Reconciliation of classification systems defining molecular subtypes of colorectal cancer: interrelationships and clinical implications

Abstract

Recently we published two independent studies describing novel gene expression-based classifications of colorectal cancer (CRC). Notably, each study stratified CRC into a different number of subtypes: one reported 3 subtypes, whereas the second highlighted 5. Given that each ascribed clinical significance, distinctive biology, and therapeutic prognosis to the different subtypes, we sought to reconcile this apparent incongruity in subtype stratification of CRC, and to interrelate the results. To do so, we each evaluated the other's data sets and analytical methods and discovered that the subtypes and their classifiers are, in fact, clearly related to each other; indeed, the 5 subtype outcomes can be coalesced into the same three. In addition to presenting this clarification, we briefly discuss how both classification methods can be viewed within the broader literature on CRC subtypes, and potentially applied.

Keywords: CIMP; MSI; cancer subtypes; cetuximab; colorectal cancer; consensus clustering; serrated pathway; therapy resistance.

Figure 1. Overview of CRC classification studies. (A and B) Graphic showing the clinical and biological characteristics and gene signatures of Colon Cancer Subtypes (CCS) (A) and CRCassigner subtypes (B). SSA, sessile serrated adenoma; TA, transit-amplifying; CR-TA, cetuximab-resistant TA; CS-TA, cetuximab-sensitive TA; DFS, disease-free survival.

Figure 2. Relationships of colorectal cancer subtype classifications. (A and B) Heatmaps depict the AMC-AJCCII-90 data set classified according to the CRCassigner signature (A) and the Sadanandam et al. core data set classified based on the CCS classification (B). Columns represent patients; rows indicate CRCassigner genes (A) or CCS classifier genes (B). Colors represent relative gene expression levels; blue signifies low expression, and brown high expression. (C and D) Heatmaps indicate association of CRCassigner subtype samples (left) with CCS group samples (top) for the AMC-AJCCII-90 set (C) and the Sadanandam et al. core data set (D). Colors indicate significance of association; green signifies a low association, and red a high association. P values are determined using hypergeometric tests.

Figure 3. Validation in the TCGA data set. (A) Heatmap depicts the TCGA data set classified by both taxonomies. Columns represent patients; rows indicate genes from the CCS classifier. Colors represent relative gene expression levels; blue signifies low expression, and brown, high expression. (B) Heatmap indicates association of CRCassigner subtype samples (left) with CCS group samples (top) for the TCGA set. Colors indicate significance of association; green signifies a low association, and red a high association. P values are determined using hypergeometric tests. (C) Graph depicts the association of clinical and (epi)genetic characteristics with both classification methods for the TCGA data set. Features that are significantly associated with both classification schemes are indicated. P values are determined using Chi-square tests.