thermo Fisher Science, 16000044). These cell lines were cultured at 37°C in an incubator (Thermo Fisher Science, 311) at 5% CO₂. All cells have undergone mycoplasma testing and cell line authentication, with detailed results provided in the supplementary materials.

Cord blood and PBMC of healthy donors were isolated by density gradient centrifugation (Cell pellets were resuspended in equal-volume normal saline and carefully layered over Ficoll-Paque (MERCK, GE17-1440-02, Hesse, Germany). Following centrifugation at 1000× g for 20 min at 20°C with the brake off, the middle white lymphocyte ring layer was harvested into a fresh tube, washed with normal saline, and centrifuged at 450× g for 5 min. After discarding the supernatant, red blood cells were lysed with RBC lysis buffer (Thermo Fisher Science, 00-4333-57) for 10 min at room temperature (RT) in the dark. Lysis was terminated with normal saline, followed by centrifugation at 450× g for 5 min at RT. The supernatant was discarded to obtain isolated lymphocytes. CD8⁺ T cells were stained with anti-CD8 (1 μg/mL; BioLegend, 980908, San Diego, CA, USA) and CD3 (1 μg/mL; BioLegend, 981012) antibodies and purified by FACS Aria III (BD Biosciences, BD FACSAria™ III, San Jose, CA, USA). On day 3 after stimulation with CD3/28 antibodies (1 μg/mL; Thermo Fisher Science, 11161D), the culture was supplemented every other day with IL-7 (5 ng/mL; PrimeGene Technology, GMP-101-07, Beijing, China) and IL-15 (5 ng/mL; PrimeGene Technology, GMP-101-15) in complete X-VIVO (5% FBS; Thermo Fisher Science). Continued expansion for in vitro experiments.

RIPA (Beyotime, P0013B, Shanghai, China) is utilized to completely disrupt cells or tissues for a duration of 30 min. After this, the samples are centrifuged at 13500× g at 4°C for 10 min. Supernatant is collected and the protein concentration is determined using the Bicinchoninic Acid (BCA) protein assay (Beyotime, P0010). For Western blot analysis, proteins are quantified to a concentration of 80 μg and subjected to electrophoresis on a 10% sodium dodecyl sulfate-polyacrylamide gel (Beyotime, P0690). The proteins are then transferred onto nitrocellulose membrane (MERCK, HATF00010). The membranes are pre-incubated with a solution of skim milk in PBST containing 5% at RT for 1 h to prevent non-specific binding.

Following the blocking step, the membranes are incubated with the diluted primary antibodies at 4°C for a period exceeding 12 h. Post-incubation, the membranes are washed three times with PBST and incubated with a horseradish peroxidase (HRP)-conjugated secondary antibody for 1 h at RT. The proteins are detected using a chemiluminescent imaging system (Tanon 5200 System, Shanghai, China), and their concentrations are quantified accordingly.

This study used the following primary antibodies: β-actin, 1:5000 dilution (ProteinTech, 66009-1-Ig, Wuhan, China), PSMD2, 1:1000 dilution (ProteinTech, 14748-1-AP), CXCL14, 1:1000 dilution (Abmart, TD12377, Shanghai, China), ERK1/2, 1:1000 dilution (Affinity, AF0155, Cincinnati, USA), MEK1/2, 1:1000 dilution (Affinity, AF6385), p-ERK1/2, 1:1000 dilution (Affinity, AF1015), and p-MEK1/2, 1:1000 dilution (Affinity, AF8035). The secondary antibody: HRP-Rabbit, 1:10000 dilution (Abcam, ab6721, Cambridge, UK), HRP-Mouse, 1:10000 dilution (Abcam, ab6728)

Total RNA is isolated from human samples, either 300 mg tissue or 3 × 10⁶ cells, employing the RNA Extraction Kit (Servicebio, G3689, Wuhan, China). Subsequently, the first strand of cDNA is synthesized utilizing the Reverse Transcription System Kit (Vazyme, R323, Nanjing, China). The 20 μL reaction mixture contained 1 μg of total RNA, and the reaction was carried out under the following conditions: 37°C for 15 min, then 85°C for 5 s. Quantitative PCR (qPCR) is conducted with the ABI Prism 7500 Sequence Detection System (Applied Biosystems, CA, USA) and ChamQ Universal SYBR qPCR Master Mix (Vazyme, Q711, Nanjing, China).

The qPCR parameter was set as follows: 95°C for 5 min, followed by 40 cycles consisting of 95°C for 10 s and 60°C for 30 s. β-actin serves as the endogenous reference gene. The relative quantification of mRNA expression is determined by applying the 2^−ΔΔCT method.

The forward and reverse primers for each target gene are as follows: β-actin, F: AGAGCCTCGCCTTTGCC, R: GGGGTACTTCAGGGTGAGGA; PSMD2, F: TGCTCGTGGAACGACTAGG, R: CAGTTTGCCATAGTGTGGACG; CXCL14, F: CGCTACAGCGACGTGAAGAA, R: GTTCCAGGCGTTGTACCAC; GLI1, F: AGCGTGAGCCTGAATCTGTG, R: CAGCATGTACTGGGCTTTGAA; STON2, F: ACCATGTGATTGCCACCCAC, R: AGCTCTCGGACTGGTCTGG; SCN4B, F: GGCTTTTGGTGGAAGAAGTGG, R: AAGACTGTGGGGGATCTGGT; POLR1C, F: CGCAATGTCCATACTACTGACTT, R: CCACACGGAAATTCTTCTCGAA; CXCL8, F: GAAGTTTTTGAAGAGGGCTGAGA, R: ACCAAGGCACAGTGGAACAA; CCL18, F: ATTCTGAAACCAGCCCCCAG, R: GGGCATAGCAGATGGGACTC; CCL14, F: CCAAGCCCGGAATTGTCTTCA, R: GGGTTGGTACAGACGGAATGG; CCL16, F: CTTATCATTACTTCGGCTTCTCGC, R: GGCCTTTCTGTATCCCACCACTA.

The 10 μg plasmid was cloned into a lentiviral vector. HEK-293T cells, cultured in serum-free DMEM in a 10 cm dish, were then transfected using 10 μL of the specified reagent (Genomeditech, GMeasy-10, Shanghai, China). Subsequently, 18 h post-transfection, the medium was replaced with fresh complete DMEM containing serum. Supernatants were collected at 24 and 48 h after the medium change and stored at −80°C. For transfection, 1 × 10⁵ 5637, T24, and MB49 cells were seeded per well in a 6-well plate. Once the cells reached approximately 80% confluence, the lentivirus at a multiplicity of infection of 5 was added for 48 h, and the cells underwent 4 μg/mL puromycin-based drug selection for two weeks. Transfection success was confirmed via qPCR and WB analysis. (Sequence details are presented in Supplementary Information, Table S1).

To assess cell proliferation, 4 × 10³ cells were seeded per well in 96-well plates. After 24 h, the EdU assay kit (APExBIO Corporation, K1075, Houston, TX, USA) was used to evaluate proliferation. The procedure was briefly as follows: Cells were incubated with 10 μM EdU solution in a cell culture incubator for 2 h. Subsequently, they were fixed, permeabilized, and washed three times with PBS. Following this, a Click reaction mixture was added, and the cells were incubated at RTin the dark for 30 min. After another PBS wash, nuclei were counterstained with Hoechst. Fluorescent images were captured with a fluorescence microscope (Nikon, TS2R-FL, Tokyo, Japan). Cell counting was performed using ImageJ software (version 1.53K).

Invasive and migratory capabilities of BCa cells were evaluated using transwell inserts (Yeasen, 84051ES, Shanghai, China), for the invasion assay, the inserts were pre-coated with 100 μL of Matrigel (1.0 mg/mL) (Yeasen, 40183ES, Shanghai, China). The insert was placed in a 24-well plate, and 2.5 × 10⁴ cells were seeded in the upper chamber with an FBS-free medium. In the lower chamber, 600 μL of medium containing 10% FBS was added. After a 24-h incubation at 37°C, the insert was washed with PBS, fixed with 4% paraformaldehyde, and cells on the upper surface were removed with a cotton swab. The cells were then stained with 0.5% crystal violet (Yeasen, 60506ES, Shanghai, China) at RTfor 5 min, rinsed with PBS, air-dried, and documented using a microscope (Nikon, TS2R-FL). For evaluating the migratory capabilities of CD8⁺ T cells, 5 × 10⁶ CD8⁺ T cells were seeded in the upper compartment of transwell inserts (Yeasen, Shanghai, China) with 600 μL FBS-free medium, while 200 μL of different tumor cell supernatants were added to the lower chamber. After 24 h of incubation at 37°C, the lower compartment was gently washed with PBS, and the mixture was centrifuged at 600× g for 5 min under RTconditions. The supernatant was removed, and the cells were stained with 0.5% crystal violet for 20 min. The results were documented using a microscope (Nikon, TS2R-FL). Cell counting was performed using ImageJ software (version 1.53K).

Human CD8⁺ T cells were subjected to phenotype analysis by flow cytometry using Abs specific to human CD3 (1 μg/mL; BioLegend, 981012), CD8 (1 μg/mL; BioLegend, 980908). Mouse CD8⁺ T cells and NK cells were subjected to phenotype analysis by flow cytometry (BD Biosciences, San Jose, CA, USA) using Abs specific to mouse CD45 (1 μg/mL, BioLegend, S18009F), CD8 (1 μg/mL, BioLegend, 53-5.8), NK1.1 (1 μg/mL, BioLegend, S17016D) and CXCL14 were subjected to phenotype analysis by flow cytometry using Abs specific to human CXCL14 (1 μg/mL; Abcam, EPR22807-28). The data were analyzed using FlowJo software (Flow V10.8.1, BD Life Sciences). All flow cytometry staining was performed at 4°C, protected from light, for 40–60 min, with >10,000 singlet events acquired per sample.

To train a deep learning model for distinguishing between tumor and non-tumor regions in BCa hematoxylin and eosin (H&E) stained whole-slide images (H&E WSIs), 70 WSIs from the TCGA dataset were randomly chosen for model training, while 30 were allocated for internal validation. For external validation, 10 WSIs each were selected from the STPH and Guangdong Second Provincial General Hospital (GD2H) datasets. Tumor areas were delineated by a pathologist with extensive experience in BCa using QuPath software (version 0.3.2, Queen’s University Belfast). These WSIs were then divided into 224 × 224 pixel patches at 20× magnification (112 μm × 112 μm) using the Python package Openslide (version 4.0.0), with patches labeled as either tumor or normal tissue. Prior to deep learning model training, edge detection was employed via OpenCV (version 4.10.0, threshold set at 0.02) to exclude patches that were primarily blank. Color normalization was then performed on all patches using the Reinhard method in OpenCV to ensure uniform color representation across datasets.

To classify tumor and normal patches, transfer learning was implemented using the ResNet50 model in PyTorch (version 2.4.1). The model training ran for 50 epochs with stochastic gradient descent (SGD), an initial learning rate of 0.1, momentum of 0.9, and weight decay set at 0.001. Cross-entropy loss was the chosen criterion, and model performance was assessed on both an internal validation set and two external validation sets from STPH and GD2H. The model performance was evaluated using accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and the receiver operating characteristic (ROC).

Patient information (patient characteristics are outlined in supplementary information Table S2), details of model construction and evaluation, subcutaneous xenograft model, immunohistochemistry (IHC) and immunofluorescence (IF) (the details of the antibodies are provided in Table S3), cell counting kit-8 assay, colony formation assay, wound healing assay, co-immunoprecipitation (Co-IP) assay, in silico virtual screening and molecular docking and statistical analysis can be found in the supplementary materials.

Quantitative data are presented as mean values ± SEM from at least three independent experiments. All statistical analyses were conducted using GraphPad Prism software (v8.0.2, GraphPad Inc, San Diego, CA, USA). Intergroup differences were evaluated using appropriate statistical tests: two-group comparisons employed unpaired Student’s t-tests, while multiple group comparisons utilized one way ANOVA followed by Tukey’s post-hoc test. A probability value of p < 0.05 was considered statistically significant for all analyses.

To investigate the genetic factors driving BCa progression, Weighted Gene Co-expression Network Analysis (WGCNA) was applied to the TCGA BCa dataset to identify gene modules most strongly associated with immune scores (derived using the ssGSEA algorithm [25,26]), molecular subtypes, and T stage. The analysis identified the MEgreen module as having the highest malignancy and strongest correlation (Fig. 1A). Tumor samples from patients with BCa at STPH were subsequently collected for RNA-seq analysis (Fig. 1B). By intersecting the 106 genes in the MEgreen module with upregulated genes in MIBC and high pathological grades from STPH, the following genes were identified: POLR1C, GLI1, STON2, SCN4B, and PSMD2 (Fig. 1C). qPCR validation of these five genes was performed using cancerous and adjacent normal tissues from 10 patients with BCa, revealing that only PSMD2 showed a statistically significant difference between the samples (Fig. 1D). Further analysis of the TCGA database confirmed that PSMD2 is highly expressed in BCa tissues (Fig. 1E), with expression levels increasing alongside pathological grade (Fig. 1F). Moreover, PSMD2 expression was significantly higher in non-papillary BCa (Fig. 1G) and correlated with progression-free survival (PFS) in patients (Fig. 1H).

To further validate these findings, several additional databases were analyzed. Consistent with the TCGA results, PSMD2 expression in BCa showed significant associations with patient age (Supplementary Fig. S1A), disease progression (Supplementary Fig. S1B), tumor grade (Supplementary Fig. S1C), cancer stage (Supplementary Fig. S1D), T category (Supplementary Fig. S1E), molecular subtypes (Supplementary Fig. S1F), response to platinum-based chemotherapy (Supplementary Fig. S1G), and survival period (Supplementary Fig. S1H).

Immunofluorescence (IF) analysis of PSMD2 expression in tissues revealed significantly elevated levels in cancerous tissues compared to adjacent normal tissues (Fig. 2A). Subsequently, PSMD2 protein levels were measured in cancerous and adjacent tissues from five patients, with WB analysis confirming that PSMD2 was markedly overexpressed in the cancerous tissues (Fig. 2B). PSMD2 expression was also examined in BCa cell lines. Using the SV-HUC-1 cell line (normal bladder cells) as a control, significantly upregulated PSMD2 mRNA levels were detected in five BCa cell lines (Fig. 2C). WB analysis further validated these findings, showing consistent overexpression of PSMD2 at the protein level (Fig. 2D). These experimental results demonstrate that PSMD2 is highly expressed in BCa tissues and cells.

To investigate the role of PSMD2, two BCa cell lines, 5637 and T24, were selected for study. Stable PSMD2 knockdown cell lines were generated, and qPCR analysis confirmed a significant reduction in PSMD2 mRNA expression in both cell lines compared to the control group (Fig. 3A). WB analysis further verified the successful knockdown of PSMD2 in 5637 and T24 cells (Fig. 3B). To assess the in vivo effects of PSMD2, tumor-bearing experiments were conducted using nude mice. Significant differences in tumor volume, weight (Fig. 3C,D), and growth rate (Fig. 3E) were observed between the PSMD2 knockdown and control groups in both cell lines. The PSMD2 knockdown group exhibited reduced tumor growth and lower tumor weight. Immunohistochemical analysis of tumors from the mice showed decreased expression of Ki67 (Fig. 3F) and PCNA (Fig. 3G), which are markers of cell proliferation. These findings collectively support the conclusion that PSMD2 plays a critical role in promoting tumor proliferation in vivo.

To further explore the function of PSMD2 in BCa cells, a series of in vitro experiments were conducted. The CCK8 assay revealed a significant reduction in the proliferative capacity of the PSMD2-sh group compared to the negative control group (Fig. 4A). Plate cloning and EdU assays were also performed to assess cell proliferation. Consistent with the CCK8 results, both assays showed a decreased number of clones (Fig. 4B) and reduced proliferation (Fig. 4C) in the PSMD2-sh group. Furthermore, scratch and transwell assays were employed to evaluate the migratory and invasive abilities of T24 and 5637 cells. PSMD2 knockdown significantly diminished both migration (Fig. 4D,E) and invasion (Fig. 4F) in these cells. These results demonstrate that PSMD2 knockdown effectively inhibits the proliferation, migration, and invasion of BCa cells.

To explore how PSMD2 promotes BCa progression, KEGG enrichment analysis was performed using RNA-seq data from the STPH cohort. The results revealed significant enrichment in BCa-related pathways, including chemokine signaling, MAPK signaling, and CD8⁺ T cell activity, with PSMD2 expression closely associated with immune infiltration levels (Fig. 5A, Supplementary Fig. S2A,B). To further assess the impact of PSMD2 on the tumor immune microenvironment in vivo, subcutaneous tumor models were established in immunocompetent C57BL/6 mice. The data showed that both tumor growth rate and final tumor weight were significantly higher in the negative control (nc) group compared to the Psmd2 knockdown (Psmd2-sh) group (Fig. 5B–D). Flow cytometry analysis revealed increased infiltration of CD45⁺CD8⁺ T cells and CD45⁺NK1.1⁺ NK cells in Psmd2-sh tumors (Fig. 5E,F), indicating a negative correlation between Psmd2 expression and immune cell infiltration. Given the critical roles of chemokines and immune cells in tumor progression, this study screened for chemokines potentially regulated by PSMD2, identifying CCL18, CCL14, CCL16, and CXCL14 as potential downstream targets (Fig. 5G). qPCR validation confirmed that only CXCL14 was significantly upregulated upon PSMD2 knockdown (Fig. 5H). Subsequent flow cytometry analysis revealed increased CXCL14 protein levels in PSMD2-knockdown BCa cell lines (Fig. 5I), suggesting that PSMD2 may transcriptionally suppress CXCL14 expression.

To further validate the functional relationship between PSMD2, immune cells, and chemokines, an in vitro co-culture killing assay was performed. Following co-culture with CD8⁺ T cells, the proportion of remaining tumor cells was significantly lower in the PSMD2-sh group compared with the nc group (Fig. S3A), while the CXCL14-si group showed increased residual tumor cells, indicating that CXCL14 deficiency impairs CD8⁺ T cell-mediated cytotoxicity. In rescue experiments, CXCL14 knockdown in PSMD2-deficient cells reversed the enhanced killing effect, leading to an increased proportion of remaining tumor cells compared with the PSMD2-sh group (Fig. 6A,B).

Consistently, transwell migration assays showed that supernatants from PSMD2-sh cells strongly promoted CD8⁺ T cell migration (Fig. S3B), whereas supernatants from the CXCL14-si group exhibited reduced chemotactic ability. In contrast, the CD8⁺ T cell migration level in the rescue group was significantly attenuated compared with that in the PSMD2-sh group (Fig. 6C,D).

WB analysis revealed that PSMD2 knockdown decreased p-MEK/p-ERK levels and upregulated CXCL14 expression (Fig. S3C). CXCL14 silencing alone did not significantly affect MEK/ERK phosphorylation, while the rescue group exhibited phosphorylation levels similar to those in the PSMD2-sh group (Fig. 6E).

In summary, PSMD2 promotes BCa progression by directly activating the MAPK pathway and by suppressing CXCL14 transcription and expression through this pathway, thereby remodeling the immune microenvironment (Refer to Fig. 6F for the schematic diagram).

To explore the correlation between PSMD2 expression and pathological tissues in patients with BCa, a ResNet50 model was developed to differentiate tumors from normal tissue patches (Fig. 7A,B). Following transfer learning, the ResNet50 model exhibited high accuracy in distinguishing tumor from normal tissue across the TCGA training set (Fig. 7C), the internal validation set (Fig. 7D), and two external validation sets (STPH and GD2H) (Fig. 7E,F). These results demonstrate the model’s robust performance and strong generalizability across both internal and external datasets.

An RF model was developed using deep learning features extracted from tumor and normal tissue patches (Fig. 7G). The RF model showed strong performance in the TCGA training set, the internal validation set, and the two external validation sets (STPH and GD2H) (Fig. 7H,I). While the model performed excellently in the training set, a slight decline in performance was observed in the validation sets, though it still maintained reasonable generalizability. Similarly, an RF model trained only on features from tumor patches exhibited a comparable trend, though it generally performed worse than the model that included both tumor and normal tissue features (Fig. 7J,K). Variability between datasets, especially in the GD2H external validation set, posed challenges for model performance.

The waterfall plot (Fig. 7L) illustrates the model’s ability to effectively distinguish PSMD2 expression levels in the training set, although its discriminative ability was somewhat reduced in the validation sets. Overall, incorporating normal tissue features improved the predictive capability and stability of the model. Both RF models successfully predicted PSMD2 expression levels based on pathological features, with the combined tumor-normal model showing superior predictive performance.

These findings highlight the critical role of PSMD2 in BCa progression, suggesting that targeting PSMD2 could be a promising therapeutic strategy. To validate this hypothesis in a preclinical setting, specific PSMD2 inhibitors were identified through rapid virtual screening of the ChEMBL [27] compound library using the TransformerCPI2.0 deep learning model. Principal component analysis (PCA) was applied to project the top 100,000 compounds into a two-dimensional chemical space, providing an overview of their molecular distribution (Fig. 8A). Murcko scaffold analysis identified 30 representative structural classes, displaying their score distributions (Fig. 8B) and chemical structures (Fig. 8C), thus revealing preliminary structure-activity relationships. To assess structural diversity, Tanimoto similarity-based Butina clustering was performed on the top 100,000 compounds, with the resulting top 300 clusters visualized (Fig. 8D). Molecular docking of these clusters identified five top-scoring compounds (CHEMBL507911, CHEMBL5188263, CHEMBL3473994, CHEMBL190852, and CHEMBL5497501), with their binding modes shown in Fig. 8E. These integrated virtual screening, scaffold analysis, and docking results successfully identified several promising lead compounds, providing a strong foundation for subsequent experimental validation and drug development.