This article is mentioned in:

Introduction

Hypertrophic cardiomyopathy (HCM) is a complex and relatively common genetic heart disease, which is characterized by unexplained asymmetric or symmetric cardiac hypertrophy, interstitial fibrosis and cardiomyocyte disarray (1). HCM is a prevalent disease that affects 0.2% of the global population (2). Patients with HCM may suffer from early sudden cardiac death which is most common in individuals <35 years of age (3,4). Therefore, it is necessary to identify effective therapeutic strategies and investigate the etiology of HCM.

The development of large scale microarray analyses has led to research at the gene level, which may be a powerful and informative means of investigating the mechanism of a disease (5). HCM is predominantly induced by mutations in genes encoding sarcomere proteins, including myosin heavy chain 7, myosin light chain 2 and myosin binding protein C, cardiac (6,7). HCM-associated genes are typically selected by analyzing differentially expressed genes (DEGs). However, a previous study identified inconsistencies between numerous gene signatures across different studies of the same disease (8). A potentially more effective method is to use pathway analysis in order to evaluate disease-associated biomarkers.

Pathway analysis has been widely used in previous studies (9–11) and is becoming the primary method of obtaining a deep insight into biological processes (12). The identification of active pathways that differ between two conditions is of increased utility compared with a list of DEGs (10). However, existing methods primarily utilize pathway repositories, including the Kyoto Encyclopedia of Genes and Genomes (13), to evaluate whether a pathway is significant by assigning a P-value to the pathway, an approach which focuses on the static regulation between genes (14,15). Analysis of dynamic regulation or network rewiring may reveal important information which may not be identified in static conditions (16). Interactions between molecules may alter in different tissues and these alterations are associated with disease progression. Therefore, novel pathway-based biomarkers may be identified by studying dynamic regulation and network rewiring among molecules, in contrast with investigations which only examine differential expression.

In the present study, a novel method was developed to identify pathway-based biomarkers in HCM, by investigating interactions between molecules associated with pathogenesis through a co-expression network strategy. HCM-associated microarray expression data were downloaded from the European Molecular Biology Laboratory-European Bioinformatics Institute (EMBL-EBI) database. The co-expression network was constructed using an empirical Bayes (EB) approach following assignment of a weight value for each pathway. A random model was constructed to define the thresholds to select differential pathways. Pathway enrichment analysis of DEGs, based on the Database for Annotation, Visualization and Integrated Discovery (DAVID) test, was implemented to verify the feasibility of the novel method.

Materials and methods

Data collection and preprocessing

The HCM-associated microarray expression data, under accession number E-GEOD-36961, were downloaded from the EMBL-EBI database (http://www.ebi.ac.uk/arrayexpress/). In gene expression profile E-GEOD-36961, there were 39 healthy subjects and 106 HCM samples. In order to assess the quality of the data, standard preprocessing of microarray probe data was performed. This preprocessing included the following successive steps: Background correction using robust multiarray analysis (17); normalization using the quartile function to eliminate the influence of nonspecific hybridization (18); perfect match correction using the MAS5.0 algorithm (19); and calculation of expression values from probe intensities. The expression value for every gene was acquired, including 37,846 genes from 145 samples (39 healthy and 106 HCM).

Pathway data

Reactome is a database which includes numerous pathways in Homo sapiens and a number of reference species, offering an infrastructure for computation across the biological reaction network (20). Human biological pathways were downloaded from the Reactome pathway database (reactome.org). A total of 1,675 pathways were obtained. In order to ensure the validity of the pathways, pathways in which the gene number was ≤2 were discarded. Finally, a total of 1,639 pathways were determined as background pathways.

Construction of a co-expression network using an EB approach

Numerous methods have been demonstrated for co-expression analysis to identify co-expression gene pairs, which include EB (21), Arabidopsis Co-expression Tool (22) and WGCNA (23). In the present study, an EB framework was implemented, which offered a false discovery rate (FDR)-controlled list of the gene pairs of interest (21).

The mean gene number (N) for the 1,639 background pathways was computed based on N=overall gene number for background pathways (73,099)/the number of background pathways (1,639); therefore, N=44.6. In order to facilitate the analysis, the mean gene number was subsequently defined as N=44.

The gene number of each pathway was termed A, and the number of intersections between each pathway and the gene expression profile was termed B. The informative pathways used in the present study were screened out according to the following conditions: B>5; and the ratio B:A>0.9. Therefore, a total of 1,074 informative pathways were selected for further investigation.

Subsequently, the EB approach developed by Dawson and Kendziorski (24) was used to identify co-expression gene pairs among the genes in each informative pathway and construct a co-expression network. The number of possible gene pairs of each informative pathway was termed C [C=A × (A-1)/2]. In the EB method, the identification of co-expression gene pairs was implemented based on the following steps: Three inputs of matrix X; the conditions array; and the pattern object required. The expression levels in an m × n matrix of X, where m=the number of genes under any informative pathway and n=the total number of representative pathways, were normalized. Subsequently, the members of the conditions array with length n took values in 1, 2, …… K, where K indicated the number of conditions. Based on matrix X and the conditions array, intra-group correlations for all gene pairs were computed, and M matrix of correlations was obtained. The Mclust algorithm (25) was used to initialize the hyper-parameters via the initializeHP function of the EBcoexpress package (26), to discover the component normal mixture model which best fitted the correlations of M. The crit.fun function of EBcoexpress package (26), was used to define a threshold by controlling the posterior probabilities of co-expression to extract particular co-expression gene pairs. Gene pairs with FDR ≤0.05 were chosen to construct the co-expression network. The number of interactions in one pathway co-expression network was termed D, and D/C was recorded as the weight value of the pathway.

Identification of differential pathways

In an attempt to identify differential pathways between HCM and normal samples, a random model consisting of G genes was constructed. G genes were randomly selected from gene expression profiles and the weight values for each pathway were calculated using the EB approach. This step was repeated 10,000 times and the weight values were listed in descending order. The FDR for the 100th weight value was set at 0.01 (weight value=0.01057). The pathways with a weight value >0.01057 were considered to be differential pathways.

Identification of DEGs

As previously demonstrated, differential gene expression levels are associated with disease severity. In the current study, the identification of DEGs between HCM and normal samples was performed using the significance analysis of microarrays (SAM) package. The samr function of the SAM package was used to extract the genes which exhibited statistically significant differential expression. Each gene was assigned a score according to the difference in gene expression compared with the standard deviation of repeated measurements for this gene. Genes with scores above the threshold were considered to be potentially significant. The percentage of falsely significant genes relative to the significant genes was defined as the FDR. In order to increase the stringency for differential gene expression, the delta value was determined using the function SAMR.compute.delta.table. DEGs in HCM were identified for further analysis with the delta threshold value of 7.48.

Pathway analyses for DEGs

In order to further investigate the biological functions of DEGs, Reactome pathway enrichment analysis was conducted through DAVID (27) by means of the Expression Analysis Systematic Explorer (EASE) test (28). The EASE score was applied to identify the significant categories. In the present study, the pathways with P<0.05 were considered to be significant pathways.

Results

Identification of differential pathways

Following pre-processing, a total of 37,846 genes were obtained and used for subsequent analysis. A total of 1,074 informative pathways were identified, based on the Reactome pathway database and filtration treatment. For each informative pathway, a weight value was calculated based on the EB co-expression network. The distribution of weight values of the pathways are presented in Fig. 1; it was observed that the majority of informative pathways exhibited weight values between 0 and 0.04, particularly between 0 and 0.01. An increased weight value indicates a more significant pathway. Therefore, the threshold criteria for differential pathways was defined. In the present study, the threshold for weight values was calculated to be 0.01057, and 447 differential pathways were selected according to the threshold. The heat-map between differential pathways and the corresponding weight values is presented in Fig. 2. The top 20 differential pathways are presented in Table I. Among these, the top 5 differential pathways were: Folding of actin by chaperonin containing T-complex protein 1 (CCT)/T-complex protein 1 ring complex (TRiC) (weight value=0.88889); purine ribonucleoside monophosphate biosynthesis (weight value=0.75556); ubiquinol biosynthesis (weight value=0.42857); synthesis of 5-eicosatetraenoic acids (weight value=0.33333); and cooperation of prefoldin and TRiC/CCT in actin and tubulin folding (weight value=0.31000). In order to further elucidate each pathway, networks were constructed using the EB method. As the gene count varied between differential pathways, and too few genes are unable to form a network, differential pathways with an increased number of genes were chosen, including cooperation of prefoldin and TRiC/CCT in actin and tubulin folding (gene number=25), and nonsense mediated decay (NMD) independent of the exon junction complex (EJC) (gene number=88). The co-expression networks of the pathways of cooperation of prefoldin and TRiC/CCT in actin and tubulin folding, and NMD independent of the EJC, are exhibited in Figs. 3 and 4, respectively. For the network of cooperation of prefoldin and TRiC/CCT in actin and tubulin folding, there were 17 genes, while there were 72 genes in the network of NMD independent of the EJC.

Figure 1. Distribution of weight values of each informative pathway.

Figure 2. Heat map between differential pathways and their weight values.

Figure 3. Co-expression network for genes in the differential pathway cooperation of prefoldin and T-complex protein 1 ring complex/T-complex protein 1 in actin and tubulin folding. Nodes, genes; edges, interactions.

Figure 4. Co-expression network for genes in the differential pathway nonsense mediated decay independent of the exon junction complex. Nodes, genes; edges, interactions.

Table I. Top 20 differential pathways identified by the empirical Bayesian analysis.

Table I.

Top 20 differential pathways identified by the empirical Bayesian analysis.

Row	Weight values	Differential pathways
1	0.88889	Folding of actin by CCT/TRiC
2	0.75556	Purine ribonucleoside monophosphate biosynthesis
3	0.42857	Ubiquinol biosynthesis
4	0.33333	Synthesis of 5-eicosatetraenoic acids
5	0.31000	Cooperation of prefoldin and TRiC/CCT in actin and tubulin folding
6	0.30072	Prefoldin mediated transfer of substrate to CCT/TRiC
7	0.28571	Defective holocarboxylase synthetase causes multiple carboxylase deficiency
8	0.28571	Defects in biotin metabolism
9	0.28571	Mitochondrial fatty acid beta-oxidation
10	0.27941	Branched-chain amino acid catabolism
11	0.25974	Cytosolic tRNA aminoacylation
12	0.22807	Citric acid cycle
13	0.22222	Role of Abl in Robo-Slit signaling
14	0.20000	Uptake and function of diphtheria toxin
15	0.19044	Nonsense mediated decay independent of the exon junction complex
16	0.17778	Zinc influx into cells by the SLC39 gene family
17	0.17191	Eukaryotic translation termination
18	0.17177	GTP hydrolysis and joining of the 60S ribosomal subunit
19	0.17114	Cap-dependent translation initiation
20	0.17114	Eukaryotic translation initiation

[i] CCT, T-complex protein; TRiC, T-complex protein 1 ring complex.

Pathway analyses for DEGs using traditional DAVID software

Based on the delta cut-off value of 7.48, a total of 344 DEGs were identified. Pathway analyses indicated that there were only 2 significant pathways based with P<0.05, including signaling in immune system and hemostasis (data not shown).

Comparison of EB and DAVID

In order to assess whether the pathway co-expression network method was feasible, the method developed in the present study was compared with traditional DAVID software. It was observed that the differential pathway of hemostasis was the common pathway obtained from the co-expression network approach and the DAVID method. However, the pathway quantity obtained from co-expression network approach was increased compared with the DAVID method; A total of 447 differential pathways were selected using co-expression network method, while only 2 significant pathways were identified using the traditional DAVID). Therefore, the present study demonstrated that the co-expression network method exhibited increased efficiency compared with DAVID (data not shown).

Discussion

In the present study, a novel method of identifying pathway-based biomarkers in HCM was developed, by investigating dynamic interactions between molecules associated with pathogenesis through a co-expression network strategy. The results of the present study identified 447 differential pathways between HCM and normal samples, including folding of actin by CCT/TRiC, purine ribonucleoside monophosphate biosynthesis, and cooperation of prefoldin and TRiC/CCT in actin and tubulin folding.

Actin, as a ubiquitous protein, serves functions in numerous cellular processes, including the maintenance of cell motility, cell shape, mitosis and intracellular transport processes (29,30). Notably, the conserved nature of the amino acid sequence of actin indicates that mutated residues may affect basic functions, including actomyosin interactions and actin-actin interactions involved in polymerization (31). In previous studies, seven recognized actin mutations have been demonstrated to be associated with HCM (30,32). Alterations may be induced by environmental factors, including diet, exercise, or by the cellular protein folding machinery. Previous studies have suggested that chaperone complexes (TRiC and CCT) may assist the folding of actin (33,34). Vang et al (35) demonstrated that protein-folding pathways serve a role in disease progression for actin mutations associated with HCM. Consistent with previous studies, the differentially expressed pathway of folding of actin by CCT/TRiC was identified in the present study. Therefore, it may be inferred that the disturbance of actin folding may be the molecular basis for the subsequent initiation of the hypertrophic pathway, resulting in the occurrence of HCM.

An additional altered pathway, purine ribonucleoside monophosphate biosynthesis, was screened out in the present study. As demonstrated in a recent study, energy flow is generated from ribose in purine ribonucleoside monophosphates (36). The inability to maintain normal ATP utilization may be the primary abnormality in HCM (37,38). Therefore, the differential pathway of purine ribonucleoside monophosphate biosynthesis may be important in the progression of HCM, which may involve a response to the disruption of energy homeostasis.

In conclusion, based on a co-expression network, differential pathways, including folding of actin by CCT/TRiC and purine ribonucleoside monophosphate biosynthesis, were successfully identified and these pathways may be involved in the pathogenic process of HCM. The results of the present study may be applied clinically for the diagnosis, prognostic management and treatment of patients with HCM.

References