Exploration of genetics commonness between bladder cancer and breast cancer based on a silcio analysis on disease subtypes.

BACKGROUND AND
OBJECTIVE: Muscle-invasive bladder cancers (MIBCs) are heterogeneous cancers and can be grouped into basal-like and luminal subtypes that are highly reminiscent of those found in breast cancer. Like basal-like breast cancers, basal-like MIBCs are associated with advanced stage and metastatic disease. However, the biological and clinical significance of molecular subtypes of MIBCs remain unclear. Therefore, we implemented a serious of bioinformatics methods to explore genetic similarities between bladder and breast cancers. METHODS AND
RESULTS: In the current study, by the application of multiple levels data analysis including random forest analysis, PPI and transcription factor regulation network construction, Gene Ontology (GO) and KEGG pathway enrichment analysis, we explored the genetics commonness between MIBC and breast cancers from the molecular heterogeneity based on the disease subtypes.
CONCLUSIONS: Our study identified some basal-related and luminal-related genes shared by two cancers. These studies can help shed light on the potential relationships between MIBC and breast cancer as a whole.

Introduction

Bladder cancer is a malignant tumor that occurs in the bladder epithelium and cause the high mortality. Bladder cancer can be divided into non-muscle invasive bladder cancer (NMIBC) and muscle invasive bladder cancer (MIBC) based on invasion into the muscularis propria of the bladder [1]. Recent studies based on the integrated genomic and protein analysis have better defined molecular subtypes of MIBCs, termed basal-like and luminal, with differences in clinical outcome [2]. Basal-like MIBCs are enriched with squamous features, with a keratinized/squamous phenotype, and showed a high level of sarcomatoid differentiation, whereas luminal MIBCs showed papillary histopathological features [3]. Compared with luminal MIBCs, basal-like MIBCs have more invasive and metastatic disease, and associated with shorter disease-specific survival as well as overall survival [2]. Therefore, the identification of MIBC subtype-specific biomarkers will help provide biological value for the MIBC clinical trials and therapy strategies.

Recently, more and more evidences showed that intrinsic subtypes of MIBC are remarkably similar to the intrinsic subtypes of breast cancer [4, 5]. Breast cancer is heterogeneous disease and can be grouped into five molecular subtypes, including claudin-low, basal-like, luminal-A, luminal-B, and HER2 enrichment, based on shared gene expression patterns [6, 7, 8]. Meanwhile, basal-like and luminal-A are two major breast cancer subtypes which have shown significant differences in terms of clinical outcome, prognosis and response to treatment. When compared with the molecular profile of several tumor types, basal-like MIBC is very similar to basal-like breast cancer, whereas luminal MIBC is similar to luminal-A breast cancer [5]. Their similar molecular profiles reflect in their similar functional roles. For example, a great amount of basal-like and luminal breast cancer specific gene signatures were enriched on the corresponding MIBC subtypes, including bona fide luminal breast cancer pathways, such as GATA3 and estrogen receptor signaling in the luminal MIBC subtype [5]. Previous studies have established that high level of Np63 characterize a lethal subset of basal-like MIBC, and Np63 pathway activation is also a characteristic feature of basal-like breast cancer, controlling basal biomarker expression and collective invasion [9, 10]. Moreover, Np63 knockdown in human bladder or breast cancer cells result in down regulated expression of basal biomarkers such as CD44, KRT5, KRT14, and CDH3 [2]. Basal-like MIBCs tumors express high levels of epidermal growth factor receptor (EGFR) and several of its ligands, which are similar to basal-like breast cancer tumors [2, 11]. Just like basal-like subtypes, luminal MIBCs and luminal breast cancer also share many biomarkers. For example, genes associated with luminal subtype of both cancers enriched on activating mutations of transcription factor, such as ESR1, TRIM24, FOXA1, and GATA3 [12]. A more in-depth understanding of the genetic basis of breast cancer in recent years has led to the development of new treatments and diagnostic tools. Comparison of MIBCs with breast cancer showed many molecular commonalities, indicating a related aetiology and similar therapeutic opportunities. Previous studies in MIBCs and breast cancer have identified some signatures associated with molecular subtypes, but there is little research on the genetics commonness between these two cancers based on their subtypes directly.

In this study, we revealed the molecular commonalities between intrinsic subtypes of MIBC and breast cancer, including basal-like and luminal (for breast cancer is luminal-A) subtypes by the application of bioinformatics method. First, the MIBC and breast cancer datasets were downloaded from published works and public database, respectively. Then, we identified differentially expressed genes (DEGs) of both cancers. Second, we used DEGs identified in first step to perform GO and KEGG enrichment analyses to explore their potential functions, and consequently found some overlapped terms shared by two cancers. Third, we selected overlapping genes extracted from the MIBC and breast cancer DEGs to perform random forest analysis. We constructed molecular subtype-related classifiers and extract some important cancer subtypes-related genes by random forest analysis. Finally, we divided the common genes shared by two cancers into four groups: all basal-related genes for two cancers, all luminal-related genes for two cancers, basal-related genes for bladder cancer whereas luminal-related for breast cancer, luminal-related genes for bladder cancer whereas basal-related for breast cancer, respectively. For these four groups of genes, we explored their relationships in protein-protein interaction (PPI) networks and transcription factor (TF) regulatory networks. The workflow of our works was shown in Fig. 1, and the detail was provided in the next section.

Figure 1.

The flowchart of our work, which includes (1) identification of bladder cancer and breast cancer subtypes-related DEGs, (2) GO and KEGG enrichment analysis for the identified DEGs, (3) random forest analysis using overlapping genes shared by two cancers (4) PPI network and TF-regulation network construction based on subtype special genes involved in the defined four groups.

Materials and methods

Data source and data preprocessing

In the current study, we performed the analysis for a set of bladder cancer subtype samples extracted from a MIBC dataset [13]. The original MIBC dataset was already normalized using a robust multichip averaging method. The dataset contained 1000 MIBC subtype related genes with high standard deviation. Twenty-two basal samples and 132 luminal samples were included in this dataset [13]. We used independent samples t-test by BRB-Array Tools to filter statistically significant DEGs that distinguish basal-like and luminal bladder cancer subtypes. In this method, repeated permutations of the data are used to determine if the expression of any gene is significant related to the phenotype. To get more information, 0.05 and false discovery rates (FDR) 0.1 is often as a popular and less stringent filter criterion to select a larger set of DEGs. We therefore also used this criterion to determine genes with various differentially expressed. Additionally, we selected breast cancer subtype-related mRNA expression profiling data (GSE45827) reported by Gruosso et al. to implement our analysis [14]. This mRNA dataset includes 41 basal-like samples and 29 luminal-A samples. For the breast cancer dataset, that has been normalized, the Limma package of R software (http://r-project.org) was used to compare the differential expression on different classes of samples. We selected the same criterion of 0.05 and FDR 0.1 to filter DEGs.

GO and KEGG enrichment analysis

In order to see whether the identified DEGs of both cancers can focus on some commonness function, we performed Gene Ontology (GO) and KEGG pathway enrichment analysis using DAVID web tool (http://david.abcc.ncifcrf.gov/). A GO term (or a KEGG pathway) with a p-value of less that 0.01 was considered to be significant. We did not perform the multiple test correction to avoid a loss of true-positive results.

Random forest analysis

To further explore relationships between MIBC and breast cancer in the view of data, we analyze those DEGs shared by two cancers using machine learning method. In this analysis, we selected the common DEGs shared by two cancers, and then applied Random forests (RF) method to construct molecular subtype related classifiers and extract the important cancer subtype related genes. RF is an ensemble classifier that consists of many decision trees and each tree depends on the values of a random vector sampled independently [15]. The method to obtain subtype-related feature gene G is permuting the expression value of gene G of out-of bag according to random forest algorithm. If gene G is a good predictor, then it will appear in a large number of split trees. We used Mean Decrease Gini (MDG) to evaluate whether gene G is a feature gene or not. MDG was the total decrease in node impurities measured by the Gini index from splitting on the variable, averaged over all trees. It provided possible ways to quantify which genes contribute most to classification accuracy. Greater MDG will indicate that the degree of impurity arising from category could be reduced farthest by gene G, and thus suggests an important feature gene [16]. The program was implemented with the randomForest package of R software.

Exploration of the potential regulation relationships of common subtype-related genes shared by MIBC and breast cancer

Among the common DEGs shared by two cancers, we defined two kinds of gene: basal-related gene when it shows higher expression in basal-like samples than in luminal (luminal-A), and luminal-related gene when it shows higher expression in luminal (luminal-A) samples than in basal-like, respectively. Under these conditions, we further divided these genes into four groups: all basal-related genes for two cancers, all luminal-related genes for two cancers, basal-related genes for MIBC whereas luminal-related for breast cancer, luminal-related genes for MIBC whereas basal-related for breast cancer, respectively. Then, we explored the potential regulation relationships of common subtype-related genes shared by MIBC and breast cancer.

Exploration of the interaction in PPI network of subtype-related genes

For four genes groups, we explored their relationships in protein-protein interaction (PPI) networks. In the practice, the PPI network is the basic method for systematically displaying protein-protein interaction information. In PPI network, the protein as a node, the interaction between two proteins as an edge, and the proteome as a whole is mapped to a system network. In this analysis, we use String database (http://string.org) to get the protein-protein pairs. Then, the Cytoscape software (version 3.4.0; http://cytoscape.org/) was used to visualize the PPI networks and to further explore the association between genes.

Exploration of transcription factor regulatory network of subtype-related genes

The transcription factor can bind to the nucleotide sequence specific upstream of the gene, thereby regulating the expression of the gene. Gene transcription regulation network can be constructed by gene transcriptional regulation data. For four groups of genes, we explored their transcription factor regulation networks. In this analysis, we selected TRANSFAC database (Release 2016) (http://www.biobase-international.com/product/transcription-factor-binding-sites) which is a manually curated database of eukaryotic transcription factors, their genomic binding sites and DNA binding profiles. The gene transcriptional regulatory network describes the relationship between transcription factors and their target genes, which can be represented by directed graphs. We used Cytoscape software to visualize these networks.

Results

Identification of DEGs for bladder cancer and breast cancer

The classic t-test in the BRB-Array Tools and Limma package of R were used to analyze the gene expression profiles of MIBC and breast cancer, and identified 741 MIBC subtype-related DEGs and 9946 breast cancer subtype-related DEGs. We described some important basal-like and luminal genes in both cancers in the Supplementary Fig. 1. We observed 9 genes overexpressed in basal-like samples whereas 3 genes overexpressed in luminal samples. Among these genes, KRT6A, KRT5 and CDH3 are reported direct DeltaNp63 transcriptional targets and can restore the expression of the basal kertains [2, 17]. Luminal subtype MIBC show features similar to those of Luminal-A breast cancer, with high mRNA expression in luminal samples than in basal-like samples, such as FOXA1 and GATA3 [8, 11, 12, 18].

We performed GO and KEGG enrichment analysis for the identified DEGs of both cancers in the first step using DAVID web tool. We used all 741 DEGs to perform enrichment analysis for bladder cancer. Notably, for breast cancer, because the number of DEGs was too great to analyze by online tool owing to overload, we chose a subset of genes (2677 DEGs) which are high significant (screening threshold: 0.01, logFC 1.2) to perform GO and KEGG enrichment analysis.

GO enrichment analysis

We listed the top ten significant GO terms for MIBC and breast cancer from Cellular Component (CC), Biological Process (BP) and Molecular Function (MF), respectively (see Fig. 2). From the GO enrichment analysis results, we found those common DEGs shared by MIBC and breast cancer focus on some significant functions. The most significant GO terms shared by both cancers were extracellular matrix organization. Extracellular matrix (ECM) is defined as a complex mixture of various proteins that provides structural and mechanical support for cells and tissues, and has an important role in the regulation of gene expression, cell division, survival, shape, and movement [19]. The altered ECM produced by the oncogene-exposed or transformed by carcinogen may play an important role in determining the ability of human bladder epithelial tumor to proliferate, relapse, or progress [20]. For breast cancer, a great amount of studies indicated that the ECM may be involved in various processes of breast tumors growth [21]. Some studies demonstrated that the ECM components were more heterogeneously at basal-like tumors when compared with the luminal subtype [22]. Interestingly, in our study we found that most of the GO terms shared by MIBC and breast cancer were related to extracellular, such as extracellular space, extracellular exosome, ECM structural constituent and so on.

Figure 2.

The GO analysis results for two cancers. The GO terms included biological process (red bars), cellular component (green bars) and molecular function (blue bars). The length of bar is positive correlation with significant level. (A) GO analysis results for MIBC. (B) GO analysis results for breast cancer.

Furthermore, we found some other co-enrichment GO terms of two kinds of cancers, such as collagen catabolic process, cell adhesion, integrin binding and chemokine activity, and all of these biochemical mechanisms play important roles in the development of the tumors.

KEGG enrichment analysis

We listed the top ten significant KEGG pathways of MIBC and breast cancer (see Fig. 3). The KEGG enrichment analysis results showed that the shared pathway by both cancers were ECM-receptor interaction and focal adhesion. Previous study reported that ECM molecules play an important role in the development of invasion, progression and metastasis in bladder cancer [20, 23]. The ECM in MIBCs may be a potential therapeutic target for treatment [23]. In the case of breast cancer, ECM is necessary for normal functional differentiation of mammary epithelia [21]. The altered ECM proteins may lead to different phenotypes of breast cancer. Moreover, an increasing number of researches indicated that the extracellular matrix and extracellular matrix receptors may be involved in controlling most of the successive stages of breast tumors, from appearance to progression and metastasis [21]. Focal adhesion kinase (FAK) is an intracellular non-receptor tyrosine kinase. FAK is a vital symbolic component that is activated by many stimuli, acting as a biosensor or an integrator to control cell motility [24]. FAK is an important regulator of bladder cancer cell invasion and migration, and it may become a potential therapeutic target [25]. For breast cancer, FAK may play a role in breast cancer by regulating breast stem cells (MaSCs) and breast cancer stem cells (MaCSCs) [24]. Emerging data suggested that FAK can be an effective therapeutic target in breast tumors, particularly in highly invasive triple-negative breast cancer (also terms basal-like breast cancer) [26]. Interestingly, we found that the amount of breast cancer related genes were enriched on bladder cancer and prostate cancer pathway.

Figure 3.

KEGG enrichment analysis results for two cancers. The length of bar is positive correlation with significant level. (A) Enriched KEGG pathway for MIBC. (B) Enriched KEGG pathway for breast cancer.

We next sought to define a set of genes that could accurately classify MIBC and breast cancer samples into the basal-like and luminal intrinsic subtypes. We extracted 460 overlapping genes shared by MIBC and breast cancer from DEGs, and then used these genes to construct MIBC type classifier and breast cancer type classifier. Two types of classifier can predict subtypes of two cancers.

Classification accuracy

The classification result showed that the MIBC and breast cancer subtype classifiers have powerful classification performance when 460 overlapping genes were used as predictor variables to classify samples. The classification accuracy of these two type classifiers were up to 98.1% and 100%, respectively. Importantly, when classified breast cancer subtype samples with MIBC type classifier, the classification accuracy is 85.7% (see Fig. 4A), indicating that the gene expression patterns that distinguish basal-like and luminal MIBC reflect the mRNA expression patterns that defined the intrinsic subtypes of breast cancer. However, the accuracy of breast cancer type classifier was only 66.2% when predict MIBC subtype samples. The reason of this case may be that we didn’t further select the subset genes with high fold change to create random forest models because of the limitation of available datasets. For this, we selected a gene subset with higher fold change (logFC 1.8) from 460 DEGs to construct another random forest classifiers. Accordingly, we used 319 genes as input to construct two cancer subtype random forest classifier models (see Fig. 4B). As a result, the accuracy of breast cancer subtype classifier improved to 67.5% when predict MIBC subtype samples.

Figure 4.

Generation of the cancer subtype classifiers. (A) Random forest analysis was performed using 460 overlapping genes. Each classifier generated by one kind of cancer predicted the other kind of cancer. (B) Random forest analysis was performed using 319 overlapping genes extracted from 460 DEGs according to the cutoff of logFC 1.8. Each classifier generated by one kind of cancer predicted the other kind of cancer.

Identification of important cancer subtype-related genes based on MDG

Specially, it is often interest to know which of genes are important in classification in Random Forests program. In this analysis, we used MDG to measure the distinguishing ability of genes. Greater MDG will indicate that the degree of impurity arising from category could be reduced farthest by a gene, and thus suggests an important gene. We listed the top 30 genes for MIBC and breast cancer respectively according to their MDG on Supplementary Fig. 2.

For MIBC genes, 17 genes were all basal-related genes whereas 9 genes were all luminal-related genes. Meanwhile, as a basal-related biomarker, KRT6B has the highest MDG. In the case of breast cancer, 26 genes were all basal-related genes whereas 4 genes were all luminal-related genes. Among these genes, PRC1 had the highest MDG. Among the top 30 MDG for each of cancers, there were some overlapped genes shared by both cancers, such as GATA3 and S100A9. The genomic gains of GATA3 were associated with the urothelial differentiation component. GATA3 is important for the proliferation of luminal breast tumors and maybe play the same role in bladder tumors presenting a molecular urothelial differentiation [5, 13, 18, 27]. The overexpression of S100A9 significantly correlated with tumor progression and worse outcome in the patients with bladder cancer, prostate cancer and breast cancer [28].

We divided 460 overlapping DEGs into four groups: all basal-related genes for two cancers (189 genes), all luminal-related genes for two cancers (130 genes), basal-related genes for bladder cancer whereas luminal-related for breast cancer (103 genes), luminal-related genes for bladder cancer whereas basal-related for breast cancer (38 genes), respectively. Then, we performed network analysis to explore the potential regulation relationships among these genes.

The relationships for genes involved in four groups respectively based on PPI network was shown in the Fig. 5. We found that some important genes have higher degree than other genes in PPI networks. As we know, basal-like subtype in both cancers is susceptible to have more aggressive and metastatic disease at the time of initial diagnosis and is associated with shorter disease specificity and overall survival. In basal-related group, genes with high degree were CCNB, CDC20, AURKA, AURKB, BUB1, KIF11, IL6 and KIF23.

Figure 5.

The relationships for genes involved in four groups respectively based on PPI network. The larger circle indicates the gene with higher degree. (A) All basal-related genes for two cancers. (B) All luminal-related genes for two cancers. (C) Basal-related genes for MIBC whereas luminal-related for breast cancer. (D) Luminal-related genes for MIBC whereas basal-related for breast cancer.

Among these genes, CCNB1 had the highest degree (degree 35). The high expression of CCNB1 can lead to uncontrolled growth of the tumor. Song et al. have shown that CCNB1 promotes cancer invasion and metastasis by enhancing epithelial to mesenchymal transition (EMT) process [29]. The overexpression of CDC20 was associated with poor prognosis of urothelial carcinoma of the human bladder [30]. In breast cancer, CDC20 is a hallmark of triple-negative breast cancer and its high expression predict high risk of death [31]. Previous studies suggested that in MIBCs patients presenting with hematuria, AURKA can be used as a diagnostic biomarker to detect bladder cancer as well as a potential therapeutic target [32]. Breast cancer with high AURKA level showed poor prognosis [33]. Overexpression of BUB1 may be a new hallmark for estimating the biological characteristics of bladder cancer [34]. In the case of breast cancer, BUB1 expression is associated with a different clinical outcome in breast cancer patients and can be a potential therapeutic target [35]. IL-6 is a clinically significant prognostic predictor and may represent a good therapeutic target of bladder cancer [36]. For breast cancer, the systemic and local expression of IL-6 represents two different partitions, with systemic levels reflecting whole body metabolic and inflammatory status, whereas local IL-6 expression may have a direct effect on cancer cell growth and metastasis [37].

In the luminal-related group, we identified 9 important genes (CCND1, FOXA1, TFF1, GATA3, SDC1, TBX3, MSX2, FGFR3, and MYH11). We know that the luminal subtype is not as intrinsically aggressive as basal-like subtypes in both cancers. The gene with the highest degree in luminal-related group was CCND1 (degree 8). The expression of CCND1 and its protein product Cyclin D1 frequently altered in human cancers and may be used to predict tumor progression [38]. Evidence from previous studies suggested that high CCND1 expression was associated with poor prognosis in bladder cancer as well as breast cancer patients [38, 39]. FOXA1 and GATA3 are luminal molecular subtype biomarkers of bladder cancer and breast cancer [2, 18, 27, 40]. Therefore, FOXA1 and GATA3 can serve as a clinical marker for the luminal subtype, and its prognostic capacity in these low-risk bladder cancer and breast cancer can prove useful in clinical treatment decisions [40]. SDC1 is a transmembrane heparan sulfate proteoglycan and it expression was associated with stage progression and poor prognosis of MIBCs and breast cancers [41]. The SDC1 expression can provide valuable information for both cancers prognosis. FGFR3 is one of the most common biomarkers of luminal subtype MIBC and breast cancer [8, 18]. FGFR3 is an obvious molecular target in luminal caners.

The transcription factor regulatory relationships based on TRANSFAC database were shown in the Fig. 6. In this study, we found some transcription factors regulate genes involved in different groups with different ways, such as FOXA1, FOSL1, FOXM1, GATA3, TFAP2A, JUN, EGR1, MYC and so on. Among these genes, all basal-related transcription factors for two cancers included FOSL1, FOXM1, MYC, whereas all luminal-related transcription factors for two cancers included FOXA1 and GATA3. These transcription factors play important roles in regulate progression of carcinogenesis. The FOXM1 transcription factor plays a key role in regulating cell proliferation, differentiation, and transformation [42]. Overexpression of FOXM1 is associated with a variety of invasive solid carcinomas, including bladder cancer and breast cancer [42]. The MYC gene encodes about 60 kD of nuclear phosphoric acid protein with DNA binding activity [43]. Abnormal regulation of MYC can lead to phenotypic transformation, abnormal cell cycle control and genomic instability [43]. Our study, and others, found that key transcription factors FOXA1 and GATA3 were up regulated in the luminal subtypes of both kinds of cancers, and both transcription factors have important roles in luminal epithelial differentiation [2, 18, 40].

Figure 6.

The transcription factor regulatory networks for genes involved in four groups respectively based on TRANSFAC database. The green circles indicate genes in each group, and the red triangles indicate transcription factors which regulate these genes. (A) All basal-related genes for two cancers. (B) All luminal-related genes for two cancers. (C) Basal-related genes for MIBC whereas luminal-related for breast cancer. (D) Luminal-related genes for MIBC whereas basal-related for breast cancer.

Discussion

MIBC is extremely aggressive disease which cause a high mortality, and the traditional treatment of its therapeutic effect is limited [2]. For an early diagnosis and appropriate targeted treatment, studies about molecular characterization diseases phenotypes and prediction of novel biomarkers are essential. MIBC tumors can be distinguished into basal-like and luminal subtypes which remarkably similar to breast cancer subtypes in molecular features and clinical outcomes [4]. In the past few years, bioinformatics research on breast cancer subtypes has yielded great results, and there is a very mature mechanism of molecular subtype classification and therapeutic options [11]. Therefore, given the potential similarities between MIBC and breast cancer at the molecular level, the abundant breast cancer knowledge base may be utilized to help make progress more rapidly in the clinical management of MIBC.

Here, we applied bioinformatics method to analyze MIBC and breast cancer mRNA datasets, and consequently found some genetics commonness between MIBC and breast cancer subtypes. According to enrichment analysis, we found that the main biological processes based on GO focus on ECM related terms. Moreover, the KEGG enrichment analyses showed that DEGs of both cancers were significantly enriched on ECM-receptor interaction pathway. It is obvious that ECM and ECM-receptor status are altered in tumors with profound functional implications which are of value for cancer diagnosis, prognosis and therapy [20]. Bergamaschi et al. demonstrated that primary breast cancer can be classified based on the differential expression of a set of ECM-related genes, and this classification has implications for clinical outcome [19]. We believe that the association study between ECM-related genes and bladder cancer subtypes can help understand MIBC pathology from molecular levels. It is no doubt that explore the relationships between ECM related genes and MIBC subtypes is extraordinary helpful in molecular subtypes classification and target therapy.

We created gene classifiers which accurately discriminate intrinsic cancer subtypes based on the common DEGs shared by two cancers. By random forest analysis, we constructed two types classifier: one is MIBC type classifier which trained by MIBC mRNA expression and the other is breast cancer type classifier which trained by breast cancer mRNA expression respectively. Subsequently, we predicted breast cancer subtype samples by MIBC type classifier and MIBC subtype samples by breast cancer type classifier. Furthermore, different subtypes of bladder cancer and breast cancer can be identified by a similar basic regulatory system, even if the phenotype and tissue origin may differ. We further identify the gene co-regulation relation in basal-like and luminal subtypes. We divided those DEGs shared by two cancers into four groups and then performed biological network analysis for each group. We identified some hub genes that play important roles in basal-like and luminal tumor development. In addition, we identified some transcription factors which regulate genes involved in four groups, such as FOS, FOXA1, GATA3, TFAP2A, FOXM1, JUN, EZH2, EGR1, FOSL1 and MYC. Abnormities in these transcription factors causing gene expression abnormities play important roles in human cancers including bladder and breast cancer.

Specially, we performed added exploratory study to see whether existing potential interior difference between two MIBC molecular subtypes as well as breast cancer subtypes. To find the potential interior difference between two molecular subtypes for each cancer, we performed cluster analysis on basal-like samples and luminal samples based on consensus cluster method, respectively. For bladder cancer, we used original 1000 genes to implement cluster analysis. For breast cancer, for the purpose of improving cluster efficiency, we selected equal numbers of genes (1000 genes) with highest standard deviation extracted from 9946 breast cancer DEGs as input to perform the analysis. CancerSubtypes package of R software (http://www.r-project.org) was used to implement this analysis. In this study, we divided each kind of cancer into two subtypes: basal-like subtype and luminal subtype. Accordingly, four types of samples were included: basal-like MIBC samples, basal-like breast cancer samples, luminal MIBC samples and luminal-A breast cancer samples. Then, for each type of samples, we clustered samples into two clusters using consensus cluster method, and calculated the corresponding silhouette scores. Silhouette score indicates the consistency within clusters of data. A higher silhouette score indicates the greater similarity within the samples involved in the same cluster. Finally, for two obtained clusters of each type of samples, we implemented paired comparison between of them. For basal-like subtype in two cancers, the paired comparison between two clusters were all significant ( 0.001 and 0.026 for MIBC and breast cancer respectively, see Supplementary Fig. 3A and C). For luminal subtype in two cancers, the paired comparison between two clusters were all not significant ( 0.91 and 0.064 for MIBC and breast cancer, respectively, see Supplementary Fig. 3B and D). The result suggested that patients with basal-like subtypes may have greater individual differences than luminal (and luminal-A) subtype, which increases the difficulty of treatment for basal-like subtype.

It also should be pointed out the limitations of our analysis. In the present study, we only selected the limited datasets to perform a series of analyses. This may cause the inadequate conclusions and generate bias in some cases. Therefore the identified candidate genes need to be validated for further experimentation. Specifically, the lack of MIBC and breast cancer subtype expression profiling datasets causes the limitations in the data analysis and these results need to be approved in the future studies when more bladder cancer and breast cancer subtype expression profiling datasets are available.

In summary, like the basal-like and luminal subtypes of breast cancer have distinct and reciprocal gene expression profiles as well as large differences in clinical characteristics, the basal-like and luminal subtypes of MIBC also reflect many different aspects of urothelial physiological development. In this study, we use bioinformatics methods to analyze mRNA expression datasets of MIBC and breast cancer and explored the genetics commonness between MIBC and breast cancer. We constructed random forest classifiers that distinguish cancer subtypes based on gene expression profiles. The genes contributing greatly to classify tumor subtypes may be important potential biomarkers for MIBC and breast cancer development and progression, and eventually become candidates for therapeutic targeting.