br Cohort A validation set the genomic location score


Cohort A validation set, the genomic location score predicted genomic left-side tumors within the anatomic right-side tumors (Figure 4C) (AUC ¼ 0.86) and genomic right-side tumors within the anatomic left-side tumors (Figure 4D) (AUC ¼ 0.89). Score values for each sample in Cohort A are found in Supplemental Table 1 (in the online version).
To show that the utility of the genomic location score is not limited to Cohort A, it was computed for samples in Cohort B using gene MCC950 values for the 7-panel genes (Table 4) assessed by RNA-seq (see Methods). The genomic location score predicted
Colon Tumor Side by Gene Expression
Table 4 Genes Used to Compute Genomic Location Score
Side of
Feature ENTREZID SYMBOL Elevation
anatomic location in Cohort B with AUC ¼ 0.92. Thus, the relationship between genomic location score and anatomic location was comparable in Cohort B and Cohort A validation set.
Discussion
Tumor location is increasingly a consideration in planning the treatment of colon cancer.18 Because anti-tumor drugs act at the molecular level, there is need for greater knowledge of the molecular differences between left- and right-side colon cancer. In this study, we analyzed the degree to which gene expression determines the side of a tumor. Using a methodology that reduces bias owing to sta-tistical methodology, we showed that gene expression predicted tumor location with 81% accuracy; however, 6% of tumors were
Figure 4 Receiver Operator Characteristic Curves Were Plotted in the Cohort A Validation Set for the 7-Probe Genomic Location Score as a Predictor of Genomic Left-side Tumors (A), Genomic Right-side Tumors (B), Genomic Left-side Tumors Within the Anatomic Right-side Tumors (C), and Genomic Right-side Tumors Within the Anatomic Left-side Tumors (D)
sensitivity
sensitivity
1 – specificity 1 – specificity D
sensitivity
sensitivity
1 – specificity
Abbreviation: AUC ¼ area under the curve.
consistently predicted to have the opposite side from anatomic location, and 13% of tumors had no clear predicted location.
The differences between left-side and right-side tumors are often attributed to the fact that the right colon is formed from the midgut and the left colon is formed from the hindgut in the embryo.3 Indeed, 2 of the probes in the 7-probe genomic location score are HOX family genes, which are expressed differentially across the embryo.19 How-ever, the hindgut and midgut join at a nonspecific location in the transverse colon about one-third of the way from the splenic flexure.2 This could account for genomic left tumors that are anatomically right, but not genomic right tumors that are anatomically left, which we discovered. We also discovered that genomic non-consensus tu-mors are almost equally distributed between the right and left colon. Apparently, colon embryology is a factor in tumor genomic location, but not the whole story.
Although the derivation of genomic location by network analysis may be cumbersome to reproduce in multiple cohorts, the genomic location score can be derived and applied in a straightforward manner. The genomic location score could be used in statistical models of treatment response in comparison with tumor location to assess the relative predictive significance of genomic location and anatomic location. A study of Variegation nature, beyond the scope of this project, would be to test the relative significance of anatomic or genomic location to predict response to anti-EGFR therapy for patients with KRAS WT metastatic colorectal cancer.
Differences in the prevalence of CMSs between tumor sides were reflected in genomic location as well as anatomic location. For example, CMS1 tumors were more prevalent in anatomic right-side than left-side colon cancer.4 In this study, tumors that were genomic right and anatomic left were more likely to be CMS1 than those that were genomic left and anatomic left (P ¼ 1.7 10 15) (Figure 3). However, the fact that significant percentages of both right-side and left-side tumors (with respect to both anatomic and genomic locations) were CMS2 and CMS4, shows a degree of in-dependence between location and CMS. Although a consensus molecular subtype is considered molecularly homogeneous to some degree, genomic location is one sign of heterogeneity within some CMS subtypes.