Gene network landscape of mouse splenocytes reveals integrin complex as the A151 ODN-responsive hub molecule in the immune transcriptome

Abstract

Homeostatic restoration of an inflammatory response requires quenching of the immune system after pathogen threats vanish. A continued assault orchestrated by host defense results in tissue destruction or autoimmunity. A151 is the epitome of synthetic oligodeoxynucleotides (ODNs) that curb the immune response by a subset of white corpuscles through repetitive telomere-derived TTAGGG sequences. Currently, the genuine effect of A151 on the immune cell transcriptome remains unknown. Here, we leveraged an integrative approach where weighted gene co-expression network analysis (WGCNA), differential gene expression analysis, and gene set enrichment analysis (GSEA) of our in-house microarray datasets aided our understanding of how A151 ODN suppresses the immune response in mouse splenocytes. Our bioinformatics results, together with experimental validations, indicated that A151 ODN acts on components of integrin complexes, Itgam and Itga6, to interfere with immune cell adhesion and thereby suppresses the immune response in mice. Moreover, independent lines of evidence in this work converged on the observation that cell adhesion by integrin complexes serves as a focal point for cellular response to A151 ODN treatment in immune cells. Taken together, the outcome of this study sheds light on the molecular basis of immune suppression by a clinically useful DNA-based therapeutic agent.

Graphical abstract

This work investigates the molecular mechanisms underlying immune suppression by a clinically useful DNA-based therapeutic agent. Immune cell adhesion by integrin complexes appears to be central to the cellular response to A151 ODN treatment. A151 ODN acts on Itgam and Itga6 to suppress the immune response in mouse splenocytes.

Introduction

The immune system constitutes an effective barrier against a wide range of pathogenic insults, but a misdirected or an overreacting immune response may pose a significant threat to the host, as in autoimmunity. In this regard, this protective system evolved to discriminate infectious non-self from noninfectious self¹ and triggers adaptive mechanisms to fine-tune the magnitude of the response.² One such mechanism is through telomeres, which cap mammalian DNA by repetitive TTAGGG motifs capable of inhibiting immune reactions.³ It has been suggested that release of inhibitory DNA from dying host cells plays a pivotal role in downregulating pathologic inflammatory and autoimmune responses.⁴ This immune suppression by repetitive TTAGGG motifs was successfully mimicked using synthetic TTAGGG₄ oligonucleotides (oligodeoxynucleotides [ODNs]).^3,5 TTAGGG₄ (referred to as A151) is now known to inhibit production of a wide range of chemokines and cytokines invoked by bacteria, including interleukin-6 (IL-6), IL-12, interferon γ (IFNγ), tumor necrosis factor alpha (TNF-α), and MIP2α.^3,6,7 Collectively, A151 is the widely accepted archetypal immunosuppressive ODN and a promising therapeutic candidate,⁴ but the regulatory effect of A151 on the immune cell transcriptome remains unknown.⁶

It can be seen in the literature that transcriptome data have been exclusively used to screen differentially expressed genes as individual entities.⁸ Considering that functionally related genes work in groups to perform biological tasks inside the cell, the degree of interconnectedness among these genes has been largely underestimated. More specifically, genes with similar expression patterns may converge functionally on a single, high-order molecular process, and identifying these co-expressed genes as clusters might better explain their biological role in the complex intracellular environment.⁹ Toward this end, gene-level expression analysis can be coupled with network-level data-mining methods to pave the way for an integrative transcriptome landscape analysis with the potential of yielding more robust, more accurate, and biologically more relevant results. Weighted gene co-expression network analysis (WGCNA)¹⁰ is one such data-mining approach that helps refine transcriptome data that would otherwise be more complex to interpret and utilize in a biological context. This integrative approach has been gaining popularity over the past few years to study the immune transcriptome in mammals, including the mouse.^11,12,13

In the current study, we analyzed the gene network landscape of mouse splenocytes to identify A151 ODN-responsive hub genes in the immune cell transcriptome. To do that, we leveraged an integrative approach in which WGCNA and differential gene expression analysis of our in-house microarray datasets aided our understanding of how A151 ODN suppresses the immune response in mice. We first constructed, using the A151 ODN dataset, a gene co-expression network to define modules specific to A151 ODN with respect to the scrambled ODN dataset. Given these treatment groups, we then established, in the relevant modules, a list of significant genes downregulated by A151 ODN. Using a set of functional enrichment platforms and interaction network analysis tools, we annotated and prioritized this list to restrict our study to a small network of candidate genes. Our experimental and machine learning-based in silico validations identified the potential targets in the A151 ODN co-expression network. We found that cell adhesion by integrin complexes serves as a focal point for cellular response of mouse splenocytes to A151 ODN treatment. Additionally, we demonstrated that a refined set of “co-hub” genes led by Itgam and Itga6 plays a pivotal role in A151 ODN-driven suppression of the immune response. The outcome of this study sheds light on the molecular basis of immune suppression by a clinically useful DNA-based therapeutic agent and may open new avenues for A151 ODN research.

Results

Construction of the gene co-expression network from the A151 ODN dataset

To conduct comparative transcriptomics analyses of mouse splenocytes treated with immunosuppressive A151 ODN, we used array-based expression data (Table 1), as described above. This approach (Figure S1) provided us insight at the gene and network level into the changes in expression patterns of gene clusters in response to A151 ODN treatment in mice. Before the downstream analyses, we first verified the replicate arrays in scrambled and A151 ODN treatment groups separately after within and between-library normalizations (Figures S2 and S4). We also performed batch correction to remove technical and other sources of unwanted or spurious variation from each dataset, which was then confirmed for uniformity and reproducibility (Figures S3 and S5). For the WGCNA algorithm to identify co-expressed gene clusters (or “modules”) and describe correlation patterns of genes in the A151 ODN dataset, the user is required to analyze network topology for a range of soft threshold (β) powers. Our results indicated, at the scale-free fit index (R²) threshold of equal to or higher than 0.9, that the optimum β value that ensures a scale-free topology for the A151 ODN network construction is 12 (Figure 1A). A hierarchical tree of 7,763 filtered genes, represented as a dendrogram, identified a total of 12 modules with a range of intermodular correlations (Figure 1B). The topological overlap matrix (TOM) plot we generated using all filtered genes as a heatmap revealed the degree of overlap among the modules (Figure 1C). Last, the network representing the adjacencies among the module eigengenes, where an eigengene refers to the first principal component of an individual module expression matrix, displayed how much a module is correlated with other modules in the network (Figure 1D). Overall, after successfully constructing the A151 ODN co-expression network, we observed that the red and turquoise modules appeared to be intertwined and highly correlated, given the network visualization results as a whole.

Table 1.

The number of arrays used for each treatment group per time point

	A151 ODN	Scrambled ODN
1 h post exposure	3	4
2 h post exposure	3	5
4 h post exposure	3	9
8 h post exposure	7	11
Total number of arrays	16	29

The total number of arrays in each group included in our time-series dataset denotes the corresponding sample size within the scope of this work.

Figure 1.

Construction of co-expressed gene modules using WGCNA in murine splenocytes treated with A151 ODN

Using transcriptome data, a total of 12 modules with a range of intermodular correlations were successfully constructed. (A) Analysis of network topology for a range of soft threshold (β) powers. Combined results from the scale-free fit index (R² threshold ≥ 0.9) as a function of β and the mean connectivity plot identified “12” as the optimum β value that ensures a scale-free topology for the A151 ODN network construction. (B) Hierarchical tree of 7,763 filtered genes represented as a dendrogram using dissimilarity measure (1-TOM). The branches and color bands display individual module assignments determined by Dynamic Tree Cut (deepSplit = 3). Branch tips represent genes. The minimum module size is 30 genes. The turquoise and red modules are intertwined, given the corresponding color bands. (C) The TOM plot of all filtered genes visualized using a heatmap. Light colors refer to low overlap, while the progressively darker red color refers to higher overlap between the genes involved (making up a module). Intermodular correlations can be visually assessed at the gene level with the help of color bands to give an idea of module merging. (D) Module eigengene adjacency visualized using a dendrogram and a heatmap. A module eigengene (ME) is the first principal component of the expression matrix of the corresponding module, which summarizes the module’s expression profile. Each column, or row, in the heatmap corresponds to a module assignment (labeled by color). The blue color means negative correlation, while the red color means positive correlation. The red and the turquoise modules (indicated with asterisks) are highly correlated.

Characterization of the modules in the A151 ODN co-expression network

We surveyed the degree of preservation of the modules in the A151 ODN network with respect to the scrambled ODN dataset to identify the most unpreserved modules that are healthy and specific to A151 ODN at the functional level. Together with the functional annotation results obtained using the GOenrichmentAnalysis() function in the R package WGCNA,¹⁴ the Zsummary and medianRank statistics from the module preservation analysis identified turquoise and red modules as candidate gene clusters for downstream analysis (Figures 2A and S6). A closer inspection of these two modules, considering our merging criteria for transcriptionally and functionally correlated modules that look intertwined in the gene dendrogram, justified the notion that turquoise and red modules belong in a larger module specific to A151 ODN at the network level. A principal-component analysis (PCA) plot showed that expression data of the genes in the merged module distinguish well between the scrambled and the A151 ODN datasets (Figure 2B). In general terms, we identified key gene clusters involved in immune response to A151 ODN treatment this way.

Figure 2.

The turquoise and red modules are the only healthy and reproducible modules in the A151 ODN network that are not preserved in the scrambled ODN dataset and can group by the type of ODN at the transcriptome level

The two immune-specific modules (turquoise and red) were found to be not preserved in the scrambled ODN dataset. The principal-component analysis (PCA) confirmed the grouping by different (i.e., scrambled or A151) ODN treatments. (A) Zsummary and medianRank statistics as a function of module size obtained from the module preservation analysis¹⁵ performed using the A151 ODN dataset as “reference” data and the scrambled ODN dataset as “test” data. The medianRank value of the modules close to zero means a high degree of module preservation in the test data, which is also corroborated by a high Zsummary score. The blue module is likely a technical artifact. The turquoise and red modules (indicated with asterisks) appear as healthy modules. The dashed blue line indicates the preservation threshold. A module was considered not preserved when it had a medianRank value greater than 7 and a Zsummary score of less than 2. (B) A PCA¹⁶ using expression data of the genes in the modules where domain-specific effects of A151 ODN on the immune cell transcriptome were functionally verified (the turquoise and the red modules combined) identified two separate clusters corresponding to the scrambled and the A151 ODN datasets. Each dot represents a microarray experiment.

Identification and functional annotation of the co-hub genes in the A151 ODN dataset

A heatmap of expression profiles of all filtered genes across replicate arrays in the scrambled and A151 ODN groups verified two distinct transcriptome profiles by treatment group before statistical analysis, implying potency of A151 ODN to regulate immune cell transcriptome (Figure 3A). After differential expression analysis of these groups (Figure S7), a volcano plot confirmed the immune-suppressive nature of A151 ODN: the differentially expressed genes (DEGs) downregulated in the A151 ODN samples have a significantly higher expression fold change after treatment than the upregulated DEGs (Figure 3B). Although the upregulated DEGs have also been given attention in the downstream analyses, they failed to converge on a common functional pathway or a network related to A151 ODN and were therefore set aside for post hoc analysis. With cutoff values at BH-corrected p value of less than 0.01 and |log₂(fold change [FC])| greater than 1.5, our analysis identified, in total, 1,060 DEGs, 523 of which were up- and 537 downregulated by A151 ODN. A Venn diagram demonstrated with high statistical significance (p < 1.66e−43) the degree of overlap between the significant genes downregulated in the A151 ODN samples and the genes in the merged module mentioned above, associating 139 (co-hub) genes with both sets (Figure 3C). We used the top 30 intramodular hub genes (IHGs) of the merged module ranked by the module membership (MM) value as a reference in prioritizing these co-hub genes. Interestingly, the functional enrichment results of the co-hub genes identified abnormal cell adhesion by integrin complexes as the topmost ontology term/pathway underlying A151 ODN-based immune modulation (Figures 3D, 3E, and S7). The only gene shared among the topmost gene sets in all enrichment sublevels is the integrin subunit alpha M, or simply “Itgam.” Itgam is also one of the IHGs discussed above. Additionally, cellular energetics pathways were found to be associated with A151 ODN treatment of immune cells (Figure S8), confirming A151 ODN-based regulation of mTOR signaling, as published previously by our group.⁶ Overall, cell adhesion by integrin complexes was identified as a focal point for cellular response of mouse splenocytes to A151 ODN treatment when network-level analysis is corroborated with differential expression analysis.

Figure 3.

Cell adhesion by integrin complexes was identified as a focal point for cellular response of mouse splenocytes to A151 ODN treatment when network-level analysis was corroborated with differential expression analysis

Combined with differential gene expression analysis,¹⁷ WGCNA yielded more refined results, identifying immune cell adhesion by integrin complexes as a key biological process in response to A151 ODN treatment. (A) A heatmap of expression profiles of all filtered genes (columns) across replicate arrays (rows) in different treatment groups. Two distinct transcriptome profiles by treatment group verify the potency of A151 ODN to induce its domain-specific effects on the immune cell transcriptome. (B) A volcano plot demonstrating the immunosuppressive nature of A151 ODN. The significant genes downregulated in the A151 ODN samples have a substantially higher expression FC after treatment than upregulated genes. NS, not significant; FC, (expression) fold change; P, p value. (C) A Venn diagram showing the degree of overlap between the significant genes identified as downregulated in the A151 ODN samples and the genes in the turquoise and red modules. The statistical significance of the overlap between two groups of genes, calculated by a hypergeometric test, is given in parentheses. (D and E) Functional enrichment^18,19 results of the 139 co-hub genes identified abnormal cell adhesion by integrin complexes as the topmost ontology term/pathway underlying the A151 ODN-based immune modulation.

Biological knowledge base and in-depth enrichment analyses of the co-hub genes

To collectively study the genuine functional potential of the co-hub genes, we leveraged the commonly used module enrichment and biological knowledge databases DAVID, Ingenuity Pathway Analysis (IPA), and STRING. The KEGG pathway²⁰ category from DAVID analysis with the co-hub genes taken as input highlighted cell adhesion molecules as a top deregulated signaling pathway in A151 ODN-based immune modulation (Table 2). More strikingly, Itgam and the integrin subunit alpha 6, or simply Itga6, appear as the most frequently observed genes in the top 10 pathways enriched in the co-hub genes. Then, these query genes with the associated differential expression results were evaluated in the IPA platform for a comprehensive knowledge analysis, which allows, also using high confidence predictions, gene-level molecular interactions with each gene’s change in direction of expression all accounted for. We observed that “cell-to-cell signaling and interaction” was the predominant term associated with the co-hub genes, given the top 5 gene-gene interaction (GGI) networks ranked by network score and listed in Table 3. Visualizing the top-scoring network on this list helped us understand the biological relevance of Itgam in immune response to A151 ODN treatment (Figure S8). Finally, a STRING PPI analysis performed to gain deeper insight into the molecular interactions of these query genes based on physical and functional associations described Itgam as a hub gene in the network (Figures 4A and S9). Included in the core of gene interactions, an essential component of the integrin complex, Itga6, surfaced again as a key element of this network. We then decided to elaborate this refined set of 23 co-hub genes in Figure 4A and, for the sake of reproducibility, plotted their expression profiles across the scrambled and A151 ODN datasets in a heatmap. Of these 23 genes, 13 (as shown in Figure 4B) were best at distinguishing between the treatment groups. As can be seen in the same figure, A151 ODN-based downregulation of these core genes is evident, confirming the differential expression analysis results. Taken together, we found that a refined set of co-hub genes led by Itgam and Itga6 seems to play a pivotal role in A151 ODN-driven suppression of immune response.

Table 2.

Top 10 “pathways” enriched in the shared co-hub genes by DAVID knowledge analysis

Analysis	Term	Count	Percentage	p Value
KEGG pathway	toxoplasmosis	Birc3, Hspa1a, H2-Oa, Itga6, Il10ra, Stat3	4.4	2.9E−3
KEGG pathway	cell adhesion molecules (CAMs)	Cd2, Cldn13, H2-Oa, Itga6, Itgam, Selplg, Sdc4	5.1	4.0E−3
KEGG pathway	B cell receptor signaling pathway	Cd79, Card11, Inpp5d, Ptpn6, Syk	3.6	4.1E−4
KEGG pathway	hematopoietic cell lineage	Cd2, Cd7, Csf1, Itga6, Itgam	3.6	8.1E−3
KEGG pathway	Jak-STAT signaling pathway	Il10ra, Il2rb, Il2rg, Ptpn6, Pim1, Stat3	4.4	1.1E−2
KEGG pathway	PI3K-Akt signaling pathway	Ddit4, Col1a1, Csf1, Gnb2, Itga6, Il2rb, Il2rg, Spp1, Syk	6.6	1.6E−2
KEGG pathway	protein processing in ER	Ero1lb, Sec61a2, Xbp1, Capn2, Calr, Hspa1a	4.4	2.0E−2
KEGG pathway	platelet activation	Rasgrp2, Col1a1, Fgb, Ptgs1, Syk	3.6	3.4E−2
KEGG pathway	focal adhesion	Birc3, Capn2, Col1a1, Itga6, Prkcb, Spp1	4.4	4.4E−2
KEGG pathway	ECM-receptor interaction	Col1a1, Itga6, Spp1, Sdc4	2.9	4.9E−2

KEGG pathways with a FET-based p value of less than 0.05 were selected. PI3K, phosphatidylinositol 3-kinase; ECM, extracellular matrix. The genes of interest are given in bold face.

Table 3.

Top 5 networks enriched in the shared co-hub genes by IPA knowledge base

Rank	Associated network functions	Score
1	cell-to-cell signaling and interaction, hematological system development/function, lymphoid tissue structure/development	32
2	hematological system development/function, lymphoid tissue structure/development, tissue morphology	29
3	inflammatory response, gastrointestinal disease, hepatic system disease	21
4	cell-to-cell signaling and interaction, hematological system development/function, dermatological diseases/conditions	21
5	cell cycle, cell death and survival	18

The network scores were calculated using FET-based p values (p score = −log10[p value]).

Figure 4.

A refined set of co-hub genes led by Itgam and Itga6 plays a pivotal role in A151 ODN-driven expression changes in the immune cell transcriptome

Confirmed using raw expression data, a set of significantly downregulated DEGs in the key modules associated with A151 ODN were identified to explain the underlying mechanism behind A151 ODN-based immune modulation. (A) Taking the whole set of 139 co-hub genes as input, an STRING PPI²¹ analysis described Itgam as a hub gene in the interaction network. Included in the core of gene interactions is another component of integrin complex, Itga6. Nodes, genes; edges, interactions. The edge thickness correlates positively with the interaction strength between nodes involved. PPI enrichment, p < 1.0E−16; minimum required interaction score, 0.400. (B) A heatmap of expression profiles of the genes mentioned in (A). Of the 23 co-hub genes in (A), 13 were best at distinguishing between the treatment groups. A151 ODN-based downregulation of these core genes is evident.

Experimental and in silico validation of the key genes in the A151 ODN co-expression network

The bioinformatics findings provided, within the scope of this study, a rich source of information for experimental validations, including likely involvement of a refined set of co-hub genes led by Itgam and Itga6 in A151 ODN-driven expression changes in the immune cell transcriptome. To separately verify this potential interplay in immune cells between A151 ODN and these two genes, we first performed a group of physical validations using qRT-PCR. Similar to the microarray datasets, which consist of mouse splenocyte samples, we stimulated BALB/c splenocytes with A151 ODN, as suggested in the literature.⁶ Potent stimulators of the immune system, LPS⁴ and CpG ODN,⁵ were also included in the experiment as controls. The results confirmed that there is a significant decrease in Itgam as well as Itga6 at the RNA level in mouse splenocytes treated with the suppressive ODN (Figure 5A), experimentally consolidating our bioinformatics findings. The same conclusion was reached when the samples were stimulated in turn with CpG and A151 ODNs (Figure S10). Also, the magnitude of this decrease for both genes is roughly in line with the differential expression results (|log₂(FC)| of 2.1 and greater than 10.0 for Itgam and Itga6, respectively). Then, we turned our attention to the in silico evaluation of the capacity of Itgam and Itga6 to distinguish between samples treated with different ODNs. Toward this end, we leveraged a machine learning-based metric assessing discrimination ability of a classifier, called the receiver operating characteristic (ROC) curve analysis. The results indicate that Itgam and Itga6 can distinguish well between samples treated with either scrambled or A151 ODNs (Figure 5B). The area under the ROC curve (the area under the curve [AUC] value) was calculated to be 0.985 for Itgam and 1.000 for Itga6, considering that an AUC value higher than 0.9 denotes an outstanding classifier for phenotype discrimination (high sensitivity and specificity). Taken together, our validation experiments showed that A151 ODN acts on Itgam and Itga6 at the transcriptome level in mouse splenocytes and that Itgam and Itga6 appear to be transcriptionally responsive to different ODN treatments in the same cell line. To add another layer of validation, we measured levels of different cytokines in activated mouse bone marrow-derived macrophages in the presence or absence of A151 ODN stimulation using ELISA. The data revealed that all measured cytokines were suppressed upon Toll-like receptor (TLR) stimulation (Figures 5C–5E). The order of A151 ODN-mediated cytokine suppression was as follows: TLR3 >> TLR9 >> TLR7 >> TLR4 ligands. All tested stimulants elicited expected levels of cytokine production, and A151 ODN subsequently suppressed all of these cytokine responses more than 50%, implying a global inhibitory effect of A151 ODN on inflammatory responses in the presence of wild-type Itgam and Itga6.

Figure 5.

A151 ODN acts on Itgam and Itga6 at the transcriptome level in mouse splenocytes

(A) Physical validation of Itgam and Itga6 downregulation by A151 ODN using qRT-PCR. A decrease in Itgam as well as Itga6 RNA levels in samples treated with suppressive ODN was confirmed experimentally. A potent stimulator of the immune system, LPS,⁴ was also included in the experiment as a control. ∗∗∗p ≤ 0.0001; ∗∗∗∗p ≤ 0.00001; ns, p > 0.05. Scrambled, control ODN. qRT-PCR data are presented as mean±s.e.m. from triplicate samples (n = 3). (B) A machine learning-based metric to assess the discrimination ability of a classifier reveals that Itgam and Itga6 can distinguish between samples treated with scrambled or A151 ODNs. The area under the ROC curve²² (the AUC value) is 0.985 for Itgam and 1.000 for Itga6. An AUC value higher than 0.9 denotes an outstanding classifier for phenotype discrimination (high sensitivity and specificity). (C–E) Physical validation of cytokine suppression by A151 ODN using ELISA in activated mouse BMDMs expressing wild-type Itgam and Itga6. A variety of stimulants was used to elicit an immune response, including poly(I:C) as a TLR3 ligand, LPS as a TLR4 ligand, R848 as a TLR7 ligand, and CpG ODN as a TLR9 ligand. The order of A151 ODN-mediated cytokine suppression is as follows: poly(I:C) >> CpG ODN >> R848 >> LPS. All tested stimulants elicited expected levels of cytokine production, and A151 ODN subsequently suppressed all of these cytokine responses, including TNF-α (C), IL-6 (D), and IFN-γ (E), more than 50% in the presence of wild-type Itgam and Itga6. Cytokine levels are presented as mean±s.e.m. from triplicate samples (n = 3).

Discussion

Telomeres that cap mammalian chromosomes by (TTAGGG)_n motif (mimicked using a synthetic ODN called A151) have been shown to have a regulatory effect on the immune system and, more specifically, suppress the immune response to pathogens by a range of immune cells.⁴ Nevertheless, the mechanistic underpinnings and the modus operandi of A151-based immune suppression at the gene level have so far remained unknown, and the alterations in immune transcriptome caused by this suppressive ODN are only partially understood.⁶ Considering the clinical potential of A151 ODN in immunosuppressive therapy as a “self-biomolecule,”^3,4,23 exploring transcriptional means of immune suppression by A151 ODN bears importance. In the current study, we analyzed the gene network landscape of mouse splenocytes to identify A151 ODN-responsive co-hub genes in the immune cell transcriptome. Technically, we took advantage of an integrative approach where network- and gene-level expression analyses were combined to dissect how A151 ODN transcriptionally controls, or curbs, the immune response.

The major finding of this work is the observation that A151 ODN acts on components of integrin complexes, Itgam and Itga6, to interfere with immune cell adhesion and therefore quench the immune response in mouse splenocytes. Accompanying A151 ODN treatment of immune cells is deregulation of a group of genes associated with immune system and cellular energetics (Figures 3D, 3E, S6, and S7). This, in turn, is in line with our previous finding that immune suppression by A151 ODN also involves energy metabolism; namely, mTOR signaling and glycolysis.⁶ In addition to that, another set of results that confirmed previous findings in this field came from the observation that Stat3 appeared in the core of gene interactions as a hub molecule of the PPI network generated using the co-hub genes (Figure 4A). STATs are major mediators of gene expression induced by many important cytokines during immune responses.²⁴ In particular, the transcriptional regulator Stat3 controls immunity through dendritic cells (DCs) and regulatory T lymphocytes (Tregs), which are among the known targets of A151 ODN.^4,24,25 Taken together, the results of this work build on, and contribute to, several strands of A151 ODN literature.

Technically, WGCNA constitutes the mainstream analysis of the expression data within the scope of this study. We used this systems biology approach to refine, at the network level, the transcriptome data that would otherwise be more complex to interpret and utilize in a biological context after differential expression analysis alone. The technical aspects of this method, together with the unique advantages it provides analytically, have been discussed elsewhere.¹⁴ We first constructed, using the A151 ODN dataset, a gene co-expression network particularly to define gene clusters, or modules, specific to A151 ODN with respect to the scrambled ODN data. Given the sample size (16 replicate arrays) to construct the network, the optimum soft threshold β value (i.e., 12) that ensures a scale-free topology was reasonable. Based on the merging criteria for transcriptionally and functionally correlated modules in the materials and methods section, we combined the intertwined and highly correlated (red and turquoise) modules to make one larger module for downstream analysis. Afterward, the module preservation analysis,¹⁵ as a part of the WGCNA method, was leveraged to identify the modules in the original A151 ODN network that are not preserved in the scrambled ODN data. This step of our analysis pipeline successfully confirmed our decision to merge two highly correlated modules based on their preservation statistics as well. Besides these, a known caveat of WGCNA is the need for empirically setting the cutoff values during the analysis, including, but not limited to, soft threshold β value, minimum module size, module merging criteria, and module preservation threshold. That is why functional annotation of each module identified serves as a quality control checkpoint after network construction and module detection. Additionally, integrating differential expression analysis with WGCNA adds another layer of confidence to the biological relevance of the outcome. This can as well be understood from the high degree of relevance of the co-hub genes (i.e., the downregulated DEGs in the merged module specific to A151 ODN) to the known functional categories associated with this suppressive ODN.

Knowing that quenching the immune response can also be done by inducing expression of genetic elements that, in turn, downregulate immune genes, such as the SOCS family of proteins,²⁶ we turned our attention to the upregulated DEGs in the merged module. Yet, these genes failed to converge on a common functional pathway or a network related to A151 ODN and were therefore set aside for post hoc analysis. Further research is underway to accumulate more evidence linking activation of immune suppressor genes with A151 ODN treatment in mouse splenocytes.

The full-size PPI network generated using the co-hub genes can be seen in Figure S9, which is a more detailed view of the core network presented in Figure 4A. The genes in the full-size network associated with the Gene Ontology (GO) terms “regulation of neutrophil activation,” “negative regulation of glycolytic process,” and “membrane rafts” are indicated using red, blue, and green colors, respectively. A clear enrichment of green nodes (i.e., genes associated with membrane rafts) in the core of gene interactions imply the involvement of transmembrane proteins in cellular response to A151 ODN treatment. Together with the observation that Itgam and Itga6 appeared in the core network as hub genes, the results of our knowledge base and functional enrichment analyses (Table 2; Figures 3D, 3E, and S7) made more biological sense. Last but not least important, it is already well established in the literature related to integrin adhesion complexes that the hallmark phenotype of genetically engineered mouse models that lack ITGAM expression is diminished neutrophil activation during inflammation.²⁷ We then tested this potential link of Itgam and Itga6 genes to A151 ODN experimentally (using qRT-PCR) and bioinformatically (using machine learning-based ROC analysis). The results indicated that A151 ODN acts on Itgam and Itga6 at the transcriptome level in mouse splenocytes and that Itgam and Itga6 appear to be transcriptionally responsive to different ODN treatments in the same cell line. Taken together, we identified integrin complexes as a focal point for cellular response of mouse splenocytes to A151 ODN treatment.

Encoding the integrin alpha M chain, Itgam (CD11b) is broadly expressed in the mouse spleen and more specifically in macrophages, DCs, and granulocytes, including all 3 subtypes: neutrophils, basophils, and eosinophils.^28,29 Interestingly, macrophages and DCs are also at the top of the list of cell types heavily regulated by A151 ODN.⁴ Also, impaired phagocytosis of CD177-PRTN3-mediated activation of TNF-primed neutrophils has been associated with Itgam,³⁰ which is in line with our findings assuming that A151 ODN acts on neutrophils through Itgam (Figures 3D and S7). Furthermore, it is now known that Itgam, together with Itgb2, is involved in various adhesive interactions of monocytes, macrophages, and granulocytes as well as in arbitrating the uptake of complement-coated particles and pathogens.^31,32 Given our current findings, these implications make more biological sense to understand the molecular mechanisms of A151 ODN-based quenching of immune activities by macrophages and, possibly, granulocytes.⁴ Finally, mice expressing a functionally less active form of Itgam have been shown to have defective DC-mediated T cell proliferation.³³ T cells are also known to be severely affected upon A151 ODN treatment in mice.³⁴ All of these observations seem to be in parallel with the current literature on the cell type specificity in human of Itgam expression and immune response to A151 ODN treatment.^4,35,36,37 Correspondingly, our results pave the way for potential therapeutic use of this suppressive ODN in clinics. On the other hand, Itga6 (CD49f), which encodes a member of the integrin alpha chain family of proteins, appears to be more promiscuously expressed in all immune cells, with a peak level of expression in macrophages and Tregs.²⁸ Specialized to the suppress immune response for maintaining homeostasis and self-tolerance as regulatory T cells, Tregs constitute a major means by which A151 ODN operates to curb the immune system.^4,28 Furthermore, Itga6 has been implicated in resistance to cancer therapy and in progression of tumorigenesis.³⁸ That a cell adhesion molecule such as Itga6 seemingly provides a selective advantage to cancer cells in immune evasion is not surprising.³⁹ Overall, the literature on Itga6, as opposed to Itgam, is rather limited.

Given that biological processes in general have an impact on sets of genes acting in concert and that even a 10% change in expression of all genes encoding members of a metabolic pathway may substantially change the flux through that pathway (which may actually be more significant than a 10-fold change in expression of a single gene involved),⁴⁰ we finally performed a gene set enrichment analysis (GSEA) to detect gene sets enriched between scrambled and A151 ODN treatment groups. One of the top enriched gene sets was “the positive regulation of cell-to-cell adhesion” (Figure S11). In other words, we confirmed, using another bioinformatics approach, a concerted effect that impairs cell-to-cell adhesion in immune cells treated with A151 ODN. Also, we showed, using a heatmap, that the genes in the leading edge, which include Itga6 as well as some other components of the integrin complex, were downregulated by A151 ODN. This, in turn, revealed that, in addition to the gene- and network-level analyses, we reached the same conclusions at the gene set level.

Materials and methods

Data sources

Our time-series microarray datasets representing raw gene expression data from two different (scrambled ODN and A151 ODN) treatment groups with a varying number of non-technical replicate arrays (29 and 16, respectively) were prepared to capture true biological variation within and between sample groups at 4 different time points (i.e., 1, 2, 4, and 8 h after exposure) and have been published before.⁶ Briefly, each replicate data file describes a dual-channel, custom-made microarray experiment with a common reference design covering the murine splenocyte genome using a set of 16,897 unique probes. In an experiment, the green signal by Cy3 is the control, and the red signal by Cy5 is the treatment. The scrambled ODN is the sequence scramble as a control for A151 ODN.

Data preprocessing

The limma¹⁷ v.3.46.0 and the genefilter⁴¹ v.1.72.1 packages in R v.4.0.5¹⁶ were used to preprocess the raw data in tabular (TSV) format. Initially, probes with low (<100) signal intensity and spots with low quality (flag == 0) and small size (<10) in an array were removed using a custom function. Identifiers for genes were replaced with the corresponding gene symbols in the annotation file downloaded together with the datasets. The background-corrected data were normalized using the lowess intensity-dependent normalization function normalizeWithinArrays() to adjust for differences in labeling intensities of the Cy3 and Cy5 dyes, which was then followed by the quantile normalization function normalizeBetweenArrays() to account for any heterogeneity in the statistical properties of the underlying distributions between arrays. Gene filtering was then performed using the genefilter() function, taking an interquartile range (IQR)-based variance filtering function as an argument to filter out probes with low (>50%) variation across sample classes. Afterward, empty or control wells were removed, and the corresponding expression values that are already log base 2 transformed were averaged over the probes mapping to the same gene to condense the datasets. Standard diagnostic plots were used before and after each step to decide whether further improvement is needed technically. Ultimately, the final datasets were batch corrected using the ComBat() function in the sva⁴² v.3.38.0 package with a non-parametric test to fit a linear regression model to the dataset for removal of technical variation associated with the input data.

Construction of the gene co-expression network

After evaluating the quality and propriety of the expression data as input for the R package WGCNA¹⁴ v.1.70.3, we performed outlier array detection. The plots generated using the pickSoftThreshold() function, which are the scale-free fit index (scale independence R2 threshold ≥ 0.9) as a function of β and the mean connectivity plot, identified “12” as the optimum β value in a range of soft thresholding powers that ensures a biologically meaningful scale-free topology for construction of the A151 ODN network. Using the blockwiseModules() function, we constructed a gene co-expression network and identified the modules as follows. (1) Co-expression similarity between each pair of genes across all arrays was calculated based on biweight mid-correlation and recorded to form an adjacency matrix. (2) The adjacencies were then transformed into a TOM. (3) Modules defining groups of co-expressed genes with a minimum module size of 30 genes were detected using hierarchical clustering of the TOM in a high-sensitivity mode (deepSplit = 3) of the branch cutting algorithm (the Dynamic Tree Cut), which was then visualized as a gene dendrogram. (4) Module eigengenes as the first PC of module expression levels were finally identified to find, based on MM and gene significance values, the IHGs. (5) When two modules were intertwined in the gene dendrogram, when their functional enrichment results were correlated at the ontology level, and when the absolute value of correlation between the corresponding eigengene values (also color coded in the eigengene network) was sufficient (>0.75), these modules were merged. Biologically relevant modules were then visualized in the Cytoscape⁴³ software v.3.0 and investigated further using the Cytohubba⁴⁴ plugin v.0.1 for an in-depth analysis of the network features.

Module preservation analysis

After construction of the A151 ODN network, we used the modulePreservation() function¹⁵ to survey the degree of preservation of the modules in this “control” network with respect to the scrambled ODN dataset, which serves as the “test” group in this analysis. The same procedures mentioned as for data preprocessing and the initial quality checks before the network analysis were applied to the test expression dataset as well. The most unpreserved modules in the test data that were healthy and specific to A151 ODN at the functional level were of particular interest. Based on the combination of two preservation statistics, medianRank and Zsummary, we identified high-quality modules that were not preserved in the scrambled ODN dataset (number of permutations, 500). A module was considered not preserved when it had a medianRank value greater than 7 and Zsummary score of less than 3.

Differential expression analysis

The data were merged by gene names to create a new, normalized count matrix representing scrambled and A151 ODN expression datasets. The R package limma was again used to fit a linear model to the gene-level expression data in this count matrix. Differential expression analysis was performed, computing empirical Bayes-moderated t statistics for the significance of class comparison for each gene. We established cutoff values at q (BH-corrected p value) less than 0.01 and |log₂(FC)| greater than 1.5 to screen DEGs. The results were visualized at the genome and module levels in R.

Identification of the co-hub genes associated with A151 ODN

We first identified, in the A151 ODN-specific modules that were not preserved in the scrambled ODN dataset, the IHGs with an MM value of greater than 0.8. Keeping the top 30 hub genes ranked by the MM value in mind as a reference, we then intersected, using a Venn diagram in R, the DEGs downregulated in the A151 ODN samples with the genes in the modules associated with A151 ODN. The statistical significance of the overlap between two groups of genes was calculated using a hypergeometric test. The shared group of genes is what we called the co-hub genes within the scope of this study.

Functional enrichment analyses of the co-hub genes

The co-hub genes were next assessed in a biological context using the functional annotation tool of DAVID v.6.8 with the mouse genome being the background.^45,46 Also, these co-hub genes with the associated differential expression results were analyzed in the IPA⁴⁷ platform v.2.4 for a comprehensive knowledge analysis using the “Core Analysis” function (Fisher’s exact test [FET] p = 1e−03). Over-representation of GO terms was evaluated using the R package ReactomePA¹⁸ v.1.28.0 or the goANA() function in the R package limma. EnrichR v.01.07.2020,¹⁹ a multi-domain gene list enrichment analysis platform (https://amp.pharm.mssm.edu/Enrichr/), with the following libraries was used in this study: MGI Mammalian Phenotype Ontology, PANTHER Pathways, and MSigDB Hallmark (2020) Gene Set Enrichment. To construct a PPI network of the query genes, STRING²¹ v.11 software (https://string-db.org/) was used with a minimal interaction score of 0.400 (FET p = 1e−03).

GSEA of the normalized raw expression data

The input data for the differential expression analysis above were fed into the GUI-based GSEA software⁴⁸ v.3.0 to detect whether any gene sets in a large set of pre-defined biological processes (namely, the msigdb.v5.2.symbols_mouse.gmt database) were enriched between scrambled and A151 ODN treatment groups. The number of permutations was set to 500 for stringent enrichment criteria (NOM p < 0.001, |NES| > 1.95).

Validation of the co-hub genes using the ROC curve

The ROC plot was generated together with the AUC calculations using the R package pROC²² v.1.18.0 to assess the discrimination ability of a classifier (in this case, the selected co-hub gene[s]) for different treatment groups. For efficacy evaluation, AUC = 0.5 means non-efficiency, 0.5 < AUC < 0.7 means a modest level of efficiency, and AUC > 0.7 means high efficiency.

qRT-PCR

Spleen tissues from 6-week-old female BALB/c mice were homogenized, and splenocytes were cultured with 3 μM scrambled ODN, 3 μM A151 ODN, 1 μM CpG ODN (1,555), or 1 μg/mL LPS. 8 h after stimulation, cells were harvested for RNA isolation. For CpG + A151 treatment, A151 ODN was added at 4 h. RNA was isolated using the Quick-RNA Microprep Kit (Zymo Research, USA). cDNA was synthesized from total RNA using ProtoScript First Strand cDNA (New England Biolabs, USA). Quantitative PCR (qPCR) was done using primers targeting Itga6 and Itgam and LightCycler 480 SYBR Green I Master Mix (Roche, Switzerland) (ITGAM_F: 5′-GCTCGACACCATCGCATCTA, ITGAM_R: 5′-TGGTACTTCCTGTCTGCGTG-3′, ITGA6_F: 5′-ATGAAAGTCTCGTGCCCGTT-3′, ITGA6_R: 5′-CCCCACTGTGATTGGCTCTT-3′). qPCR was performed using the LightCycler 96 instrument and analyzed with the manufacturer-supplied software (Roche, Switzerland).

IL6, TNF-α, and IFN-γ ELISA

In an attempt to further validate that A151 ODN treatment quenches the immune response, we treated mouse bone marrow-derived macrophages (BMDMs; 2.5 × 10⁵/well) with 4 different TLR ligands (30 μg/mL of poly(I:C) [a TLR3 ligand], 1 μg/mL of LPS [a TLR4 ligand], 1 μg/mL of R848 [a TLR7 ligand], and 1 μM of CpG ODN [a TLR9 ligand]) in the presence or absence of 1 μM A151 ODN. 24 h later, culture supernatants were collected for assessment of IL-6, TNF-α, and IFN-γ levels using mouse cytokine ELISA kits from Mabtech.

Code availability

The analysis pipeline and custom-made scripts presented in this manuscript are available from the corresponding author upon reasonable request.

Statistics

A p value or false discovery rate (FDR) of less than 0.05 was considered significant throughout this study unless stated otherwise. The type of data distribution and validity of required assumptions were verified before statistical analysis in this work. For the experimental validations, ΔCt was calculated by subtracting Ct measured for gene of interest from Ct measured for ACTB. For each treatment group and gene, ΔΔCt was calculated by subtracting ΔCt (A151 ODN) from ΔCt (scrambled ODN). FCs, 2^(−ΔΔCt), were then log2 transformed. Statistical testing was done using 2^(−ΔCt) values. Non-parametric t test (Mann-Whitney) was used to calculate significance between treatment groups in GraphPad Prism v.8 software (GraphPad, San Diego, CA, USA).⁶

Data availability

The custom-made microarray datasets have been deposited in the NCBI’s Gene Expression Omnibus (GEO) database under accession GSE184994. The gene lists described in this work can be viewed as supplemental information files.