Novel machine learning methods for computational chemistry

The experimental assessment of absorption, distribution, metabolism, excretion, toxicity and related physiochemical properties of small molecules is counted among the most timeand cost-intensive tasks in chemical research. Computational approaches, such as machine learning methods, represent an economic alternative to predict these properties, however, the limited accuracy and irregular error rate of these predictions restrict their use within the research process. This thesis introduces and evaluates new ideas to enhance the acceptance and usage of kernel-based machine learning models in chemical research. The first part of the thesis investigates different approaches to improve the quality of machine learning predictions in drug discovery. By taking the precise chemical application into account we derive a new virtual screening algorithm, StructRank, which enables to focus on the correct ranking of compounds with high binding affinities. Then, the limits of single and ensemble learning methods are analyzed in the context of hERG inhibition. Since the drug discovery process often requires the assessment of new chemical series different to previously examined structures, we introduce and evaluate a clustered cross-validation scheme that stresses the extrapolation capacity of models. We present a local bias correction to incorporate new measurements efficiently and without the need for model retraining. The second part of the thesis is concerned with two different approaches to assess the reliability and interpretability of kernel-based prediction models. The first approach builds on the visual interpretation of predictions based on the most relevant training compounds. A compact method to calculate the impact of training compounds on single predictions is derived and the resulting visualizations are evaluated in a questionnaire study. The second approach addresses interpretability in terms of chemical features. Here, local gradients are employed to measure the local influence of specific chemical features on a predicted property. The capacity of this approach to identify local as well as global trends in Ames mutagenicity data, and, to reveal unique characteristics of compound classes such as steroids is depicted. Finally, we show that the potential of the developed methods extends beyond drug discovery by using local gradients to enhance the assessment of reaction rates in transition state theory. While computational chemistry remains a challenging field of application for machine learning, the present work introduces methods to improve and assess the quality of machine learning predictions in order to increase the usage of these methods in chemical research.