A supervised machine learning approach identifies gene‐regulating factor‐mediated competing endogenous RNA networks in hormone‐dependent cancers

Abstract

Competing endogenous RNAs (ceRNAs) have become an emerging topic in cancer research due to their role in gene regulatory networks. To date, traditional ceRNA bioinformatic studies have investigated microRNAs as the only factor regulating gene expression. Growing evidence suggests that genomic (e.g., copy number alteration [CNA]), transcriptomic (e.g., transcription factors [TFs]), and epigenomic (e.g., DNA methylation [DM]) factors can influence ceRNA regulatory networks. Herein, we used the Least absolute shrinkage and selection operator regression, a machine learning approach, to integrate DM, CNA, and TFs data with RNA expression to infer ceRNA networks in cancer risk. The gene‐regulating factors‐mediated ceRNA networks were identified in four hormone‐dependent (HD) cancer types: prostate, breast, colorectal, and endometrial. The shared ceRNAs across HD cancer types were further investigated using survival analysis, functional enrichment analysis, and protein–protein interaction network analysis. We found two (BUB1 and EXO1) and one (RRM2) survival‐significant ceRNA(s) shared across breast‐colorectal‐endometrial and prostate–colorectal–endometrial combinations, respectively. Both BUB1 and BUB1B genes were identified as shared ceRNAs across more than two HD cancers of interest. These genes play a critical role in cell division, spindle‐assembly checkpoint signalling, and correct chromosome alignment. Furthermore, shared ceRNAs across multiple HD cancers have been involved in essential cancer pathways such as cell cycle, p53 signalling, and chromosome segregation. Identifying ceRNAs' roles across multiple related cancers will improve our understanding of their shared disease biology. Moreover, it contributes to the knowledge of RNA‐mediated cancer pathogenesis.

1. INTRODUCTION

Historically, the role of protein‐coding genes has been studied extensively in the development and occurrence of cancer. However, in recent years, the investigation of noncoding RNAs (ncRNAs) has increased attention. For example, long noncoding RNAs (lncRNAs) and microRNAs (miRNAs) are two types of ncRNAs that play crucial roles in transcriptional and posttranscriptional gene regulation. ¹ miRNAs are one of the most‐studied regulating factors on gene networks, and their aberrant expression link with the development of several diseases, including cancer. ² , ³ , ⁴ The competing endogenous RNA (ceRNA) hypothesis postulates that messenger RNAs (mRNAs) and other RNA transcripts, such as lncRNAs and pseudogenes, can act as natural miRNA sponges. ⁵ These RNAs influence each other's expression levels by competing for the same pool of miRNAs through miRNA response elements (MREs) on their target transcripts, thereby modulating gene expression and protein activity. ⁶ Crosstalk between ceRNAs through shared MREs represents a novel layer of gene regulation that plays an important role in the physiology and development of complex human diseases such as cancers.

Existing ceRNA network analyses have described many risk‐associated ceRNAs in common cancers. ⁷ , ⁸ , ⁹ , ¹⁰ , ¹¹ Nevertheless, these bioinformatic analyses have considered miRNAs the only factor regulating gene expression. Additional gene‐regulating factors, namely DNA methylation (DM), transcription factors (TFs), and copy number alteration (CNA), also affect ceRNA regulatory networks. ¹² DM plays a critical role in regulating gene expression through epigenetic control of various biological processes and diseases, including cancer. TFs are commonly deregulated in the pathogenesis of human cancer and are a significant class of cancer cell dependencies. ¹³ The CNAs are somatic changes to chromosome structure that results in gain or loss in copies of sections of DNA and are prevalent in many types of cancers. Therefore, current ceRNA‐based computational analyses should be accountable for these genetic and epigenetic effects. ¹²

Herein, we used a supervised machine learning approach, ‘Cancerin (Cancer‐associated ceRNA interaction networks)’, ¹² to identify regulation factors‐mediated ceRNA networks in a group of genetically related hormone‐dependent (HD) cancers, using genetic data derived from The Cancer Genome Atlas (TCGA). Cancerin follows three main steps to infer genome‐wide ceRNAs in cancer risk as follows: (i) retrieving putative miRNA–mRNAs and miRNA–lncRNA interactions; (ii) selecting miRNAs that contribute to the RNA expression level; and (iii) inferring ceRNA associations through the hypergeometric test, Pearson correlation analysis and the sparse partial correlation analysis. We identified common ceRNA pairs (mRNA–mRNA or lncRNA–mRNA pairs that have shared the same miRNA[s]) across multiple HD cancers. All cross‐cancer ceRNA pairs obtained from the analysis were involved in survival analysis, functional enrichment analysis, and protein–protein interaction (PPI) network analysis. Identification of the function and mechanism of ceRNAs improves our understanding of RNA‐mediated cancer pathogenesis. Moreover, finding common ceRNAs across HD cancers will significantly contribute to our understanding of the shared biology of HD cancers.

2. MATERIALS AND METHODS

2.1. Data sets

The genomic and clinical data of five HD cancer types (reposited in TCGA were extracted using the Genomic Data Commons data portal (https://gdc.cancer.gov/access-data/gdc-data-portal). ¹⁴ Cancer types considered in this study were breast invasive carcinoma (BRCA), prostate adenocarcinoma (PRAD), colon adenocarcinoma (COAD), rectum adenocarcinoma (READ) and uterine corpus endometrial carcinoma (UCEC). The PRAD, BRCA, COAD, READ and UCEC consist of 499/52 (cases/controls), 1109/113, 480/41, 167/10 and 552/35, respectively. We retrieved HTSeq‐counts and isoform quantification for RNA and miRNA expression data. Putative miRNA–mRNA and miRNA–lncRNA interactions were downloaded from the starBase (https://starbase.sysu.edu.cn/) and miRcode (http://www.mircode.org/) databases. ¹⁵ , ¹⁶ The miRcode ¹⁵ database facilitates mRNA–miRNA and lncRNA–miRNA target predictions using a broad searchable map that contains 10,419 lncRNAs. The starBase ¹⁶ includes miRNA–mRNA interactions predicted by probing 108 CLIP‐seq data sets. The TF–gene associations were extracted from the Dorothea R package, ¹⁷ covering 1395 TFs and 20,244 genes with 486,676 interactions.

2.2. Differential expression analysis

First, samples with duplicated IDs were removed from the retrieved TCGA data set. Then, metastatic samples were excluded as solid tumours and adjacent normal groups were compared in the differential expression analysis. The low‐expressed genes (log counts per million < 1 in more than 50% of the samples) were excluded to increase the sensitivity and precision of the differential expression analysis. After excluding low‐expressed genes, RNA and miRNA raw counts data were re‐normalized using the trimmed mean of M values method implemented in the edgeR R package. ¹⁸ Then data were standardized using the voom method in the linear modelling for microarrays (limma) R package. ¹⁹ Differentially expressed (DE) mRNAs, lncRNAs and miRNAs were identified using a linear modelling method (lmFit function) followed by empirical Bayes (eBayes) moderation. The eBayes method borrows information across all the genes to obtain more precise estimates of gene‐wise variability. According to linear modelling results, |logFC| > 1 and false discovery rate < 0.01 were considered thresholds to identify statistically significant mRNAs, lncRNAs and miRNAs. ²⁰ The RNA and miRNA expression levels of DE mRNAs, lncRNAs and miRNAs were selected as independent (X)/dependent (Y) variables for machine learning models.

2.3. CNA and DM

We obtained gene‐level copy number values and gene‐level DM values on CpG sites (i.e., β values) using the Linked Omics database (http://www.linkedomics.org/login.php), ²¹ a publicly available resource for multi‐omics data from all 32 cancer types in TCGA. The β value was estimated as the ratio of the methylated probe intensity to the overall intensity (sum of methylated and unmethylated intensities). Therefore, the β value ranges between 0 and 1, with 0 being hypomethylated and 1 being hypermethylated. Linked Omics ²¹ combines COAD and READ copy number values and methylation β values. Therefore, we integrated COAD and READ RNA and miRNA expression samples into a single analysis of colorectal cancer (COLCA). We use the ‘COLCA’ term to explain results in the combined analysis of COAD and READ data.

2.4. Cancerin: a supervised machine learning approach to identify ceRNA networks in cancer risk

We used a supervised machine learning approach, Cancerin, ¹² to identify genome‐wide ceRNAs in cancer risk. The Cancerin pipeline requires DE RNA expression data, gene‐level copy number values, and gene‐level methylation β values from tumour samples. The Cancerin pipeline consists of three main steps as follows:

• 1.Retrieving putative miRNA–mRNAs and miRNA–lncRNA interactions.

  The putative interactions between DE miRNA and DE mRNA (obtained from the differential expression analysis step) were retrieved from the starBase ¹⁶ and miRcode ¹⁵ databases, where DE miRNA and DE lncRNA associations were extracted from the miRcode. ¹⁵

• 2.Selecting miRNAs that contribute to the RNA expression level.
  We used a regularized regression method, the least absolute shrinkage and selection operator (LASSO), to evaluate the effect of miRNA regulators on RNA expression levels. ¹² LASSO regression is a type of linear regression that uses shrinkage where data values are shrunk towards a central point such as the mean. The mathematical equation of LASSO regression is given in Equation 1.

  $$\sum _{i=1}^n{\left(y_i-\sum _j{x}_{\mathit{ij}}{\beta }_j\right)}^2+\lambda \sum _{j=1}^p|{\beta }_j|$$ (1)

  The blue‐coloured part implies the residual sum of squares calculation, λ denotes the amount of shrinkage and the green‐coloured section defines the sum of the absolute value of the magnitude of coefficients. The LASSO models are equivalent to the linear regression models when λ = 0. Increasing/decreasing λ value cause increased biasedness/increased variance, respectively. We have used the 10‐fold cross‐validation approach to find the optimal λ value for each LASSO model.

  The LASSO models are constructed considering transcription‐mediating factors such as DM, TF and CNA. Two model types are fitted for DE mRNAs (see Equation 2) and DE lncRNAs (see Equation 3).

  $$\mathrm{mRNA}\mathrm{expression}~\mathrm{CNA}+\mathrm{DM}+\mathrm{miRNA}\mathrm{expression}+\mathrm{TF}\mathrm{expression}$$ (2)

  $$\mathrm{lncRNA}\mathrm{expression}~\mathrm{miRNA}\mathrm{expression}$$ (3)

  We executed every LASSO model 100 times and non‐zero coefficients more than 75 times were selected as LASSO predictors. Unlike multiple linear regression models, LASSO‐selected independent variables are not associated with any statistical significance test. Therefore, we followed the bootstrap procedure for constructing confidence intervals for frequently selected predictors. The subsequent analysis selected the statistically significant mRNA–miRNA and lncRNA–miRNA pairs from LASSO models.

• 3.Inferring ceRNA associations through the hypergeometric test, Pearson correlation analysis, and the sparse partial correlation analysis.

The Cancerin pipeline introduces three steps to infer ceRNA associations as follows: (i) Pearson correlation analysis to filter strongly positive correlated ceRNA candidate pairs; (ii) hypergeometric test to identify lncRNA–mRNA sharing the significant number of miRNAs; (iii) using the bnlearn R package ²² to calculate sensitivity correlation (scor) for each candidate ceRNA pair. The scor value in Step iii does not account for a combinatorial effect of multiple miRNAs. Subsequently, strong ceRNAs mediated by multiple moderate miRNA regulators cannot be detected. Therefore, we utilized an extension of scor, the multiple scor (mscor) method, which has been implemented in the Sparse Partial correlation ON Gene Expression R/Bioconductor package. ²³ We used three thresholds as follows: the hypergeometric test p < 0.05, the Pearson correlation coefficient of lncRNA–mRNA or mRNA–mRNA pairs > 0.40, and the adjusted p of mscor < 0.05 to infer ceRNA candidates of each HD cancer. We have extensively described these methods in our previous conventional ceRNA network analyses paper. ⁶

Identifying ceRNAs common among multiple HD cancer types may assist in understanding their shared molecular pathogenesis and drug repositioning. Therefore, we evaluated shared ceRNA pairs (lncRNA–mRNA or mRNA–mRNA) among HD cancer types. The statistically significant ceRNA pairs (RNA A–RNA B) in each HD cancer were compiled into a single table. Then, we coded a new variable combining ceRNA candidate pairs, such as ‘RNA A ensemble ID_RNA B ensemble ID.’ Then, the one‐way table was created to identify shared values of the derived variable across all four HD cancer types. Suppose one‐way frequency equals 4 for a given RNA A–RNA B combination. In that case, a ceRNA association is classified into ‘the shared ceRNA network of HD cancers.’ Similarly, if frequency equals 2 or 3, these ceRNAs have been associated with two or three cancer types, respectively.

Shared ceRNAs across multiple HD cancer types were applied in survival analysis, functional enrichment analysis, and PPI network analysis to evaluate their predictive ability and biological functions in cancer. We used survival ²⁴ and clusterprofiler ²⁵ R packages for survival and functional enrichment analyses. The shared ceRNAs among HD cancer combinations were checked using the Kaplan–Meir (K‐M) survival curves in survival analysis. We studied whether a given ceRNA exhibits prognosis ability in individual cancers included in combinations. For example, assume that Gene A is a common ceRNA for BRCA and PRAD, then we check whether Gene A is statistically significant from both BRCA and PRAD survival analyses. The PPI network analyses were conducted using the STRING version 11.5. ²⁶

3. RESULTS

After the quality‐control process, we retrieved 495/52 (cases/controls), 1091/113, 456/41, 166/10, and 543/35 samples from PRAD, BRCA, COAD, READ, and UCEC. In the differential expression analysis, we used 15,509, 15,244, 14,771, 14,866, and 15,197 genes from PRAD, BRCA, COAD, READ, and UCEC after removing those that were low‐expressed.

3.1. Differential expression analysis results

The number of DE RNAs, DE miRNAs, DM, CNA, TF–target interactions and miRNA–lncRNA/mRNA interactions used in each HD cancer‐specific ceRNA analysis are given in Table 1.

Table 1.

The count of DE RNAs and other regulator factors involved in the machine learning‐based ceRNA network analysis

Cancer		RNA	miRNA	DM	CNA	TFs–target interactions	miRNA–lncRNA/mRNA interactions
PRAD		1701	61	19,708	24,079	454,505	7,963,270
BRCA		2934	158	20,106	24,776	454,505	16,608,714
UCEC		3837	245	20,118	24,776	454,505	27,124,806
COLCA
COAD		3187	339	20,113	24,776	454,505	17,449,352
READ		3193	279	20,113	24,776	454,505	18,422,800

Abbreviations: BRCA, breast invasive carcinoma; ceRNA, competing endogenous RNA; CNA, copy number alteration; COAD, colon adenocarcinoma; COLCA, colorectal cancer; DE, differentially expressed; DM, DNA methylation; lncRNA, long noncoding RNA; miRNA, microRNA; mRNA, messenger RNA; PRAD, prostate adenocarcinoma; READ, rectum adenocarcinoma; TFs, transcription factors; UCEC, uterine corpus endometrial carcinoma.

In the next analytical steps (machine‐learning approach and inferring ceRNAs), DE mRNA/miRNA expression data of COAD and READ combined as colorectal cancer (COLCA) analysis.

3.2. Machine‐learning‐based ceRNA network analyses results

In LASSO models, we used DE RNA expression data, DM β estimates, CNA values, miRNA–lncRNA/mRNA and TF–target interactions. The statistically significant LASSO predictors (estimated by the bootstrap procedure) were used to identify potential ceRNA pairs. After that, we applied three user‐defined thresholds (see Section 2 for more details) to infer ceRNA interactions. Table 2 describes the number of significant ceRNA associations found in each HD cancer and shared among more than two cancer types. Please see the Supporting Information file for ceRNAs in individual/shared HD cancer(s).

Table 2.

The number of ceRNAs in individual HD cancer networks/shared among cancer combinations

Cancer or cancer combinations	Count of statistically significant ceRNA combinations (number of unique ceRNAs)
PRAD	1802 (280)
BRCA	9340 (963)
COLCA (COAD $\cup$READ)	2858 (750)
UCEC	3165 (604)
PRAD $\cap$BRCA	37 (35)
PRAD $\cap$COLCA	5 (9)
PRAD $\cap$UCEC	15 (16)
BRCA $\cap$COLCA	70 (72)
BRCA $\cap$UCEC	202 (88)
COLCA $\cap$UCEC	77 (60)
PRAD $\cap$BRCA $\cap$COLCA	0
PRAD $\cap$BRCA $\cap$UCEC	2 (4)
BRCA $\cap$COLCA $\cap$UCEC	9 (16)
PRAD $\cap$COLCA $\cap$UCEC	2 (3)
PRAD $\cap$BRCA $\cap$ COLCA $\cap$UCEC	0

Note: $\cap$ denotes the intersection between given data sets. For instance, PRAD$\cap$BRCA implies the set of shared competing endogenous RNAs among prostate and breast cancers.

Abbreviations: BRCA, breast invasive carcinoma; ceRNA, competing endogenous RNA; COAD, colon adenocarcinoma; COLCA, colorectal cancer; HD, hormone‐dependent; PRAD, prostate adenocarcinoma; READ, rectum adenocarcinoma; UCEC, uterine corpus endometrial carcinoma.

As shown in Table 2, most ceRNAs were unique for individual cancer types. BRCA and UCEC have shared the highest ceRNA candidates compared with other cancers. All significant ceRNAs have been derived from mRNA–mRNA pairs and none of the lncRNA–mRNA pairs was observed across multiple HD cancers.

3.3. Survival analysis

The ceRNAs common among HD cancer combinations were checked for the prognosis in each cancer (included in the combination). Table 3 describes significant ceRNAs from each HD cancer combination with their hazard ratios and p values. The hazard ratio is defined as the slope of the survival curve, which measures how rapidly the event (death) has occurred. The hazard ratio compares two groups, here low and high expressed. The hazard ratios and p values in blue/red fonts indicate that the gene's low/high expressed level is associated with the survival of cancer in interest. In Table 3, the significant ceRNAs from COAD‐READ combined analysis (COLCA) were checked for prognostic in COAD/READ.

Table 3.

Survival significant ceRNAs in the combinations of HD cancers

Note: The hazard ratios and p in blue/red fonts indicate that the gene's low/high expressed level is associated with the survival of cancer in interest.

Abbreviations: BRCA, breast invasive carcinoma; ceRNA, competing endogenous RNA; COAD, colon adenocarcinoma; COLCA, colorectal cancer; HD, hormone‐dependent; PRAD, prostate adenocarcinoma; READ, rectum adenocarcinoma; UCEC, uterine corpus endometrial carcinoma.

As described in Table 3, most survival‐significant ceRNAs have been shared among BRCA and UCEC. Previous studies have also shown the shared genetic susceptibility of BRCA and UCEC. ¹⁴ , ²⁷ Two ceRNAs (BUB1 and EXO1) from the BRCA‐COAD‐UCEC and the RRM2 from PRAD‐COAD‐UCEC were prognostic in all three cancer types included in the combination of interest.

The K‐M survival curves for BUB1 (in BRCA, COAD and UCEC), EXO1 (BRCA, COAD and UCEC), and RRM2 (PRAD, COAD and UCEC) are illustrated in Figure 1.

Figure 1.

Kaplan–Meier survival curves for the shared competing endogenous RNAs (ceRNAs) across three hormone‐dependent (HD) cancers. The BUB1 (top) and EXO1 (middle) genes are prognostic in breast (A), colon (B) and endometrial (C) cancers. The RRM2 (bottom) is survival significant in prostate (A), colon (B) and endometrial (C) cancers.

In BRCA (A) and UCEC (C), low‐expressed BUB1 and EXO1 show better survival. Nevertheless, high‐expressed BUB1 and EXO1 are associated with a higher survival probability in COAD (B). The low‐expressed RRM2 gene has shown the highest survival probability (hazard ratio = 9.13) in PRAD (A) compared with other cancers, COAD (B) and UCEC (C).

3.4. Functional enrichment analysis

We conducted the functional enrichment analysis for shared ceRNAs in each HD cancer combination. The genes shared across two/three HD cancers (listed in Supporting Information) was used for the functional enrichment analysis. Table 4 reports the top‐five components of Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways for each two‐HD cancer combination‐associated ceRNAs. Please see the Supporting Information file for list of genes in each GO/KEGG pathway.

Table 4.

Functional enrichment analyses result for shared ceRNAs across two HD cancers combinations

Cancer combination	Terms (GO/KEGG pathway)	FDR
PRAD∩BRCA	Mitotic sister chromatid segregation (GO)	1.44E − 19
	Mitotic nuclear division (GO)	1.44E − 19
	Chromosome segregation (GO)	7.54E − 19
	Organelle fission (GO)	8.46E − 18
	Regulation of chromosome segregation (GO)	4.41E − 15
	Cell cycle (KEGG)	2.26E − 05
	Oocyte meiosis (KEGG)	0.008609
PRAD∩COLCA	Condensed chromosome (GO)	0.001566
	Mitotic chromosome condensation (GO)	0.002994
	Mitotic sister chromatid segregation (GO)	0.002994
	Condensed nuclear chromosome (GO)	0.004676
	DNA packaging complex (GO)	0.004676
PRAD∩UCEC	Chromosome segregation (GO)	2.38E − 18
	Sister chromatid segregation (GO)	7.4E − 17
	Nuclear division (GO)	2.19E − 15
	Mitotic nuclear division (GO)	2.21E − 15
	Regulation of chromosome segregation (GO)	4.29E − 13
	Cell cycle (KEGG)	0.042667
BRCA∩COLCA	Nuclear division (GO)	4.61E − 20
	Chromosome segregation (GO)	5.05E − 19
	Mitotic nuclear division (GO)	5.54E − 19
	Mitotic sister chromatid segregation (GO)	3.59E − 16
	Chromosomal region (GO)	3.38E − 12
	Cell cycle (KEGG)	1.97E − 12
	Oocyte meiosis (KEGG)	2.05E − 05
	Progesterone‐mediated oocyte maturation (KEGG)	5.45E − 05
	p53 signalling pathway (KEGG)	0.000205
	Cellular senescence (KEGG)	0.00052
BRCA∩UCEC	Chromosome segregation (GO)	2.41E − 27
	Chromosomal region (GO)	5.22E − 26
	DNA replication (GO)	1.5E − 22
	Organelle fission (GO)	8.05E − 22
	Nuclear division (GO)	8.05E − 22
	Cell cycle (KEGG)	3.06E − 14
	DNA replication (KEGG)	7.94E − 06
	Mismatch repair (KEGG)	0.001704
	p53 signalling pathway (KEGG)	0.002887
	Cellular senescence (KEGG)	0.004181
COLCA∩UCEC	Chromosome segregation (GO)	8.31E − 18
	Chromosomal region (GO)	1.12E − 16
	Mitotic nuclear division (GO)	2.86E − 16
	Nuclear division (GO)	5.24E − 16
	DNA replication (GO)	1.42E − 15
	Cell cycle (KEGG)	2.88E − 10
	DNA replication (KEGG)	0.000148
	p53 signalling pathway (KEGG)	0.001661
	Fanconi anaemia pathway (KEGG)	0.007581
	Oocyte meiosis (KEGG)	0.007581

Abbreviations: BRCA, breast invasive carcinoma; ceRNAs, competing endogenous RNAs; COAD, colon adenocarcinoma; COLCA, colorectal cancer; FDR, false discovery rate; GO, Gene Ontology; HD, hormone‐dependent; KEGG, Kyoto Encyclopedia of Genes and Genomes; PRAD, prostate adenocarcinoma; READ, rectum adenocarcinoma; UCEC, uterine corpus endometrial carcinoma.

Enrichment analyses result of ceRNAs shared among three‐HD cancers are illustrated in Figure 2.

Figure 2.

Top‐five Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genome (KEGG) pathway components for shared ceRNAs across prostate‐breast‐endometrial (A), breast‐colorectal‐endometrial (C), and prostate‐colorectal‐endometrial (C).

3.5. PPI network analysis results

PPI network analyses for ceRNAs shared among HD cancer combinations were conducted using the STRING online platform. ²⁶ The minimum required interaction score of the PPI networks was adjusted to the highest confidence level (0.900) to improve the strength of the interaction network. Table 5 describes the PPI network analyses results.

Table 5.

PPI network analysis result for each ceRNAs set shared among HD cancer combinations

Cancer combinations	Number of nodes	Number of edges	PPI enrichment p
PRAD∩BRCA	35	141	<1.0E − 16
PRAD∩COLCA	9	7	6.68E − 10
PRAD∩UCEC	16	67	<1.0E − 16
BRCA∩COLCA	72	193	<1.0E − 16
BRCA∩UCEC	88	332	<1.0E − 16
COLCA∩UCEC	60	150	<1.0E − 16
PRAD∩BRCA∩UCEC	4	6	1.12E − 09
BRCA∩COLCA∩UCEC	16	10	2.04E − 09
PRAD∩COLCA∩UCEC	3	1	0.0354

Abbreviations: BRCA, breast invasive carcinoma; ceRNAs, competing endogenous RNAs; COLCA, colorectal cancer; HD, hormone‐dependent; PPI, protein–protein interaction; PRAD, prostate adenocarcinoma; UCEC, uterine corpus endometrial carcinoma.

4. DISCUSSION

The aetiology of complex diseases such as cancer is often related to aberrant gene expression at the transcriptional and posttranscriptional levels. Over the last decade, ceRNAs have emerged as an important class of posttranscriptional regulators that alter gene expression through a miRNA‐mediated mechanism. This study has extended the knowledge of ceRNA networks in HD cancers (PRAD, BRCA, COLCA and UCEC) by integrating other molecular effects such as DM, TF and CNA, which influence the ceRNA mechanism. Moreover, it expands the knowledge of HD cancers shared biology by investigating and comparing gene regulatory networks of more than two HD cancer types.

First, we investigated regulating factors‐mediated ceRNAs in individual HD cancers. After that, we identified ceRNAs shared across two/three/four HD cancers. However, none of the ceRNAs was shared among the four cancers of interest. The shared ceRNAs across two/three HD cancers were involved in survival analysis and two downstream analyses, functional enrichment analysis and PPI network analysis.

We found 16 ceRNAs shared across BRCA, COLCA, and UCEC. Among them, BUB1 and EXO1 were associated with patient survival of all three cancers of interest. BUB1 gene encodes a serine/threonine‐protein kinase that plays a central role in mitosis. It also involves DNA damage response and cell cycle regulation. ²⁸ The BUB1 gene has been previously reported in BRCA, COLCA, and multiple digestive tract cancers but not in UCEC. ²⁹ Herein, BUB1 appears for the first time in UCEC risk. The BUB1 β (BUB1B) was recognized as a shared ceRNA among PRAD, BRCA and UCEC. These BUB1 and BUB1B genes are central components of the mitotic checkpoint for spindle assembly (SAC). ²⁹ The SAC plays a fundamental role in maintaining genome stability by ensuring the timely segregation of the genetic material at every cell division. Mistakes in the cell division process can lead to rearrangement, loss or gain of chromosomes, genome instability and cancers. ³⁰ Therefore, the molecular function of BUB1 and BUB1B genes in HD cancers should be further examined. Several experimental and bioinformatic studies have shown the importance of EXO1 in replication, DNA repair pathways, cell cycle checkpoints, and its association to cancer. ³¹ Genome‐level studies have identified specific mutations in the EXO1 gene as risk alleles for different types of cancers. ³² , ³³

Three genes, DEPDC1B, BUB1 and RRM2, were identified as shared ceRNAs among PRAD, COLCA and UCEC. DEPDC1B has been identified as a prostate cancer metastasis oncogene. It was positively correlated with metastasis status, high Gleason score, advanced tumour stage and poor prognosis. Moreover, mechanistic investigations found that DEPDC1B induced epithelial–mesenchymal transition (EMT) and enhanced proliferation by binding to Rac1 and enhancing the Rac1‐PAK1 pathway. ³⁴ EMT plays an important role in empowering cancer cells to adapt and survive at the start of the metastatic stage. The role of DEPDC1B in COLCA and UCEC has not been described in previous bioinformatics or functional studies. According to the survival analysis result, the RRM2 gene was a prognostic biomarker in PRAD, COLCA and UCEC. The well‐established role of RRM2 is to maintain the balance of deoxynucleoside triphosphate pools for DNA synthesis and DNA repair. ³⁵ The tumour‐promoting feature of RRM2 is associated with inducing activities of various oncogenes, including those encoding nuclear factor‐κB, Myc proto‐oncogene protein, tyrosine‐protein kinase transforming protein Fes and ornithine decarboxylase. ³⁶ The overexpression of RRM2 correlates with cellular invasiveness, metastasis and tumourigenesis of cervical ³⁷ and gastric ³⁸ cancers. Aberrant RRM2 expression has also been described in ovarian ³⁹ and prostate ⁴⁰ cancers. Liu et al. ⁴¹ have described that RRM2 serves as a prognostic biomarker in colorectal cancer. None of the ceRNAs was shared among PRAD, BRCA and COLCA, and hence common ceRNAs were not found across four HD cancer types (PRAD, BRCA, COLCA and UCEC) included in this study.

The survival analyses of shared ceRNAs among two HD cancer combinations have shown that 12 (KIF4A, KPNA2, TPX2, TUBA1C, RAD54L, MTFR2, ANLN, RACGAP1, FAM83D, KNSTRN, CCNE1 and DSCC1) out 88 and 6 (KPNA2, CCNF, CKAP2L, TTK, DEPDC1B and MCM2) out of 77 ceRNAs shared among BRCA‐UCEC and COLCA‐UCEC pairs, respectively, have exhibited the survival significance in both cancers included in the combinations. These 18 genes have been previously described in HD studies due to their tumour‐suppressive/oncogenic/cancer‐driven nature. ⁴²

We conducted a functional enrichment analysis for common ceRNAs across two/three HD cancers combinations. The shared ceRNAs across HD cancers have mainly been involved in mitosis (mitotic sister chromatid segregation, mitotic nuclear division and mitotic chromosome condensation) and chromosome related GO pathways (chromosome segregation, regulation of chromosome segregation, condensed chromosome, condensed nuclear chromosome and chromosomal region), which play an important role in cancer. The shared ceRNAs across PRAD, COLCA and UCEC were involved in five unique KEGG pathways, drug metabolism‐other enzymes, pyrimidine metabolism, glutathione metabolism, oxidoreductase activity, acting on CH or CH2 groups, and ferric iron‐binding. Experimental studies might be required to investigate these KEGG pathways' involvement in the shared biology of PRAD, COLCA and UCEC.

The PPI network describes the physical interaction between identified shared ceRNAs in this study, mediating the assembly of proteins into protein complexes, for example, mediating signalling/regulation and transport events in the cell. According to PPI network analysis results, ceRNAs lists of each two/three HD cancer combinations have interacted with stronger statistical evidence. Five out of nine HD cancer combinations' PPI network enrichment p was <1.0E − 16. Four and 16 ceRNAs shared among PRAD‐BRCA‐UCEC (p = 1.12E − 09) and BRCA‐COLCA‐UCEC (p = 2.04E − 09) exhibited strong PPI network connecting all the ceRNA nodes. Therefore, these gene sets should be further studied to evaluate their combined effect on HD cancers.

The major limitation of this study is restricting miRNA–mRNA and miRNA–lncRNA selection (in Step 1). We have chosen both predicted and experimentally validated miRNA–target interaction only from two miRNA–mRNA/lncRNA databases, starBase and miRcode. We followed this limitation to get a substantial set of miRNA–mRNA/lncRNA interactions. Consideration of more possible ceRNA associations will improve the machine‐learning outcome by increasing the number of predictors. Moreover, we have not included circRNAs and pseudogenes in the ceRNA network analyses. This study retrieved only a limited number of statistically significant miRNAs from the differential expression analysis of miRNA‐seq data in TCGA‐PRAD. Therefore, the number of miRNAs (independent variables) involved in PRAD‐LASSO models was comparatively lower. Nevertheless, selections of TFs‐target interactions and CNAs were independent of PRAD differential expression analyses.

Our previous bioinformatic study identified shared ceRNAs (both lncRNAs and mRNAs) across PRAD, BRCA, COLCA and UCEC, considering miRNAs as the only gene regulator in the ceRNA network. ⁶ This study has integrated conventional ceRNA networks with DM, TF and CNA data using a machine‐learning approach, ‘Cancerin.’ According to our results, DM, TF and CNA significantly have cut‐off shared ceRNAs across HD cancers found from the conventional methods. According to LASSO models, none of the lncRNAs was identified as a shared ceRNA among two/three HD cancers of interest. These results imply that other regulating factors such as DM, TF and CNA play a vital role in the ceRNA regulatory network apart from miRNAs. Therefore, existing statistical and computational methods should be improved to interpret these gene‐regulating factors‐mediated ceRNAs in diseases of interest.

Our machine learning‐based study of omics data provides a set of pivotal ceRNAs associated with two or three HD cancers of interest. According to functional enrichment analyses, shared ceRNAs across two/three HD cancers are involved in hallmarks of cancer, such as cell cycle, DNA replication/repair, chromosomal segregation, mitotic checkpoint, nuclear division, p53 signalling, glutathione metabolism, pyrimidine metabolism and ferric iron binding. Most of these pathways are directly/indirectly associated with RNA‐processing steps that influence RNA regulatory networks, including ceRNAs. A recent study found the complex biology of glutamine (precursor of glutathione) metabolism in driving BRCA growth. ⁴³ We found glutathione metabolism as a shared pathway among multiple HD cancers. Therefore, future functional studies are required to investigate the biology of glutathione in HD cancers. Dysfunction of pyrimidine metabolism is closely related to cancer progression. Moreover, various drugs targeting pyrimidine metabolism have been approved for multiple cancer types, including BRCA. ⁴⁴ This study has introduced bioinformatics analyses into the mechanistic understanding of HD cancers. A natural progression of this study is to identify the link between shared ceRNAs and respective molecular pathways. Nevertheless, these findings will bring novel insights to explore the underlying shared molecular mechanism of HD cancers.

5. CONCLUSIONS

It is noteworthy that, to our knowledge, this is the first study that integrates genomic (CNA), transcriptomic (mRNA, lncRNA, miRNA and TF) and epigenetic (DM) regulatory factors to infer genome‐wide ceRNA interactions shared across a group of cancers. These cross‐HD cancer ceRNAs have shown tumour‐suppressive/tumour‐promoting and survival‐prognosis properties in previous HD cancer studies. The main bottleneck of this machine learning approach is the lack of RNA, DM and CNA data from a single study, especially for other diseases, except for cancers. Besides these limitations, cross‐cancer shared ceRNAs may help identify potential unexpected targets applied for a subset of cancers (e.g., HD). Furthermore, it will help to understand the shared disease biology of HD cancers.

AUTHOR CONTRIBUTIONS

Dulari K. Jayarathna, Neha S. Gandhi, Miguel E. Rentería and Jyotsna Batra designed the study. Dulari K. Jayarathna performed data acquisition, data analysis and data interpretation. Dulari K. Jayarathna wrote the original manuscript. Neha S. Gandhi, Miguel E. Rentería and Jyotsna Batra revised the manuscript critically. All authors read and approved the final manuscript.

CONFLICT OF INTEREST

The authors declare no conflict of interest.

ETHICS STATEMENT

This study was performed in line with the principles of the Declaration of Helsinki. Approval was granted by the Ethics Committees of Queensland University of Technology (19 December 2019/1900001147) and QIMR Berghofer Medical Research Institute (23 August 2019/P1051).

Supporting information

Supplementary information.

Jayarathna DK, Rentería ME, Batra J, Gandhi NS. A supervised machine learning approach identifies gene‐regulating factor‐mediated competing endogenous RNA networks in hormone‐dependent cancers. J Cell Biochem. 2022;123:1394‐1408. 10.1002/jcb.30300

DATA AVAILABILITY STATEMENT

The author confirms that data from the public database for this study have explained the source of the data in detail in the manuscript and provided a link address.