Using transcriptomics to understand cancer progression and predict response to therapy

Abstract

Transcriptomics data provide useful information to better understand molecular phenotypes in cancer. Epithelial-to-mesenchymal transition (EMT) is one of these molecular phenotypes that is hijacked by cancer cells to obtain mobile mesenchymal characteristics which may assist cells to intravasate into blood stream, generate circulating tumour cells (CTCs) and metastasize to distant organs. CTCs also have heterogeneity in their molecular phenotypes and it is of utmost importance to understand these variations to be able to understand differences in their therapy response and use them to monitor treatment outcome. Using transcriptomics, we can also explore and predict molecular phenotypes associated with sensitivity to different therapeutic regimen.

Although EMT is a single molecular phenotype, it can be regulated through different underlying molecular mechanisms, leading to differences in response to therapies. To identify samples with TGFβ-driven EMT, I derive a gene expression signature of EMT induced by TGFβ using metaanalysis and transcriptomics data integration. This signature is able to identify transcriptional profiles arising during TGFβ-driven EMT, and yields highly consistent results in multiple independent pan-cancer cell lines and patients data. Samples fitting this signature show lower number of mutations in elements of TGFβ signalling, poorer overall survival outcome and preferential response to certain drugs.

Meta-analysis and data integration such as the above require careful attention to batch effects in datasets. I apply different batch correction methods in order to perform general normalisation or obtain differentially expressed genes (DEGs) in integrated transcriptomics data sets. Further, to classify the fit of individual samples to a gene signature, I apply existing single-sample scoring methods. However, these methods all use information borrowed from the whole set of samples, meaning they are not truly single sample scores. To address this, I developed a rank-based scoring method, called singscore, which generates more stable scores that are independent from sample size and composition in a dataset.

CTCs are integral to cancer progression, but while these cells are extremely rare in blood, they have great potential to provide a real-time representation of cancer progression and treatment efficacy. I perform an assessment of current markers for enrichment and/or detection of CTCs, and then, introduce new CTC markers, including general, epithelial and mesenchymal markers obtained by analysing multiple breast cancer and blood data sets. I then assess their expression in publically available CTC data and a number of in-house patient samples.

Finally, I use pharmacogenomics data in breast cancer cell lines and the singscore method to predict drug response outcome for 90 drugs based on gene expression data, which have been shown to be the most predictive molecular feature in breast cancer. I derive drug sensitivity signatures by quantifying associations between gene expression and drug response and evaluate the utility of these gene signatures using cell lines, PDX models and patient data and show consistent pattern of response across independent data sets. Further associations between drug sensitivity scores and EMT phenotype are assessed.