Inferring effective cancer combination therapies using network-based multi-omics data integration

Abstract

Despite advances in cancer treatments, disease recurrence and treatment resistance remain a leading challenge for patients. Genetic and epigenetic changes enable tumour cells to continuously proliferate and evade apoptosis and growth suppression, causing resistance to monotherapy-based interventions. Drug combination therapies hold the promise of higher efficacy by simultaneously targeting compensatory mechanisms in tumours and reducing toxicity through the use of lower drug doses. However, identifying effective combination therapies and the biological contexts in which they could be beneficial is a challenge. The work presented in this Thesis aimed to implement network modelling to predict effective drug combinations utilising a comprehensive drug combination dataset and well-characterised cancer cell line genomic, transcriptomic, proteomic and methylation data. We hypothesised that (i) drug efficacy as measured by cancer cell viability is adequate, and synergy between drugs is not necessarily required for a clinically effective outcome; (ii) network-based approaches better inform drug modulation on protein-protein interaction networks and, therefore, improve drug combination response prediction in cancer cell lines. To provide background and contextualise my research, in Chapter 1 of this Thesis, I provided an overview of cancer biology, biomarkers and therapeutics, with a focus on network modelling and machine learning. In Chapter 2, I presented the drug combination and cancer cell line data, and the computational methods applied. Through data integration, this Thesis created one of the most extensive drug combination datasets, including the highest number of cancer cell lines screened to date. In Chapter 3, I detailed the overall data landscape and reproducibility analysis of data quality. Furthermore, I integrated cell line-specific multi-omics data to model differential drug responses. In Chapter 4, I described the linear regression pipeline and outcome of biomarker analyses and pinpointed specific combination therapy-associated robust biomarkers which improved efficacy over single-agent activity. Chapter 5 described my work to reconstruct drug sub-networks through a random walk with restart algorithm, mimicking drug modulation on the human protein-protein interaction network and identifying the topological relationships between drugs. In Chapter 6, I detailed the generation of a machine-learning random forest model incorporating network-based approaches to predict drug combinations and highlighted the importance of biological information on model generalisability. In summary, in this Thesis, I developed a new approach to identifying clinically meaningful drug combinations, moving beyond classical synergy and demonstrating the promise of systems-level information through network approaches to improve the prediction of drug combination activity. Ultimately, this study advanced our capabilities in drug response prediction to improve patient treatment. In the Appendices, I briefly described my development of Python-based software, BEstimate, for the gRNA design and an in silico annotation tool for CRISPR base editing.