Comprehensive analysis of N⁶-methyladenosine regulators with the tumor immune landscape and correlation between the insulin-like growth factor 2 mRNA-binding protein 3 and programmed death ligand 1 in bladder cancer

All data were obtained from The Cancer Genome Atlas (TCGA)-Urothelial Bladder Carcinoma (BLCA) database (https://​portal.​gdc.​cancer.​gov/​) and the Genotype Tissue Expression project (GTEx) database (https://​gtexportal.​org/​), including RNA sequencing data (fragments per kilobase of transcript per million fragments sequenced (FPKM) value) of gene expression, copy number variation (CNV), somatic mutation, and clinical information. The GTEx database includes over 10,000 bulk RNA-seq samples representing 53 different tissues (corresponding to 30 organs) obtained from 635 pre-healthy individuals, to link the influence of genetic variants on gene expression levels via quantitative trait loci analysis (eQTL) [23]. The expression levels of normal samples from TCGA and GTEx databases were integrated and used for comparison with tumor samples. The Cancer Cell Line Encyclopedia (CCLE) database was used to evaluate the expression levels of the m⁶A regulator in several cell types, and the correlation with PD-L1 at the bladder cancer cell level. The protein levels of m⁶A regulators were determined using the Human Protein Atlas database (https://​www.​proteinatlas.​org/​). The Oncomine database was used to determine the insulin-like growth factor 2 mRNA-binding protein 3 (IGF2BP3) expression levels in several cancer types (https://​www.​oncomine.​org/​). The gene expression profiling interactive analysis (GEPIA) database was used to evaluate the prognostic value of IGF2BP3 (http://​gepia.​cancer-pku.​cn/​). The workflow of this study is shown in Additional file 1: Fig. S1.

A total of 24 m⁶A RNA methylation regulators were extracted from the database according to the relevant m⁶A studies [22, 24], including nine writers [METTL3, METTL14, METTL16, RNA-binding motif protein (RBM15)-15, RBM15B, WT1-associated protein (WTAP), KIAA1429, Cbl proto-oncogene like 1(CBLL1), and zinc finger CCCH-type containing 13 (ZC3H13)], two erasers (ALKBH5 and FTO), and 13 readers [YTH domain containing YTHDC1, YTHDC2, YTHDF1, YTHDF2, YTHDF3, IGF2BP1, IGF2BP2, IGF2BP3, heterogeneous nuclear ribonucleoprotein A2/B1 (HNRNPA2B1), heterogeneous nuclear ribonucleoprotein C (HNRNPC), FMR1, leucine rich pentatricopeptide repeat containing (LRPPRC), and ELAV-like RNA binding protein 1 (ELAVL1)]. The extracted information was used for further analyses.

Unsupervised clustering of BCa samples was conducted to identify different m⁶A regulatory patterns based on their mRNA expression levels and classify the patients for further analysis. The “ConsensuClusterPlus” R package was used to perform the consensus clustering algorithm, which could determine the number of clusters and assess their stability according to the m⁶A regulator expression levels [25]. “k” was used to represent the number of subgroups. Principal component analysis (PCA) was conducted to verify the grouping results, and the set of protein in PCA is the total of 24 m⁶A RNA methylation regulators.

The “limma” R package was applied to identify the differentially m⁶A expressed genes between normal and tumor samples [26]. The significance criteria for determining DEGs was set as the P value < 0.05, and |log2FC|> 1. The Search Tool for the Retrieval of Interacting Genes/Proteins (STRING) database and Cytoscape software was used to retrieve and construct a PPI network of the m⁶A regulator network.

To investigate the changes in pathway activity between different groups, GSVA was performed using the “GSVA” R packages. GSVA is a non-parametric, unsupervised method for estimating the variation in gene set enrichment and biological process activity of samples from an expression dataset [27]. In the present study, the gene sets of “c2.cp.kegg.v7.1.symbols” were downloaded from the Molecular Signatures Database (MsigDB) for running the GSVA analysis. We defined an adjusted P value < 0.05, and |log2FC|≥ 0.08, as statistically significant.

The single-sample gene-set enrichment analysis (ssGSEA) algorithm was used to quantify the relative abundance of immune signatures in the TME. The enrichment levels of 29 immune signatures were quantified for each sample. The gene sets representing each immune signature are shown in Additional file 1: Table S1, including B cells, Th2 cells, NK cells MHC class I cells, CD8 + T cells, and so on [28]. The “sparcl” R package was employed to divide the samples into three groups, including immunity high, medium and low for further analyses. ESTIMATE was used to evaluate TME cell infiltration level (including immune and stromal scores) and tumor purity for each sample [29].

Univariable and multivariable Cox regression analyses were used to assess the prognostic value of m⁶A regulators, and hazard ratios (HRs) > 1 or < 1 were regarded as risk and protective genes, respectively. The least absolute shrinkage and selection operator (LASSO) Cox regression algorithm was used to construct the optimal prognostic model out of the m⁶A regulators using the “glmnet” package in R. The LASSO analysis performed predictor selection, minimized over-fitting, selected genes to reduce bias, and developed the best survival-associated risk pattern [30]. After testing for collinearity, the sum of the Cox coefficient and gene values were calculated using the risk score based on the following formula:

Each patient was assigned a risk score based on integrative m⁶A regulator expression patterns. Tree-fold cross-validation and 1000 iterations were conducted to reduce the potential instability of the results.

To explore the relationship between m⁶A regulators involved in the risk score pattern and the infiltrating immune cells, we utilized the Tumor IMmune Estimation Resource (TIMER) web tool (https://​cistrome.​shinyapps.​io/​timer/​) to calculate the correlation coefficients between m⁶A regulator expression and mutations with infiltrated immune cells including B cells, CD4 + T cells, CD8 + T cells, macrophages, neutrophils, and dendritic cells [31, 32].

Based on the results of the multivariable Cox regression model, a nomogram based on independent prognostic factors was constructed to predict 3- and 5-year OS. The nomogram provides a graphic representation linking individual patient factors to predict the survival probability of BCa patients [33]. In addition, a bootstrapped resample with 1000 iterations was applied to verify the accuracy of the nomogram. Furthermore, the performance of the prognostic models was evaluated by receiver operating characteristic (ROC) analyses, and the concordance index (c-index) was measured to quantify the nomogram discrimination. A scale of 1.0, represents perfect predictions, and 0.5, the equivalent of a coin toss. The calibration of the model was assessed by comparing the observed survival with the predicted survival by plotting a calibration curve. The 45° line indicates a perfect calibration. Any deviation above or below the 45° line indicates underprediction or overprediction, respectively. Due to the limited conditions, no extra BCa cohort could be used as an externally validated database to evaluate the efficacy of model validation and prediction. Thus, only internal validation was conducted to evaluate the nomogram model.

This study was approved by the Medical Ethics Committee of the Shandong University School of Clinical Medicine. Twenty human BCa tissues were collected at the Qilu Hospital of Shandong University. Informed consent was obtained from all patients.

T24, 5637, and UMUC3 cell lines were purchased from the American Type Culture Collection (ATCC). T24 and 5637 cells were cultured in the Roswell Park Memorial Institute (RPMI)-1640 medium (Gibco; 11875093). UMUC3 cells were cultured in Dulbecco’s modified Eagle’s medium (DMEM; Gibco; 11995065). All media were supplemented with 10% fetal bovine serum (Gibco; 10099-141C). The cells were incubated at 37 °C in a humidified atmosphere with 5% CO₂.

Stable IGF2BP3 knockdown, overexpression cell lines, and their controls were generated as described previously [34]. Lentiviruses were purchased from GeneChem Inc. (Shanghai, China).

Western blotting was performed as previously described [34]. The primary antibodies included anti-IGF2BP3 (Abcam; ab177477), anti-PD-L1 (Proteintech; 28076-1-AP), and anti-GAPDH (Proteintech; 60004-1-Ig).

Extraction of total RNA and RT-PCR were performed as previously described [34]. Primers used were purchased from Sangon Biotech (Shanghai, China), and primer sequences are shown in Additional file 1: Table S2.

Flow cytometric assays were performed as previously described [35]. Briefly, single-cell suspensions were freshly prepared from the inidicated cells. Cells were washed once with PBS and stained with anti-PD-L1 antibody (Biolegend, 329705) for 30 min. The samples were analyzed on a FACSCanto II (BD Bioscience, USA) using FlowJo 7.6.5.

All analyses were conducted using the R v.4.0.0 and SPSS v.20.0 software. The “RCircos” R package was used to plot the CNV landscape of m⁶A regulators. The somatic mutation landscape was assessed using the “maftools” R package to plot the mutation summary, waterfall, and gene cloud figures. The correlations among different m⁶A regulators were computed by “corrplot” R package. “Survival” R package was adopted to analyze Kaplan–Meier curve analysis. The “forestplot” R package was conducted to visualize the univariable and multivariable prognostic analysis for risk score. The specificity and sensitivity of risk score were assessed using the ROC curve and the area under the curve (AUC), which were quantified by the “pROC” R package. Data from the two groups were evaluated using a two-tailed unpaired Student’s t-test. Categorical data were analyzed using the chi-square test. The correlation between continuous variables was assessed using the Spearman’s correlation analysis. Survival analysis was performed using log-rank test. Statistical significance was set at P < 0.05.

A total of 24 m⁶A regulators in BCa were used in the present study. We first clarified the incidence of CNV and somatic mutations in m⁶A regulators in BCa. Among the 412 samples, 116 (28.2%) experienced somatic mutations in m⁶A regulators (Fig. 1A). METTL3 showed the highest mutation frequency among the m⁶A regulators (approximately 4%), while two writers (RBM15B and METTL16) and two readers (HNRNPC and FMR) did not exhibit any somatic mutations in BCa. Correlation analyses revealed that most m⁶A somatic mutations did not exhibit a co-occurrence relationship, except for FMR1 and YTHDF2, YTHDF1 and KIAA1429, WTAP and METTL3, ZC3H13, and RBM15 (Fig. 1B). Next, we summarized the CNV mutation frequency among the m⁶A regulators, and KIAA1429, YTHDF1, YTHDF3, and IGF2BP2 had a widespread frequency of CNV amplification, while METTL16 and ALKBH5 showed high CNV deletion frequency (Fig. 1C). We also explored the CNV mutation in normal tissues, and only 7 m⁶A regulators had a CNV mutation burden, with an extremely low frequency (Additional file 1: Fig. S2). The location of the CNV mutation in m⁶A regulators on different chromosomes is shown in Fig. 1D.

After exploring the mutation of m⁶A regulators, we investigated the mRNA expression levels of m⁶A regulators in normal bladder samples and tumor samples. The GTEx and TCGA datasets were merged for further analysis, with 28 normal samples and 411 tumor samples. As shown in Fig. 2A, B, mRNA expression levels of CBLL1, ELAVL1, HNRNPA2B1, IGF2BP1, IGF2BP2, IGF2BP3, LRPPRC, RBM15, RBM15B, YTHDF1, YTHDF2, and YTHDF3 were significantly higher in tumor samples than in healthy samples, while the expression levels of FTO, METTL14, METTL16, WTAP, YTHDC1, YTHDC2, and ZC3H13 were decreased in tumor samples. Due to the functional similarity or complementation, the comprehensive landscape of m⁶A regulator connections was depicted by Spearman correlation analysis, STRING website, and Cytoscape software, METTL3 and YTHDF3 showed the strongest positive correlation, while METTL3 and IGF2BP2 showed the strongest negative relevance (Fig. 2C, D). Not only did the m⁶A regulators with similar functions show a significant correlation, but a remarkable interaction was shown among writers, erasers, and readers. Moreover, correlations between writers and erasers were investigated to determine whether tumors with high eraser expression levels exhibited low writer expression levels. The results revealed that tumors with high expression of CBLL1 and METTL14 showed a high expression of FTO, while the high expression of CBLL1 and METTL14 showed low expression of ALKBH5. Tumors with high expression of ZC3H13 and WTAP showed high expression of FTO. However, ZC3H13 and WTAP did not interfere with ALKBH5 expression. The remaining writer genes did not affect the eraser genes ALKBH5 and FTO (Additional file 1: Fig. S3A). Immunohistochemistry staining images of m⁶A regulator proteins were retrieved from the Human Protein Atlas, revealing cellular sublocalization and intensity (Fig. 2E and Additional file 1: Fig. S3B). The above results revealed that cross-talk among m⁶A regulators might construct important modification patterns.

The empirical cumulative distribution function was plotted to analyze the optimal k value at which the cluster model achieved maximum stability (Fig. 3A and Additional file 1: Fig. S4A). The results showed that k = 2 gained the most powerful clustering efficacy, and the samples were divided into two subclusters using unsupervised clustering (Fig. 3B and Additional file 1: Fig. S4B-F). PCA analysis was used to judge the classification, and the two clusters could gather together (Fig. 3C). Prognostic analysis for the two clusters did not show a statistically significant difference, but a trend in overall survival (OS) (Fig. 3D). We explored PCA analysis and prognostic analysis for k = 3 and 4, and no significant benefits in OS were found (data not shown). However, the correlation analysis showed that clustering was associated with the tumor grade (Fig. 3E). The m⁶A expression profiles showed that all m⁶A regulators were upregulated in cluster 2, except for IGF2BP1 (Additional file 1: Fig. S4G). GSVA enrichment analysis was performed to explore the biological behaviors of the two clusters. As shown in Fig. 3F–G, cluster 1 presented enrichment pathways associated with metabolism, such as linoleic acid, arachidonic acid, retinol, drug, and xenobiotic metabolism. Cluster 2 was remarkably enriched in DNA damage, including mismatch repair, DNA replication, cell cycle, nucleotide excision repair, and spliceosome.

To further explore the prognostic value of m⁶A regulators in BCa, we first performed a univariable Cox regression analysis on the expression levels of m⁶A regulators. The results showed that high expression of IGF2BP3 (hazard ratio [HR] = 1.2, 95% confidence interval [CI] = 1.1–1.3) and LRPPRC (HR = 1.0, 95% CI = 1.0–1.1) had worse survival outcomes in patients with BCa, whereas YTHDC1 (HR = 0.9, 95% CI = 0.9–1.0) and WTAP (HR = 1.0, 95% CI = 0.9–1.0) were regarded as protective markers for BCa (Fig. 4A). However, high expression of ZC3H13 seemed to have worse survival outcomes in patients with BCa regardless of P value.

LASSO Cox regression analysis was performed to determine the optimal genes for selecting predictors and building the most regularized and parsimonious risk score pattern, and we only chose the prognostic value of m⁶A regulators P < 0.1 for further analysis. The genes and their coefficients obtained from the LASSO analysis were used to calculate the risk scores for individual patients (Fig. 4B, Additional file 1: Fig. S5A, Table S3). The final LASSO model with the best optimal lambda included six m⁶A regulators (IGF2BP3, LRPPRC, YTHDC1, YTHDF2, and WTAP). To investigate the prognostic value of the risk score pattern, BCa patients were divided into low-and high-risk groups, and the Kaplan–Meier curve revealed that the patients in the high-risk group had a worse survival than patients in the low-risk group (Fig. 4C). As shown in Fig. 4D, E, green (low risk or alive) and red (high risk or dead) dots demonstrated significant differences between the low-and high-risk groups. ROC analysis showed a risk score pattern with AUC = 0.64, indicating that the risk score could predict the OS of patients with BCa (Fig. 4F). Next, we explored the correlation between the risk score pattern and clinical characteristics, and the risk score pattern was related to tumor grade, tumor stage, T status, M status, and N status (Fig. 4G and Additional file 1: Table S4). Moreover, the expression of risk genes (IGF2BP3, LRPPRC, and FTO) was higher in high-rsik patients, while YTHDC1, YTHDF2, and WTAP tended to be expressed in the low-risk group. Univariable and multivariable Cox analyses were performed to determine the independent prognostic value of the risk score pattern. Patient age, tumor T status, N status, and risk score were independent prognostic predictors in patients with BCa (Fig. 4H).

To better understand the function of the risk score pattern, we analyzed the GO analysis of DEGs based on expressions in low-and high-risk score groups. GO analysis indicated that upregulated genes in the high-risk group were enriched in malignancy-related biological processes, including extracellular matrix organization, extracellular structure organization, and antimicrobial humoral response, and downregulated genes were enriched in hormone metabolic processes and terpenoid metabolic processes (Additional file 1: Fig. S5B, C). GSVA enrichment analysis was conducted to explore the different pathways between the two groups. The results revealed that the high-risk group was significantly related to the malignant pathways, including gap junction, focal adhesion, and ECM receptor interaction (Additional file 1: Fig. S5D, E). Then, we investigated the distribution differences of somatic mutation between low and high risk score by using “maftools” R package. As shown in Additional file 1: Fig. S5F–K, the high-risk score group exhibited more somatic mutation burden than the low-risk score group, and missense mutation was the most common variant classification; the most common variant type was SNP, and C>T transversion was the most common type of SNV class. Moreover, the top three mutated genes were TP53, TTN, and KDM6A in the low-risk group and TP53, TTN, and KMT2D in the high-risk group, respectively. Taken together, the risk score pattern based on m⁶A regulators could be regarded as an independent prognostic factor in patients with BCa, and the high-risk group gained more malignant behaviors and more mutation burden.

To explore the potential relationship between immunity and risk score pattern, we first divided samples into three clusters, immunity low, median and high using the ssGSEA score to quantify the immune cell types, functions and pathways, and the differences in 29 immune-associated gene sets were shown in three distinct immunity clusters (Additional file 1: Fig. S6A, B). Next, we investigated the correlation between the immune landscape and the risk score pattern. As shown in Fig. 5A, the enrichment of the immune landscape in the high-risk group was higher than that in the low-risk group. Moreover, the percentage of low immunity samples in the high-risk group was significantly lower than that in the low-risk group and more median immunity samples in the high-risk group than that in the low-risk group (Fig. 5B). In addition, comparing the stromal score, immune score, tumor purity, and ESTIMATE score between the two distinct risk score groups, we found that the high-risk group had significantly higher stromal scores, immune scores, and ESTIMATE scores, and lower tumor purity (Fig. 5C–F). Taken together, these results suggest that the risk score pattern has a strong relationship with the immune landscape, and the potential mechanisms of m⁶A regulators in tumorigenesis and progression may be associated with tumor immunity.

Next, we explored the immune characteristics of independent m⁶A regulators in the risk group using the TIMER database to investigate the correlation between m⁶A regulator expression and immune cells, including B cells, CD8 + T cells, CD4 + T cells, macrophages, neutrophils, dendritic cells, and tumor purity. As shown in Fig. 5G and Additional file 1: Fig. S6C, the expression of IGF2BP3 was positively correlated with macrophages, neutrophils, and dendritic cells and negatively correlated with tumor purity. The expression of LRPPRC was positively correlated with B cells, CD8 + T cells, neutrophils, and dendritic cells, and negatively correlated with CD4 + T cells. As for FTO, B cells, CD8 + T cells, macrophages, neutrophils, and dendritic cells were identified as significant co-expression cells. The expression of WTAP was negatively correlated with tumor purity, but positively correlated with CD8 + T cells, CD4 + T cells, neutrophils, and dendritic cells. The expression of YTHDC1 was only correlated with tumor purity, B cells, and macrophages. As for YTHDF2, tumor purity, B cells, CD8 + T cells, and neutrophils showed a strong correlation. The SCNA module, which was defined by GISTIC 2.0, was conducted to provide a comparison of immune infiltration levels in BCa with different somatic copy number alterations for m⁶A regulators. As shown in Fig. 5H and Additional file 1: Figure S6D, IGF2BP3 amplification was associated with dendritic cells.YTHDC1 deletion was related to B cells, while amplification was related to CD8 + T cells, neutrophils, and dendritic cells. Moreover, FTO amplification had a connection with B cells and macrophages, and deletion had a connection with CD8 + T cells. Interestingly, LRPPRC deletion and amplification were both associated with CD4 + T cells, neutrophils, and dendritic cells. The YTHDF2 mutation was associated with immune cells, except for CD8 + T cells and macrophages. The WTAP mutation is only related to CD4 + T cells and neutrophils. Furthermore, we investigated the co-expression of m⁶A regulators in the risk score model and several immune checkpoints (Fig. 5I). The results indicate that m⁶A regulators are correlated with most immune checkpoints, including PD-L1 (also known as CD274). In summary, these results strongly indicate that the risk score pattern based on m⁶A regulators is significantly correlated with the tumor immune landscape.

A nomogram was established based on the independent factors using a multivariable Cox regression model to predict OS in patients with BCa (Fig. 6A). The AUCs of the nomogram for predicting the 3- and 5-year OS were 0.69 and 0.70, respectively (Fig. 6B, C). The c-index of the nomograms for OS in the training set was 0.68. As shown in Fig. 6D, E, calibration plots were generated to validate the similarities between the actual survival rate and the survival prediction by the nomogram, and the results demonstrated that the 3- and 5-year survival rates predicted by the nomogram were closely corresponded with the actual survival rates in the training set.

Moreover, 30 percent of patients with BCa were selected in the internal validation set. The AUCs in the validation set for predicting the 3- and 5-year OS were 0.75 and 0.72, respectively (Additional file 1: Fig. S7A, B). The c-index of the nomogram in the validation set was 0.688. The results of the calibration plot suggested that the predicted 3- and 5-year survival rates were consistent with the actual survival rate within an acceptable margin of error in patients with BCa (Additional file 1: Fig. S7C, D).

Because of the important role of IGF2BP3 in risk score patterns, we used TCGA, GTEx CCLE, and Oncomine datasets to further understand IGF2BP3 in normal and tumor tissues. As shown in Fig. 7A, the expression of IGF2BP3 was higher in BCa, cholangiocarcinoma (CHOL), colon adenocarcinoma (COAD), esophageal carcinoma (ESCA), head and neck squamous cell carcinoma (HNSC), kidney chromophobe (KICH), kidney renal clear cell carcinoma (KIRC), kidney renal papillary cell carcinoma (KIRP), liver hepatocellular carcinoma (LIHC), lung adenocarcinoma (LUAD), lung squamous cell carcinoma(LUSC), stomach adenocarcinoma (STAD), uterine corpus endometrial carcinoma (UCEC), and low expression in thyroid carcinoma (THCA), compared to their corresponding normal tissues (Fig. 7A). Moreover, the CCLE dataset was used to evaluate the expression levels of IGF2BP3 in various tumor cell lines. The results showed that the top three expression levels in tumor cell lines were liver cancer, lymphoma, and medulloblastoma. IGF2BP3 seemed to be positively associated with PD-L1 expression in BCa cell lines (Fig. 7B, C).

The Oncomine database was used to determine the expression level of IGF2BP3. And as shown in Fig. 7D, IGF2BP3 in most cancer types showed high expression levels, except for kidney cancer and myeloma, which were opposite to the TCGA database. Furthermore, the correlation between IGF2BP3 and PD-L1 in patients with BCa showed a trend similar to that of bladder cancer cell lines from CCLE (Fig. 7E). Next, we investigated the prognostic value of IGF2BP3 in BCa using the GEPIA website, and the results revealed that patients with high expression of IGF2BP3 had worse prognosis in BCa (Fig. 7F). The GETx database indicated that the expression level of IGF2BP3 in male and female bone marrow was significantly high. To compare the expression differences between males and females, and there was no difference in the expression of IGF2BP3 in most female and male tissues, except for blood vessels, brain, breast, and lung (Fig. 7G–J). Taken together, these results reveal that the expression of IGF2BP3 is high in various tumors and is associated with PD-L1, which may be a potential target for anti-PD-L1 immunotherapy.

Given that IGF2BP3 had a strong correlation with PD-L1 analyzed using a public database, we examined the expression levels of IGF2BP3 and PD-L1 in vitro. Stable IGF2BP3 overexpression and knockdown of T24, 5637, and UMUC3 cells were established, and the results revealed that overexpression of IGF2BP3 significantly increased, while knockdown of IGF2BP3 decreased both the protein and mRNA levels of PD-L1 in BCa cells (Fig. 8A–F). Further, flow cytometric assay showed that overexpression of IGF2BP3 significantly enhanced membrane-bound PD-L1 expression, and knockdown of IGF2BP3 decreased membrane-bound PD-L1 expression in T24 cells (Fig. 8G–H). The correlation between IGF2BP3 and PD-L1 expression was analyzed using BCa specimens. As shown in Fig. 8I, the positively correlated expression between IGF2BP3 and PD-L1 was found in 14/20 (70%) tumor specimens. Taken together, these data demonstrate that IGF2BP3 regulates both total and membrane-bound PD-L1 expression levels in BCa.