2.1 Selection of gene expression datasets and human biological pathways

The expression datasets containing the CRC and their adjacent normal tissue samples were searched in the Gene Expression Omnibus database (GEO, https://www.ncbi.nlm.nih.gov/geo/). Three datasets, namely, GSE110223, GSE110224, and GSE113513, were selected based on the following criteria: (i) tissue samples must be human CRC and normal samples, (ii) total sample size in each dataset should be >20, (iii) tissue with no drug treatment or knocking out a gene. The detailed information about the samples can be found in Table (Table 1). All the selected datasets were preprocessed independently using the robust multi-array average (RMA) algorithm.¹⁷ We considered the common genes across all three selected datasets for further study. Limma (linear models for microarray data) package in the R programming software¹⁸ was used to analyze the fold change (log2FC) in gene expression between normal and CRC states with a statistical significance with a p-value < 0.01.¹⁹

In this study, we selected the pathway information from the KEGG (Kyoto Encyclopedia of Genes and Genomes, https://www.genome.jp/kegg/pathway.html) and Reactome (https://reactome.org/) databases. Pathways, including the gene sets, were retrieved from the C2: Canonical Pathway collection of the MSigDB (Molecular Signature Database, https://www.gsea-msigdb.org/gsea/msigdb). The pathways were further selected based on the following criteria: (i) pathways with a minimum of five genes appeared in the selected expression datasets, (ii) exclude the pathways belonging to human disease and drug development to reduce the pathway redundancy, and (iii) at least two overlapping genes among two pathways. The pathway analysis was performed independently for both the KEGG and Reactome pathways.

Besides the above three datasets, GSE24514, GSE28722, GSE41258, and TCGA RNA-Seq were used to further analyze and cross-validate the prognostic significance of the selected genes. The TCGA RNA-Seq dataset comprised 50 paired normal and CRC samples obtained from the Genomic Data Commons Data Portal (https://portal.gdc.cancer.gov/). Normalization and analysis of log2FC for RNA-seq data were conducted using the DESeq2 method.

2.2 Construction of state-dependent pathway cross-talk networks

As pathways as well, genes interact with each other under certain conditions; therefore, we constructed the pathway cross-talk networks for normal and CRC states separately. For this, we integrated the state-dependent gene co-expression with the selected pathways. The gene co-expression networks were independently constructed for normal and CRC states in all the selected datasets. The Pearson correlation coefficient (PCC) was calculated for gene pairs within each state using the HMISC package in R (https://hbiostat.org/R/Hmisc/). This study considers the gene pairs with PCC ≥ |0.7| and p-value < 0.05 as co-expressed genes. The final co-expression network included gene pairs that exhibit a similar correlation pattern, either positive or negative, across all three datasets. Then, co-expressed gene pairs were mapped with the selected pathways to construct the state-dependent pathway cross-talks.

We considered all possible pairs of selected pathways to construct the pathway cross-talk networks. The pathway cross-talk in the network is denoted as G(V, E), nodes (V) represent the set of selected pathways, and edges (E) represent the pathway-pathway pair guided by the co-expressed gene pairs. Certain genes are involved in multiple pathways, so a single gene pair might contribute to false positive cross-talk between pathways. Therefore, we considered the two pathways, for example, PathwayA and PathwayB, to be involved in cross-talk if they fulfilled at least one of the following criteria as given below and illustrated in Figure 2:

(i) $$$ \left\{{g}_{\mathrm{i}},{g}_{\mathrm{j}}\right\}\ \mathrm{and}\ \left\{{g}_{\mathrm{p}},{g}_{\mathrm{q}}\right\}\in $$$ co-expressed gene pairs, where (g_i $$$ \wedge $$$ g_p $$$ \in $$$ PathwayA) $$$ \wedge $$$ (g_i $$$ \wedge $$$ g_p $$$ \notin $$$ PathwayB) $$$ \wedge $$$ (g_j $$$ \wedge $$$ g_q $$$ \in $$$ PathwayB) $$$ \wedge $$$ (g_j $$$ \wedge $$$ g_q $$$ \notin $$$ PathwayA).

(ii) $$$ \left\{{g}_{\mathrm{k}},{g}_{\mathrm{l}}\right\}\ \mathrm{and}\ \left\{{g}_{\mathrm{l}},{g}_{\mathrm{m}}\right\}\in $$$ co-expressed gene pairs, where (($$$ {g}_{\mathrm{k}}{g}_{\mathrm{k}}\in $$$ PathwayA) $$$ \wedge $$$ ($$$ {g}_{\mathrm{k}}{g}_{\mathrm{k}}\notin $$$ PathwayB)) $$$ \wedge $$$ ($$$ {g}_{\mathrm{l}} $$$ $$$ {g}_{\mathrm{l}}\in $$$ PathwayA $$$ \wedge $$$ PathwayB) $$$ \wedge $$$ (($$$ {g}_{\mathrm{m}}{g}_{\mathrm{m}}\notin $$$ PathwayA) $$$ \wedge $$$ ($$$ {g}_{\mathrm{m}}{g}_{\mathrm{m}}\in $$$ PathwayB)).

This pathway cross-talk construction methodology is inspired by the review article authored by Azad et al..¹ An in-house Python script was used to construct the pathway cross-talk network. This results in constructing four networks, consisting of pathway cross-talk and co-expressed genes involved in inter-pathway communication in both normal and CRC states. Cytoscape v. 3.10.0²⁰ software was used to visualize and analyze the constructed networks in the present study.

2.3 Comparative analysis of dependent pathway cross-talk networks state-dependent pathway cross-talk networks

The pathway cross-talk network of the normal state was compared with the CRC state to understand the importance of the cross-talk among pathways in the development of CRC. As genes play crucial roles in the cross-talk among pathways at the molecular level, therefore, the gene pairs connecting two different pathways were also compared and analyzed to identify the crucial genes that might be playing an important role in the pathway cross-talk required for the CRC progression. An in-house Python script was used for comparing the path-ways and gene co-expression network from normal to CRC state. The gene pairs helping in the construction of pathway cross-talk networks in the CRC state were further analyzed to identify the key genes involved in pathway cross-talk as well as in the CRC progression. As each gene pair corresponds to the pathway–pathway pairs, some genes might be involved in regulating more than one pathway. Therefore, we calculated the betweenness centrality for each gene in the co-expressed gene networks in both states to identify such crucial genes. The betweenness centrality was calculated for each gene using a Cytoscape plugin CytoNCA.²¹ The genes were ranked according to their centrality score, and the top-ranked genes based on their centrality score were defined as key genes in the present study.

2.4 Knowledge-driven analysis of all the genes involved in pathway cross-talks

To understand the significance of the gene pairs involved in the pathway cross-talks, their association with the disease, cancer hallmarks, and genetic alterations analysis was performed. For this, the DisGeNET database²² was used to know the disease association of the selected genes, and only the curated association was considered to check the disease association. The genetic alteration was also checked as it plays an important role in the regulation of gene expression as well as in disease development. For this, the 220 samples with mutation data of TCGA, Firehose Legacy of the cBioPortal (https://www.cbioportal.org/) server were analyzed. With the advancement of sequencing data, researchers are also exploring the significance of co-occurrence or mutual exclusive mutation patterns in cancer as it can provide information about the pathways and their mechanisms in tumor development.²³ Therefore, we also performed the mutual exclusivity analysis using the cBioPortal server. The association of the selected genes and pathways was checked with the cancer hallmarks using the CHG database (http://www.bio-bigdata.com/CHG/index.html).²⁴

2.5 Identification of potential drug candidates

To identify potential therapeutic drugs for CRC and reverse the expression of key genes involved in CRC progression, we explored the L1000CDS ² database (https://maayanlab.cloud/L1000CDS2/#/index). This publicly available database contains over 1 million gene expression profiles of chemically perturbed human cell lines. The database utilizes the characteristic direction (CD) scoring method, which prioritizes genes exhibiting minimal changes but moves together with a large group of other genes.²⁵ Genes from identified clusters showing significant expression changes (log2FC) from normal to CRC states across three datasets were selected for this analysis.

2.6 Calculation of dysregulated score for genes

Gene co-expression networks tell the potential functional association between the genes, but they do not provide information on regulatory relationships such as which gene regulates which gene activities. Therefore, we also analyzed the gene regulatory network (GRN) integrated with gene expression data. As interactions among genes are context-specific, a change in their interaction can lead to disease development. Therefore, we calculated the dysregulation score for each gene based on its change in interaction strength with other genes from the normal state to the CRC state by integrating the gene expression information with the GRN. The directed gene interactions were sourced from the study conducted by Huo et al.²⁶ and includes interactions from KEGG, Reactome, Panther, CellMap, and NCI Pathway Interaction Databases. The gene interactions containing the predicted interaction were excluded from the study.

Here, d_i represents the dysregulation score of gene i, n_i is the number of first neighbors of gene i (interacting genes) in the gene regulatory network. The dysregulated score was calculated for each dataset separately, and to compare the dysregulation strength of genes among all three datasets, the dysregulated score was converted to Z-score = ((dysregulation score − average of all dysregulation score)/standard deviation of all dysregulation score).

2.7 Prognostic Analysis of the selected key genes and cross-validation

where, Exp_i and β_i are the expression levels of gene i in a particular patient sample and coefficients of gene i in the multivariate Cox regression analysis. Then, the CRC samples were categorized into low-risk and high-risk groups based on the median value of the risk score. The prognostic performance of these gene groups was evaluated by generating Kaplan–Meier plots and assessing the statistical significance using a log-rank test with a significance p-value < 0.05.

3.1 State dependent KEGG pathway cross-talks network construction

Pathway cross-talk networks were systematically constructed based on the criteria mentioned in the material and methods sections (2.2). The example of a pathway cross-talk network guided by a state-dependent gene co-expression network was illustrated in Figure 2B,C. This figure illustrated the pathway cross-talk between focal adhesion (FA, KEGG ID: hsa04510) and vascular smooth muscle contraction (VSMC, KEGG ID: hsa04270) pathways for both normal and CRC states. Research has also reported the role of the focal adhesion pathway in smooth muscle cell contractility, which is aligning with our analysis.³⁰ The figure also showed the co-expressed genes involved in intra-pathway cross-talk. A comparison between inter-pathway cross-talk in normal and CRC states reveals alterations in the interactions of many genes during the transition from normal to CRC state. This highlights the significance of cross-talk at both the pathway and gene levels in maintaining normal cellular function and its role in CRC development. The gene pairs involved in the pathway cross-talks were considered potential key genes in this study.

In normal state, the pathway cross-talk network consists of 109 pathways with 710 interactions (Figure 3A). The corresponding gene co-expression involved in the pathway cross-talk is termed as PathGeNet (Pathway cross-talk interface gene co-expression network) in this study. In the normal state, the PathGeNet consists of 744 genes with 2024 edges (Figure S1). At the gene level, we observed a well-connected component consisting of 567 genes and many discrete groups of genes. The key genes were also identified by analyzing the PathGeNet, and the top 10 genes based on the betweenness score include PLNA, ITGA7, MYL9, GNA11, FEN1, FLNC, CHP2, HSD11B2, CDK1, and MYLK.

In the CRC state, the pathway cross-talk networks were composed of 124 interactions among 49 pathways (Figure 3B), and PathGeNet consisted of 273 edges among 182 genes (Figure 4A and Figure S2). At the pathway level, we observed that the gap junction (KEGG ID: hsa04540) pathway acts as a connecting link between two clustered of pathways. At the gene level, we observed two distinct largest connected components (LCCs) of genes in the PathGeNet, namely, LCC 1 and LCC 2. LCC 1 consists of 81 genes, and the genes of LCC 1 belong to the immune system and cell communication-related pathways, whereas 48 genes comprise LCC 2 and mainly belong to the cell cycle. The topology-based analysis of these LCCs led us to identify 10 key genes CCL2, CCR7, CD3D, CD48, FCER1G, LY96, PTPRC, RAB31, TUBB6, and TYROBP in ClusterK1 and 10 key genes AURKA, CCNA2, CDK1, CUL1, FEN1, MCM7, PLK1, POLD2, RFC2, and RNASEH2A in the ClusterK2. All the identified key genes were found to be interconnected as shown in Figure 4C,D. At the gene expression level, we observed a variation in the fold change (log2FC) of the key genes belonging to ClusterK1 in all three datasets, whereas genes from ClusterK2 were found to be up-regulated in all three datasets (Figure 4C,D). The consistent expression pattern (log2FC) was observed for these groups of genes in the three independent datasets (Figure S3). Among the identified key genes from the normal state, only two genes, ITGA7 and MYH10, were found to be involved in the pathway cross-talk in the CRC state. The identified key genes from different clusters were observed to be well interconnected; therefore, their combined prognostic potential was assessed, revealing a significant prognostic power (Figure 4C,D). In all the analyzed independent datasets, the survival probability was observed to be significantly decreased over time in the high-risk group compared to the low-risk group (Figure S4). The Wilcoxon rank-sum analysis revealed that deceased individuals had significantly higher risk scores than those who were alive, indicating an association between high-risk scores and mortality risk in CRC patients (Figure S4). An analysis of risk score distributions over time revealed that the high-risk group had a greater number of CRC patients who experienced death compared to the low-risk group (Figure S4). This suggests the potential ability of these key gene clusters as prognostic biomarkers for CRC patients.

3.2 State dependent comparative analysis of interaction at KEGG pathway and gene level

Network comparison was made at both the pathway and gene level to identify the pathways and genes that undergo significant changes in cross-talk during the transition from a normal to a CRC state. In the normal state, 637 interactions were identified among 109 pathways and 744 genes with 1961 interactions specifically involved in pathway cross-talk unique to the normal state when compared to the CRC state (Figure S5). In the CRC state, we observed 23 cross-talks among 18 pathways and 213 interactions among 163 genes at the gene level only in the CRC state compared to the normal state, as shown in Figure 5. These cross-talks were mainly observed among the pathways related to the immune system, signaling pathways such as chemokine, MAPK, and JAK–STAT signaling pathway.

In the comparison of pathway cross-talk networks, it was observed that 101 cross-talks among 44 pathways overlapped in both states, as shown in Figure 6A. At the gene level, the overlapping interactions among 55 genes were observed in both states (Figure 6B). Among these 55 genes, 25 genes were found clustered together and mainly associated with the cell cycle-related pathways, and the expression of all these genes was found to be up-regulated in the CRC state (Figure 6C). Likewise, nine genes related to cell–cell communication pathways were observed in one cluster, and all were down-regulated in the CRC state, as shown in Figure 6D. Survival analysis was also performed on conserved gene clusters, and it was observed that both the clusters could differentiate between patients at higher risk and those at lower risk in all the three independent datasets, as shown in Figure 6 and Figure S4. We called these clusters as ClusterK3 and ClusterK4 based on the number of genes, respectively.

3.3 Knowledge-driven analysis of the genes involved in KEGG pathway cross-talks

The disease-gene association analysis revealed that among 182 genes in pathway cross-talk in the CRC state, 95 genes were associated with at least one type of cancer (Figure S6), and 22 genes were found to be associated with the CRC (Figure 4B). Four key genes CCL2, CCR7, LY96, and PTPRC, from ClusterK1 and five genes AURKA, CDK1, CCNA2, FEN1, and PLK1 from ClusterK2, were associated with at least one type of cancer. LY96, FEN1, and AURKA key genes were known CRC-associated genes. The genetic alterations of these genes in 220 CRC patient samples were assessed using the cBioPortal, and we found out that, except for 20 genes, all the genes were showing alterations. Figure 7A shows the genes' genetic alteration observed in at least 5% of the CRC patients. The COL6A3 gene was found to be altered in 13% of the samples, whereas the DOCK2 gene was altered in 10% of the samples, as shown in Figure 7A. The co-occurring mutation patterns were observed among the 328 gene pairs, and the DOCK2 gene showed co-occurrence with 32 genes, including COL6A3 gene (Figure 7B). The co-occurring mutation pair is reported to generally activate different complementary biological pathways, resulting in different cancer characteristics with a few exceptions in the tumorigenesis.²³ Among the identified key genes, five genes, AURKA, CCR7, FEN1, PTPRC, and RNASEH2A, showed co-occurring mutations with other genes. Cancer hallmarks analysis results revealed that 23 pathways and 128 genes were associated with at least one type of cancer hallmark, as shown in Figure 7C,D. These results highlight the significance of genes participating in pathway cross-talk in the CRC.

3.4 State dependent Reactome pathway cross-talks network analysis

The state-dependent pathway cross-talk network was constructed for the 1475 Reactome pathways using specified selection criteria and network construction methods outlined in the materials and methods section (2.2). A consistent rewiring pattern in pathway cross-talk was observed in Reactome and KEGG pathways analyses. In the normal state, 1,57,071 cross-talks were identified among 1347 pathways whereas 53,100 cross-talks were observed among 909 pathways in the CRC state. While some pathways lost connectivity in CRC, others maintained, and some formed new interactions (Figure S7). The immune-related pathways, followed by signaling pathways such as chemokine signaling, were observed to be the most rewired pathways (Figure S7).

At the gene level, PathGeNet analysis in the CRC state identified two distinct largest connected components (LCCs) of genes, which include LCC1, consisting of 557 genes, and LCC2, comprising 265 genes. Analyzing the network topology of these LCCs revealed 28 key genes in LCC1 and 14 key genes in LCC2, named as ClusterR1 and ClusterR2, respectively. Survival analysis of these key genes clustered showed their prognostic significance in distinguishing low-risk from high-risk groups (Figures S10, S11, and S12), indicating their potential impact on CRC progression (Figure 8A,B). These findings highlight the importance of these genes in pathway cross-talk and their role in CRC prognosis.

The comparison of PathGeNet among normal and CRC states identified the three different conserved sub-networks: sub-network1, consisting of 116 genes; sub-network2, comprised of 67 genes; and sub-network 3, which included 11 genes. The sub-network1 and sub-network2 were further analyzed with Cytoscape plugin jActiveModule³¹ to identify the smaller clusters of genes showing significant changes in gene their expression patterns, and they were termed as ClusterR3 comprising 25 genes and ClusterR4 consisting of 14 genes. The sub-network3 with 11 genes showed a significant change in the log2FC from normal to CRC; therefore, it was considered as ClusterR5. These all three clusters showed their potential impact on CRC progression by distinguishing between low- and high-risk groups (Figures S11 and S12).

A comparative analysis of PathGeNet in the CRC state showed a partial overlap of key genes as well as clustered genes such as MCM7, AURKA, FEN1, CDK1, POLD2, FCER1G, PT-PRC, CCL2, RAB31, CD48, TYROBP, and LY96 identified in both the KEGG and Reactome analysis. The independent analysis of Reactome identified a unique gene cluster, ClusterR5, which was not present in the KEGG analysis (Figures S8 and S9). The ClusterR5 mainly involved in the ion transport channels, cell–cell communication and digestion and absorption-related Reactome pathways. CA4 gene of the ClusterR5 is a known diagnostic and prognostic biomarker for CRC.^{32, 33} The analysis of disease-gene associations revealed that 10 genes (FEN1, GCG, ACTG2, LY96, TPX2, AURKA, GPNMB, ITGA7, LIG1, and SCNN1B) part of the identified clustered from Reactome analysis (Figure S12). This highlights the importance of conducting individual pathway analyses to capture specific gene clusters that might be missed in the combined analyses.

3.5 Identification of potential drug candidates

All the identified clustered showing significant association with CRC progression were searched against the L1000CDS² database, resulting in 36 unique small molecules with a score >0.5 showing reverse changes in gene expression (Table S1). Seven of these molecules, including Palbociclib (an approved drug for breast cancer treatment), were found to reverse gene expression in HT29 (human colorectal adenocarcinoma) cell lines (Table 2). The chemical structure of these seven drug candidates was shown in Figures S17 and S18.

3.6 Gene regulatory network analysis

The dysregulated score was calculated for all the genes using directed gene regulatory networks in all three datasets. The 24 genes were found to be common among all the three datasets with Z-score ≥5 (Figure 9A,B) and considered as key dysregulated genes in the CRC. These dysregulated scores were calculated in the three independent datasets comprising normal and CRC samples, revealing that all 24 identified dysregulated genes overlapped with gene Z-scores ≥4, except for two genes (PRKACG and HRAS) in the RNA-seq datasets (Figures S13 and S14). This analysis highlights the potentially significant role of these dysregulated genes in CRC development. Excluding the pathway associated with disease, pathway enrichment analysis results revealed that these 24 genes were mainly enriched in Relaxin signaling (hsa:04926), Chemokine signaling (hsa:4062), thyroid hormone signaling (hsa:04915) (Table 3). Among 24 genes, nine genes were reported as transcription factors CTNNB1, EP300, JUN, MYC, NFKB1, RELA, SP1, STAT1, and TP53 based on the TTRUST database (https://www.grnpedia.org/trrust/) of gene regulation. Disease gene association analysis revealed that all 24 key dysregulated genes except GNAI3, PRKACG, RPS27A, and UBB were associated with at least one type of cancer (Figure S14), whereas eight genes (MYC, TP53, CTNNB1, JUN, RELA, EP300, SRC, and NFKB1) were associated with the CRC as shown in Figure S15. Among all the identified 24 dysregulated genes, only the CTNNB1 gene showed a poor survival probability with the altered gene expression, as shown in Figure S15. We also checked the CTNNB1 prognostic significance in other cancers, and we found that CTNNB1 showed a significant survival probability (p-value < 0.05) in the nearby organs of CRC, such as the bladder, liver, and pancreas, but did not show any significant results in the distant organs such as breast and lung tissue (Figure S15).