A six-gene prognostic model predicts overall survival in bladder cancer patients

We downloaded DNA methylation, gene transcriptome and clinical survival data of BC patients from TCGA [16]. There were 437 samples with DNA methylation data (21 normal and 416 cancer), 430 samples with gene transcriptome data (19 normal and 411 cancer), and 404 patients with valid survival data. These data are an open resource, and no ethical issues were involved.

First, we applied the Limma package in R to extract the DNA methylation data. Next, we used the edgeR package to obtain the gene expression data. A comprehensive analysis was performed to obtain the following three data matrices: DNA methylation (normal, cancer) and gene expression. Subsequently, we used MethylMix [13, 17] to compare DNA methylation of cancer with that of normal tissue to detect specific genes, particularly BC-specific hyper/hypomethylation genes, and the methylation level of these genes was described as ‘transcriptionally predictive’. A mixture model of each gene was built, and Wilcoxon rank tests were computed with the following parameters: set as logFC > 0, P < 0.01, and Cor < − 0.3.

Metascape [18] integrates several authoritative data resources, such as GO, KEGG, UniProt and DrugBank, so that it can execute pathway enrichment and biological process annotation to provide comprehensive and detailed information for each gene [19]. GO enrichment and KEGG pathway analyses of the genes identified by MethylMix were performed. Only terms with P < 0.01, a minimum count of 3 and an enrichment factor > 1.5 were considered significant.

First, the genes identified by MethylMix were applied to a univariate Cox regression. Second, we used a LASSO regression to narrow the range of target genes because the predictor variable was much larger than the sample content in the gene expression data. A strong correlation often exists between the variables, which is suggestive of high dimensionality and collinearity, and this method could decrease the characteristic dimension [20]. Then, we built a multivariate Cox regression model to select the genes that were most tightly associated with survival [21]. In addition, we validated this model in subgroups based on different characteristics. The following 12 subgroups based on different clinical characteristics and 9 subgroups based on different mRNA subtypes and mutational signatures [5] were subjected to further tests: high grade (n = 381), low grade (n = 20), stage I (n = 2), stage II (n = 128), stage III (n = 139), stage IV (n = 133), muscle-invasive (n = 368), non-muscle-invasive (n = 4), no distant metastasis (n = 193), distant metastasis (n = 11), lymph node metastasis (n = 169), no lymph node metastasis (n = 235), or Msig 1 (n = 28), Msig 2 (n = 220), Msig 3 (n = 99), Msig 4 (n = 55), basal squamous (n = 137), luminal (n = 26), luminal infiltrated (n = 77), luminal papillary (n = 140), and neuronal (n = 20).

The sensitivity and specificity of the model in the diagnosis of BC were analysed by a time-dependent ROC curve.

Furthermore, an OS-associated nomogram including the risk score and clinical factors was designed using the rms [22] and the Hmisc [23] packages in R. Calibration curves were drawn, and the concordance index (C-index) was computed to assess the efficiency of the nomogram.

Kaplan–Meier curves were used to distinguish the connection between these genes and prognosis. A subgroup analysis was performed by dividing the patients based on clinical characteristics. A methylation/methylation site and gene expression matched survival analysis was carried out to explore the prognostic significance of these genes individually. The relationships between gene expression and methylated sites were additionally examined.

All data analyses were performed with R [24]. Student’s t-test was used to evaluate the differences between two groups. The log-rank test was applied in the Kaplan–Meier survival examination.

In total, 167 genes were identified (Fig. 1; Additional files 1, 2) by applying MethylMix to the three matrices, and a mixture model of each gene was constructed (Fig. 2). Intuitively, the relationship between the peak curve and the black bar indicates whether a gene is hyper- or hypomethylated.

The Metascape analysis shows the top 20 clusters of enriched sets (Fig. 3). These genes were enriched in the molecular function (MF) categories structural constituents of muscle and RNA polymerase II distal enhancer sequence-specific DNA binding. For biological process (BP), these genes showed enrichment in anterior/posterior pattern specification, chordate embryonic development, intrinsic apoptotic signalling pathway in response to DNA damage and so on (Additional file 3). The KEGG pathway data were enriched in Glutathione metabolism and Cardiac muscle contraction.

The results of the univariate Cox regression analysis of 167 genes were used in the LASSO regression to identify robust markers. A set of twelve genes (DAPP1, TCEAL7, PAXIP1-AS1, TDRD1, NUPR1, ARHGDIB, LINC00526, IDH2, ARL14, KLHDC7A, GSTM2, and LURAP1) and their coefficients were computed (Fig. 4a, b). Then, multivariate Cox regression analyses were performed, and a six-gene model was constructed according to their methylation levels and coefficients. Risk score = (ARHGDIB * 4.533910954) + (LINC00526 * 1.999499891) + (IDH2 * − 2.048441591) + (ARL14 * 0.779318158) + (GSTM2 * − 1.375204374) + (LURAP1 * − 1.504186188).

The risk score of each BC patient was computed, and the patients were assigned to the low-risk (n = 202) or high-risk (n = 202) group based on the median cut-off value (Additional file 4, Fig. 4c). Intuitively, the number of deaths was significantly higher in the high-risk group (Fig. 4d). The distribution of the six genes across all samples showed that the patients in the low-risk group were likely to have a higher degree of methylation of IDH2, GSTM2 and LURAP1. In contrast, the patients in the high-risk group were inclined to have higher methylation of ARHGDIB, LINC00526, and ARL14 (Fig. 4e). The Kaplan–Meier analysis of all patients (Fig. 4f) indicated that the survival of the patients in the low-risk group was significantly better than that of the patients in the high-risk group (P = 1.679e−05). The AUC of the survival assessment model of the six methylation-driven genes was 0.698 at 3 years of OS (Fig. 4g).

We further tested the survival assessment model by Kaplan–Meier analysis in subgroups. Of the 12 subgroups classified by clinical characteristics, there were no enough cases for stage I (n = 2), non-muscle-invasive (n = 4), and distant metastasis (n = 11), and all patients in the low grade group (n = 20) were alive, the test in the remaining 8 groups showed the same results as in all patients (Fig. 5a–h). Although the P-value in the stage III group was not statistically significant (Fig. 5h), these patients all showed the same predictive trends. Of the 9 subgroups (Additional file 4) classified by different mRNA subtypes or mutational signatures of BC [5]. The Kaplan–Meier curves (Additional file 5) show that this model is still effective in the Msig2, Msig3, luminal infiltrated, luminal papillary, and neuronal groups. Thus, the model has a certain reliability and practicability in evaluating prognosis.

We designed a nomogram to predict the survival probability of each patient. In the nomogram, each predictor was assigned a score. Based on the Cox analysis results (Table 1), six genes were integrated in the nomogram to predict the survival probability of BC patients (Additional file 6). The hypermethylation of ARHGDIB, LINC00526, and ARL14 is a risk factor for OS. Similarly, we carried out an analysis of the risk score and five clinical factors (Table 2; Fig. 6a). Based on the univariate Cox analysis, four factors (race, age, gender, and stage) and the risk scores were included in the multivariate Cox analysis (the factor ‘grade’ was not suitable for further analysis according to R). We constructed a nomogram to predict the OS probability (Fig. 6b). The C-index of this model was 0.694 (Fig. 6c). The predicted survival rate is close to the actual survival situation, and the prediction accuracy is similar to the ROC curve.

The survival status evaluation of 6 genes was computed by the Survival package in R. IDH2 and ARL14 were identified as independent prognostic indicators (Fig. 7a, b), and the hypomethylation of IDH2 and hypermethylation of ARL14 were related to worse prognosis in BC patients. The methylation/methylation-site and gene expression matched evaluation was additionally carried out to discover the prognostic value. We found that a high expression and hypomethylation of ARHGDIB and ARL14 were meaningfully correlated with better prognosis (Fig. 7c, d). Among the six genes, 6 methylated sites in ARHGDIB (Fig. 8a–f), 1 methylated site in ARL14 (Fig. 8g), 3 methylated sites in GSTM2 (Fig. 8h–j), 2 methylated sites in LINC00526 (Fig. 8k, l) and 2 methylated sites in LURAP1 (Fig. 8m, n) were significantly associated with BC prognosis. The hypermethylation of 5 sites in LURAP1 and GSTM2 is associated with better prognosis; in contrast, the hypermethylation of another 9 sites in ARHGDIB, LINC00526 and ARL14 is associated with poor prognosis. This result is consistent with the results shown in Figs. 4e and 5a. The hypermethylation of IDH2, LURAP1, and GSTM2 may act as a protective factor in BC patients. Other three genes, i.e., ARHGDIB, LINC00526, and ARL14, may have the opposite effect. Additionally, several abnormally methylated sites were identified as linked to gene expression (Table 3; Additional file 7).