Chromatin-regulatory genes served as potential therapeutic targets for patients with urothelial bladder carcinoma

2.1 Bladder cancer data from Genomic Data Commons (GDC)

There were totally 412 cases of the urothelial bladder carcinoma from TCGA project in the GDC data portal (Grossman et al., 2016). The clinical information was retrieved using TCGA biolinks (Colaprico et al., 2016). The Level 3 data of transcriptome profiling from Illumina RNA-Seq platform, processed mutation files using MuTect2 from exome sequencing (Cibulskis et al., 2013) and DNA methylation levels from Illumina Human Methylation 450 were directly downloaded using the GDC client successively in early November 2017.

2.2 Mutated genes involving in chromatin modification

The genes participating in modifying or regulating chromatin were obtained from the Reactome Knowledgebase (Croft et al., 2014; Fabregat et al., 2017). Jupe (2014) defined the pathway “Chromatin Modifying Enzymes,” which contained 241 different proteins. The mutation information of the genes encoding these 241 proteins was extracted from the Mutation Annotation Format (MAF) file download from the GDC. After removing the synonymous mutations, the mutations were categorized into missense mutations, in-frame indels, frame shift indels, mutations in splice regions, nonsense mutation, and mutations at the translation start site. The genes with mutations in at least 10 patients were visualized using OncoPrint (Cerami et al., 2012; Gao et al., 2013) and plotted with R package ComplexHeatmap (Gu et al., 2016).

2.3 Subtypes of bladder cancer

The counts of genes based on HTSeq were retrieved from GDC and the counts per million were calculated for each gene. The normalization procedures were executed by EDASeq to remove with-lane and between-lane differences. The expression levels and mutational profiles of the genes involving in the chromatin were utilized altogether to figure out whether there existed subtypes of bladder cancer. Similarity network fusion (SNF) was used to construct a fusion patient similarity network by integrating the patient similarity obtained from mutational and transcriptional profiles, which included the similarity matrix calculated by a cross-network diffusion process and a clustering step (Wang et al., 2014). These procedures were carried out using the R package CancerSubtypes (Xu et al., 2017). The number of nearest neighbors was set as 30 and the variance for local model was 0.5. The number of clusters was determined by the average silhouette width. If the number of clusters was set as 3, compared with other number between 2 and 10, the average silhouette width was maximized.

To identify whether the clustered groups were correlated with clinical parameters, different statistical methods were used including analysis of variance tests for age at diagnosis and the packs smoked, log-rank test for survival status (Harrington & Fleming, 1982) and Fisher's exact test for all other histoclinical parameters (Agresti, 2002). To visualize the survival status, the Kaplan–Meier curves were plotted.

2.4 DNA methylation data preprocessing

Considering that the DNA methylation is one important mechanism modifying chromatin and the methylation datasets from Illumina Human Methylation 450 were available, the methylation levels was calculated for each genes with R package MethylMix (Gevaert, 2015). First, the samples with over 50% of unavailable data were eliminated and the left unknown cells were imputed based on the other data. Next, each probe of DNA methylation was mapped to a gene based on the annotation of Illumina methylation arrays. Probes with single-nucleotide polymorphisms (SNPs) were removed. Finally, for each gene, all its CpG sites were clustered together using hierarchical clustering and Pearson's correlation. Then the hierarchical tree was cutoff at a Pearson's correlation threshold of 0.4 to define CpG site clusters and single CpG sites when they do not correlate with other sites, resulting in potentially multiple CpG site clusters representing a single gene.

2.5 Genes probably regulated by chromatin structure

The genes, which were differentially expressed in one clustered group or two with R package EBSeq, were identified (Leng et al., 2013). The genes with adjusted p value (false discovery rate [FDR]) lower than 0.05 were considered as significant. Biological functional analysis was finished using DAVID bioinformatics resources (Huang da et al., 2009, 2009). Combined with their DNA methylation levels, we obtained genes whose messenger RNA (mRNA) expression levels were also negatively correlated with the methylation levels.

3.1 The mutational profiles of chromatin-regulatory genes

Fifty of the 241 chromatin-regulating genes had nonsilent mutations in at least 10 patients, which were shown in the OncoPrints (Figure 1). Individual genes were represented as rows, and individual cases were represented as columns. The SNPs, including truncated mutations, in-frame deletion or insertion and missense mutations, were shown as different colors.

The genes had mutating frequency over 5% included histone demethylases (like KDM6A), histone methyltransferases (like KMT2A, KMT2C, and KMT2D), histone acetylases (like CREBBP and EP300), members of the SWI/SNF chromatin remodeling complex (like ARID1A and ARID2), and histone demethylase (like KDM6A), which were listed in Table 1. The mutation counts of different types were also calculated. Mutations in some genes were predominantly inactivating. For example, KMT2D, KDM6A, CREBBP, ARID1A, and ARID2 had more inactivating mutations than nonsynonymous mutations or in-frame indels. Compared with other genes, nonsilent mutations in genes involving in the chromatin modification were significantly enriched. Such findings suggested that these genes were functionally relevant with the cancer progression.

The inactivating mutations on KMT2D and KDM6A were highly exclusive, where only eight patients had both inactivating mutations and 158 patients had only inactivating mutations in one gene. Similar patterns were presented between EP300 and CREBBP. The possible reasons include that mutations in the two genes have redundant downstream effects or the combined loss cause synthetic lethality.

3.2 The subtypes of bladder cancer based on chromatin-regulatory genes

SNF clustering was performed with a cluster number as three. The patients were divided into three groups using the mutation information and the mRNA expression levels of 241 chromatin-regulatory genes. And the similarity between patients were represented by the proportion of clustering runs in which two patients were group together, which was named as consensus values (Figure 2a). In the plot, patients belonging to the same cluster were adjacent to each other and color-coded gradient represented the similarity, which showed that the patients were clearly clustered into different groups.

Among 399 cases where both the SNP information and the mRNA expression profiles were available, 326, 23, and 50 patients were clustered into Groups 1, 2, and 3, respectively. We checked whether there existed differences among groups based on pathologic stage, gender, race, age, or the number of packs the patients smoked every year (Table 2). However, there was no significant relationship. Only the number of packs seemed to have poor correlation with a p value of 0.06. On the other hand, the survival analysis suggested that the patients in these three groups had a different prognosis. The p value of the log-rank test was 0.0025. The patients in Group 2 had significantly low survival time while those in Group 3 had better prognosis (Table 3).

3.3 Differentially expressed genes among clustered groups

R package EBSeq was utilized to obtain the genes that were significantly expressed in at least one group. The obtained p values were adjusted by FDR and 0.05 was used as the cutoff for significance. Thirty-one genes were identified as differentially expressed finally. The boxplots of top six were shown as Figure 3.

Next, the correlations between DNA methylated levels and the mRNA expression levels were checked but there showed no correlation. It suggested that the variety of mRNA expression levels of these genes were not regulated by the DNA methylation but other factors. Chromatin structure was one possible factor, where the aberrant status of chromatin-regulatory genes affected the expression levels of the differentially expressed genes.

The GO term enrichment showed that the differentially expressed genes were enriched in the immune response especially in the complement activation.