Novel systems biology tools for the identification of biomarkers and drug targets in cancer research

Abstract

Cancer, the second cause of death all around the world, can only be beaten by a complete understanding of its underlying biology. Cancer is characterized by a complex reprogramming of key cellular processes in order to support growth and proliferation, and evade the immune system. The advent of high-throughput (omics) technologies have enabled to investigate cancer from a more holistic point of view, leading to the development of the field of Systems Biology, a vital part of the current fight against cancer. The main objective of this doctoral thesis is to provide the community with novel system biology tools for the identification of drug targets and biomarkers in cancer. This doctoral thesis can be divided into two distinct parts. The first part is focused on constraint-based modeling and the development of computational tools for identifying therapeutic targets using genome-scale metabolic models and transcriptomics data. First, the robust Metabolic Transformation Analysis (rMTA) approach is presented. rMTA proposes a more robust version of a previously developed algorithm, MTA, which searches for genetic/drug perturbations that transform a disease metabolic phenotype back to the healthy situation. Second, we introduce gMCStool, an automated tool to predict metabolic vulnerabilities in cancer. gMCStool exploits the concept of genetic Minimal Cut Sets (gMCSs), a network-based approach to synthetic lethality, previously developed in our group. gMCStool adapts previous algorithms to RNA-seq data and Human1, the most recent human metabolic reconstruction. Proof-of-concept of gMCStool is presented for multiple myeloma (MM). We elucidated and in vitro validated the dependence on CTPS1 of a subgroup of MM patients. The second part of this doctoral thesis focuses on the search of biomarkers that guide patient stratification according to prognosis or response to therapy. Firstly, we present a method that uses RNA-seq experiments in a nonconventional way for predicting the gene promoter activity. We identify a subset of gene promoters that improves the prognosis of MM patients in combination with some of the stablished high-risk genetic biomarkers. Secondly, we present a novel machine learning technical for linear regression, called BOSO (Bilevel Optimization Selector Operator), which provides a more elegant solution to the feature selection problem. BOSO improves existing methodologies in datasets where the number of variables is higher than the number of samples, which are frequent in biomedical research. We provide a proof of concept of BOSO for the prediction of drug sensitivity in cancer cell lines. Detailed computational analysis and experimental validation is conducted for methotrexate, a well-studied drug targeting cancer metabolism. Finally, all the research, models and tools discussed in this dissertation have been published in open repositories, including the COBRA toolbox, GitHub, CRAN or Docker Hub.