Two-step models for tumour-drug response using heterogeneous high-dimensional assays
- Zweischrittmodelle zur Beschreibung der Pharmakodynamik von Tumormedikamenten mit heterogenen, hochdimensionalen Assays
Kusch, Nina; Schuppert, Andreas (Thesis advisor); Mitsos, Alexander (Thesis advisor)
Aachen : RWTH Aachen University (2021, 2022)
Dissertation / PhD Thesis
Dissertation, Rheinisch-Westfälische Technische Hochschule Aachen, 2021
Cancer constitutes a major challenge to public health worldwide. It comprises a large group of heterogeneous diseases, which are caused by the complex process of tumourigenesis that induces extensive genomic, epigenomic and transcriptomic alterations in cells. Identifying crucial links between specific profiles of alterations and either disease progression or therapeutical outcome would aid in the treatment of individual patients within the framework of personalised medicine. Hence, there is a need for powerful and robust multi-omics models of cellular sensitivity to antineoplastic drugs that can identify stable candidates for computational biomarkers. Such models need to be able to accurately integrate structurally heterogeneous molecular predictors from distinct data sources and leverage the complementary response-related information contained therein. In this thesis, we propose a novel pan-cancer multi-omics modelling approach for drug sensitivity that directly addresses three pivotal challenges arising in this context: firstly, we devised a powerful two-step multi-omics modelling framework for classifying cellular drug sensitivity that successfully integrates structurally heterogeneous predictors stemming from distinct high-dimensional data sources. Secondly, the resulting models are equipped with a predictor preprocessing scheme that enables them to fit predictor weights accurately and in an unbiased fashion, counteracting well-documented predispositions towards continuous-valued predictors. Thirdly, these models are easily and intuitively interpretable and return a wealth of additional information that allows for comprehensive posterior evaluations of the impact of diverse predictors on the model performance. Therefore, the two-step modelling framework enables users to identify promising candidates for both simple and complex drug-specific computational biomarkers and to assess dependencies and redundancies in information content between them. The two-step approach was evaluated on a well-known publicly accessible data base in order to demonstrate that these goals have indeed been met. In particular, the predictive performance of the resulting models was compared extensively to both an internal and external standard and found to be comparable in a majority of the studied cases.