A six‑gene support vector machine classifier contributes to the diagnosis of pediatric septic shock

Introduction

Septic shock (SS) is caused by an uncontrolled inflammatory immune response to pathogens (for example, bacteria, fungi, parasites and viruses) and occurs when sepsis results in dangerous hypotension and abnormal cellular metabolism (1). SS can lead to multiple organ failure and death, and in 2018 the mortality rate of SS was 25–50% (2). Innate and adaptive immune responses of hosts of different ages have various effects against sepsis (3,4), with the survival rate of pediatric septic shock (PSS) being particularly low compared with other age groups (5). PSS is one of the main causes of mortality in critically ill children in pediatric intensive care units worldwide (6). Adult and pediatric septic shock differ in terms of pathophysiology, clinical presentation and therapeutic approaches (7). Hyperdynamic shock syndrome or warm shock occurs in ~90% of adult patients, while severe hypovolemia often characterizes PSS (7). Activated protein C is the preferred treatment for adult patients, whereas plasma exchange is preferred for the treatment of pediatric patients with thrombocytopenia-associated multiple organ failure (7). The therapy regimens widely used in adults with SS have been shown to have little therapeutic effect on children (8). Therefore, investigation into the molecular mechanisms of PSS for the diagnosis and therapeutic management of patients with PSS is required.

Nuclear factor erythroid 2-related factor 2 (Nrf2)-linked genes are dysregulated in PSS and may affect fatty acid metabolism, peroxisome proliferator-activated receptors (PPARs) and retinoic acid receptor-α families, which are related to intermediary metabolism and oxidative stress in PSS (9). A recent study revealed a list of differentially expressed genes (DEGs) that are important for SS diagnosis and are implicated in the immune response, chemokine-mediated signaling, neutrophil chemotaxis and chemokine activity (10). Olfactomedin-4 controls sepsis heterogeneity and may be a biomarker of a pathogenic neutrophil subset associated with organ failure and mortality caused by PSS (11). Moreover, there is evidence that serum propionic acid has diagnostic value for septic shock (12). Despite the aforementioned findings, the mechanisms contributing to PSS are largely unknown.

Support vector machine (SVM) classifiers are gaining significance as a robust classification tool in cancer genomics (13) and have been used for the diagnosis of various diseases, including chronic kidney disease (14) and acute coronary syndrome (15). In the present study, it was hypothesized that an SVM classifier based on optimal feature genes of PSS could facilitate the diagnosis of the disease. The present study aimed to accurately identify the key genes in PSS from DEGs between PSS and control samples, and constructed a SVM classifier for distinguishing patients with PSS from normal controls. The results from the present study could aid the development of appropriate treatment strategies for PSS.

Materials and methods

Data source and pre-processing

Using the Gene Expression Omnibus (GEO; www.ncbi.nlm.nih.gov/geo) database, ‘septic shock’ and ‘pediatric’ were used as keywords to search for relevant datasets. The criteria for eligible datasets were as follows: i) Included gene expression data of blood samples; ii) age information was available to ensure that the subjects were children; iii) there were both PSS samples and normal control samples; and iv) the total number of samples was ≥100 and the number of samples that could be used for analysis was ≥50. Finally, four microarray datasets based on the GPL570 (HG-U133_Plus_2) Affymetrix Human Genome U133 Plus 2.0 Array, including GSE26378 (9), GSE26440 (16), GSE13904 (17) and GSE4607 (18), were selected (Table I). The GSE13904 and GSE4607 datasets contained other samples, therefore, only the control and septic shock samples were extracted for analysis in the present study.

Table I. Information of the four microarray datasets.

Table I.

Information of the four microarray datasets.

Dataset	Platform	Samples (n)	Control	Septic shock
GSE26378	GPL570 (HG-U133_Plus_2) Affymetrix Human Genome U133 Plus 2.0 Array	103	21	82
GSE26440	GPL570 (HG-U133_Plus_2) Affymetrix Human Genome U133 Plus 2.0 Array	130	32	98
GSE13904a	GPL570 (HG-U133_Plus_2) Affymetrix Human Genome U133 Plus 2.0 Array	227	18	67
GSE4607a	GPL570 (HG-U133_Plus_2) Affymetrix Human Genome U133 Plus 2.0 Array	123	15	42

a These datasets contained other samples; only the control and septic shock samples were extracted from these datasets for analysis in the present study.

For pre-processing the raw data in the four datasets, data formats were converted using the oligo package (19) (version 1.40.2; www.bioconductor.org/packages/oligo.html) in R and normalization of gene expression values was performed using the unit-scale normalization algorithm (20).

Meta-analysis to identify the consistent DEGs in the four datasets

The four datasets included in the present study involved experimental tests of samples from different patients and thus they may have displayed different degrees of bias. To resolve this issue, the MetaQC package (21) (version 0.1.13; www.cran.r-project.org/web/packages/MetaQC/index.html) in R was adopted to perform quality control on the datasets. The quality control standards in the MetaQC package contained accuracy quality control (AQC) g and AQCp, external quality control, consistency quality control (CQC) g and CQCp and internal quality control. Combined with the two-dimensional diagram of the principal component analysis (PCA) and the standardized mean rank, these datasets were further assessed and screened.

Following the quality control analysis, DEGs between PSS and normal control samples were analyzed using the MetaDE.ES method in the MetaDE package (22) (version 1.0.5; www.cran.r-project.org/web/packages/MetaDE). To ensure that gene expression was consistent across the four datasets, homogeneity test parameters were set as tau2=0 and Qpval >0.05. A false discovery rate (FDR)<0.05 was selected as the significance threshold for screening of the DEGs. Moreover, the genes with log2 fold-change (FC)>0.5 in at least one of the datasets were used for analysis.

Weighted gene co-expression network analysis (WGCNA)

WGCNA is an algorithm based on high-throughput expression data, which is utilized for the construction of a co-expression network (23). In WGCNA, the GSE26440 dataset, which contained a relatively large number of samples, was taken as the main analysis dataset and the other three datasets were considered as the secondary analysis datasets. Using the WGCNA package (23) (version 1.61; www.cran.r-project.org/web/packages/WGCNA/) in R, all genes in the GSE26440 dataset were analyzed and screened for disease-associated modules and genes. The requirements were: Number of module genes ≥80, cutHeight=0.995 and P<0.05. Combined with the three secondary analysis datasets, significantly stable modules across the four datasets were screened. Module preservation across the four datasets was analyzed using the module preservation function of the WGCNA package. Using clinical information of the samples in the GSE26440 dataset, the correlations between each significantly stable module and clinical information were calculated using the WGCNA cor function (version 1.68; 127.0.0.1:13239/library/WGCNA/html/cor.html) and WGCNA corPvalueStudent function (version 1.68; 127.0.0.1:13239/library/WGCNA/html/corPvalueStudent.html).

The consistent DEGs were mapped into the significantly stable modules. Significant enrichment parameters of target genes in the modules were calculated using the hypergeometric algorithm (24): f(k,N,M,n)=C(k,M)xC(n-k,N-M)/C(n,N), where N stands for the number of total genes participating in WGCNA network analysis, M stands for the number of genes included in each module, n stands for the number of genes identified by the MetaDE method and k stands for the number of DEGs mapped to a module.

The thresholds of significant enrichment distribution were selected as P<0.05 and fold enrichment >1. For the consistent DEGs included in the significantly stable modules, Gene Ontology (GO; www.geneontology.org) and Kyoto Encyclopedia of Genes and Genomes (KEGG; www.genome.jp/kegg) enrichment analyses were conducted based on the Database for Annotation, Visualization and Integrated Discovery tool (25) (version 6.8; ww.david.ncifcrf.gov). P<0.05 was selected as the screening threshold.

Selection of the optimal gene combination and construction of the SVM classifier

In order to further narrow down the range of SS-related genes and accurately identify important feature genes, the GSE26440 dataset was used as the training dataset and the other three datasets were taken as the validation datasets to optimize the previously identified DEGs. Recursive feature elimination (RFE) is an integrated machine learning method, which regards the selection of subsets as an optimization problem and evaluates gene combinations (26). From the training dataset GSE26440, the combination of optimal feature genes were selected using the RFE method in the R caret package (27) (version 6.0–76; www.cran.r-project.org/web/packages/caret). In 10-fold cross validation, the gene combination with the highest accuracy and the lowest Root Mean Square Error (RMSE) was selected as the optimal gene combination.

SVM is a supervised classification algorithm for machine learning, which uses the eigenvalues of features in each sample to discriminate sample types and estimate the probability that a sample belongs to a certain class (28). Using the SVM method in the e1071 package (28) (version 1.6–8; www.cran.r-project.org/web/packages/e1071) in R, the SVM classifier (Core; Sigmoid Kernel; Cross; 100-fold Cross validation) based on the optimal gene combination was built. Based on the pROC package (29) (version 1.12.1; www.cran.r-project.org/web/packages/pROC/index.html) in R, the area under the receiver operating characteristic (AUROC) was used to evaluate the efficiency of the SVM classifier in the training and validation datasets. Values of AUROC were distributed between 0.5 and 1, and the closer the AUROC value was to 1, the higher the efficiency of the classifier.

Results

Meta-analysis to identify the consistent DEGs across four datasets

The expression data in the four datasets were standardized; the curves before and after standardization are shown in Fig. 1A and B, respectively. After standardization, the gene expression levels of each dataset were distributed between −1 and 1, and the peak expression level was ~0, which was uniform (Fig. 1B). The results of the quality control analysis (Table II) and the PCA diagram (Fig. 1C) suggested that the distributions of the four datasets were balanced and all indexes met the quality test standards. Therefore, all datasets could be included in the subsequent analyses.

Figure 1. Distribution curves and a two-dimensional diagram of the PCA for the four datasets. Distribution curves (A) before and (B) after normalization. (C) Two-dimensional diagram of the PCA. Distribution curves for the GSE4607, GSE13904, GSE26378 and GSE26440 datasets are represented in black, green, red and blue, respectively. PCA, principal component analysis; IQC, internal quality control; EQC, external quality control; CQCg, consistency quality control gene; CQCp, consistency quality control pathway; AQCg, accuracy quality control gene; AQCp, accuracy quality control pathway.

Table II. Results of quality control analysis of the four microarray datasets.

Table II.

Results of quality control analysis of the four microarray datasets.

Dataset	IQC	EQC	CQCg	CQCp	AQCg	AQCp	SMR
GSE26378	5.350	4.453	237.678	107.220	41.268	114.658	2.839
GSE26440	6.089	3.784	246.767	102.538	33.131	94.224	3.561
GSE13904	6.448	4.579	74.753	76.039	37.205	106.445	4.101
GSE4607	6.649	3.141	127.552	114.078	24.009	74.664	4.202

[i] IQC, internal quality control; EQC, external quality control; CQCg, consistency quality control gene; CQCp, consistency quality control pathway; AQCg, accuracy quality control gene; AQCp, accuracy quality control pathway; SMR, standardized mean rank.

Through comprehensive analysis using the MetaDE package, the P, FDR, tau2, Q pval and Q values for each gene, as well as the log2FC value in each dataset, were calculated. According to the pre-set thresholds, a total of 2,699 consistent DEGs were identified from the four datasets. The heat map suggested that the differential expression patterns of the 2,699 DEGs were consistent in the four datasets (Fig. 2).

Figure 2. Heat map of the 2,699 consistent differentially expressed genes. Black and white sample strips represent pediatric septic shock and control samples, respectively. Red, green and black dots represent upregulated, downregulated and unchanged genes, respectively. CTRL, control; SS, septic shock.

WGCNA

To ensure that the gene expression levels in each dataset were comparable, all gene expression values in the four datasets were analyzed for expression level consistency. The expression level and connection level correlations between all combinations of pairs within the four datasets were positive and the P-values were significant (P<1×10−200), indicating that the datasets were comparable (Fig. S1).

WGCNA needed to satisfy the pre-condition of scale-free network distribution. Therefore, the value of the weighting parameter ‘power’ was firstly explored. Based on the GSE26440 dataset, the square values of the correlation coefficients between log(k) and log[p(k)] corresponding to different power values were calculated. Finally, the ‘power’ value was selected to be 28 when the square value of the correlation coefficient reached 0.9 (Fig. 3A). Under the parameter of ‘power’=28, the mean gene connectivity was statistically analyzed. The mean connectivity was 1, which confirmed the small-world property of the scale-free connection network (Fig. 3B).

Figure 3. Selection graph of the weighting parameter ‘power’ and schematic diagram of mean connectivity. (A) Selection graph of the weighting parameter ‘power’. The red line is the standard line when the square of the correlation coefficient reaches 0.9. (B) Schematic diagram of the mean RNA connectivity under different power parameters; when the power=28, the mean connectivity is 1.

After the co-expression network was constructed using the GSE26440 dataset as the main analysis dataset, a total of 11 modules were selected (Fig. 4A). The other three datasets including GSE26378 (Fig. 4B), GSE13904 (Fig. 4C) and GSE4607 (Fig. 4D) were also constructed with module partition. Meanwhile, module stability was evaluated and 10 significantly stable modules across the datasets were obtained (Table III). The number of genes present in each significantly stable module are shown in Fig. 5A. Using clinical information of the samples in the GSE26440 dataset, the correlations between each significantly stable module and clinical information were calculated using the WGCNA cor and corPvalueStudent packages (Fig. 5B).

Figure 4. Tree diagrams for the partitioning modules. Tree diagram for (A) GSE26440, (B) GSE26378, (C) GSE13904 and (D) GSE4607. Each color represents a different module.

Figure 5. Pie chart and clinical information correlation heat map of significantly stable modules. (A) Pie chart displaying the number of genes contained in each of the 10 significantly stable modules. (B) Correlation heat map between each significantly stable module and the clinical information of patients, including age, disease status and outcome. The data are presented as the correlation coefficient (P-value). Correlations were calculated using WGCNA cor and corPvalueStudent packages. The color from blue to orange indicates the correlation from −1 to 1, respectively.

Table III. Preservation and enrichment information of the 11 modules and the module genes.

Table III.

Preservation and enrichment information of the 11 modules and the module genes.

			Preservation information			Enrichment information
Module ID	Color	Module size	Z-score	P-value	DEGs (n)	Enrichment fold (95% CI)	Phyper
Module 1	Black	95	11.823	1.78×10−4	0	–	–
Module 2	Blue	139	9.292	1.18×10−2	1	0.157 (0.004–0.898)	3.13×10−2
Module 3	Brown	137	20.037	1.80×10−3	1	0.159 (0.004–0.912)	3.09×10−2
Module 4	Green	115	21.756	1.57×10−10	3	0.569 (0.115–1.727)	4.90×10−1
Module 5	Grey	3,907	8.083	1.00×10−200	55	0.307 (0.224–0.415)	2.20×10−16
Module 6	Magenta	85	14.990	3.17×10−6	12	3.079 (1.509–5.766)	1.30×10−3
Module 7	Pink	91	20.320	1.10×10−10	2	0.479 (0.057–1.803)	4.40×10−1
Module 8	Purple	82	22.134	6.00×10−4	26	6.915 (4.186–11.103)	2.67×10−12
Module 9	Red	105	21.649	1.57×10−6	8	1.662 (0.691–3.452)	1.64×10−1
Module 10	Turquoise	288	36.143	7.00×10−38	67	5.074 (3.714–6.868)	2.20×10−16
Module 11	Yellow	128	23.321	3.20×10−34	62	10.559 (7.458–14.849)	2.20×10−16

[i] The higher the stability parameter (Z-score), the higher the stability of the module. In general, 5<Z<10 indicates stability and Z>10 indicates high stability. Phyper, P-value calculated by the hypergeometric algorithm; DEGs, differentially expressed genes; CI, confidence interval; -, not applicable.

The consistent DEGs were compared to the genes within the significantly stable modules, resulting in the identification of 237 overlapping genes. Enrichment and distribution situations of the overlapping genes in each significantly stable module are presented in Table III. The consistent DEGs were significantly enriched in four stable modules, including the magenta, purple, turquoise, and yellow modules, which contained 12, 26, 67 and 62 genes, respectively (P<0.05). Afterwards, enrichment analysis for the genes involved in the four stable modules was performed and 18 GO biological process terms, including ‘translational elongation’ (P=6.45×10−16) and ‘translation’ (P=2.16×10−13), as well as nine KEGG pathways, including ‘ribosome’ (P=1.55×10−14) and ‘Fc gamma R-mediated phagocytosis’ (P=1.89×10−3) were acquired (Table IV).

Table IV. GO biological process terms and KEGG pathways enriched for the genes involved in the four stable modules.

Table IV.

GO biological process terms and KEGG pathways enriched for the genes involved in the four stable modules.

A, GO analysis
Biological process term	Count	P-value
GO:0006414: Translational elongation	17	6.45×10−16
GO:0006412: Translation	23	2.16×10−13
GO:0030029: Actin filament-based process	10	3.51×10−4
GO:0030036: Actin cytoskeleton organization	9	1.06×10−3
GO:0046907: Intracellular transport	15	2.55×10−3
GO:0042254: Ribosome biogenesis	6	5.09×10−3
GO:0022613: Ribonucleoprotein complex biogenesis	7	6.01×10−3
GO:0006413: Translational initiation	4	7.88×10−3
GO:0006928: Cell motion	11	1.13×10−2
GO:0016192: Vesicle-mediated transport	12	1.58×10−2
GO:0034621: Cellular macromolecular complex subunit organization	9	1.62×10−2
GO:0007010: Cytoskeleton organization	10	1.80×10−2
GO:0006886: Intracellular protein transport	9	2.08×10−2
GO:0034613: Cellular protein localization	9	3.39×10−2
GO:0070727: Cellular macromolecule localization	9	3.52×10−2
GO:0015031: Protein transport	13	4.44×10−2
GO:0001667: Ameboidal cell migration	3	4.44×10−2
GO:0045184: Establishment of protein localization	13	4.70×10−2
B, KEGG analysis
Pathway	Count	P-value
hsa03010: Ribosome	15	1.55×10−14
hsa04666: Fc gamma R-mediated phagocytosis	5	1.89×10−3
hsa04810: Regulation of actin cytoskeleton	6	8.13×10−3
hsa05110: Vibrio cholerae infection	3	1.22×10−2
hsa04062: Chemokine signaling pathway	4	3.30×10−2
hsa04130: SNARE interactions in vesicular transport	2	3.40×10−2
hsa04670: Leukocyte transendothelial migration	3	3.66×10−2
hsa04360: Axon guidance	3	4.09×10−2
hsa00190: Oxidative phosphorylation	3	4.13×10−2

[i] GO, Gene Ontology; KEGG, Kyoto Encyclopedia of Genes and Genomes.

Selection of the optimal gene combination and construction of SVM classifier

With the GSE26440 dataset as the training dataset, the optimal feature genes were further identified from the genes involved in the four stable modules. Under the optimal parameters (min RMSE=0.0849 and max accuracy=0.9262), six optimal feature genes [cysteine rich transmembrane module containing 1 (CYSTM1), S100 calcium binding protein A9 (S100A9), solute carrier family 2 member 14 (SLC2A14), stomatin (STOM), uridine phosphorylase 1 (UPP1) and utrophin (UTRN)] were selected.

Based on the six optimal feature genes, an SVM classifier was constructed within the GSE26440 dataset. Subsequently, the efficiency of the SVM classifier in the training and validation datasets was assessed. All the precision rates, calculated as true positive/(true positive + false positive), (Table V) and AUROC values (Fig. 6) were >0.9, suggested that the SVM classifier could accurately discriminate PSS samples from control samples. In addition, the expression of the six optimal genes in the four datasets was analyzed. The results suggested that the expression differences of the six optimal genes were consistent in the four datasets. The UTRN gene was significantly downregulated in the PSS samples (P<0.005), whereas the other five genes were significantly upregulated in the PSS samples (P<0.005; Fig. 7).

Figure 6. ROC curves based on the six optimal feature genes. ROC curves for GSE26440, GSE26378, GSE13904 and GSE4607 are displayed in black, red, blue and green, respectively. ROC, receiver operating characteristic.

Figure 7. Expression of the six optimal genes in the four datasets. Column chart for (A) GSE26440, (B) GSE26378, (C) GSE13904 and (D) GSE4607. ***P<0.005 vs. the control group. SS, septic shock; CTRL, control; CYSTM1, cysteine rich transmembrane module containing 1; S100A9, S100 calcium binding protein A9; SLC2A14, solute carrier family 2 member 14; STOM, stomatin; UPP1, uridine phosphorylase 1; UTRN, utrophin.

Table V. Indexes for assessing the efficiency of the support vector machine classifier in the training and validation datasets.

Table V.

Indexes for assessing the efficiency of the support vector machine classifier in the training and validation datasets.

Dataset	Precision rate	Sensitivity	Specificity	PPV	NPV	AUROC
GSE26440	0.977	0.989	0.939	0.979	0.969	0.994
GSE26378	0.961	1.000	0.840	0.951	1.000	0.990
GSE13904	0.976	1.000	0.900	0.970	1.000	0.983
GSE4607	0.947	0.976	0.875	0.952	0.933	0.951

[i] Precision rate was calculated as follows: True positive/(true positive + false positive). PPV, positive predictive value; NPV, negative predictive value; AUROC, area under receiver operating characteristic.

Discussion

In the present study, 2,699 consistent DEGs were screened from four datasets and 10 significantly stable modules across the datasets were obtained, based on WGCNA. Subsequently, the consistent DEGs were found to be enriched in the four stable modules, including the magenta, purple, turquoise, and yellow modules, which contained 12, 26, 67 and 62 genes, respectively. Moreover, six optimal feature genes (CYSTM1, S100A9, SLC2A14, STOM, UPP1 and UTRN) were identified from the genes included in the four stable modules. Additionally, an effective SVM classifier based on the six optimal genes was constructed. The results from the ROC curve analysis showed that the SVM classifier had high sensitivity and specificity in discriminating patients with PSS from normal subjects. To the best of our knowledge, an SVM classifier for SS diagnosis has not been reported previously. Therefore, the SVM classifier of six optimal genes developed in the present study may aid in the early identification of patients with PSS in clinical practice.

S100A9, constitutively expressed in neutrophils, is a member of the alarmins family, and exhibits several immune functions, including immunological defense and homeostasis (30). Upregulation of S100A9 at the mRNA level in SS is related to the occurrence of hospital-acquired infections following SS and may contribute to the early identification of patients at high risk of infection (31). Moreover, S100A8/S100A9 alarmins compromise the suppression of the immune system by myeloid-derived suppressor cells, a specific inflammatory monocyte population, by repressing their expansion, thus preventing the development of SS in neonates (32). Endotoxin tolerance (ET) is a critical immune dysfunction related to SS (33). Elevated S100A8 and S100A9 expression has been shown to be induced in ex vivo models of ET, suggesting that these two genes may serve as promising biomarkers of ET and therefore providing valuable information for immunotherapy of patients with SS (34). The present study suggested that S100A9 is closely related to the development and diagnosis of PSS. CYSTM1 confers tolerance and stress responses to heavy metals (35) and is a novel biomarker in Huntington's disease (36). The role of CYSTM1 in PSS is not completely understood. The results of the present study indicated that CYSTM1 might have a role in the mechanisms of PSS by interacting with other genes in the stable modules.

Previous studies have revealed that genetic variations in SLC2A14 are involved in the development and progression of chronic diseases, including inflammatory bowel disease (IBD) and Alzheimer's disease (37–39). The facilitated glucose transporter 14 encoded by SLC2A14 promotes the development of IBD and may be applied for precision intervention of IBD (38). SLP-2 belongs to the stomatin protein family, plays a critical role in T cell activation and is a candidate target for immunomodulation (40). STOM expression is dysregulated between septic children and healthy controls and may be a diagnostic marker for pediatric sepsis (41). The present study revealed that SLC2A14 and STOM might be involved in the pathogenesis of PSS by mediating the inflammatory immune response.

UPP1 functions in the homeostatic regulation of intracellular uridine concentrations and the activation of fluoropyrimidine nucleoside chemotherapeutic agents (42). Uridine displays anti-inflammatory action during lung inflammation (43). UTRN upregulation is implicated in the immune reaction in Duchenne muscular dystrophy mouse models (44). UTRN upregulation induced by proinflammatory factor-associated post-transcriptional mechanisms exhibits an antidystrophic effect (45). Therefore, the present study suggested that UPP1 and UTRN might act in the development and progression of PSS by affecting inflammation and immune reactions as well.

To the best of our knowledge, the present study is the first to report the diagnostic value of the aforementioned six genes for PSS. The six genes may be useful biomarkers for the early detection of PSS. However, in the present study the genes were not confirmed experimentally, therefore, further investigation is required to validate the results of the present study.

In conclusion, 2,699 consistent DEGs from four GEO datasets were analyzed in the present study. Furthermore, a SVM classifier based on six optimal genes was constructed for the accurate diagnosis of PSS, which may assist in the early identification of PSS and provide useful guidance for clinical interventions.

Supplementary Material

Supporting Data

Acknowledgements

Not applicable.

Funding

No funding was received.

Availability of data and materials

The datasets analyzed during the current study are available in the Gene Expression Omnibus database (www.ncbi.nlm.nih.gov/geo) with the following accession numbers: GSE26378, GSE26440, GSE13904 and GSE4607.

Authors' contributions

GL analyzed and interpreted the microarray datasets. CY designed the study and majorly contributed to writing the manuscript. All authors read and approved the final manuscript.

Ethics approval and consent to participate

Not applicable.

Patient consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests.

References