Modelling biodegradation of pesticides using regulatory data
The persistence of a chemical, defined as its DT50 or half-life, is well known to show orders of magnitude of variability, not just in the environment as a whole, but already within the same environmental compartment. This variability is problematic in the context of risk assessment, especially in the case of frequently emitted chemicals like pesticides. Although different environmental conditions are clearly the reason such variability is observed for a given pesticide, current predictive models of pesticide degradation mostly purely rely on structural descriptors and rarely incorporate environmental conditions as predictors. This thesis therefore addresses the need for considering environmental conditions as explanatory factors in DT50 variability, using data from regulatory studies of pesticide aerobic biodegradation in soil. In a first step the data, which consist of experimental pesticide half-lives and corresponding metadata on environmental conditions, were pre-processed, explored, and characterised. Bivariate Correlation Analysis was performed for compounds and groups of compounds to explore individual environmental condition-DT50 relationships while probing the hypothesis that pesticides of similar structure and/or reactivity should show similar dependencies between DT50 and environmental condition when considered together as a structural/reactivity class. This hypothesis was only partially supported by the results. In contrast, the hypothesis that different bioavailabilities contribute to the observed half-life variability for individual pesticides was confirmed by the results. Finally, Quantitative Environmental-condition Biodegradability Relationships (QEBRs) were developed for individual compounds in the forms of models obtained using Multiple Linear Regression, Generalised Additive Modelling, and Partial Least Squares Regression. The models’ abilities to explain DT50 variance ranged from 15-81%, and showed varying degrees of precision and accuracy which were attributable to disparate levels of data quality. Soil pHCaCl2 and Temperature were found to be the strongest predictors in these models, and interpretation of the model coefficients yielded a Q10 value of 2.48. The feasibility and limitations of the QEBRs presented in this thesis are discussed.