De novo transcriptomic subtyping of colorectal cancer liver metastases in the context of tumor heterogeneity

Abstract

Background Gene expression-based subtyping has the potential to form a new paradigm for stratified treatment of colorectal cancer. However, current frameworks are based on the transcriptomic profiles of primary tumors, and metastatic heterogeneity is a challenge. Here we aimed to develop a de novo metastasis-oriented framework.

Methods In total, 829 transcriptomic profiles from patients with colorectal cancer were analyzed, including primary tumors, liver metastases, and non-malignant liver samples. High-resolution microarray gene expression profiling was performed of 283 liver metastases from 171 patients treated by hepatic resection, including multiregional and/or multi-metastatic samples from each of 47 patients. A single randomly selected liver metastasis sample from each patient was used for unsupervised subtype discovery by nonnegative matrix factorization, and a random forest prediction model was trained to classify multi-metastatic samples, as well as liver metastases from two independent series of 308 additional patients.

Results Initial comparisons with non-malignant liver samples and primary colorectal tumors showed a highly variable degree of influence from the liver microenvironment in metastases, which contributed to inter-metastatic transcriptomic heterogeneity, but did not define subtype distinctions. The de novo liver metastasis subtype (LMS) framework recapitulated the main distinction between epithelial-like and mesenchymal-like tumors, with a strong immune and stromal component only in the latter. We also identified biologically distinct epithelial-like subtypes originating from different progenitor cell types. LMS1 metastases had several transcriptomic features of cancer aggressiveness, including secretory progenitor cell origin, oncogenic addictions, and microsatellite instability in a microsatellite stable background, as well as frequent RAS/TP53 co-mutations. The poor-prognostic association of LMS1 metastases was independent of mutation status, clinicopathological variables, and current subtyping frameworks (consensus molecular subtypes and colorectal cancer intrinsic subtypes). LMS1 was also the least heterogeneous subtype in comparisons of multiple metastases per patient, and tumor heterogeneity did not confound the prognostic value of LMS1.

Conclusions We report the first large study of multi-metastatic gene expression profiling of colorectal cancer. The new metastasis-oriented subtyping framework showed potential for clinically relevant transcriptomic classification in the context of metastatic heterogeneity, and an LMS1 mini-classifier was constructed to facilitate prognostic stratification and further clinical testing.