Integrative bioinformatics and graph-based methods for predicting adverse effects of developmental drugs

Adverse drug effects are complex phenomena that involve the interplay between drug molecules and their protein targets at various levels of biological organisation, from molecular to organismal. Many factors are known to contribute toward the safety profile of a drug, including the chemical properties of the drug molecule itself, the biological properties of drug targets and other proteins that are involved in pharmacodynamics and pharmacokinetics aspects of drug action, and the characteristics of the intended patient population. A multitude of scattered publicly available resources exist that cover these important aspects of drug activity. These include manually curated biological databases, high-throughput experimental results from gene expression and human genetics resources as well as drug labels and registered clinical trial records. This thesis proposes an integrated analysis of these disparate sources of information to help bridge the gap between the molecular and the clinical aspects of drug action. For example, to address the commonly held assumption that narrowly expressed proteins make safer drug targets, an integrative data-driven analysis was conducted to systematically investigate the relationship between the tissue expression profile of drug targets and the organs affected by clinically observed adverse drug reactions. Similarly, human genetics data were used extensively throughout the thesis to compare adverse symptoms induced by drug molecules with the phenotypes associated with the genes encoding their target proteins. One of the main outcomes of this thesis was the generation of a large knowledge graph, which incorporates diverse molecular and phenotypic data in a structured network format. To leverage the integrated information, two graph-based machine learning methods were developed to predict a wide range of adverse drug effects caused by approved and developmental therapies.