A sialyltransferases-related gene signature serves as a potential predictor of prognosis and therapeutic response for bladder cancer

The study was performed with dataset downloaded from The Cancer Genome Atlas (TCGA, https://​portal.​gdc.​cancer.​gov/​) with gene expression profile, copy number variation, single-nucleotide variant and clinicopathological data (age, gender, TNM stages, and prognostic data). The TCGA-BLCA cohort, containing 414 tumor tissue samples and 19 normal bladder tissue samples, was utilized as the training group, while GSE13507 dataset, based on GPL6102 platform downloaded from Gene Expression Omnibus (GEO, https://​www.​ncbi.​nlm.​nih.​gov/​geo/​), was applied as the validation group. Because of the public availability of bladder cancer data from online databases, no ethical approval or informed consent was required from patients in this study.

Based on the expression levels of the 20 SiaTs (ST3GAL1-6, ST6GAL1-2, ST8SIA1-6, ST6GALNAC1-6), consensus clustering analysis was applied to clarify patients into different sialyltransferases-related clusters (SiaTs_Cluster) with k-means algorithms by using R package of “ConsensusClusterPlus”. Gene set variation analysis (GSVA, c2.cp.kegg.v7.5.symbols and c5.go.bp.v7.5.symbols) was performed to investigate the biological functions between SiaTs_Clusters with R package “GSVA”.

Relationships between SiaTs_Clusters, clinicopathological features (age, gender, grade, TNM stage) and patients’ outcomes were explored to elucidate the significance of clusters generated by consensus clustering analysis. The comparison of the overall survival probability between SiaTs_Clusters was determined using Kaplan–Meier analysis with R packages of “survival” and “survminer”.

The infiltration levels of 22 kinds of immune cells were computed with CIBERSORT algorithm, which was also analyzed with a single sample gene set enrichment analysis (ssGSEA) algorithm.

Univariate Cox regression analysis for SiaTs_Clusters related DEGs was performed to identify the overall survival-associated DEGs. TCGA-BLCA patients were divided into distinct sialyltransferases gene clusters (gene_Clusters) based on the expression level of DEGs analyzed with consensus clustering analysis by using R package “ConsensusClusterPlus”. Then, Kaplan–Meier analysis was applied to compare the overall survival in different gene_Clusters.

Expression of the DEGs from different SiaTs_Clusters were standardized across bladder cancer specimens and the intersect genes were obtained. The univariate Cox regression analysis was carried out, and the survival-associated genes were retained for further analysis. Principal component analysis (PCA) was conducted to generate sialyltransferases-related gene score (SRGs_score) with the following algorithm: SRGs_score (risk score) = expression of a gene [1] × corresponding coefficient [1] + expression of a gene [2] × corresponding coefficient [2] + expression of a gene [3] × corresponding coefficient [3] + expression of a gene [4] × corresponding coefficient [4] + expression of a gene [5] × corresponding coefficient [5] + expression of a gene [6] × corresponding coefficient [6] + expression of a gene [7] × corresponding coefficient [7].

After the prognostic scoring system was established, the median value of the predicted SRGs_score was set as the cut-off. Patients with bladder cancer was divided into high- and low-risk group. Then, the comparison of the overall survival probability between high- and low-risk groups was conducted using Kaplan–Meier analysis with R packages “survival” and “survminer”. The 1-, 3- and 5-years’ ROC curve analysis was performed with R package “timeROC”, and the corresponding area under the curve (AUC) was calculated.

The RNA-seq data of TCGA-BLCA was applied to evaluate the abundance of 22 types of immune cells. Correlations between immune cells infiltration and SRGs_score were analyzed with Spearman’s correlation analysis, and the expression profile of immune checkpoint genes was performed between different SRGs_score groups. We also extracted the mutation annotation format from TCGA data to identify the mutational landscape of bladder cancer patients in SRGs_score subgroups, and the tumor mutation score for each patient was calculated and the mutation annotation format was created. The tumor stem cell features were extracted and stem cell-like characteristics of the tumors were evaluated with the transcriptome and epigenetics of the samples from RNAss file downloaded from TCGA database.

The clinicopathological features, including age, gender, grade, and TNM stages, were acquired from TCGA. To individualize the predicted survival probability for bladder cancer patients, a nomogram was established using clinical features and SRGs_score to assess the predictive accuracy, including 1-, 3-, and 5-years overall survival probability. We further utilized calibration curve analysis to confirm the reliability of the predictive nomogram we have established.

The semi-inhibitory concentration values (IC50) of various chemotherapeutic agents were evaluated with “pRRophetic” package. The lower-imputed drug sensitivity represented more sensitivity to the agents, whereas the higher-imputed represented low sensitivity.

R software (version 4.1.2), as well as relevant packages, was applied to make the statistical analysis. Correlations were evaluated with Spearman’s correlation analysis. Differences between two groups were analyzed with the independent sample t-tests or Mann–Whitney U tests, whereas differences between three or more groups were performed with one-way ANOVA or Kruskal–Wallis tests. The survival evaluation was carried out with Kaplan–Meier analysis. P < 0.05 was set as a significant.

The SiaTs family consists of 20 members (ST3GAL1-6, ST6GAL1-2, ST8SIA1-6, ST6GALNAC1-6), and abnormal expression of the SiaTs has been observed in multiple cancers. However, the relationship between the SiaTs and bladder cancer remains to be clarified. Here, the somatic mutations and copy number variations (CNV) of the SiaTs were explored with TCGA-BLCA dataset. The comprehensive landscape of the interactions between the 20 SiaTs is illustrated in Fig. 1A. As shown, the CNV alteration frequency was prevalent among the 20 SiaTs. ST6GALNACs (1–2, 4, 6), ST3GALs (1–3, 5), ST6GALs (1–2), and ST8SIAs (2, 6) represented a general frequency of CNV gain, while ST3GAL4, ST8SIAs (3–5) had a widespread frequency of CNV loss (Fig. 1B). Among the 412 bladder cancer samples, 42 (10.19%) experienced mutation in the 8 SiaTs which were ST6GALNAC1, ST3GAL6, ST8SIA2, ST6GAL2, ST6GAL1, ST6GALNAC5, ST6GALNAC3, and ST3GAL5 (Fig. 1C). Locations of the CNV alteration of the 20 SiaTs on chromosomes are shown in Fig. 1D. When compared to normal, 9 SiaTs showed decreased expression in tumor, including ST8SIAs (1, 6), ST6GALNACs (3, 5–6), ST6GALs (1–2), and ST3GALs (3, 5) (Fig. 1E), and the interactions between the 9 differentially expressed SiaTs was constructed by using STRING (Fig. 1F). The genetic heterogeneity and altered expression profile of the SiaTs suggested that SiaTs might play critical roles in the occurrence and progression of bladder cancer.

The regulator network showed the interactions between the 20 SiaTs, regulator relation, as well as their prognostic value for bladder cancer (Fig. 2A). Patients were classified into two subgroups (SiaTs_Cluster A and SiaTs_Cluster B) analyzed with the ConsensusClusterPlus R package based on the expression level of the 20 SiaTs with the optimal clustering variable of two (k = 2) (Fig. 2B–D). The principal component analysis indicated that the SiaTs-based classification pattern could classify patients with bladder cancer into these two subgroups (Fig. 2E). SiaTs_Cluster A had a particularly outstanding overall survival benefit compared with SiaTs_Cluster B (P = 0.005) (Fig. 2F). Moreover, a heatmap was generated representing the correlations between the SiaTs_Clusters and the clinical features of bladder cancer patients (age, gender, grade, TNM stage) (Fig. 2G), as well as the GSVA enrichment analysis of the biological differences (Fig. 2H). SiaTs_Cluster A was mostly enriched in the pathogenic Escherichia coli infection, progesterone-mediated oocyte maturation, oocyte meiosis, glioma, melanoma, regulation of actin cytoskeleton, focal adhesion, nod like receptor signaling pathway, prion diseases, systemic lupus erythematosus, viral myocarditis, chemokine signaling pathway, and leishmania infection; whereas SiaTs_Cluster B was enriched in the steroid hormone biosynthesis, retinol metabolism, metabolism of xenobiotics by cytochrome P450, glycerophospholipid metabolism, ether lipid metabolism, linoleic acid metabolism, and alpha linoleic acid metabolism.

To further analyze the differential level of the immune infiltration between the two clusters, the abundance of 22 kinds of the immune cells were included with the ssGSEA analysis. Among them, 21 types of the immune cells showed statistically difference between SiaTs_Cluster A and SiaTs_Cluster B, with significant differences in the immune cell compositions (Fig. 3A). SiaTs_Cluster A exhibited a greater enrichment of activated B cells (Fig. 3E), activated CD4 T cells (Fig. 3F), activated CD8 T cells (Fig. 3G), natural killer T cells (Fig. 3H), macrophages (Fig. 3I), and neutrophils (Fig. 3J), etc.; while CD56dim natural killer cells (Fig. 3B), monocytes (Fig. 3C) and type 17 T helper cells (Fig. 3D) were enriched in SiaTs_Cluster B. According to these findings, the SiaTs represented crucial relationship with immune infiltration, which might serve as a potential contributor to bladder cancer.

After the construction of SiaTs_Clusters and correlation analysis with biological functions, 372 DEGs were identified between SiaTs_Cluster A and SiaTs_Cluster B (∣log2fold change∣≧1, P < 0.05). The functional enrichment analysis was performed, and the top 10 enriched GO terms of biological process (BP), cellular component (CC), and molecular function (MF) were shown (Fig. 4A). It was found that the DEGs enriched in GO pathways were mainly associated with extracellular matrix organization, extracellular structure organization, collagen-containing extra cellular matrix and extracellular matrix structural constituent. In addition, KEGG enrichment analysis was also conducted, suggesting that the DEGs were mostly enriched in pathways related to extracellular matrix organization, extracellular structure organization, and negative regulation of immune system process (Fig. 4B). Then, the consensus clustering algorithm was utilized to classify patients into distinct subtypes (gene_Cluster A and gene_Cluster B) according to the expression of 372 DEGs, with the optimal clustering variable of two (k = 2) (Fig. 4C–E). Gene_Cluster B demonstrated better outcome than gene_Cluster A (Fig. 4F). The expression of the 20 SiaTs in these two gene_Clusters were further explored. We found that the expression of 17 kinds of the SiaTs showed significant, among which 14 were elevated (ST8SIA6, ST8SIA4. ST8SIA2, ST8SIA1, ST6GALNAC6, ST6GALNAC5, ST6GALNAC4, ST6GALNAC3, ST6GALNAC2, ST6GAL2, ST6GAL1, ST3GAL6. ST3GAL3, and ST3GAL2) in gene_Cluster A while 3 were increased (ST3GAL5, ST4GAL4, and ST3GAL1) in gene_Cluster B (Fig. 4G). The relationships between gene_Clusters and the clinical manifestations were investigated, and different characteristic genes were illustrated (Fig. 4H).

The Sankey diagram was applied to visualize the correlation between the SiaTs_Cluster, gene_Cluster, risk score (SRGs_score), and the outcomes (Fig. 5A). It was found that the risk score was significantly higher in SiaTs_Cluster A (Fig. 5B) and gene_Cluster A (Fig. 5C), supporting our previous findings that SiaTs_Cluster B and gene_Cluster B had better outcomes. In order to eliminate the influence of overfitting, least absolute shrinkage and selection operator (LASSO) Cox regression analysis was performed, and the best prognostic genes were identified, which were CD109, TEAD4, FN1, TM4SF1, CDCA7L, ATOH8, and GZMA (Fig. 5D, E). The risk score was calculated based on the expression of seven genes, as well as their coefficients, and the formula was established: Risk score = (0.222 × expression of CD109) + (0.216 × expression of TEAD4) + (− 0.147 × expression of ATOH8) + (0.097 × expression of FN1) + (0.106 × expression of TM4SF1) + (− 0.364 × expression of GZMA) + (0.124 × expression of CDCA7L). Patients with bladder cancer were divided into high- and low-risk categories based on the median risk score. According to the Kaplan–Meier plot, patients with low risk tended to have a better prognosis and longer survival time (Fig. 5F). Then, time-dependent ROC curves were constructed to evaluate the predictive ability of the constructed prognostic model in predicting 1-, 3-, and 5-years survival rates for bladder cancer patients, with AUC of 0.698 after 1 year, 0.685 after 3 years, and 0.700 after 5 years (Fig. 5G), suggesting the model was a good predictor. The distribution of the risk score and the patient survival outcomes revealed that the mortality rate of patients was elevated as the risk score increased (Fig. 5H). We also evaluated the expression of the seven genes in the risk group which found that five of them (CD109, TEAD4, FN1, TM4SF1, and CDCA7L) were upregulated while two (ATOH8 and GZMA) were downregulated in the high-risk group (Fig. 5I). These results demonstrated that the SRGs-related prognostic model could effectively predict the prognosis of patients with bladder cancer. Moreover, a nomogram based on the risk score and the clinical characteristics consisting of age, gender, grade, and TNM stages for predicting the patients’ overall survival (OS) was generated. As indicated, the risk score, age, gender, T stage and M stage were significantly correlated with the outcome of bladder cancer patients (Fig. 5J), and the calibration curve for internal validation of the nomogram showed good consistency between the nomogram-predicted OS and the observed actual OS of 1, 2, and 3 years (Fig. 5K).

ssGSEA algorithm was utilized to estimate the infiltration levels of immune cells. Differential abundance of immune infiltration in high- and low-risk groups is shown in Fig. 6A. Results indicated that the high-risk group had more immune cells present. The plasma cells (Fig. 6B), T cells CD4 memory activated (Fig. 6C), T cells CD8 (Fig. 6D), T cells regulatory (Tregs) (Fig. 6E) and monocytes (Fig. 6G) were negatively correlated with the risk score. On the contrary, the macrophages M0 (Fig. 6F) and neutrophils (Fig. 6H) were positively correlated with the risk score. The relationships between the seven genes and the infiltration levels of 22 types of immune cells were detected (Fig. 6I), as well as the correlations between the risk score and the expressions of 25 kinds of immune checkpoint genes (ICIs) (Fig. 6J). A total of 22 ICIs were differentially expressed, among which the majority of them were upregulated in the high-risk group (TNFRSF9, CD27, TNFRSF18, CTLA4, CD244, ICOS, TNFSF4, CD48, NRP1, CD276, TIGIT, TNFSF9, PDCD1, CD40, HAVCR2, TNFSF14, CD70, ADORA2A, CD40LG, TNFRSF4, and LAIR1), whereas only TNFRSF14 was downregulated.

As reported, the tumor mutation burden (TMB) was a catalyst of tumor progression and had a close linkage with the tumor outcome, however, we did not observe any significant difference between high- and low-risk groups (Fig. 7A). Moreover, no obvious correlations were detected between the risk score and TMB (Fig. 7B), as well as the RNA stemness (RNAss) (Fig. 7C). The mafttools package was chosen for the analysis of the distribution of the somatic mutations in the risk groups, which indicated that the high-risk group experienced a little bit increased TMB (95.02%) than low-risk group (92.61%), and the most significant mutated genes were TP53 and TTN. TP53 was the gene with the highest mutation rate (low/high = 35%/57%), while TTN was the gene with mutation rate of 40% in the low-risk group and 38% in the high-risk group (Fig. 7D, E).

The chemotherapeutic agents were screened based on the GDSC drug susceptibility database (https://​www.​cancerrxgene.​org/​), and the relationships between the SRGs-related prognostic signature and the responsiveness to various anticancer drugs were evaluated by calculating the half-maximal inhibitory concentration (IC50) of the samples. The high-risk group revealed more sensitive to bortezomib (Fig. 8A), cisplatin (Fig. 8B), docetaxel (Fig. 8C), gemcitabine (Fig. 8D), paclitaxel (Fig. 8E) and pazopanib (Fig. 8F), while the low-risk group was more sensitive to lapatinib (Fig. 8G), methotrexate (Fig. 8H), temsirolimus (Fig. 8I), MK2206 (Fig. 8J), SB590885 (Fig. 8K) and VX702 (Fig. 8L). In light of these findings, we believed that the risk scores based on the SRGs-related genes might be promising indicators to guide the selection of the clinical medical therapies.