Predicting Mechanisms of Toxicity for Drug Development

Abstract

The aim of this thesis is to predict different mechanisms of toxicity from the metabolomic response of HepG2 liver cells. In order to utilize the metabolomic data the a semisupervised machine learning approach is investigated, namely the cluster-then-label approach. The research focuses on the unsupervised part due to the centrality to this method. The dose-dependency within the data is modelled by clustering the dose-response curves according to their shape and transforming the feature space to a categorical one. This dataset is then clustered with the K-Modes algorithm. The analysis of the experimental data has shown that it is possible to distinguish toxic from non-toxic compounds on individual dose level though mechanisms can not clearly be distinguished. The proposed method is not able to clearly distinguish between toxic and non-toxic compounds or between the mechanisms of toxicity. It is hypothesized that the lack of mutually exclusive labels makes the prediction harder. Furthermore, the model could benefit from a more fine-grained dose levels in the identified range.