Date of Award

Spring 2015

Document Type


Degree Name

Doctor of Philosophy (PhD)


Mathematics, Statistics and Computer Science

First Advisor

Merrill, Stephen J.

Second Advisor

Povinelli, Richard J.

Third Advisor

Sem, Daniel S.


Computational toxicology is the development of quantitative structure activity relationship (QSAR) models that relate a quantitative measure of chemical structure to a biological effect. In silico QSAR tools are widely accepted as a faster alternative to time-consuming clinical and animal testing methods for regulatory risk assessment of xenobiotics used in consumer products. However, different QSAR tools often make contrasting predictions for a new xenobiotic and may also vary in their predictive ability for different class of xenobiotics. This makes their use challenging, especially in regulatory applications, where transparency and interpretation of predictions play a crucial role in the development of safety assessment decisions. Recent efforts in computational toxicology involve the use of in vitro data, which enables better insight into the mode of action of xenobiotics and identification of potential mechanism(s) of toxicity. To ensure that in silico models are robust and reliable before they can be used for regulatory applications, the registration, evaluation, authorization and restriction of chemicals (REACH) initiative and the organization for economic co-operation and development (OECD) have established legislative guidelines for their validation. This dissertation addresses the limitations in the use of current QSAR tools for regulatory risk assessment within REACH/OECD guidelines. The first contribution is an ensemble model that combines the predictions from four QSAR tools for improving the quality of predictions. The model presents a novel mechanism to select a desired trade-off between false positive and false negative predictions. The second contribution is the introduction of quantitative biological activity relationship (QBAR) models that use mechanistically relevant in vitro data as biological descriptors for development of computational toxicology models. Two novel applications are presented that demonstrate that QBAR models can sufficiently predict carcinogenicity when QSAR model predictions may fail. The third contribution is the development of two novel methods which explore the synergistic use of structural and biological similarity data for carcinogenicity prediction. Two applications are presented that demonstrate the feasibility of proposed methods within REACH/OECD guidelines. These contributions lay the foundation for development of novel mechanism based in silico tools for mechanistically complex toxic endpoints to successfully advance the field of computational toxicology.