Improving the Accuracy and Interpretability of Machine Learning Models for Toxicity Prediction

Humans are exposed to a multitude of chemicals (e.g. pharmaceuticals and cosmetics) and the safety of these needs to be demonstrated. Quantitative structure-activity relationship (QSAR) models provide an alternative to undesired animal studies for this purpose. However, in practice their use is often limited either due to insufficient model accuracy or due to a lack of model interpretability. This thesis addresses current limitations of QSAR models used for toxicity prediction. Firstly, it was investigated whether multi-task and imputation modelling yield more accurate models compared to standard single task QSAR models. Secondly, attempts were made to improve the interpretability of neural networks used for QSAR modelling. In particular, a method was developed to extract information about chemical features learned in the hidden layers of neural networks. While no significant differences in performance were found between single task models and traditional multi-task models (using only chemical descriptors for test compounds), multi-task imputation models (using experimental data labels of related assays for test compounds) were found to clearly outperform single task models on in vitro toxicity datasets. Imputation is therefore a promising tool to improve the performance of QSAR models for toxicity prediction. The novel method developed to interpret neural network models, called IG_hidden, makes use of integrated gradients to identify neurons relevant for individual predictions. Then, substructures found to be relevant for activation of these neurons are used to visualise which atoms of a compound are responsible for the model predictions. IG_hidden was compared to an established method for interpreting neural networks (i.e. applying integrated gradients to input features) using Lhasa’s Derek alerts for mutagenicity as a ground truth. The overall performance of IG_hidden was found to be comparable to the published method in terms of the quality of the model explanations that were found. However, the approaches were complementary with each method performing better on certain subsets of the dataset