The Genomic, clinical and mutational data of bladder cancer were downloaded from the public portal of The Cancer Genome Atlas (TCGA) (https://tcga-data.nci.nih.gov/tcga/). The data were screened according to the following criteria: a diagnosis of bladder cancer and the availability of whole-exome sequencing (WES) data. A Perl script (version 5.40.0) was used to structure the data into a matrix representing the expression of genes related to bladder cancer in different samples. This script was also used to divide the original dataset into mRNA and lncRNA level matrices for structural clarity during subsequent computational steps.

Ten key genes involved in migrasome-related biological processes were obtained from previous studies. Firstly, the expression profiles of these specific genes were extracted from the mRNA matrix of bladder cancer using the R environment (v4.4.2). Then, we intersected the migrasome-related genes with the lncRNA expression data previously extracted to obtain the expression patterns of migrasome-related lncRNAs. The co-expression between migrasome-related mRNA and lncRNA was evaluated and visualized by Sankey diagrams.

Univariate Cox regression analysis was performed to screen the lncRNAs related to prognosis (P < 0.05). Then, related lncRNAs were further identified by using least absolute shrinkage and selection operator (LASSO) Cox algorithm. Based on these lncRNAs, a prognostic model of migrasome-related signature was constructed. The RiskScore is calculated as follows: riskScore = ∑[Exp(lncRNA) × coef(lncRNA) (Equation 1)]. Subsequently, samples were randomly divided into test group and training group. According to the median risk score, patients were divided into high-risk and low-risk groups. ROC curve, area under the curve (AUC), survival curves, univariate and multivariate Cox regression analyses were used to evaluate the reliability and accuracy of the prediction model. In addition, a nomogram integrating clinical data was also established to better validate the model's accuracy.

Principal component analysis (PCA) was performed on the signature lncRNAs, palmitoylated genes and lncRNAs, as well as the total gene set to evaluate the efficiency of the lncRNA-based model in distinguishing high-risk samples from low-risk ones.

To further verify the predictive accuracy, we constructed a nomogram and calibration plots based on clinicopathological parameters including age, sex, tumor grade, and TNM stage using the R package "RMS" (R v4.4.2). This methodology enabled estimation of cumulative survival probabilities for bladder cancer patients at 1-, 3-, and 5-year time points, with subsequent generation of Calibration plots to evaluate the alignment between nomogram-predicted outcomes and observed survival data.

We first conducted Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analyses on the differentially expressed genes between risk groups. The functional enrichment results were classified into three categories: molecular function (MF), cellular component (CC), and biological process (BP). We focused on the genes that participated in pathways identified by GO and KEGG analyses. Gene Set Enrichment Analysis (GSEA) was used to identify pathways from the KEGG database and differentially expressed genes between the two risk groups. Significance for enrichment was defined as P < 0.05 and FDR < 0.25.

The R packages "survminer" and "survival" were employed within the R environment (v4.4.2) to analyze pre-organized clinical data. Survival analysis was conducted with respect to age, gender, stage, and grade. Based on a previously established predictive model, the study cohort was stratified into two risk groups using the median value of the calculated RiskScore as the cutoff. Subsequently, survival probabilities for both high-risk and low-risk cohorts were visualized and compared using the organized clinical information. The analytical results indicated that when P > 0.05, different clinical characteristics did not exert a statistically significant impact on the risk stratification model.

To explore the differences in TMB among different risk levels, we used the “limma” package in R to test for differences and calculate p values. We then performed survival analysis of TMB to determine whether there are significant differences in patient survival between high and low groups of TMB, which can verify that TMB is a useful prognostic predictor.

The TME was evaluated using the R package “estimate” (v4.4.2) to calculate the ImmuneScore and StromalScore for each specimen, which were then combined into a single value called ESTIMATEScore per sample. High and low concentrations of different cells in samples were determined based on these scores. The ESTIMATEScore represents the total amount of stroma and immune components within the tissue being analyzed. The three indices were compared statistically between high- and low-risk groups, and their variations were plotted. These profiles were used to estimate how sensitive tumor cells might be to treatment with immune checkpoint inhibitors.

The CIBERSORT algorithm was used to analyze the distribution of various immune cell types, using the "preprocessCore" and "limma" packages in R. A self-written R script (v4.4.2) was used to calculate the proportion content of immune cells for each sample. At the same time, a P value was calculated to determine the accuracy of the previous relative content. Therefore, we considered that the samples with P < 0.05 had more reliable estimation results. Then, box plots were drawn to show the differences in the content of immune cells in samples that met the " P < 0.05" standard.

High- and low-risk mutation data files were organized by running a perl (5.40.0) script. The column header of the file was sample name, and row header was various information about mutations including mutation site and type. Waterfall plots were drawn to visualize mutated samples, genes, and their frequencies. By comparing two waterfall plots, we could check which genetic alterations occurred in individual samples between high-risk and low-risk groups, and also see how frequently a certain gene is mutated within the same group of samples. As a result, it allowed us to screen genes that might have prognostic value.

We used the oncoPredict R package to estimate IC₅₀ values for a variety of therapeutic agents. Both CTRP and GDSC data were included in our analysis. We applied a stringent filtering threshold of P < 10^-9 to identify drugs with highly divergent sensitivity profiles across risk groups.

A total of 10 migrasome-related genes were identified. We then performed co-expression analysis and identified 808 migrasome-related lncRNAs, which are represented in the Sankey diagram [Figure 1A]. The different colors represent the different migrasome genes, with lines drawn to connect corresponding lncRNAs. It is noteworthy that the EPCIP co-expressed lncRNAs account for more than 90% of all lncRNAs in the pool of identified migrasome-related lncRNAs.

To begin with, the forest plot was used to screen out specific migrasome-related lncRNAs by univariate Cox regression modeling [Figure 1B]. Next, related lncRNAs were obtained through Lasso regression analysis, such as LINC02345, AL031716.1, LINC01126, AC027243.2, AC023494.1, AC019080.5, AC009951.6, BX284668.6 and FRMD6-AS1 [Figure 1C and D].

To determine whether the model lncRNAs have functional relevance to migrasome biology, we performed correlation analysis between them and core migrasome genes. The correlation analysis of lncRNAs and migrasome genes was used as part of the model building process [Figure 1E]. Of note, the migrasome gene EPCIP showed the strongest correlations with AC023494.1, AC027243.2, and LINC01126 (all P < 0.001).

The bladder cancer study population was divided into a training set and test set at a ratio of 1:1. The survival analyses for all samples, the training cohort, and the test cohort are shown in Figure 1F-H, respectively. The median riskScore was calculated according to Equation 1. Then, the subjects were divided into high-risk and low-risk groups based on this median value. A heatmap of gene expression in each sample was drawn for both groups. Survival analysis including riskScore distribution, as well as vital status and survival time were performed subsequently. The expression heatmap, risk score distribution map and survival state plot of the three groups are as follows: All samples [Figure 2A-C], results of the training set [Figure 2D-F] and test set [Figure 2G-I]. In the survival state plot, it can be seen that with the increase of risk score, the proportion of dead samples increased significantly.

We evaluated the prognostic accuracy of clinical features and riskScore by several methods. Both the clinical feature information and the riskScore were subjected to univariate cox analysis, in which each clinical feature was used as a prognostic indicator alone, and multi-variable Cox regression model, in which multiple factors interact with each other [Figure 3A and B]. The p value of riskScore was always less than 0.05, indicating that it had prognostic significance. Other significant prognostic factors included age and staging. Hazard ratio greater than 1 indicates that this factor is a risk factor and may worsen the disease.

We generated calibration curves [Figure 3C] using the clinical factors and riskScore. As shown in Figure 3C, the probability of a patient surviving for more than one year was 0.952, for more than three years was 0.839, and for more than five years was 0.804. In the C-index plot, the vertical axis represents the value of the C-index, while the horizontal axis indicates the predicted survival time. The magnitude of the C-index value reflects the level of accuracy of the factor. As shown in Figure 3D, the riskScore showed the highest C-index value, providing further evidence for the predictive reliability of our model. Furthermore, Figure 3E validates the validity of the model by verifying the estimated probabilities of survival at 1-, 3-, and 5-year intervals as produced by our signature.

The survival Sankey diagrams of age (≥ 65 years and < 65 years) as the classification standard in the high-risk and low-risk groups are shown Figure 4A and B. The survival Sankey diagrams of grade as the classification standard in the high-risk and low-risk groups with different tumor grades are shown in Figure 4C and D. Figure 4E-H show gender and staging as classification criteria, respectively. P < 0.05 indicates that the model can effectively distinguish and predict high and low risk for these characteristics. However, the p values for “low grade” and “Stage 1-2” were greater than 0.05, indicating that the model does not apply to this characteristic, but the overall accuracy of the model is very high.

PCA plots of lncRNAs, migrasome-related genes and migrasome-related lncRNAs used for model construction are shown in Figure 4I-K, respectively, which were applied to assess the ability of these variables as classification criteria for distinguishing high-risk group from low-risk group. The selected lncRNAs in final model showed better performance on discriminating high-risk samples from low-risk ones, indicating that our 7-lncRNA signature could capture most relevant transcriptomic variation associated with migrasome-related high-risk phenotype.

Figure 4L shows the Kaplan-Meier analysis of Progression-Free survival (PFS) with PFS in y-axis. As time goes by, patient survival decreases and high-risk group declines more rapidly and reaches a plateau at about five years while low-risk cohort remains stable after approximately eight years of follow up. The receiver operating characteristic (ROC) analysis for risk score yielded an AUC of 0.685 [Figure 4M and N]. Of note, this value is higher than AUC for all clinical traits, indicating robustness of the model.

The x-axes of Figure 5A and B represent the “Enrichment Score” and “Counts”, respectively, showing the results of GO and KEGG enrichment analysis. Pathways with smaller p values are considered to be more significant. These include extracellular space, cornified envelope, and structural molecule activity. The genes involved in these pathways are shown in Figure 5C, where most of the genes were located in the extracellular region.

In addition, we performed GSEA enrichment analysis using the KEGG database. Figure 5D and E shows five pathways with the highest enrichment scores in high-risk vs. low-risk groups. The following pathways were enriched in the high-risk group: Arrhythmogenic right ventricular cardiomyopathy, regulation of actin cytoskeleton, ECM receptor interaction, Focal adhesion. Interestingly, the pathways that were enriched in the high-risk group were completely different from those in the low-risk group.

Importantly, the pathways enriched in the high-risk group, namely “Focal adhesion”, “Regulation of actin cytoskeleton”, and “ECM-receptor interaction”, are all closely associated with cell migration, mechanical force transduction, and invasion. These biological processes are directly connected to migrasome function. When combined with the strong co-expression observed between the model lncRNAs (AC023494.1, AC027243.2, LINC01126) and the core migrasome gene EPCIP, these enrichment results offer convergent evidence that our lncRNA signature captures a cellular state conducive to migrasome-mediated migration and invasion.

The comparative proportions of 22 immune cell types in the analyzed cohort are shown in Figure 6A, and differentially expressed immune cells are highlighted in Figure 6B. Compared with the low-risk cohort, resting mast cells, M0 macrophages, and resting memory CD4+ T cells were more prevalent in the high-risk group. In contrast, activated dendritic cells, regulatory T cells (Tregs), and CD8+ T cells were increased in the low-risk population. The remaining immune cells demonstrated no statistically significant differences among the risk subgroups. This distribution pattern, which is characterized by a decrease in cytotoxic CD8+ T cells and an increase in resting immune cell types in the high-risk group, suggests microenvironmental immunosuppression.

Using the R package “maftools”, we calculated the statistical results of tumor mutation load and evaluated the probability and type of gene mutations in our samples. The waterfall plots [Figure 6C and D] intuitively show the alteration rate of genes in two risk groups, indicating that patients in the high-risk group have a significantly lower TMB than those in the low-risk group. It can also be seen that the larger the TMB value, the weaker the immune escape ability, and consequently, the better the anticipated response to immunotherapy. In addition, there were significant differences in the frequency of variation between the two groups for TP53 and KMT2D. As far as TME was concerned, Figure 6E showed that the stroma component in high-risk patients was significantly higher than that in low-risk patients (P < 0.001). However, no significant difference was found in the ESTIMATE score or the overall proportion of immune cells in different risk groups. The increased stromal scores observed in high-risk samples indicate an increase in cancer-associated fibroblasts (CAFs) activity and extracellular matrix remodeling, which are known to promote tumor invasion and metastasis. This is consistent with the pro-invasive phenotype of the migrasome. The p value obtained by analysis of variance of TMB [Figure 6F] on high and low risk groups was 0.00057, less than 0.05, and the result was considered statistically significant.

The samples were further grouped by TMB size, and the high-TMB and low-TMB groups were used for survival analysis [Figure 6G]. The results showed that the prognosis of the low-risk subgroup was significantly better than that of the high-risk cohort. The p value of the model constructed with TMB was calculated to be less than 0.001, indicating a good accuracy. In addition, we combined the TMB score with the risk score of the previously established model. As shown in Figure 6H, patients in the high TMB and high-risk group (the gray line) had the worst prognosis, suggesting that poor prognosis is related to high risk and low TMB. The p value was calculated to be less than 0.001, which again confirmed the reliability of our previous risk score grouping. These data indicate that the poor clinical outcome of high-risk patients may be due to biologically aggressive, metastasis-related phenotypes rather than a high neoantigen load and that these patients are unlikely to respond well to immunotherapy alone^[15].

Drug sensitivity analysis using the OncoPredict algorithm revealed 11 pharmaceuticals with significant differences in sensitivity between high-risk and low-risk participants (P < 10^-9). A lower half-maximal inhibitory concentration (IC₅₀) represents higher sensitivity to the drug. Of particular interest, nilotinib showed the most pronounced difference in response, where the high-risk cohort was significantly more sensitive (as indicated by a lower IC₅₀ level) than the low-risk population (P < 10^-9, Figure 6I). In contrast, KU-55933 showed the exact opposite pattern, where the low-risk cohort was more sensitive (reflected by decreased IC₅₀ values) than the high-risk group (P < 10^-9, Figure 6J). The results indicate different therapeutic response profiles between risk subgroups and suggest that patients could be treated differently based on their migrasome-associated lncRNA risk score. This goes beyond pure prognostication and suggests potential personalized or tailored treatment approaches, indicating that the migrasome-related high-risk phenotype may harbor specific druggable weaknesses. Table 1 shows the drug sensitivity analysis of 11 drugs.

In conclusion, this study is the first systematic investigation of a migrasome-related lncRNA signature for predicting clinical outcomes in patients with bladder cancer. In addition to evaluating the immune cell landscape in bladder cancer, we compared treatment response between the two risk cohorts to identify more effective therapeutic strategies. Our findings extend beyond a bioinformatics signature by providing a new mechanistic framework based on migrasome biology that can be used to understand disease progression and guide prognosis, microenvironment regulation, and personalized therapy.