Toxicity prediction using multi-disciplinary data integration and novel computational approaches Public Deposited

Downloadable Content

Download PDF
Last Modified
  • March 22, 2019
  • Low, Yen Sia
    • Affiliation: Gillings School of Global Public Health, Department of Environmental Sciences and Engineering
  • Current predictive tools used for human health assessment of potential chemical hazards rely primarily on either chemical structural information (i.e., cheminformatics) or bioassay data (i.e., bioinformatics). Emerging data sources such as chemical libraries, high throughput assays and health databases offer new possibilities for evaluating chemical toxicity as an integrated system and overcome the limited predictivity of current fragmented efforts; yet, few studies have combined the new data streams. This dissertation tested the hypothesis that integrative computational toxicology approaches drawing upon diverse data sources would improve the prediction and interpretation of chemically induced diseases. First, chemical structures and toxicogenomics data were used to predict hepatotoxicity. Compared with conventional cheminformatics or toxicogenomics models, interpretation was enriched by the chemical and biological insights even though prediction accuracy did not improve. This motivated the second project that developed a novel integrative method, chemical-biological read-across (CBRA), that led to predictive and interpretable models amenable to visualization. CBRA was consistently among the most accurate models on four chemical-biological data sets. It highlighted chemical and biological features for interpretation and the visualizations aided transparency. Third, we developed an integrative workflow that interfaced cheminformatics prediction with pharmacoepidemiology validation using a case study of Stevens Johnson Syndrome (SJS), an adverse drug reaction (ADR) of major public health concern. Cheminformatics models first predicted potential SJS inducers and non-inducers, prioritizing them for subsequent pharmacoepidemiology evaluation, which then confirmed that predicted non-inducers were statistically associated with fewer SJS occurrences. By combining cheminformatics' ability to predict SJS as soon as drug structures are known, and pharmacoepidemiology's statistical rigor, we have provided a universal scheme for more effective study of SJS and other ADRs. Overall, this work demonstrated that integrative approaches could deliver more predictive and interpretable models. These models can then reliably prioritize high risk chemicals for further testing, allowing optimization of testing resources. A broader implication of this research is the growing role we envision for integrative methods that will take advantage of the various emerging data sources.
Date of publication
Resource type
Rights statement
  • In Copyright
  • Tropsha, Alexander
  • Doctor of Philosophy
Graduation year
  • 2013

This work has no parents.