Identification of Potential Biomarkers Using Integrative Approach: A Case Study of ESCC

Abstract

This paper presents a consensus-based approach that incorporates three microarray and three RNA-Seq methods for unbiased and integrative identification of differentially expressed genes (DEGs) as potential biomarkers for critical disease(s). The proposed method performs satisfactorily on two microarray datasets (GSE20347 and GSE23400) and one RNA-Seq dataset (GSE130078) for esophageal squamous cell carcinoma (ESCC). Based on the input dataset, our framework employs specific DE methods to detect DEGs independently. A consensus based function that first considers DEGs common to all three methods for further downstream analysis has been introduced. The consensus function employs other parameters to overcome information loss. Differential co-expression (DCE) and preservation analysis of DEGs facilitates the study of behavioral changes in interactions among DEGs under normal and diseased circumstances. Considering hub genes in biologically relevant modules and most GO and pathway enriched DEGs as candidates for potential biomarkers of ESCC, we perform further validation through biological analysis as well as literature evidence. We have identified 25 DEGs that have strong biological relevance to their respective datasets and have previous literature establishing them as potential biomarkers for ESCC. We have further identified 8 additional DEGs as probable potential biomarkers for ESCC, but recommend further in-depth analysis.

Introduction

For a critical disease of interest, the knowledge of differentially expressed genes (DEG) are a crucial step toward biomarker identification. This is achieved through Differential Expression Analysis (DEA) which monitors the behavior of each gene in isolation over normal and disease conditions and streamlines the search for biomarkers by providing a candidate list of these discriminative candidate genes. DNA microarray and RNA Sequencing (RNA-Seq) are indispensable methods for DEA. Previously, microarray technology was the most widely used approach. However, there are inherent limitations such as the pre-requisite prior knowledge of the sequence for the array design or the fact that cross-hybridization makes it difficult to analyze highly correlated sequences. Furthermore, the lack of sensitivity to highly or lowly expressed genes as well as the lack of reproducibility across laboratories and platforms pose major challenges. These limitations are overcome by RNA-Seq technology.

Numerous DEA methods have been developed to serve both microarray and RNA-Seq technologies. Furthermore, a large number of datasets related to critical diseases that are compatible with both technologies are widely available. Keeping in mind the fact that most methods developed for these technologies are not effective for all cases, we propose a consensus-based integrative approach that ensembles a selected few of these methods with the aim to achieve improved performance. In other words, in this paper, we present a consensus-based approach that employs a few chosen microarray DEA methods (Limma [1], SAM [2] and EBAM [3]) and RNA-Seq (Limma+Voom [4], edgeR [5], DESeq2 [6]) to present an unbiased list of DEGs as candidates for potential biomarkers. As the primary focus of our work is on the critical disease ESCC, we validate our results on two ESCC microarray datasets (GSE20347 and GSE23400) datasets and one ESCC RNA-Seq (GSE130078) dataset.

The rest of the paper is organized as follows. Section “Related Work” provides a brief overview of related work. Section “Proposed DE Framework” describes our proposed framework for biomarker identification of critical disease(s) employing three microarray and three RNA-Seq methods. We also introduce our consensus function for unbiased integration of DEGs individually detected by previously mentioned methods. Section “Analysis” reports a detailed experimental analysis and validation of our method on two benchmark microarray gene expression datasets and one RNA Sequencing (RNA-Seq) dataset. Section “Discussion” presents a detailed analysis and discussion on candidate genes that have been identified as potential biomarkers for ESCC. In this section, we also present a comparison of our algorithm with two recent works with similar approaches. Finally, the concluding remarks are given in Section “Conclusion”.

Related Work

A number of statistical approaches are used by Limma [1] to achieve effectiveness in large-scale expression studies. Limma takes advantage of the flexibility of linear models and fits one to each row (gene) of data in a gene expression matrix where columns correspond to samples. Limma has the inherent ability to model correlations between samples through analysis of the entire dataset as one entity. Significance Analysis of Microarrays (SAM) [2] assimilates a set of gene-specific t-tests to identify genes that exhibit statistically significant changes in expression. For each gene, based on the changes in gene expression in terms of standard deviation for repeated measurements, SAM assigns a score. Potentially significant genes are identified using a threshold for these scores. Empirical Bayes Analysis of Microarrays (EBAM) [3] employs the removal Bayes rule to obtain the posterior probability that a gene was affected or unaffected under the various conditions. EBAM makes multiple testing a possibility by establishing a connection between prior probabilities and local false discovery rate (local fdr) in turn handling the issues that arise from simultaneous tests.

Voom [4] works on the idea that commonly used microarray-based statistical methods can be applied to read counts of corresponding RNA-Seq through robust and non-parametric estimation of mean-variance. In other words, Voom incorporates the mean-variance trend into the empirical Bayes procedure of Limma. edgeR [5] and DESeq2 [6] are estimations of gene-wise dispersion by conditional maximum likelihood, conditioning on the total count for the gene. edgeR effectively borrows information within genes to shrink the dispersion towards a consensus value through the use of empirical Bayes. It adapts for overdispersed data and incorporates an exact test to assess each gene. DESeq2 accurately estimates the expected dispersion value for genes of a given expression strength, which is then used to conform the gene-wise dispersion towards the predicted values . It also accounts for gene-specific variations and makes it possible to estimate fitted curves and testing even in settings with less information.

Proposed DE Framework

The proposed framework aims to work with microarray and RNA Sequencing data. We have chosen three methods that work on micro-array (Limma, SAM and EBAM) and three on RNA-Seq (Limma+voom, DESeq and EdgeR). First, both types of data require pre-processing. For microarray, pre-processing consists of the removal of unwanted and redundant information, normalization of the dataset, missing value estimation while for RNA-Seq data, we perform removal of low read counts, normalization, and transformation.

Pre-processing is followed by DE analysis that results in differentially expressed genes (DEG). For each data type, we employ a consensus function that filters out all common DEGs for each dataset. In other words, depending on the type of the input dataset, the DE analysis unit detects DEGs using three corresponding methods, followed by a consensus function that filters the DEGs common to all three methods as well as identify other relevant DEGs. Our consensus function is given by

$$\begin{array}{c}\hfill {\text{DEGs}}_{\text{relevant}}={\text{DEGs}}_{\text{common}}\cup {\text{DEGs}}_{\text{others}}\end{array}$$ 1

where

$$\begin{array}{cc}& {\text{DEGs}}_{\text{common}}=\hfill \\ \hfill & \left\{\begin{array}{cc}{\text{DEGs}}_{\text{limma}}\cap {\text{DEGs}}_{\text{SAM}}\cap {\text{DEGs}}_{\text{EBAM}},\hfill & \text{for Microarray},\hfill \\ {\text{DEGs}}_{\mathrm{limma}-\mathrm{voom}}\cap {\text{DEGs}}_{\text{edgeR}}\cap {\text{DEGs}}_{\text{DESeq2}},\hfill & \text{for RNA}\text{-}\text{Seq},\hfill \end{array}\right)\hfill \end{array}$$

and

$${\text{DEGs}}_{\text{others}}=\left\{\begin{array}{ccc}\text{DEGs}\notin {\text{DEGs}}_{\text{common}}\hfill & q\text{- value}\le \alpha ,& \text{for RNA - Seq},\\ \text{DEGs}\notin {\text{DEGs}}_{\text{common}}\hfill & local.fdr\le \beta ,& \text{for Microarray}\end{array}\right)$$

Here, $\alpha$ and $\beta$ are q-value and local.fdr significance values that are chosen according to their relevance to the experiment. Through multiple iterations of implementation, we observed that consideration of only genes common to all three methods leads to information loss. Thus, to overcome this we introduced q-value into the consensus function. The main idea behind this is that while a p-value of 0.05 gives the implication that 5% of the tests will be false positive (FP), q-value, which is an FDR adjusted p-value, implies that 5% of the test found to be significant will be FP. q-value requires a very important adjustment for multiple tests on the same data sample. Our consensus function considers all genes common to all three methods with $p=0.05$ as detected DEGs. Furthermore, all genes that are not among the common genes but have a $q=0.05$, i.e. $\alpha$ (Eq. 1), are added to the list of DEGs. However, in the microarray datasets, we implement the proposed consensus function given by Eq. 1 to start off by taking the DEGs common to all three methods. Unlike RNA-Seq, instead of q-value, we incorporate its useful counterpart local fdr ($\beta$ in Eq. 1). Local fdr is a measure of the posterior probability that the null hypothesis is true. We use local fdr since Limma and SAM calculate p-value. EBAM, on the other hand, estimates the posterior probability and local fdr. It is worth mentioning that posterior probability and p-value are not interchangeable. However, local fdr can be estimated from p-values. Our consensus function (Eq. 1) considers all genes common to all three methods with $p=0.05$ as detected DEGs. Furthermore, all genes that are not among the common genes but have a $local.fdr=0.05$ ($\beta$) are added to the list of DEGs.

The relevant DEGs are then taken as input to the DCE analysis unit. The idea behind performing DCE analysis is that it leads to the creation of biologically relevant modules which are easier for further analysis and validation. The DCE unit identifies differentially co-expressed modules and performs preservation analysis on these modules to identify biologically relevant modules. This is followed by the identification of hub genes in these modules using WGCNA [7] intramodular connectivity.

We validate our results using several approaches. First, we consider all relevant DEGs detected by the DE analysis unit in isolation and perform Gene Ontology (GO) and KEGG pathway enrichment analysis to validate biological relevance. We consider all the hub genes in the biologically relevant modules identified by the DCE unit as biomarker candidates. Furthermore, all DEGs that are annotated with the most enriched GO term in all three databases as well as the most enriched KEGG pathway are also considered as candidates for biomarkers. We term such genes as Top Enriched DEGs (TEDs). Secondly, we further analyze the biomarker candidates through observation of their interactions and behavioral changes. Finally, we trace literature evidence for the relevant genes in other scientific sources and works for further validation of these identified genes as possible biomarkers.

Analysis

Our focus is on ESCC, a cancer very common in developing countries, especially in North-East India, and is highly attributed to tobacco and betelnut chewing, alcohol consumption as well as poor diet. Two microarray datasets GSE20347 and GSE23400 and an RNA-Sequencing dataset GSE130078 were chosen to validate our proposed DE framework (see Fig. 1). Details of each dataset is provided in Table 1. All three datasets examined gene expression in tumor and matched normal adjacent tissue. The test platform is a DELL workstation with Intel(R) Xeon(R) W-2145 with 3.70GHz processor, 64 GB RAM running Windows 10 Pro for workstations.

Fig. 1.

Proposed framework for differential expression analysis

Table 1.

Datasets

Dataset	No. of genes	Normal samples	Tumor samples
GSE20347${}^{\text{a}}$	22,278	17	17
GSE23400${}^{\text{b}}$	22,283	53	53
GSE130078${}^{\text{c}}$	57,783	23	23

${}^{\text{a}}$https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE20347

${}^{\text{b}}$https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE23400

${}^{\text{c}}$https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE130078

Preprocessing

RNA-Seq dataset GSE130078 has 57,783 genes and 46 samples. Large datasets tend to add complications to the analysis and as such, we filter out genes with low read counts. We achieve this by calculating the counts per million (CPM) for each sample for each gene and keep only those genes that have CPM $>1$ for at least two samples. This reduces the dataset size from 57,783 to 22,270. We then follow up by normalization of the dataset. We also consider two microarray datasets GSE20347 and GSE23400 for analysis. The inputs to these datasets are expression values of genes across samples. First, we pre-process the data through the removal of unwanted and redundant genes, missing value estimation, and normalization. However, for both GSE20347 and GSE23400, there are no missing values and as such we proceed further down the pipeline.

DE Analysis

For the microarray datasets, Limma takes the pre-processed dataset as input and outputs the equivalent DEGs with a significance of 5% ($p\text{-value}\le 0.05$) and False Discovery Rate (FDR) of 0.05. On the other hand, for the other two methods SAM and EBAM, we employ findDelta with FDR $=0.05$ fiving us an estimate of the delta values at which FDR is closest to 0.05 and chose accordingly. In SAM, delta is the distance between the observed and the expected test scores, whereas in EBAM, delta is the probability that a gene with a specific test score is differentially expressed. Table 2 summarizes the DEGs detected by all three methods on all three datasets.

Table 2.

Summary of detected DEGs by the three RNA-Seq methods and the three microarray methods

Dataset	Method	No. of DEGs with $p\le 0.05$	Common DEGs
GSE20347	Limma	8689	7706
	SAM	10,642
	EBAM	9565
GSE23400	Limma	13,558	3431
	SAM	14,301
	EBAM	3,431
GSE130078	Limma +Voom	6858	2765
	edgeR	12,623
	DESeq2	12,766

In the case of the RNA-Seq dataset, the pre-processed data are the input to all three methods, i.e., Limma+Voom, edgeR and DESeq2. However, it is to be noted that while DESeq2 directly takes the count data as input, the other two methods require the count data to be transformed into a DGEList (Digital Gene Expression Data) object. All the methods perform multiple tests on all the 22,270 genes in the dataset across 46 samples. We consider a significance of 5%, i.e., $p\text{-value}\le 0.05$ and the corresponding DEGs detected by the 3 these methods are summarized in Table 2.

We implement the proposed consensus Eq. 1 to identify the common genes detected by these three methods. First, we consider the DEGs detected by all three methods, i.e. common genes. In GS20347, there are such 7706 DEGs. So as not to bypass crucial information, we use $\beta$ in Eq. 1, i.e., the consensus function. With $local.fdr=0.05$ ($\beta$) another 662 genes are considered DEGs resulting in a list of 8,368 DEGs. Similarly, in GSE23400, Limma, SAM, and EBAM find 3,431 common DEGs. With $local.fdr=0.05$ ($\beta$), another 4,066 genes are considered as DEGs, resulting in a list of 7,497 DEGs. In the case of GSE130078, the three methods Limma+Voom, edgeR, and DESeq2 discover 2,765 common DEGs and a q-value ($\alpha$) adds another 9,945 genes resulting in a list of 12,710 DEGs.

DCE Analysis

To analyze the interactions among the DEGs as well as the variations in behavior under normal and disease circumstances, we construct co-expression networks (CEN) using WGCNA [7]. The pipeline for DCE analysis is to detect DEGs by our method is as follows.

1. Divide the dataset into separate datasets: All genes with only normal samples and all genes with only disease samples

2. Choose the soft thresholding power to which co-expression similarity is raised to calculate an adjacency matrix. Soft thresholding power is based on the criterion of approximate scale-free topology.

3. Construct two separate CENs in the form of an adjacency matrix: normal and disease.

4. Transform adjacency matrices into topological overlap matrices (TOM [8]) to minimize the effects of noise and spurious associations

5. Extract all connections that correspond to the subset of DEGs from both CENs.

6. Extract normal and disease modules using hierarchical clustering.

7. Merge modules through eigenvector module selection and MEDissThres threshold merging.

8. Identify modules extracted in the normal dataset that are non-preserved in disease dataset and vice versa through preservation analysis [9]. We consider such modules as modules of interest for further downstream analysis.

9. Identify the top 20 hub genes using intramodular connectivity [7] in all modules of interest.

We start DCE analysis by clustering the samples using the hierarchical approach to detect outlier samples. We remove the outlier samples with the aim of creating a more robust CEN. For GSE23047, as seen in Fig. 2a, b, we find a single outlier in the case of normal samples with a tree cut at height h = 70 (Blue). However, in disease samples, there are two outliers with a cut at h = 130 (Red). Similarly, in the case of GSE23400, as seen in Fig. 2c, d, a tree cut at height h = 105 (Blue) and at h = 95 (Red) removes one and two outliers from normal and disease samples, respectively. In the case of GSE130078, a cut at h = 1,500,000 (Blue) and h = 2,000,000 (Red) removes one normal (Fig. 2e) and one disease Fig. 2f sample.

Fig. 2.

Outlier hierarchical trees for normal and disease samples for all three datasets

In GSE20347, hierarchical Clustering and tree cut results in 50 and 75 normal and disease modules, respectively. Figure 3a shows the dendrogram while the first strip of colors below represents the corresponding module colors for the normal dataset. Similarly, Fig. 3b shows the dendrogram for the disease dataset. To merge modules, we choose a height cut of 0.25, corresponding to a correlation of 0.75. Merging of the modules with tree cut at h = 0.25 further reduces the number of modules to 38 and 61 for normal and disease datasets, respectively. The second color strip in Fig. 3a, b shows the colors for the merged normal and disease modules respectively. In GSE23400, hierarchical clustering results in 9 normal (the first color strip in Fig. 3c) and 13 (the first color strip in Fig. 3d) disease modules, which are then reduced to 8 normal (the second color strip in Fig. 3c) and 11 disease (second color strip in Fig. 3d) modules after merging. Finally in GSE130078, hierarchical clustering results in 65 normal (the first color strip in Fig. 3e) and 40 disease (the first color strip in Fig. 3f) modules, which are then reduced to 21 normal (the second color strip in Fig. 3e) and 24 disease (the second color strip in Fig. 3f) modules after merging.

Fig. 3.

Dendrograms where the first strip of colors represents the colors initially assigned to corresponding modules while the second strip of colors represents the colors assigned to corresponding merged modules for the ESCC dataset: GSE20347 (a, b); GSE23400 (c, d); GSE130078 (e, f)

We follow module extraction by module preservation analysis with the aim of analyzing the distinction between preserved and non-preserved modules. According to Langfelder et al. [9], while the preserved modules retain a majority of their co-expressed connections (or edges between two genes), the same cannot be perceived from non-preserved modules. According to Langfelder et al. [9], a module with $Z_{\text{summary}}<2$ is considered non-preserved [9]. It is noteworthy that, GSE23400 due to its inherent nature, extracts a smaller number of modules with significantly larger sizes and higher densities (Fig. 4). There are no non-preserved modules with $Z_{\text{summary}}<2$ and most modules are either moderately preserved or highly preserved. We take into consideration moderately preserved modules with $Z_{\text{summary}}<10$ [9].

Fig. 4.

$Z_{\text{summary}}$ on ESCC dataset GSE20347 (a, b); $Z_{\text{summary}}$ on ESCC dataset GSE23400 (c, d); $Z_{\text{summary}}$ on ESCC dataset GSE130078 (e, f). All modules below the red line ($Z_{\text{summary}}<2$) are non-preserved, all modules between the red and blue lines ($2<Z_{\text{summary}}<10$) are weak to moderately preserved and all modules above the blue line ($Z_{\text{summary}}>10$) have strong evidence of being preserved

Table 3 summarizes the preservation analysis for non-preserved modules in all three datasets. The second column highlights the module preservation reference and test networks. For example, the table reading for module pink in Normal/Disease subset of dataset GSE20347 can be interpreted module pink of size 278 detected in the normal network that is non-preserved in disease network with a $Z_{\text{summary}}$ value of – 0.118842161.

Table 3.

Preservation analysis of modules in all three datasets

	Ref/Test	Module	Size	$Z_{\text{summary}}$
GSE20347	Normal/Disease	pink	276	– 0.118842161
		bisque4	62	0.84895892
		orangered4	82	1.10844007
		grey	17	1.38809779
	Disease/Normal	grey	3	– 0.11691952
		greenyellow	149	0.21296495
		brown2	39	0.40355693
		darkgreen	201	0.57638054
		lightpink4	61	0.58347786
		white	99	0.72904092
		lightyellow	122	0.88046344
		darkolivegreen4	40	1.13783059
		antiquewhite4	58	1.19766058
		lightsteelblue1	143	1.42400619
		mediumpurple3	77	1.74769518
		black	775	1.79866799
		skyblue1	51	1.79952166
		lavenderblush3	60	1.83004964
		lightgreen	123	1.88961457
GSE23400	Normal/Disease	magenta	45	5.63609823
	Disease/Normal	magenta	231	5.59355067
		salmon	44	5.64755802
		greenyellow	172	6.47180995
		purple	225	9.42312232
GSE130078	Normal/Disease	magenta	248	– 1.80627625
		skyblue2	37	– 0.62266325
		bisque4	1000	0.01985689
		maroon	82	0.58940588
		grey	70	1.90096828
	Disease/Normal	lightyellow	240	– 1.18378716
		red	759	– 0.18883575
		lightcyan	321	0.68868388
		steelblue	145	0.92648735
		skyblue3	104	1.01960350
		violet	142	1.38497020

We only consider non-preserved modules of substantial size ($size>100$) as modules of interest for further downstream analysis and validation. To find the hub genes for each module of interest extracted previously we employ WGCNA intramodular-connectivity proposed by Langfelder et al. [7]. Intramodular-connectivity calculates the connectivity of a node to other nodes in the same module.

Validation

Enrichment Analysis of Modules

For a module of interest to be regarded as Gene Ontology (GO) or pathway enriched, at least one gene in the module must be assigned to an enriched GO term or pathway, respectively with a significance of 5% (i.e., $p\le 0.05$). To perform functional enrichment analysis, we use the easily available online tool DAVID [10, 11]. Table 5 summarizes the percentages of genes in the modules of interest annotated to enriched GO terms as well as enriched KEGG pathways. We observe that all modules of interest are GO and pathway enriched.

Table 5.

Percentages of genes in each module that are annotated to the Gene Ontology (GO) databases (BP: Biological Processes, CC: Cellular components or MF: Molecular function) and KEGG pathways

Dataset	Module	Size	GO_BP (%)	GO_MF (%)	GO_CC (%)	KEGG (%)
GSE130078	lightyellow	240	78.9	80.5	78.9	38.3
	red	759	83.3	87.6	84.8	38.9
	skyblue3	104	82.6	87.2	85.3	46.8
	steelblue	145	73.1	76.6	71.9	26.9
	violet	142	82.0	84.7	80.0	37.3
	lightcyan	321	69.1	70.7	69.4	27.7
	bisque4	1000	87.5	91.4	89.1	40.5
	magenta	249	84.4	89.9	84.9	31.0
GSE20347	pink	276	90.1	96.0	95.2	47.2
	greenyellow	149	98.4	98.4	97.7	49.2
	darkgreen	201	95.6	96.7	95.6	55.0
	lightyellow	122	95.0	96.0	97.0	53.5
	lightsteelblue1	143	95.2	94.4	94.4	49.2
	black	775	94.9	97.0	96.1	52.9
	lightgreen	123	90.7	98.1	92.6	56.5
GSE23400	greenyellow	172	97.7	98.3	96.5	63.4
	magenta	231	94.8	96.5	93.4	44.5
	purple	225	96.6	98.1	96.3	58.2

Candidate Genes

As mentioned earlier, we select DEGs as candidates for potential biomarkers based on the following two criteria:

1. All hub genes detected by the DCE analysis unit of our framework in all modules of interest are candidate genes.

2. DEGs that have been annotated to the most enriched GO terms in all three GO databases (BP, CC and MF) and are also annotated to the most enriched pathway after GO and Pathway enrichment analysis on the entire dataset are also considered as potential biomarkers. We rename these DEGs as TEDs (Top Enriched DEGs)

Thus, alongside all DEGs that are among Top 20 hub genes in modules of interest (as summarized in Table 4), our second criterion adds 22, 18 and 11 TEDs to the list of candidate genes in GSE20347, GSE23400 and GSE130078, respectively. We summarize these DEGs (TEDs) in Table 6. The numbers of candidate genes for GSE20347,GSE23400 and GSE130078 increase from 140, 60 and 160 to 162, 78 and 171, respectively. We perform the enrichment analysis on the entire dataset or in more specific terms the list of all genes in the dataset. This leads to the observation that as the lists of genes in GSE20347 (22,278 genes) and GSE23400 (22,283 genes) are almost the same, the list of top enriched genes (35 genes) extracted are the same. However, the differences in TEDs are seen (Table 6) due to the fact that there are DEGs identified in one dataset that might not be detected in the other.

Table 4.

Top 20 hub genes for each extracted module of interest in all three datasets using WGCNA [7] intramodular connectivity

Hub genes with strong literature evidence of association to esophageal squamous cell carcinoma (ESCC) are marked in Red while hub genes with evidence of association with five other SCCs namely, Oral, Tongue, Head and Neck, Tongue or Laryngeal squamous cell carcinoma are marked in Blue

Table 6.

DEGs that are annotated to most enriched GO term in all three GO databases (BP, CC and MF) as well as the most enriched pathway

DEGs with strong literature evidence of association to Esophageal squamous cell carcinoma (ESCC) are marked in Red while hub genes with evidence of association to five other SCC namely, Oral, Tongue, Head and Neck, Tongue or Laryngeal Squamous Cell Carcinoma are marked in Blue

Biological Analysis

To establish the biological relevance of the candidate genes detected by our method, we use functional enrichment analysis and the construction of a gene regulatory network (GRN). Transcription Factors (TF) have remarkable diversity as well potency as drivers of cell transformation. Bhagwat et al. [12] justify the continued pursuit of TFs as potential biomarkers across many forms cancer by the prevalent deregulation of the same. We observe that 26 (hub genes:21, TEDs:5), 11 (hub genes:6, TEDs:5) and 23 (hub genes:23, TEDs:0) candidate genes detected by our method in GSE20347, GSE23400 and GSE130078, respectively are TFs. These TFs exhibit regulatory behavior in their respective modules, establishing their biological relevance. For easy visualization, we extract a manageable subset of hub genes from the non-preserved modules detected by our method (Figs. 5a–f and 6a–f). We construct a Gene Regulatory Network (GRN) with these hub genes and associated Transcription Factors (TFs) so as to observe the regulatory behavior of the corresponding genes. The resulting GRN is in the form of an adjacency list with weighted directed edges from TFs to other target genes (TGs).

Fig. 5.

Gene Regulatory Network (GRN) on the subset of hub genes detected by our method for modules pink, greenyellow, darkgreen, lightsteelblue1, black and magenta. Regulatory behavior is represented by a weighted directed edge from a Transcription Factor (TF) to a Target Gene (TG)

Fig. 6.

Gene Regulatory Network (GRN) on the subset of hub genes detected by our method for modules purple, greenyellow, blue, lightyellow, violet and steelblue. Regulatory behavior is represented by a weighted directed edge from a Transcription Factor (TF) to a Target Gene (TG)

As in the the case of validation of modules, we employ DAVID [10, 11] to perform functional enrichment analysis of all candidate genes detected by our method. A candidate gene can be regarded as GO enriched considering a GO database (GO_BP, GO_CC, GO_MF) if it is annotated to at least one GO term in that database with significance of 5% ($p\le 0.05$). Tables 7, 8 and 9 summarize the candidate genes annotated to the top 3 GO terms in each GO database in GSE20347, GSE23400 and GSE130078, respectively. Similarly, a candidate gene is KEGG pathway enriched if it is annotated to at least one KEGG pathway term with significance of 5%. Table 10 summarizes the candidate genes annotated to top 3 enriched KEGG pathways in GSE20347, GSE23400 and GSE130078.

Table 7.

Summary of candidate genes detected by our method in the microarray dataset, GS20347 that are annotated to top 3 GO terms in the three GO databases

	GO Term	Annotated Candidate Genes
GO_BP	GO:0007165 signal transduction	PRKACB, BCR, AR, LYN, APPL1, STAT1, SHC1, NFKB2, PIK3CD, PIK3CB, PIK3R2, PIK3R1, EXT2, RAC2, PPFIA1, RAF1, GNB5, GSK3B, MAPK10, KRAS, RARA, MAP2K1, TXNRD1, FLT3LG, FADD, RANBP1, FAS, MAPK1
	GO:0045944 positive regulation of transcription from RNA polymerase II promoter	AR, HOXC11, STAT1, PITX1, NFKB2, TP63, PIK3R2, PIK3R1, PTMA, RAF1, MED1, ZNF148, RARA, FADD, CXCR3
	GO:0016032 viral process	ITSN2, POM121, POM121C, LYN, BRD2, PSMB5, STAT1, ABI2, SHC1, PIK3R1, FADD, RANBP1, MAPK1
GO_CC	GO:0005829 cytosol	ITSN2, PRKACB, BCR, CASP10, MSRA, RAC2, PTMA, RAF1, UBA7, HOMER3, RARA, MAP2K1, MAP3K20, MCL1, MCM7, UBASH3A, AR, LYN, APPL1, STAT1, KPNA2, PIK3CD, PIK3CB, PIK3R2, PIK3R1, TRIO, RPL22, NEDD4L, PNO1, ABI2, NEB, SARS1, TXNRD1, FADD, CLIC4, PRMT1, FAS, SENP5, EPB41L1, HOXC11, SHC1, ODF2, CEP290, EEF1B2, NABP1, GNB5, KRAS, HAUS7, PSAT1, FLT3LG, MAPK1, PSMD4, RUFY3, PSMB5, TCOF1, PSMC2, NFKB2, JPT2, EXOSC4, CCT4, PPFIA1, SBF1, KIF4A, GSK3B, MAPK10, HPRT1, ITPKC, ENOX2, SPAG5, MAPRE1, RANBP1
	GO:0005654 nucleoplasm	PRKACB, CIAPIN1, TP63, MSRA, PTMA, UBA7, SLC3A2, RRP7A, RARA, ORC3, MCL1, MCM7, TIMELESS, NONO, UBASH3A, AR, STAT1, KPNA2, PIK3CB, DBF4, MED1, NEDD4L, PNO1, ABI2, DNMT3B, TXNRD1, PRMT1, SENP5, POP7, MAGOHB, HOXC11, NABP1, TERF1, NSD2, RFC2, NPIPB3, RAB8B, MAPK1, PSMD4, UQCRC2, PSMB5, TCOF1, PSMC2, NFKB2, IMP4, EXOSC4, CCT4, KIF4A, GSK3B, MAPK10, TRRAP, POM121, ZNF148, BRD2, ESF1, NMD3
	GO:0016020 membrane	TFRC, BCR, ITGB7, CEP290, NEU1, EXT2, RAC2, GNAQ, SLC3A2, MAN1C1, KRAS, MCL1, MCM7, FLT3LG, NONO, LSS, RUFY3, APPL1, PSMC2, KPNA2, PIK3CD, PIK3CB, ENTPD7, PIK3R1, CD52, CD48, SBF1, KIF4A, DDX24, MED1, LST1, NMD3, CLIC4, FAS
GO_MF	GO:0005515 protein binding	TFRC, PRKACB, BCR, ZKSCAN5, NEU1, MSRA, RAC2, RAF1, RALY, CD2, HOMER3, RRP7A, APOOL, RARA, MAP2K1, MCL1, MCM7, TIMELESS, UBASH3A, AR, APPL1, STAT1, PITX1, PIK3CD, PIK3CB, HSD17B10, EPHA2, PIK3R2, PIK3R1, RCN1, MED1, RPL22, NEDD4L, PNO1, ABI2, DNMT3B, TXNRD1, GMCL2, FADD, ITM2A, CLIC4, PRMT1, FAS, TGFB2, POP7, EPB41L1, MAGOHB, HOXC11, CNPY2, SHC1, CEP290, MAN1C1, KRAS, TERF1, RFC2, PSAT1, FLT3LG, RAB8B, VGLL4, MAPK1, UQCRC2, CKS2, TCOF1, NFKB2, IMP4, EXOSC4, KIF4A, GSK3B, MAPK10, TNS1, RANBP1
	GO:0042802 identical protein binding	TFRC, PSMD4, APPL1, STAT1, CEP290, TP63, CD3D, RAF1, RALY, CD2, KRAS, HOMER3, RPL22, HPRT1, TERF1, ABI2, PSAT1, FADD, PRMT1, MAPRE1, FAS, TIMELESS, NONO, MAPK1
	GO:0003723 RNA binding	POP7, TFRC, PSMD4, MAGOHB, TCOF1, KPNA2, HSD17B10, CCT4, NABP1, DDX24, SLC3A2, RALY, RPL22, RRP7A, PNO1, EBNA1BP2, MAP3K20, SARS1, ENOX2, ESF1, TNS1, NMD3, PRMT1, MAPRE1, PUS7, NONO

Table 8.

Summary of candidate genes detected by our method in the microarray dataset GSE23400 that have been annotated to top 3 GO terms in the three GO databases

Data base	GO Term	Annotated Candidate Genes
GO_BP	GO:0007165 signal transduction	BCR, AR, APPL1, STAT1, STAT2, CXCL10, CXCL11, PIK3CD, PIK3R1, RAC2, CD53, IL15, TYROBP, RAF1, GSK3B, PRKCB, HIF1A, RARA, MAP2K1, VRK1, TXNRD1, FADD
	GO:0045944 positive regulation of transcription from RNA polymerase II promoter	AR, STAT1, STAT2, CXCL10, PIK3R1, RAF1, HIF1A, RARA, FADD, TFEC
	GO:0045893 positive regulation of transcription	PRKCB, AR, HIF1A, RARA, STAT1, MAP2K1, CD86
GO_CC	GO:0005829 cytosol	BCR, MTHFD1, SCO2. SRM, GZMB, RAC2, RAF1, PRKCB, IFIH1, RARA, MAP2K1, IFIT1, IFIT3, OAS3, AR, APPL1, STAT1, STAT2, HNMT, PIK3CD, UBE2L6, PIK3R1, LAPTM5, CD163, TXNRD1, MNDA, FADD, DLGAP5, CDKN3, IL15, VRK1, EIF2S1, PSMA3, PSMB8, PSMB9, PSMC1, IFI35, ISG15, GSK3B, HIF1A, TUBG1
	GO:0005654 nucleoplasm	MNAT1, PRKCB, RARA, OAS3, AR, STAT1, STAT2, HNMT, UBE2L6, TXNRD1, MNDA, IL15, NASP, VRK1, PSMA3, PSMB8, PSMB9, PSMC1, ISG15, GSK3B, HIF1A, TFEC
	GO:0016020 membrane	BCR, MTHFD1, ITGB2, GZMB, RAC2, PLXNC1, HLA-C, HLA-F, HLA-G, EIF2S1, OAS3, APPL1, PSMC1, PIK3CD, IFI35, ENTPD1, PIK3R1, TAP1, CD163
GO_MF	GO:0005515 protein binding	BCR, SCO2, MNAT1, SRM, RAC2, FCER1G, RAF1, PRKCB, RARA, MAP2K1, AR, APPL1, STAT1, STAT2, PIK3CD, UBE2L6, ENTPD1, PIK3R1, GLRX5, CD163, TXNRD1, FADD, IL15, TIMM9, PLXNC1, VRK1, EIF2S1, FCGR2A, IFI35, GSK3B, HIF1A, TUBG1
	GO:0042802 identical protein binding	APPL1, STAT1, STAT2, SRM, IFI35, CD53, FCER1G, TYROBP, RAF1, IFIH1, IFIT3, HLA-G, FADD, TUBG1, C1QB
	GO:0003723 RNA binding	PSMC1, IFIH1, EBNA1BP2, IFIT1, IFIT3, EIF2S1

Table 9.

Summary of candidate genes detected by our method in GS130078 that have been annotated to top 3 GO terms in the three GO databases

Data base	GO Term	Hub Genes
GO_BP	GO:0007165 signal transduction	PIP5K1A, RHBDL1, RRAS2, CXCL16, HPGDS, PIK3CD, PIK3CB, VEGFC, RIT1, RALB, CDC42SE1, HIF1A, GPI, PGF, CDK4, PDE1B, PDE2A, PDE3A, TENM1, TYMP, PDE9A, CSF2RB, PI4KA, PI4KB, PDE3B, PDE4A, PDE4D, PPP2R5A, PDE5A, GNA13, LIMK1
	GO:0000122 negative regulation of transcription from RNA polymerase II promoter	MEF2A, H1-2, ZNF239, ZNF85, OSR2, CAV1, NFIB, ZNF568, PDE2A
	GO:0045944 positive regulation of transcription from RNA polymerase II promoter	TBK1, MEF2A, LRP6, OSR2, ZNF91, HIF1A, NFIB, ESRRG, LMO4, SENP1
GO_CC	GO:0005886 plasma membrane	FEZ1, KCNH3, ARL4A, SDE2, RIT1, RALB, PI4K2A, COA6, TENM1, CDH22, CDH23, RRAS2, CXCL16, PIK3CD, PIK3CB, LRP6, EPHA3, CAVIN1, CARMIL1, PDE2A, PDE9A, PDE4A, PDE4D, PIP5K1A, ADGRL3, EXO1, SYT15, CACNA1B, CACNA1E, CAV1, GPI, CSF2, HRH1, PI4KA, GNA13, ZAN, PCDH10, ANO4, CDC42SE1, CDH4, MUCL1
	GO:0005829 cytosol	ARFGAP1, ARL4A, SDE2, PI4K2A, ELP6, TYMP, NEXMIF, SAMD4A, HPGDS, PIK3CD, PIK3CB, DTL, EPHA3, CAVIN1, CHPF, PPP6C, PDE1B,CARMIL1, PDE2A, PDE3A, PDE9A, PDE3B, PDE4A, PDE4D, PDE5A, TBK1, PIP5K1A, CRLF1, GMNN, IGF2BP2, MON2, GPI, POLR2D, PANK2, HRH1, PI4KA, PI4KB, PPP2R5A, RBMS3, GNA13, LIMK1, PPP4R4, MEF2A, THUMPD3, KIF15, CCT2
	GO:0005654 nucleoplasm	MSH2, ARL4A, SDE2, ZNF85, ERCC3, NFIB, COA6, NEXMIF, HPGDS, PIK3CB, SUMO3, DTL, EPHA3, CAVIN1, APPBP2, RNF112, ESRRG, PPP6C, RBBP5, CARMIL1, PDE9A, CTDSPL2, PDE4A, SENP1, TBK1, PIP5K1A, LIN52, ZNF470, GMNN, EXO1, UBE2T, GPI, ZNF367, DHRS2, RFC5, POLR2D, DNA2, MEF2A, ZNF232, CLUAP1, HIF1A, CDK4
GO_MF	GO:0005515 protein binding	ORAI2, KCNH3, MSH2, PIWIL1, SDE2, RALB, XIRP1,GSE1, TYMP, RRAS2, HPGDS, MFHAS1, PIK3CD, PIK3CB, DTL, LRP6, EPHA3, CRELD2, APPBP2, RNF112, ESRRG, RBBP5, PDE1B, PDE2A, PDE3A, IGFN1, PDE9A, SOHLH1, PDE3B, PDE4A, PDE4D, PDE5A, ACAN, PIP5K1A, LIN52, CRLF1, TMBIM4, VEGFC, CACNA1B, UBE2T, GNS, GPI, RFC5, DNA2, PI4KA, PI4KB, PPP2R5A
	GO:0042802 identical protein binding	TBK1, CAV1, PGF, EMILIN3, LRP6, PCSK1, CAVIN1, ESRRG, PDE2A, PDE9A
	GO:0004712 protein serine/threonine/tyrosine kinase activity	TBK1, CDK18, CDK4, EPHA3, LIMK1

Table 10.

Summary of candidate genes detected by our method in all three databases that have been annotated to the top 5 KEGG enriched pathways in these two microarray datasets

Data set	KEGG Pathways	Annotated Candidate Genes
GSE20347	hsa05200:Pathways in cancer	BCR, AR, APPL1, CKS2, RARA, STAT1, MAP2K1, NFKB2, PIK3CD, PIK3CB, TXNRD1, FLT3LG, PIK3R2, PIK3R1, FADD, RAC2, FAS, RAF1, TGFB2, GNAQ, GNB5, GSK3B, MAPK1, MAPK10, KRAS
	hsa04010:MAPK signaling pathway	PRKACB, FLT3LG, EPHA2, RAC2, FAS, MAP2K1, RAF1, MAP3K20, TGFB2, NFKB2, MAPK1, MAPK10, KRAS
	hsa05169:Epstein-Barr virus infection	PSMD4, PIK3R2, PIK3R1, FADD, LYN, CD3D, STAT1, PSMC2, FAS, NFKB2, PIK3CD, PIK3CB, MAPK10
	hsa04151:PI3K-Akt signaling pathway	FLT3LG, EPHA2, PIK3R2, PIK3R1, MAP2K1, RAF1, MCL1, GNB5, GSK3B, PIK3CD, PIK3CB, ITGB7, MAPK1, KRAS
	hsa05171:Coronavirus disease - COVID-19	RPL22, PIK3R2, PIK3R1, STAT1, PIK3CD, PIK3CB, MAPK1, MAPK10
GSE23400	hsa05200:Pathways in cancer	PRKCB, BCR, AR, HIF1A, APPL1, RARA, STAT1, STAT2, MAP2K1, PIK3CD, TXNRD1, PIK3R1, FADD, RAC2, IL15, RAF1, GSK3B
	hsa04010:MAPK signaling pathway	PRKCB, CD14, RAC2, MAP2K1, RAF1
	hsa04151:PI3K-Akt signaling pathway	PIK3R1, MAP2K1, RAF1, GSK3B, PIK3CD
	hsa05169:Epstein-Barr virus infection	ENTPD1, HLA-C, HLA-F, HLA-G, PIK3R1, FADD, STAT1, STAT2, PSMC1, TAP1, ISG15, CXCL10, PIK3CD, OAS3
	hsa05171:Coronavirus disease - COVID-19	C3AR1, PRKCB, PIK3R1, IFIH1, C1QB, C1QA, STAT1, STAT2, ISG15, CXCL10, PIK3CD, OAS3, FCGR2A
GSE130078	hsa01100:Metabolic pathways	PIP5K1A, PIGT, HPGDS, PIK3CD, PIK3CB, ST6GALNAC5, ST6GAL1, MOGS, ARSB, NAT8B, PI4K2A, GNS, GPI, CHPF, PANK2, PDE1B, PDE2A, PDE3A, TYMP, PDE9A, PI4KA, PI4KB, PDE3B, PDE4A, PDE4D, PDE5A, SGPP1
	hsa05200:Pathways in cancer	HIF1A, PGF, MSH2, CDK4, PIK3CD, PIK3CB, LRP6, VEGFC, CSF2RB, RALB, GNA13
	hsa04144:Endocytosis	PIP5K1A, CAV1, ARFGAP1, FOLR3
	hsa04010:MAPK signaling pathway	RRAS2, VEGFC, PGF, CACNA1B, CACNA1E
	hsa05165:Human papillomavirus infection	TBK1, CDK4, PIK3CD, PIK3CB, PPP2R5A

Literature Trace

Zhu et al. [13] highlight that prothymosin alpha (PTMA) expression was up-regulated in ESCC tissues, thus presenting PTMA as a potential candidate for ESCC. Tang et al. [14] indicated that the expression of PTPRF interacting protein alpha 1 is significantly increased and is related to some malignant clinical features and poor outcomes in ESCC patients, thus establishing it as a valuable biomarker for early detection, treatment formulation and prognosis evaluation for ESCC. Jiang et al. [15] suggested that downregulation of VGLL4 was very important in the progression of ESCC, and restoring the function of VGLL4 might be a promising therapeutic strategy for ESCC. In Shen et al. [16], homer scaffolding protein 3 (HOMER3) is one of the three genes presented as candidate cancer-associated genes and may play a tumorigenic role in ESCC. Ma et al. [17] summarized that upregulation of Proteasome 26S subunit non-ATPase 4 (PSMD4) promotes the progression of ESCC through the reduction of ERS-induced cell apoptosis. Chen et al. [18] found that overexpression of DNA methyltransferase 3b (DNMT3b) is responsible for more aggressive tumor growth and resistance to treatment in ESCC and is linked to activated STAT3 signaling. Liu et al. [19] concluded that Phosphoserine Aminotransferase 1 (PSAT1) expression was elevated in ESCC tissues compared to normal esophageal tissues and increase in the same is significantly associated with stage of disease, lymph node metastasis, distant metastasis and poor prognosis. Findings by Cheng et al. [20] suggested that through activation of the Akt and Erk1/2 signaling pathways, Non-POU Domain Containing Octamer Binding (NONO) plays a potent role in multiple biological aspects of ESCC. Wada et al. [21] highlighted the clinically important implications associated with Transferrin Receptor (TFRC) and concluded that it offers an independent prognostic factor. By employing Cox regression He et al. [22] demonstrated the prognostic value of Canopy FGF Signaling Regulator 2 (CNPY2) for ESCC. Yu et al. [23] demonstrated that Myeloid cell leukemia 1 (MCL-1) contributes to the development of ESCC. Yang et al. [24] concluded that a lower expression of Processing Of Precursor 7 (POP7) predicts a worse prognosis in esophageal cancer. Qiu et al. [25] suggested that through activation of AKT1/mTOR signaling pathway, maintenance complex component 7 (MCM7) promotes tumor cell proliferation, colony formation and migration of ESCC cells. Choy et al. [26] and [27] further suggested MCM7 as a more sensitive proliferation markers for evaluation and for predicting various clinical outcomes of ESCC respectively. Miyazaki et al. [28] concluded that ephrin receptor A2 (EphA2) overexpression appears to be related to poor degree of tumor differentiation and lymph node metastasis in ESCC. Ma et al. [29] suggested that Karyopherin $\alpha$ 2 (KPNA2) protein levels were high in ESCC tumors, and siRNA against KPNA2 could inhibit the growth of ESCC cells, suggesting it may be a new potent marker and therapeutic target for ESCC. Sakai et al. [30] further concluded that KPNA2 expression is associated with poor differentiation, tumor invasiveness, and tumor proliferation in ESCC. Wang et al. [31] identified kinesin family member 4A (KIF4A) as a facilitator of proliferation, cell cycle, migration, and invasion of ESCC in vivo and in vitro. Similarly, Sun et al. [32] stated that through the Hippo signaling pathway, KIF4A regulates the biological function of ESCC cells thus promoting ESCC cell proliferation and migration. Kita et al. [33] demonstrated that the expression of cyclin-dependent kinase subunit 2 (CKS2) in ESCC was elevated relative to levels in normal tissue, and that CKS2 overexpression is associated with the depth of tumor invasion, lymphatic invasion, clinical stage, distant metastasis and poor prognosis. Zheng et al. [34] found that the expression of cytokine induced apoptosis inhibitor 1 (CIAPIN1) was statistically correlated with the degree of differentiation, depth of invasion, and lymph node metastasis of ESCC and thus has been considered as a valuable prognostic indicator in ESCC. Zhao et al. [35] highlighted that Protein arginine methyltransferase 1 (PRMT1) activates and maintains esophageal TICs by mediating transcription alteration through histone H4 arginine methylation. Zhou et al. [36] highlighted that PRMT1 activates Hedgehog signaling and up-regulated the expression of target genes downstream of Hedgehog signaling thus taking an oncogenic role of PRMT1 in the progression of ESCC.

Zhang et al. [37] provided evidence that Human Leukocyte Antigen-F (HLA-F) antigen expression was associated with survival in patients with ESCC. Yie et al. [38] established that Human Leukocyte Antigen-G (HLA-G) expression has a strong and independent prognostic value in human ESCC. According to Sato et al. [39], high  chemokine (CXC motif) ligand 10 (CXCL10) expression is an independent prognostic factor and has the potential to serve as a clinically useful marker of the need for adjuvant chemotherapy after surgery in patients with advanced thoracic ESCC. Yuan et al. [40] suggested the tumor promotion role of Interferon-stimulated gene 15 (ISG15) in ESCC via c-MET/Fyn/$\alpha$-catenin pathway. Yu et al. [41] and Wang et al. [42] identified that Cyclin-dependent kinase inhibitor 3 (CDKN3) regulates tumor progression through activation of AKT signaling pathway in ESCC. Liu et al. [43] further suggested that CDKN3 acted as an oncogene in human ESCC and may accelerate the G1/S transition by affecting CyclinD-CDK4 complex via regulating pAKT-p53-p21 axis and p27 independent of AKT. Preliminary studies by Hu et al. [44] suggested that disks large-associated protein 5 (DLGAP5) promotes cell proliferation in ESCC. According to Liu et al. [45], vaccinia-related kinase (VRK) serine/threonine kinase 1 promotes CDDP resistance through c-MYC by activating c-Jun and potentiating a malignant phenotype in ESCC. Liu et al. [46] provided a potential target for the immuno-oncology effect of roteasome alpha-subunit 3 (PSMA3) in ESCC therapy. Wang et al. [47] detected the major role of Pleckstrin-2 (PLEK2) in driving metastasis and chemoresistance in ESCC by regulating LCN2. Qu et al. [48] found that Component 3a Receptor 1 (C3AR1) might be the cause of an immunosuppressive microenvironment by affecting the polarization of macrophages to M2 phenotype and lead to the progression of ESCC. Zhang et al. [49] suggested that Signal Transducer and Activator of Transcription-1 (STAT1) is a tumor suppressor in ESCC. According to Shao et al. [50], Hypoxia-inducible factor 1$\alpha$ (HIF-1$\alpha$), p53, and vascular endothelial growth factor (VEGF) are important factors that facilitate tumor progression. The results from the study conducted by Hu et al. [51] indicated that HIF-1$\alpha$ promotes metastasis of ESCC by targeting SP1 in a hypoxic microenvironment. Bolidong et al. [52] suggested that via cyclin D1/CDK4-mediated cell cycle progression Glycogen synthase kinase 3$\beta$ (GSK3$\beta$) has a tumor promoting role in ESCC. According to Gao et al. [53], GSK3$\beta$ expression promotes ESCC progression through STAT3 in vitro and in vivo, and GSK3$\beta$-STAT3 signaling could be a potential therapeutic target for ESCC treatment.

According to Kato et al. [54] and [55] caveolin-1 (CAV1) is a biomarker for ESCC. Lu et al. [56] and Shu et al. [57] indicated that insulin-like growth factor 2 mRNA-binding protein 2 (IGF2BP2) serves a major carcinogenic role in ESCC. According to [58], the expression of vascular endothelial growth factor C (VEGF-C) correlates with lymph node metastasis and poor prognosis. Similarly, [59] suggests that VEGF-C expression in ESCC may play a great key role in lymphatic spread. Feng et al. [60] indicated that ras-like without CAAX1 (RIT1) displays tumor-suppressing functions in ESCC, and these functions were carried out by inhibiting MAPK and PI3K/AKT signaling pathway, inhibiting EMT, and downregulating cancer stemness of ESCC cells. Zhou et al. [61] demonstrated that Dehydrogenase/reductase member 2 (DHRS2) had an important part in ESCC development and progression. Li et al. [62] suggested a tumor suppression function for RNA Binding Motif Single Stranded Interacting Protein 3 (RBMS3) gene in ESCC. According to Wang et al. [63], UBE2T is involved in the development of ESCC, and gene signatures derived from UBE2T-associated genes are predictive of prognosis in ESCC. Gao et al. [64] demonstrated that High Mobility Group Box 3 (HMBG3) may be a potential molecular marker for predicting the prognosis of ESCC patients. According to Huang et al. [65], cyclin-dependent kinase 4 (CDK4) amplification was identified as an independent prognostic factor for survival, which could be incorporated into the tumor-node-metastasis staging system to refine risk stratification of patients with esophageal squamous cell carcinoma. Ling et al. [66] suggested that MutS Homolog 2 (MSH2) methylation in the plasma would be a good predictor of DFS for these ESCC patients before oesophagectomy. Xu et al. [67] identified Estrogen-related receptor gamma (ESRRG) as one of four molecular markers that may be helpful in the diagnosis and treatment of ESCC. Chen et al. [68] demonstrated that silencing EphA3 in KYSE410 cells triggers epithelial–mesenchymal transition, and promoted cell migration and invasion in ESCC. Luo et al. [69] found that knockdown of Insulin-like growth factor binding protein-3 (IGFBP-3) confers resistance to the cell killing effects of IR on ESCC in vitro and in vivo. Zhao et al. [70] indicated that the increased ESCC chemosensitivity might be dependent on IGFBP-3 upregulation through EGFR-dependent pathway. Furthermore, according to Luo et al. [71], high level of IGFBP-3 expression in ESCC associates with early clinical stages and are predictive for favorable survival of the patients treated with radiotherapy. According to Wang et al. [72], carbohydrate sulfotransferase 15 (CHST15) promotes the proliferation of TE-1 cells via multiple pathways in ESCC.

Discussion

In Table 11, we give a detailed summary of all DEGs that have been identified by our method as candidates for potential biomarkers for ESCC. In our method, we consider strong literature evidence for association with ESCC and five other SCCs related to ESCC as the necessary criterion for a candidate gene to be a potential biomarker, and the findings from literature are summarized in Table 11. In the table, we also highlight the enriched GO terms and pathways to which the candidate genes has been annotated. Furthermore, it also details whether the same is a hub gene, a transcription factor (TF) or whether it is upregulated or down-regulated. A DEG is upregulated if $logFC>0$ and downregulated when $logFC<0$. We take into consideration logFC values calculated by limma for the microarray datasets, and edgeR in the RNA-Seq dataset.

Table 11.

Summary of potential biomarkers identified by our proposed framework (Fig. 1)

	Detected hub gene	Enriched GO Term	Enriched cancer pathway(s)	HG or TED	TF?	+/–	Literature Evidence
GSE20347	PTMA	BP, CC, MF	Nil	Yes	Yes (Fig 5a)	+	ESCC [13], HNSCC [73]
	PPFIA1	BP, CC	Nil	HG	No		ESCC [14], HNSCC [74]
	PIK3CB	BP, CC, MF	hsa05200,hsa05169,hsa04151,hsa05171,hsa05165, hsa05417,hsa05161,hsa04510,hsa04015,hsa05162, hsa05164,hsa05205,hsa05166,hsa04210,hsa05418, hsa05167,hsa05163,hsa04380,hsa05146,hsa05222, hsa04933,hsa05135,hsa05215,hsa05017, hsa04722, hsa04722,hsa05170,hsa04810,hsa04066,hsa01100, hsa05131,hsa05010,hsa04024,hsa04014,hsa05210, hsa05221,hsa05160,hsa05212,hsa04218,hsa04630, hsa05203	HG	No		OSCC [75]
	VGLL4	BP, CC, MF	Nil	HG	Yes (Fig. 5a)	+	ESCC [15]
	HOMER3	BP, CC, MF	hsa04068,hsa04724	HG	No	+	ESCC [16]
	SHC1	BP, CC, MF	hsa04510,hsa04722,hsa04014,hsa04072,hsa04935, hsa04910,hsa01521,hsa04062,hsa05220,hsa05225, hsa05226,hsa04917,hsa05224,hsa04915,hsa04012, hsa05214,hsa04926,hsa01522,hsa04650,hsa05100	HG	No	+	LSCC [76]
	PSMD4	BP, CC, MF	hsa05169,hsa05017,hsa05022,hsa05010,hsa05020, hsa05016,hsa05014,hsa05012,hsa03050	HG	No	+	ESCC [17]
	CLIC4	BP, CC, MF	Nil	HG	No	+	HNSCC [77]
	DNMT3B	BP, CC, MF	hsa01100	HG	Yes (Fig. 5b)	+	ESCC [78], HNSCC [79], OSCC
	MYO1B	BP, CC, MF	hsa05130	HG	No	+	HNSCC [80]
	PSAT1	CC, MF	hsa01100,hsa01200	HG	No	+	ESCC [19]
	SLC3A2	BP, CC, MF	hsa04150,hsa04216	HG	NO	+	OSCC [81]
	NONO	BP, CC, MF	Nil	HG	Yes (Fig. 5c)		ESCC [20]
	TFRC	BP, CC, MF	hsa04066,hsa04145,hsa04640,hsa04144,hsa04216	HG	No	+	ESCC [21]
	SENP5	BP, CC	Nil	HG	No	+	OSCC [82]
	RCN1	BP, CC, MF	Nil	HG	No	+	OSCC [83]
	CNPY2	BP, CC, MF	Nil	HG	No	+	ESCC [22]
	SBF1	BP, CC, MF	Nil	HG	No	–	HNSCC [84]
	MCL1	BP, CC, MF	hsa04151,hsa04210,hsa04630	HG	No	+	ESCC [23], OSCC [85]
	POP7	BP, CC, MF	Nil	HG	No	+	ESCC [24]
	MCM7	BP, CC, MF	hsa04110,hsa03030	HG	Yes (Fig. 5d)	+	ESCC [25–27], OSCC [86]
	POM121	BP, CC, MF	hsa03013,hsa05014	HG	No		OSCC [87]
	PSMC2	BP, CC, MF	hsa05169,hsa05017,hsa05022,hsa05010,hsa05020, hsa05016,hsa05014,hsa05012,hsa03050	HG	Yes (Fig 5d)	+	OSCC [88]
	EPHA2	BP, CC, MF	hsa04010,hsa04151,hsa04015,hsa04014,hsa04360	HG	No	+	ESCC [28] ,LSCC [89], HNSCC [90, 91]
	KPNA2	BP, CC, MF	hsa05164,hsa03013,hsa05207	HG	No	+	ESCC [29, 30], OSCC [92]
	KIF4A	BP, CC, MF	Nil	HG	No	+	ESCC [31, 32], OSCC [93]
	CCT4	BP, CC, MF	Nil	HG	Yes (Fig. 5e)	+	HNSCC [94]
GSE20347	CKS2	BP, CC, MF	hsa05200,hsa05222	HG	No	+	ESCC [33], TSCC [95]
	CIAPIN1	BP, CC, MF	Nil	HG	No	+	ESCC [34]
	PRMT1	BP, CC, MF	hsa04068,hsa04922	HG	No	+	ESCC [35, 36], HNSCC [96]
	HPRT1	BP, CC, MF	hsa01100	HG	No	+	OSCC [97], HNSCC [98]
	NDD4L	Nil	Nil	HG	No	+	OSCC [99]
	KRAS	BP, CC, MF	hsa05200,hsa04010,hsa04151,hsa05165,hsa05417, hsa05161,hsa04015,hsa05205,hsa05166,hsa04210, hsa05167,hsa05163,hsa04933,hsa05215,hsa05022, hsa04722,hsa05170,hsa04810,hsa04066,hsa05010, hsa04014,hsa05210,hsa05221,hsa05160,hsa05212, hsa04218,hsa05203,hsa04072,hsa04919,hsa04935, hsa04660,hsa04068,hsa04910,hsa01521,hsa04062, hsa05211,hsa05220,hsa05225,hsa04625,hsa05226, hsa05223,hsa05230,hsa04917,hsa05224,hsa04540, hsa04915,hsa04550,hsa05218,hsa04012,hsa05214, hsa04071,hsa04720,hsa04926,hsa05213,hsa05219, hsa05208,hsa05235,hsa04140,hsa04662,hsa04912, hsa05231,hsa01522,hsa04650,hsa04929,hsa04664, hsa04360, hsa04730,hsa04916,hsa04371,hsa04921, hsa05216,hsa04725,hsa04213,hsa04211,hsa04137, hsa04960,hsa04370,hsa04150,hsa05207,hsa04726	TED	No	–	LSCC [100]
	PIK3R2	BP, CC, MF	Nil	TED	No	–	LSCC [101]
GSE23400	TAP1	BP, CC, MF	hsa05169,hsa05163,hsa05170,hsa04145,hsa04612	HG	No	+	TSCC [102]
	HLA-F	BP, CC, MF	hsa05169,hsa05165,hsa05166,hsa05163,hsa05167, hsa05170,hsa05203,hsa04218,hsa04145,hsa04144, hsa04612,hsa05416,hsa05320,hsa04940,hsa05332, hsa04514,hsa05330	HG	No	+	ESCC [37]
	IFIT3	BP, CC, MF	Nil	HG	No	+	OSCC [103, 104]
	HLA-G	BP, CC, MF	hsa05163,hsa05167,hsa05170,hsa05203,hsa04218, hsa04145,hsa04144,hsa04612,hsa05416,hsa05320, hsa04940,hsa05332,hsa04514,hsa05330	HG	No	+	ESCC[38] , HNSCC [105, 106], OSCC [107]
	IFI44L	BP, CC, MF	Nil	HG	No	+	OSCC [108]
	CXCL10	BP, CC, MF	hsa05169,hsa05171,hsa05164,hsa05160,hsa04060, hsa04062,hsa04668,hsa04620,hsa04622,hsa04657, hsa04061,hsa04623	HG	No	+	ESCC [39], TSCC [109], HNSCC [110]
	IFIT1	BP, CC, MF	hsa05160	HG	No	+	OSCC [103], HNSCC [111]
	ISG15	BP, CC, MF	hsa05169,hsa05171,hsa05165, hsa04622	HG	No	+	ESCC [40] , OSCC [112, 113]
	CDKN3	BP, CC, MF	Nil	HG	No	+	ESCC [41–43]
	DLGAP5	BP, CC, MF	Nil	HG	No	+	ESCC [44]
	VRK1	BP, CC, MF	Nil	HG	No	+	ESCC [45], HNSCC [114]
	PSMA3	BP, CC, MF	hsa05017,hsa05022,hsa05010,hsa05020,hsa05016, hsa05012,hsa05014,hsa03050	HG	No	+	ESCC [46], OSCC [115]
	PLEK2	BP, CC	Nil	HG	Yes (Fig. 5f)	+	ESCC [47], HNSCC [116]
	CD163	BP, CC, MF	Nil	HG	No	+	HNSCC [117], OSCC [118, 119]
	ITGB2	BP, CC, MF	hsa04015,hsa05166,hsa05152,hsa05146,hsa04810, hsa04145,hsa04610,hsa05134,hsa05140,hsa04613, hsa05416,hsa04650,hsa04670,hsa04390,hsa05133, hsa05323,hsa05144,hsa04514	HG	No	+	OSCC [120]
	FCGR2A	BP, CC, MF	hsa05171,hsa05152,hsa04380,hsa05135,hsa04145, hsa05130,hsa04611,hsa04666,hsa05140,hsa04613	HG	No	+	ESCC [121] , HNSCC [122, 123]
	C3AR1	BP, CC, MF	hsa05171,hsa04936,hsa04080,hsa04610	HG	No	+	ESCC [48]
	HIF1A	BP, CC, MF	hsa05200,hsa05205	TED	Yes 6a	+	ESCC[50, 51], OSCC [124–126], TSCC [127]
	CAV1	BP, CC, MF	hsa04144,hsa04510,hsa05205	HG	No	+	ESCC[54, 55], OSCC[128], HNSCC[129], TSCC[130]
GSE130078	IGF2BP2	BP, CC, MF	Nil	HG	No	+	ESCC[56, 57], OSCC[131, 132]
	VEGFC	BP, CC, MF	hsa05200,hsa04010,hsa04151,hsa04015,hsa04510, hsa04020,hsa04926	HG	No	+	ESCC[58, 59], OSCC[133, 134]
	HIF1A	BP, CC, MF	hsa05200,hsa05205	HG	Yes 6d	+	ESCC[50, 51], OSCC [124–126], TSCC [127]
	GNA13	BP, CC, MF	hsa05200,hsa04270	HG	No	+	LSCC [135]
	RIT1	BP, CC	Nil	HG	No	-	ESCC [60]
	DHRS2	BP, CC	Nil	HG	No	+	ESCC [61]
	RBMS3	BP, CC, MF	Nil	HG	No	+	ESCC [62], LSCC [136]
	RFC5	BP, CC, MF	Nil	HG	No	+	LSCC [137]
	UBE2T	BP, CC, MF	Nil	HG	No	+	ESCC [63]
	HMGB3	BP, CC, MF	Nil	HG	Yes 6f	+	ESCC [64]
	LMO4	BP, CC, MF	Nil	HG	Yes 6f	+	HNSCC [138], TSCC[139]
	CDK4	BP, CC, MF	hsa05200,hsa05165,hsa04151,hsa04530,hsa05224, hsa04110	HG	No	+	ESCC [65], HNSCC[140, 141]
	EXO1	BP, C C, MF	Nil	HG	No	+	HNSCC[142]
	MSH2	BP, CC, MF	hsa05200	HG	No	+	ESCC[66], HNSCC [143]
	TBK1	BP, CC, MF	hsa05165	HG	No	+	HNSCC [144]
	HPDL	CC, MF	Nil	HG	No	+	TSCC [145]
	LRP6	BP, CC, MF	hsa05200,hsa05224	HG	No	+	OSCC [146]
	ESRRG	BP, CC, MF	Nil	HG	Yes 6e	–	ESCC [67], LaSCC [147]
	EPHA3	BP, CC, MF	Nil	HG	Yes	+	ESCC [68]
	TRAM2	CC	Nil	HG	Yes	+	OSCC [148]
	IGFBP3	BP, CC, MF	Nil	HG	Yes	+	ESCC [69–71], OSCC[149, 150]
	CLPTM1L	BP, CC	Nil	HG	No	+	OSCC[151, 152]
	CHST15	CC	Nil	HG	No	+	ESCC[72]
GSE20347 & GSE23400	TXNRD1	BP, CC, MF	hsa05200,hsa05225	TED	No	+	OSCC [153], HNSCC [154]
	FADD	BP, CC, MF	hsa05200,hsa05169,hsa05165,hsa05161,hsa05162, hsa05164,hsa04210,hsa05167,hsa05152,hsa05163, hsa05022,hsa05170,hsa04936,hsa05010,hsa05160, hsa05132,hsa05142,hsa04668,hsa04620,hsa05130, hsa01524,hsa04622,hsa04657,hsa04217,hsa04621	TED	No	+	OSCC [155], HNSCC [156, 157]
	STAT1	BP, CC, MF	hsa05200,hsa05169,hsa05171,hsa05165,hsa05161, hsa05162,hsa05164,hsa05167,hsa05152,hsa04380, hsa04933,hsa05145,hsa05160,hsa05212,hsa04630, hsa04919,hsa04935,hsa04062,hsa04625,hsa04620, hsa04659,hsa04917,hsa04658,hsa05140,hsa05235, hsa04217,hsa05321,hsa04621	TED	Yes 6b	+	ESCC [49], HNSCC [158, 159]
	AR	BP, CC, MF	hsa05200,hsa05215,hsa05207	TED	Yes 6c	-	OSCC [160, 161]
	RAF1	BP, CC, MF	hsa05200,hsa04010,hsa04151,hsa05165,hsa05161, hsa04510,hsa04015,hsa05164,hsa05205,hsa04210, hsa05167,hsa05152,hsa05163,hsa05215,hsa05022, hsa04722,hsa05170,hsa04810,hsa05010,hsa04024, hsa04014,hsa05210,hsa05221,hsa05160,hsa05212, hsa04218,hsa04630, hsa04072,hsa04919,hsa05132, hsa04935,hsa04660,hsa04068,hsa04910,hsa01521, hsa04062,hsa04022,hsa05211,hsa04928,hsa05220, hsa05225,hsa04625,hsa05226,hsa05223,hsa05230, hsa04917,hsa05224,hsa04540,hsa04915,hsa04550, hsa05218,hsa04012,hsa05214,hsa04071,hsa04720, hsa04926,hsa05213,hsa04666,hsa05219,hsa05208, hsa05235,hsa04140,hsa04662,hsa04270,hsa04912, hsa05231,hsa01522,hsa04650,hsa04613,hsa04929, hsa04664,hsa04360,hsa04730,hsa04916,hsa04371, hsa04921,hsa04370,hsa04150,hsa05207,hsa04726	TED	No	–	OSCC [162]
GSE20347 & GSE23400	GSK3B	BP, CC, MF	hsa05200,hsa04151,hsa05165,hsa05417,hsa04510, hsa05162,hsa05167,hsa05163,hsa05135,hsa05215, hsa05022,hsa04722,hsa05131,hsa05010,hsa05210, hsa05160,hsa04919,hsa04935,hsa04660,hsa04910, hsa01521,hsa04062,hsa05225,hsa05226,hsa04110, hsa04917,hsa05224,hsa04932,hsa04550,hsa04012, hsa04728,hsa04934,hsa05020,hsa05415,hsa05213, hsa04931,hsa04662,hsa04657,hsa04360,hsa04916, hsa04390,hsa04310,hsa04150	TED	No	+	ESCC [52, 53], OSCC [163, 164]
	MAP2K1	BP, CC, MF	hsa05200,hsa04010,hsa04151,hsa05165,hsa05161, hsa04510,hsa04015,hsa05164,hsa05205,hsa05166, hsa04210,hsa05167,hsa05163,hsa04380, hsa05135,hsa05215,hsa05022,hsa04722,hsa05170, hsa04810,hsa04066,hsa05010,hsa04024,hsa04014, hsa05210,hsa05221,hsa05160,hsa05212,hsa04218, hsa04072,hsa04919,hsa05132,hsa04935,hsa04660, hsa04068,hsa04910, hsa01521,hsa04062,hsa04022, hsa05211,hsa04928,hsa05220,hsa04668,hsa05225, hsa04620, hsa05226,hsa05223,hsa05230,hsa04917, hsa05224,hsa04540,hsa04915,hsa04550,hsa05218, hsa04012,hsa04934,hsa05214,hsa04071,hsa04720, hsa04926,hsa05213,hsa04666,hsa05219,hsa05208, hsa05235,hsa04140,hsa04662,hsa04270,hsa04912, hsa05231,hsa01522,hsa04650,hsa04613,hsa04929, hsa04664,hsa04360,hsa04730,hsa04916,hsa04371, hsa04921,hsa04370,hsa05216,hsa04725,hsa04150, hsa05207,hsa04726	TED	No	–	HNSCC [165]

HG hub gene, TED Top Enriched DEG, DEG annotated to most enriched GO term in all three databases as well as to the most enriched pathway, TF transcription factor, + Upregulated DEG and – Downregulated DEG

The biological relevance of a candidate to its respective dataset is considered based on three criteria:

1. Annotated to at least one GO term in 2 of 3 GO databases with $p\mathrm{value}\le 0.05$,

2. Annotated to at least one KEGG pathway with $p\mathrm{value}\le 0.05$, and

3. It’s a TF and thus exhibits regulatory behavior towards other DEGs in the network.

For a candidate gene to be considered a potential biomarker, we consider following four cases.

• Case 1: Strong literature evidence of association with ESCC and biologically relevant to its dataset based on all three criteria a,b and c,

• Case 2: Strong literature evidence of association with ESCC and biologically relevant to its dataset based on criteria a and b,

• Case 3: Strong literature evidence of association with ESCC and biologically relevant to its dataset based on criteria a or b,

• Case 4: Biologically relevant to its dataset using all three criteria a,b and c, and has literature evidence of association with previously mentioned 5 SCCs related to ESCC, namely, oral SCC, Lung SCC, Tongue SCC, Head and Neck SCC and Laryngeal SCC.

All candidate genes that fall under Case 1 and Case 2 are considered potential biomarkers for ESCC because of existing evidence of association with ESCC in the form of other works while our biological validation of these genes establishes their relevance to their respective datasets. For candidate genes that fall under Case 3, although there is strong literature evidence of association with ESCC, we have weak evidence of their biological relevance to their datasets. On the other hand, for candidate genes that fall under Case 4, although we strongly validate their biological relevance to their datasets, there is only literature evidence of association with other SCCs related to ESCC. For both these cases, the candidates can be considered probable potential biomarkers, but need further in-depth analysis.

Top Enriched DEGs (TEDs), STAT1 and HIF1A detected in both microarray datasets (GSE20347 and GSE23400) and GSE130078, respectively, belong to Case 1. In GSE20347, two candidates DNMT3B and MCM7 also belong to Case 1. Thus, STAT1, HIF1A, DNMT3B and MCM7 are potential biomarkers for ESCC. GSK3B is a TED detected in both microarrays, and belongs to Case 2. In dataset GSE20347, 9 candidate genes HOMER3, PSMD4, PSTAT1, TFRC, MCL1, EPHA2, KPNA2, CKS2 and PRMT1 belong to Case 2, and thus are potential biomarkers for ESCC. Similarly, 7 candidate genes HLA-F, HLA-G, CXCL10, ISG15, PSMA3, FCGR2A and C3AR1 are potential biomarkers for ESCC as they fall under Case 2. Four candidate genes in the RNASeq dataset GSE130078, CAV1, VEGFC, CDK4 and MSH2 fall under Case 2 and are potential biomarkers for ESCC.

Three candidates genes in GSE20347, PTMA, VGLL4 and NONO fall in Case 3. In other words, although there are other works that establish their role as potential biomarkers for ESCC, the biological relevance to their respective datasets is not that strong. However, they can still be regarded as probable potential biomarkers for ESCC, but need further in-depth validation. Similarly in GSEE23400 and GSE130078, one (PLEK2) and 2 (HMGB3 and ESRRG) genes fall under Case 3. PSMC2 detected in GSE20347, on the other hand falls under Case 4. We validate its strong association with the dataset as this candidate gene has been annotated to GO terms in all three GO databases as well as several enriched pathways. They further exhibit regulatory behavior in a GRN, but there are no previous works that relate the same to ESCC. However, its worth mentioning that there is literature evidence that identify PSMC2 as potential biomarker for OSCC. Similarly, the TED identified in the two microarray datasets, AR, also falls under Case 4. Both, PSMC2 and AR are probable potential biomarkers for ESCC, but need further in-depth analysis.

In Table 12, we put forward a comparison between our work and two recent works presented by Patowary et al. [166] and Hu. et al. [44] that perform DE analysis by employing approaches and methods similar to our work.

Table 12.

Comparison of our method with two recent works that employ DE analysis on ESCC datasets

Parameter	Our Method	Patowary et al. [166], 2020	Hu. et al. [44], 2020
Datasets	Microarray: GSE20347 and GSE23400, RNA-Seq: GSE130078	Microarray: GSE20347 and GSE23400, RNA-Seq: SRP064894	Microarray: GSE20347 and GSE26886
Identification of DEGs	DEA methods: Limma, SAM, EBAM, limma-voomm, edgeR and DESeq2 with $p\le 0.05$	DEA methods: Limma, SAM, EBAM, limma-voomm, edgeR and DESeq2 with $p\le 0.01$	Limma; $|log2foldchange(FC)|)>2$ and $adjustedP<0.05$
Consensus function	(a) Common genes given by the methods with $p\le 0.05$ (b) For each method, DEGs not in common genes with $q\le 0.05$(RNA-Seq) or $localfdr\le 0.05$ (Microarray)	(a) Common genes given by the methods with $p\le 0.01$ and (b) top ranked DEGs (other than common) given by each method with $p\le 0.001$	Up- and downregulated DEGs common to both microarray datasets
Module Extraction	Heiarchical Tree Classification; Eigenmodule selection and MEDissThres threshold merging	Hierarchical Tree Classification; Eigenmodule selection and MEDissThres threshold merging	(a) PPI network; Nodes with edge of $>20$ as hub genes; (b) WGCNA and Heiarchical Tree Classification
Preservation Analysis	Yes	Yes	No
Modules of Interest	All non-preserved modules of size larger that 100 nodes	Least Preserved module	Relevance between each module and hub genes identified in PPI network $<0.5$
hub gene	Intramodular Connectivity	Intramodular Connectivity; Degree and confidence in PPI Network	Nodes with edge of $>20$ in PPI network
Candidate Identification	Top 20 hub genes in all modules of interest; DEGs that are annotated to most enriched GO term in all three GO databases as well as the most enriched pathway	DEG with highest intramodular connectivity in each module of interest	Top hub gene in module of interest
Enrichment Analyses	DAVID; GO and KEGG pathway	DAVID; GO and pathway	DAVID; GO and pathway and GSEA
Diseases considered for literature trace	ESCC and five other SCCs, namely oral SCC, Lung SCC, Tongue SCC, Head and Neck SCC, Laryngeal SCC	ESCC and all other cancers	ESCC and all other cancers
Identified potential biomarkers	STAT1,HIF1A,DNMT3B, MCM7,GSK3B,HOMER3, PSMD4,PSTAT1,TFRC, MCL1,EPHA2,KPNA2, CKS2,PRMT1,HLA-F, HLA-G,CXCL10,ISG15, PSMA3,FCGR2A,C3AR1, CAV1,VEGFC,CDK4 and MSH2	COL27A1,SOX11,BAG6, TOP3,CDC6,EZH2, COL7A1,G6PD and AKR1C2	DLGAP5

Conclusion

All six methods, three microarray and three RNA-sequencing, employed by the proposed integrative approach based differential expression (DE) Analysis framework were found effective in extracting differentially expressed genes (DEGs) with a p-value of 0.05. We further proposed a consensus function that takes into account the information loss due to the DEGs common to all three respective methods and further employs local fdr (for microarray) and q-value (for RNA-Seq). Through differential co-expression (DCE) and preservation analysis, we studied the behavioral changes among the DEGs under normal and disease circumstances. All non-preserved modules of reasonable sizes are considered modules of interest and analyzed further down the pipeline. DEGs are considered candidates for potential biomarkers for ESCC when they are either: (a) hub genes in the modules of interest, or (b) Top Enriched DEG (TED), i.e., a DEG annotated to the most enriched GO term in all three GO databases as well as the most enriched KEGG pathway in their respective datasets. Our proposed framework was validated on two microarray datasets (GSE20347 and GSE23400) and one RNA-Sequencing dataset (GSE130078). From, 7, 3 and 8 modules of interest in GSE20347, GSE23400 and GSE130078 respectively, 124, 59 and 160 hub genes were identified. The consideration of 22, 18 and 16 TEDs detected by GSE20347, GSE23400 and GSE130078 respectively results in 146, 77 and 176 as candidates for potential biomarkers of ESCC. Biological relevance for each candidate to their respective datasets is analysed based on (a) annotation to enriched GO terms in the GO databases, (b) annotation to enriched KEGG pathways, and (c) if the candidate gene is a transcription factor in a gene regulatory network. Another very important criterion that we considered for a candidate gene to be a potential biomarker is previous literature that has either (a) established them as potential biomarkers for ESCC itself, or (b) established them as potential biomarkers for 5 other SCCs related to ESCC, namely, Oral SCC, Tongue SCC, Lung SCC, Head and Neck SCC and Laryngeal SCC.

Our method identified 4 candidate genes, STAT1, HIF1A , DNMT3B and MCM7, that are transcription factors (TFs), have strong biological relevance to their respective datasets and have previous literature works that establish their role as potential biomarkers in ESCC. Our method identified GSK3B , detected as DEG by both microarray datasets (GSE20347 and GSE23400), as a TED and has both strong literature evidence as potential biomarker of ESCC and we established its strong biological relevance to both microarray datasets. Similarly, nine (HOMER3 , PSMD4, PSTAT1, TFRC, MCL1, EPHA2, KPNA2, CKS2 and PRMT1), seven (HLA-F, HLA-G, CXCL10, ISG15, PSMA3, FCGR2A and C3AR1) and four (CAV1, VEGFC, CDK4 and MSH2) candidates genes in GSE20347, GSE23400 and GSE130078 are established as potential biomarkers for ESCC. We further identified 3 (PTMA, VGLL4 and NONO), 1 (PLEK2) and 2 (HMGB3 and ESRRG) TFs in GSE20347, GSE23400 and GSE130078 respectively, that have strong literature evidence as potential biomarkers of ESCC, but have moderate evidence for biological relevance to their respective datasets, and thus can be regarded as probable potential biomarkers for ESCC. On the other side of the spectrum, the transcription factor AR, a TED that is identified as a DEG in both microarray datasets and PSMC2 have strong biological relevance to their datasets, but have been identified as potential biomarkers for other SCC related to ESCC. They can also be considered probable potential biomarkers for ESCC, but need further in-depth analysis. For future work, there is scope for improvement in the framework in general and consensus function specifically, towards detection of a smaller number of DEGs with minimum loss of relevant information.

Declarations

Conflict of interest

On the behalf of all authors, the corresponding authors states that there is no conflict of interest.