Biomarker correlation network in colorectal carcinoma by tumor anatomic location

Colorectal carcinoma evolves through a multitude of molecular events including somatic mutations, epigenetic alterations, and aberrant protein expression, influenced by host immune reactions. One way to interrogate the complex carcinogenic process and interactions between aberrant events is to model a biomarker correlation network. Such a network analysis integrates multidimensional tumor biomarker data to identify key molecular events and pathways that are central to an underlying biological process. Due to embryological, physiological, and microbial differences, proximal and distal colorectal cancers have distinct sets of molecular pathological signatures. Given these differences, we hypothesized that a biomarker correlation network might vary by tumor location. We performed network analyses of 54 biomarkers, including major mutational events, microsatellite instability (MSI), epigenetic features, protein expression status, and immune reactions using data from 1380 colorectal cancer cases: 690 cases with proximal colon cancer and 690 cases with distal colorectal cancer matched by age and sex. Edges were defined by statistically significant correlations between biomarkers using Spearman correlation analyses. We found that the proximal colon cancer network formed a denser network (total number of edges, n = 173) than the distal colorectal cancer network (n = 95) (P < 0.0001 in permutation tests). The value of the average clustering coefficient was 0.50 in the proximal colon cancer network and 0.30 in the distal colorectal cancer network, indicating the greater clustering tendency of the proximal colon cancer network. In particular, MSI was a key hub, highly connected with other biomarkers in proximal colon cancer, but not in distal colorectal cancer. Among patients with non-MSI-high cancer, BRAF mutation status emerged as a distinct marker with higher connectivity in the network of proximal colon cancer, but not in distal colorectal cancer. In proximal colon cancer, tumor biomarkers tended to be correlated with each other, and MSI and BRAF mutation functioned as key molecular characteristics during the carcinogenesis. Our findings highlight the importance of considering multiple correlated pathways for therapeutic targets especially in proximal colon cancer.


Background
Colorectal cancer evolves through a progressive accumulation of genetic and epigenetic alterations that are influenced by the environment and host immunity. The interplay of molecular alterations forms biological interaction networks in colorectal cancer [1]. For example, epigenetic changes are known to be associated with genetic changes, as exemplified by the well-established link between MLH1 promoter hypermethylation and microsatellite instability (MSI) [2,3]. Additionally, microRNAs including MIR21 and MIR155 are reported to regulate gene expression [4][5][6], resulting in abnormal loss and overexpression of proteins [7]. In the tumor microenvironment, tumor-host interactions manifest as lymphocytic reactions directed at the tumor [8]. Given this multitude of molecular interactions, it is challenging to comprehensively understand the perturbations of the complex biological system and identify the key, underlying molecular events that drive colorectal carcinogenesis. Network analysis is recognized as an emerging approach to evaluating multidimensional tumor molecular data with the aim of revealing critical molecular events and pathways [9,10].
Numerous lines of evidence indicate differences between proximal colon cancer, and distal colon and rectal cancer [11][12][13]. When compared with distal colon and rectal cancer, proximal colon cancer exhibits a higher prevalence of microsatellite instability (MSI) and highlevel CpG island methylator phenotype (CIMP-high) [11]. The proximal colon and distal colorectum have different embryologic origins; the former is derived from the midgut with blood supply from superior mesenteric artery, while the latter is derived from the hindgut with inferior mesenteric artery. Moreover, epithelial cells in the proximal colon are exposed to different bowel contents, including microbiota, compared with the distal colon and rectum [13]. Based on these differences in molecular alterations and tumor-host interactions with the microenvironment, the underlying etiology may differ between proximal and distal colorectal cancer. Accordingly, the structure of the biomarker correlation network may differ by tumor location. A better understanding of the connectivity and correlation of molecular events in biomarker networks can provide new insights into colorectal cancer etiology and pathogenesis, potentially advancing prevention and treatment strategies. In cancer biological networks, highly connected molecular markers, called hubs, play essential roles in carcinogenesis [14,15]. We hypothesized that correlation network structures and hub biomarkers might differ between proximal and distal colorectal cancer. Utilizing colorectal cancer databases of the Nurses' Health Study (NHS) and the Health Professionals Follow-up Study (HPFS), we examined the status of key driver mutations, MSI, CIMP, expression of selected proteins in tumor cells, and immune reactions to colorectal cancer.

Comparison of biomarker networks by tumor location
Within the NHS and the HPFS, we identified 1591 colorectal cancer patients with analyzed molecular features of colorectal tumors (Table 1). Biomarkers in this analysis included somatic oncogenic mutations, epigenetic features, protein expression levels, and host immune reactions in colorectal carcinoma. We conducted network analyses stratified by tumor location in a subset (n = 1380) of these patients. This subset included 690 patients with proximal colon cancer and 690 patients with distal colorectal cancer, matched by age and sex. Biomarker availability differed by tumor location only for the following three biomarkers: the intratumoral periglandular reaction (95.9% available in proximal colon, 91.6% in distal colorectum), peritumoral lymphocytic reaction (95.8% in proximal colon, 91.2% in distal colorectum), and tumor infiltrating lymphocytes (TIL) (95.9% in proximal colon, 91.6% in distal colorectum), with the significance level of P = 0.05 after the Bonferroni correction. The demographic characteristics of patients according to tumor location are described in Table 2. Patients with proximal colon cancer tended to have tumors with higher TNM stage and poor differentiation compared with patients with distal colorectal cancer.
In the correlation network analysis, a node represented a tumor tissue biomarker, and an edge was defined as a correlation between two nodes based on Spearman correlation analysis. Compared with the distal colorectal cancer network, the proximal colon cancer network had more edges and greater median degree (Table 3, Fig. 1). Nodes in the proximal colon cancer network tended to have a higher degree than those in the distal colorectal cancer network (P = 0.043 based on a K-S test, Fig. 2). The median degree value was 3.0 in the proximal colon cancer network and 2.0 in the distal colorectal cancer network. The total number of edges was significantly greater in the proximal colon cancer network (n = 173) than in the distal colorectal cancer network (n = 95) in a permutation test where tumor location variable was permuted among 1380 patients (P < 0.0001). Also, we observed significant difference in the total number of edges in another permutation test where values of each biomarker were permuted within proximal colon cancer patients and distal colorectal cancer patients separately (P < 0.0001). In addition, the value of the average clustering coefficient indicated that the nodes in the proximal colon cancer network tended to cluster together more than those in the distal colorectal cancer network (Table 3). We defined hubs as nodes with a high degree centrality (high connectivity with other nodes) based on the overall colorectal cancer network that pooled both proximal colon and distal colorectal cancer. Nodes with degree centrality greater than the 80th percentile were considered as hubs. We found hubs in the proximal colon cancer network, including MSI and MLH1 methylation, but not in the distal colorectal cancer network (Table 3). For these biomarkers, degree centrality was computed as the fraction of nodes to which a node was connected. The values of degree centrality in the proximal colon cancer network and the distal colorectal cancer network were 0. 39   To identify biomarkers that had higher connectivity particularly in the proximal colon cancer network compared with the distal colorectal cancer network, we calculated Cook's distance (see Methods) for each highly-connected marker (those at or above the 80th percentile of the degree distribution in the overall colorectal cancer network) (Additional file 1: Figure S1). In Table 4, MSI showed the largest Cook's distance (0.12) and was highly connected with other markers in the proximal colon cancer network (degree = 21), but not in the distal colorectal cancer network (degree = 9). In the proximal colon cancer network, MSI was positively correlated with These were the subset of biomarkers positively correlated with MSI in the proximal colon cancer network, implying a partly shared mechanism by which MSI promotes perturbation of cellular and molecular functions. Moreover, the methylation markers of MLH1, IGF2, and CACNA1G had a higher degree in the proximal colon cancer network compared with the distal colorectal cancer network. As shown in Fig. 1, negative correlations were observed between KRAS mutation and other biomarkers in the proximal colon cancer network, but not in the distal colorectal cancer network. In the proximal colon cancer network, KRAS mutation was negatively correlated with BRAF mutation (ρ = −0.47), MLH1 methylation (ρ = −0.40), MSI-high

Biomarker networks constructed in non-MSI-high colorectal cancer
The higher frequency of MSI-high in the proximal colon cancer network compared with the distal colorectal cancer network could potentially result in the higher connectivity of the former. Therefore, as a secondary analysis, we restricted our analysis to patients with non-MSI-high cancer, and constructed networks by tumor location. After matching by age and sex, there were 246 patients with non-MSI-high proximal colon cancer and 246 patients with non-MSI-high distal colorectal cancer (Additional file 2: Figure S2). Compared with the distal colorectal cancer network, we observed a slightly higher connectivity and more hubs in the proximal colon cancer network. The total numbers of edges was 64 in the proximal colon cancer network and 56 in the distal colorectal cancer network (Table 3).
Highly connected biomarkers (with degree centrality) were BRAF mutation (0.19) and methylation-related markers including CDKN2A (0.17), IGF2 (0.17), RUNX3 (0.17), CACNA1G (0.15), CRABP1 (0.15), and NEUROG1 (0.15) in the proximal colon cancer network. In the distal colorectal cancer network, two biomarkers were hubs including NEUROG1 methylation (degree centrality = 0.15) and RUNX3 methylation (degree centrality = 0.15). Among non-MSI-high cases, BRAF mutation was the most distinct marker that was strongly correlated with other markers in the proximal cancer network (degree = 10), but not in the distal cancer network (degree = 3) (Cook's distance = 0.17). In the proximal colon cancer network, KRAS mutation tended to be mutually exclusive only with BRAF mutation, indicating that the observed The % numbers indicate the fraction of cases with a given feature among total cases, proximal colon cancer cases, or distal colorectal cancer cases a P value was calculated using a t-test for age and chi-squared tests for categorical variables negative correlations of KRAS mutation with methylationrelated markers and other biomarkers might be confounded by MSI-high in our earlier analysis.

Biomarker networks with equal edge counts
To examine whether the difference in the number of edges across the two networks affected hub identification (designation of nodes as hubs), we modified each network by retaining 100 of the largest Spearman correlation coefficients in absolute values regardless of their P-values, resulting in two networks with 100 edges (Additional file 3: Figure S3). The median degree was 4.0 in the proximal colon cancer network and 1.0 in the distal colorectal cancer network, indicating that edges were confined to Markers with degree centrality at or above the 80th percentile in the colorectal cancer network b CDKN2A promoter hypermethylation MSI microsatellite instability, TIL Tumor infiltrating lymphocytes particular biomarkers in the proximal colon cancer network. When defined as markers with degree centrality at or above the 80th percentile in the colorectal cancer network, there were 11 hubs including methylation-related markers, MSI, BRAF mutation, and TIL in the proximal colon cancer network (Table 3). In contrast, only three methylation-related markers were identified as hubs in the distal colorectal cancer network. This finding supports the characteristics of the proximal colon cancer network observed in the original network analysis, where MSI and methylation-related markers were associated with perturbation of other molecular events.

Sensitivity analyses
Across different significance levels in the Spearman correlation analyses, the total number of edges in the proximal colon cancer network was consistently greater than in the distal colorectal cancer network, supporting a more highly correlated biomarker network in proximal colon cancer (Fig. 3). In addition, we constructed biomarker networks based on the Spearman correlation analyses using binary variables for all markers and still found that the proximal colon cancer network was denser than the distal colorectal cancer network (Additional file 4: Table S1). Moreover, the proximal colon cancer remained denser when the analysis included only biomarkers with missing data in less than 20% of the patients (Additional file 4: Table S1).

Discussion
Within a tumor molecular dataset including major mutational events, MSI, epigenetic features, protein expression status, and host immune reactions in 1380 colorectal cancer patients, we conducted a network analysis to evaluate biomarker network structure in proximal colon cancer and distal colorectal cancer. We found that in proximal colon cancer there were many highly correlated biomarkers, leading to an overall denser network as compared with distal colorectal cancer. In the distal colorectal cancer network, biomarkers had fewer connections and were less clustered, resulting in fewer hubs. This finding indicates that carcinogenic events in distal colorectal cancer tended to occur independently from each other. In the proximal colon cancer network, MSI was a hub and had differential connectivity with other biomarkers compared with the distal colorectal cancer network. In both the proximal colon and distal colorectal cancer networks, we observed common biomarkers that positively correlated with MSI including methylation-related markers, BRAF mutation, and immune reactions, while overexpression or loss of protein expression associated with MSI were further found only in the proximal colon cancer network. In addition to MSI, CpG island methylation in MLH1, IGF2, and CACNA1G genes were highly connected to other biomarkers mainly in proximal colon cancer. Among non-MSI-high patients, BRAF mutation status was the most distinct marker that was strongly associated with other molecular events in the proximal colon cancer network, but not in the distal colorectal cancer network. The sensitivity analyses consistently showed a denser network in proximal colon cancer when compared with the distal colorectal cancer network. Our results indicate that many molecular events occur in relation to MSI in proximal colon cancer, and that MSI and BRAF mutation played important roles in the carcinogenic process of proximal colon cancer.
In colorectal cancer and other types of cancer, previous biomarker network studies have shown the importance of interdependence of mutation, methylation, and protein expression [16,17]. In a biological network, hubs (highly connected nodes) are considered to be drivers playing an essential role during tumorigenesis [14,15]. Previous studies showed that MLH1 promoter methylation causes mismatch repair deficiency and MSI [2,3]. In sporadic cancer, it is not well known whether MSI causes mutations in oncogenes and tumor-suppressor genes, and aberrant protein expressions. Earlier studies reported that genomic instability conferred by somatic mutational alterations further generates mutations in oncogenes or anti-oncogenes [18,19]. Our results suggest that the role of MSI differ by tumor anatomic location; in proximal colon cancer, MSI might lead to perturbations of many molecular events, and in distal colorectal cancer molecular events are more likely independent from MSI. Consistent with our findings, MSI and CIMP-high were highly correlated and occurred more frequently in proximal colon cancer than in distal colorectal cancer [11]. Methylation signatures  measured in our study were CIMP-specific markers which were correlated with each other. Thus, higher frequencies of MSI and CIMP-high might contribute to the higher connectivity of the proximal colon cancer network. However, even when we excluded MSI cancer patients from the analyses, the higher connectivity persisted in the proximal colon cancer network. In addition, when we modified the two networks such that each had 100 edges, corresponding to the 100 most positive or negative correlation coefficients, the proximal colon cancer network remained to have more hubs and greater median degree than the distal colorectal cancer network. Our findings provide new insights into the carcinogenesis of proximal colon cancer, which possibly exhibits highly interactive biological mechanisms.
BRAF oncogenic mutations are observed in 10% to 20% of colorectal cancer patients and are associated with MSI-high and CIMP-high phenotypes [19,20]. The serine/threonine-protein kinase BRAF is involved in the mitogen-activated protein kinase (MAPK) pathway, which is associated with proliferation, cell growth, and differentiation [20]. In our study, CIMP-high was associated with BRAF mutation in both proximal colon and distal colorectal cancer. Previous studies observed CIMP-high and BRAF mutations in an early-stage colorectal neoplasm, and acquisition of BRAF mutation was considered to be mediated by DNA hypermethylation of several genes, including IGFBP7 and BMP3 [18]. The close relationship between CIMP and BRAF mutation may drive carcinogenesis regardless of tumor location.
A large body of literature demonstrated that colorectal cancer patients with non-MSI-high and mutant BRAF were associated with the highest mortality, whereas patients with MSI-high/BRAF-wild-type showed the lowest mortality when both MSI and BRAF mutation status were considered [21][22][23][24]. Among proximal colon cancer patients, the non-MSI-high/BRAF-mutant subtype was associated with a worse prognosis [25]. Together with these previous studies, our findings suggest the critical role of BRAF mutation in the absence of MSI-high especially in the proximal colon carcinogenesis.
Our results showed the strong negative correlation between BRAF and KRAS mutations specifically in proximal colon cancer. Both BRAF and KRAS oncogenes encode proteins involved in the MAPK pathway, and BRAF mutation has been reported to be mutually exclusive with KRAS mutation in colorectal cancer [26,27]. BRAF is more frequently mutated in the proximal colon than distal colorectum, while the frequency of KRAS mutation is more similar throughout the colorectum except for the cecum [11]. The biological correlation pattern between BRAF and KRAS might depend on tumor location.
The distal colorectal cancer network was less densely connected than the proximal colon cancer network. Although we did not find biomarkers with high degree centrality, the promoter methylations, BRAF mutation, and TIL tended to be correlated with other molecular events. Distal carcinoma was previously reported to be more responsive to drug regimens, including 5fluorouracil-based chemotherapeutics and single-agent anti-epidermal growth factor receptor (EGFR) antibodies, while proximal carcinoma tended to show more resistance to those treatments [12,26]. Our results from the network analysis found that oncogenic processes are relatively independent from one another in distal colorectal cancer. Taken together, our data support a biological difference between colorectal carcinoma by tumor location.
We recognize that there are several limitations to our study. Although we chose widely recognized pathological biomarkers in colorectal cancer, our data did not cover all molecular pathological markers that have been reported in the literature, such as the mutation status of SMAD4 and PTEN [28,29]. Our selection of the biomarkers might have resulted in a greater connectivity of the biomarker network in the proximal colon. Measurement errors could have existed in protein expression analyses using immunohistochemistry (IHC). Nonetheless, based on reviews by two independent pathologists, most of the IHC markers showed agreement levels generally considered to be reasonable in pathology research (kappa coefficient > 0.6). We conducted network analyses focusing on the proximal colon and distal colorectum, but not on more detailed segments (the cecum, ascending colon, hepatic flexure, transverse colon, splenic flexure, descending colon, sigmoid colon, rectosigmoid junction, and rectum). The importance of examining these segments warrants further investigations with even larger sample sizes. These results contribute to our understanding of colorectal carcinogenesis, suggesting a different mechanism by tumor location. Although our findings were corroborated by sensitivity analyses, future validation studies with independent data sets are necessary to confirm these findings.
Despite the limitations, our study has several strengths that distinguish it from previous studies. First, we measured different kinds of tumor biomarkers including somatic mutations, methylation levels, MSI status, immune reactions, and protein expressions. These measurements provided a reasonably comprehensive view of molecular interplay in the networks and enabled the identification of important carcinogenic pathways. Second, our overall sample size with 1380 cases enabled us to conduct robust network analyses using multiple biomarkers. Third, we collected colorectal tumors from hospitals and pathology laboratories across the U.S. (rather than one or a few hospitals), which increases generalizability of our findings.

Conclusions
Biomarkers in proximal colon cancer possessed higher connectivity while those in distal colorectal cancer tended to be independent from each other. In proximal colon cancer, MSI-high and BRAF mutation occurred in relation to many other tumor features, indicating their important roles during carcinogenesis. Our findings highlight the necessity of a systems therapeutic approach that can target both specific and multiple pathways of the proximal colon cancer network.

Study population and design
Initiated in 1976, the NHS enrolled 121,701 female nurses aged 30-55 years. Established in 1986, the HPFS enrolled 51,529 male health professionals including dentists, optometrists, osteopaths, pharmacists, podiatrists, and veterinarians aged 40-75 years. Within these two cohorts, colorectal cancer patients were ascertained by biennial questionnaires, the National Death Index (for unreported lethal cases), and reviews of medical records. Study physicians reviewed medical and pathological records to retrieve detailed information on colorectal tumors including TNM stage, differentiation, and bowel subsites. For bowel subsite sections, the proximal colon included the cecum, ascending colon, hepatic flexure, and transverse colon; the distal colorectum included the splenic flexure, descending colon, sigmoid colon, rectosigmoid junction, and rectum.
We collected formalin-fixed paraffin-embedded (FFPE) archival tumor tissue blocks from pathology laboratories in U.S. hospitals where patients underwent tumor resection. A centralized pathological review of hematoxylin and eosin-stained tissue sections was conducted by a single pathologist (S.O.) for all colorectal carcinoma cases. Tumor differentiation was categorized into well to moderate or poor, based on the extent of glandular areas. Based on the availability of tumor tissue blocks, 1591 cases were selected for the molecular and pathological analyses. Patients with proximal colon cancer were matched by age and sex to patients with distal colorectal cancer by randomly selecting a patient diagnosed with proximal colon cancer from the same sex and age category (<50, 50-59, 60-69, or ≥70 years). Subsequently, out of the 1591 patients in the original dataset, 1380 patients (690 patients with proximal colon cancer and 690 patients with distal colorectal cancer) were included in the network analysis dataset. For each marker, patients were excluded from an analysis when data on that marker were not available. Percentages of patients with available biomarkers are shown in Table 1.

Molecular pathological analysis
In the NHS and the HPFS, we utilized a colorectal cancer database which contained biomarker data on molecular alterations reported to play important roles in carcinogenic pathways in colorectal cancer. Tumor molecular and pathological markers in the database included mutational events, epigenetic features, protein expression levels, and host immune reactions in colorectal carcinoma. The analyses were performed as previously described. In FFPE tissues, DNA extraction and pyrosequencing were conducted for BRAF (codon 600), KRAS (codons 12, 13, 61, or 146), and PIK3CA (exons 9 and 20) [30][31][32]. The analysis of MSI was performed using 10 microsatellite markers (D2S123, D5S346, D17S250, BAT25, BAT26, BAT40, D18S55, D18S56, D18S67, and D18S487) [24]. We defined MSI-high as instability in ≥30% of the markers and MSI-low/microsatellite stability (MSS) as instability in <30% of the markers [24]. To quantify the amount of Fusobacterium nucleatum DNA, quantitative polymerase chain reaction (PCR) assay was performed using the Taq-Man primer/probe sets (Applied Biosystems) for the nusG gene of Fusobacterium nucleatum and for SLCO2A1 as a reference human gene [33]. Amplification and detection of DNA was performed with the StepOnePlus Real-Time PCR Systems (Applied Biosystems).
RNA was extracted from the colorectal tumor tissue and adjacent non-tumor tissue within FFPE samples, and cDNA was synthesized as previously described [4]. The expression levels of MIR21 and MIR155 were quantified using miScript PCR System (Qiagen, Valencia, CA) with the StepOnePlus Real-Time PCR Systems (Applied Biosystems, San Diego, CA) [65].

Statistical and network analysis
To compare demographic characteristics of proximal colon cancer patients with distal colorectal cancer patients, a chi-squared test was used to compare two or more categorical variables by location, and a t-test was conducted to compare age by location.
In the correlation network analysis [66], a node represented a tumor tissue biomarker. The structure of correlations was captured by placing an edge between any two nodes whose associated biomarkers exhibited statistically significant Spearman correlation, where the significance level for correlations was chosen to be 3.5 × 10 −5 (= 0.05/ 1431, based on the Bonferroni correction). The pairwise correlation analysis was conducted as a complete case analysis in which patients without either biomarker information were excluded from the analysis. In the correlation analyses, we used continuous (for MSI and markers measuring levels of methylation, T-cell densities, miRNA expression, and Fusobacterium nucleatum), ordinal (for markers measuring tumor protein expression levels), and binary (for markers assessing mutation status) variables. The degree of a node was defined as the number of edges adjacent to the given node. We used a Kolmogorov-Smirnov (K-S) test to evaluate the distance between the cumulative degree distributions of the proximal colon cancer network and the distal colorectal cancer network. For each marker, degree centrality was computed as the fraction of nodes to which a node was connected (i.e., degree of a node divided by the number of all nodes in the network). A network node (marker) showing a high degree of connectedness, often referred to as a hub, is more likely to play an essential role in the disease network [14]. In the current study, hubs were defined as nodes with degree centrality greater than the 80th percentile based on overall colorectal cancer network that pooled both proximal colon and distal colorectal cancer. In each network, average clustering coefficient quantifies the overall clustering of nodes, indicating the tendency of markers to clustering together in the network [67]. For each node, a clustering coefficient was computed as the proportion of directly connected neighbors, and then an average of clustering coefficients was calculated in the network. To identify the markers that show distinct correlation patterns in proximal colon cancer versus distal colorectal cancer, we computed Cook's distance based on linear regression analysis [68]. In this analysis, the degree of each node in the distal colorectal cancer networks was regressed on the degree of each node in the proximal colon cancer networks.
To assess the robustness of our results from the network analysis, we conducted permutation tests and sensitivity analyses. First, to evaluate the null hypothesis of no difference in network edge counts by tumor location, we permutated the dataset 10,000 times by randomly assigning tumor location in 1380 patients. Each time, networks were constructed by tumor location, and the difference in the total number of edges was computed. The two-sided P value was obtained as the proportion of random permutations that resulted in a difference in edge counts that was greater than or equal to the observed difference. Similarly, another permutation test was conducted by permuting values of each biomarker in 690 patients with proximal colon cancer and in 690 patients with distal colorectal cancer separately. In this process, we broke the linkage between each pair of biomarkers but retained imbalanced data (e.g., difference in mutation frequency) between tumor location. Second, across different significance levels of the Spearman correlation, we compared the number of edges in the proximal colon cancer network with that in the distal colorectal cancer network. Third, we used binary variables for all markers in the Spearman correlation analysis, and we constructed network models by tumor location. Fourth, we conducted a sensitivity analysis to include only selected markers that were available for more than 80% of the patients (BRAF, KRAS, PIK3CA, MLH1 methylation, CDKN2A methylation, CACNA1G methylation, CRABP1 methylation, IGF2 methylation, NEUROG1 methylation, RUNX3 methylation, SOCS1 methylation, MSI, LINE-1 methylation level, PTGS2, peritumoral lymphocytic reaction, intratumoral periglandular reaction, and TILs).
All the statistical analyses were carried out with SAS software (version 9.4, SAS Institute, Cary, NC). Network analysis was carried out with Python (version 2.7). All P values were two-sided.

Additional files
Additional file 1: Figure S1. The biomarker network in colorectal cancer. A node represents a molecular feature, and an edge specifies the Spearman correlation between two markers with a significance level of