Applications of intertumoural, intratumoural and intermolecular heterogeneity for personalised medicine in colorectal cancer

Abstract

Colorectal cancer (CRC) is a heterogeneous disease, both at the molecular level and in the context of patients' responses to treatment. Few biomarkers are currently in place that can help to stratify patients in the clinical setting. This thesis begins by describing inter- and intratumoural transcriptomic heterogeneity in CRC, before extending to the integration of multiomics data for a system-wide view of the pathways active in this disease. Initially, taking previously described gene expression subtypes of CRC that have prognostic indications and potential associations with patient outcomes/drug responses, I redefined their gene expression signatures to a smaller gene set (measurable on a platform that has previously been approved for clinical use) using a consensus of statistical gene selection and class prediction methods, thus enabling future subtype-based prospective clinical trials. Subtyping with this new gene set and platform was highly accurate the previous standard, and has the additional benefit that it can also be applied to large archives of formalin-fixed paraffin-embedded tissues for retrospective analyses. Furthermore, I explored the intratumoural heterogeneity of these subtypes using machine learning techniques and single-cell data, concluding that they co-exist in the vast majority of tumours. Using these subtype sub-populations, I was able to significantly improve prognostic power in survival models versus traditional "bulk" subtyping, and identify subsets of patients that respond best to already-available therapies (as well as those who could be spared unnecessary toxicities). For example, early stage patients whose tumours were deemed to have a high stem-like subpopulation by computational deconvolution had significantly poorer prognosis than those with a low subpopulation, while no prognostic difference was observed between patients with bulk stem-like versus other bulk subtype tumours. In addition, TA subtype sub-populations were significantly higher in patients and pre-clinical models of CRC who responded/were sensitive to cetuximab. Finally, I have used a Bayesian latent variable machine learning framework to integrate multi-omics data (including gene, miRNA and protein expression, methylation, copy number and mutations) and clinicopathological variables form the TCGA CRC database. In this way, I found patterns of co-expression across molecular levels that relate to complex interactions between clinically interpretable covariates. The results from this analysis included novel biomarkers that had significant and context-specific prognostic implications. Overall, in this thesis, I present several characterisations of CRC's multi-faceted heterogeneity. I demonstrate how the existing transcriptomic CRCAssigner intertumoural subtypes can be profiled in a clinically-practicable manner, expand on our understanding of these subtypes by quantifying their co-existence within individual tumours, and move beyond transcriptomics to delineate CRC heterogeneity on a panmolecular scale.