The applicability domain of EPI Suite ™ for screening phytotoxins for potential to contaminate source water for drinking

Background Toxins produced by plants constitute a potential threat to water supplies in Europe, but have not been widely considered in systematic risk assessments. One way to begin to address potential risks of phytotoxins is to conduct screening-level assessments of known phytotoxins for their potential to contaminate source water for drinking due to persistence (P) and mobility (M). Chemical properties relevant for such an assessment (octanol–water partition coefficient K OW and biodegradation half-life) can be estimated from the structure of phytotoxins with quantitative structure–property relationship (QSPR) models found in the United States Environmental Protection Agency’s Estimation Program Interface (EPI Suite ™ ) software, but predictions must be considered critically since these models have been developed using data for anthropogenic chemicals and many phytotoxins could lie outside their applicability domain. Results We analyzed two EPI Suite ™ models—KOWWIN and BIOWIN5—by evaluating the quality of property predictions for their validation sets as a function of Euclidean distances d E to the centroid of descriptor space of the models’ training sets. We identified model-specific applicability domain boundaries as local maxima in plots of the difference between root mean square error (∆RMSE) of modeled property values of validation set compounds within and outside applicability domain boundaries defined by a continuum of possible boundaries. And, we also evaluated four generic boundaries that have been suggested in literature. The ∆RMSE between validation set compounds outside and inside applicability domain boundaries had positive values for all but one of the possible boundaries we considered, indicating that properties of chemicals with d E inside the boundaries were better predicted. With our proposed model-specific boundaries, 21% of 1586 phytotoxins produced by plants found in Switzerland were out of domain of KOWWIN, and 46% of were out of domain of BIOWIN5. Conclusions Estimates of Log K OW and biodegradation half-life of phytotoxins that lie outside the domain of applicability of the QSPR models should be viewed as extrapolations that are subject to unquantified and potentially large errors. Phytotoxins outside the domain of applicability of QSPR models should be prioritized for property measurements as a basis to expand the training sets of QSPR models and to support hazard identification for better management of drinking water quality in Europe.