You are here:
QSAR Evaluation and Development for the Prediction of Acute Responses in Fish to Exposure to Pesticides and their Degradates
Citation:
Vitense, K., R. Kolancyzk, W. Eckel, G. Elonen, T. Dawson, AND D. Hoff. QSAR Evaluation and Development for the Prediction of Acute Responses in Fish to Exposure to Pesticides and their Degradates. SETAC, Louisville, KY, November 12 - 16, 2023. https://doi.org/10.23645/epacomptox.24794583
Impact/Purpose:
N/A
Description:
When little or no observed toxicity data are available for ecotoxicological hazard assessments of pesticides or their degradates, predictive toxicity tools such as Quantitative Structure Activity Relationship (QSAR) models may be used to assess the potential acute toxicity of chemicals in fish. While existing models for several chemical classes are available, the domain of data used to develop these tools often do not represent the broad range of pesticidal structures and modes-of-action that contribute specifically to pesticide toxicity. We evaluated the predictive accuracy of acute (LC50s) responses in fish for three existing models (ECOSAR, TEST, FishTox) specifically for pesticide parent structures and their degradates. Through the course of evaluation, we developed a fourth model using a random forest (RF) machine learning algorithm using strictly pesticide data for 266 chemicals and 17 fish species. Candidate predictors for the RF model included a targeted set of physicochemical properties, three levels of the ClassyFire structure-based taxonomic classification system (superclass, class, subclass), pesticide target species, and test fish species and exposure type. A minimal set of predictors needed to maintain top RF prediction accuracy was: solubility, ClassyFire subclass, n-octanol/water distribution coefficient (LogD), target pesticide species, and test fish species. The RF model outperformed the other models across predictive metrics (mean absolute error [MAE], mean squared error, bias), whether summarized across individual chemical-species observations (MAE=4.3-fold cross validation [CV] error), or averaged across fish species response for each chemical (MAE=3.9-fold CV error). FishTox had the second-best performance but exhibited substantial bias (predicted > observed LC50 on average). ECOSAR outperformed TEST when assessed on the full dataset, but TEST had the best performance across models when the data were restricted to TEST’s target species, fathead minnow. We further evaluated the top performing RF model on independent sets of 17 parent chemicals (MAE=6.4-fold error for mean response) and 32 degradates (5.0-fold error for mean response). The RF model likely had top predictive performance because: 1) the model was trained using only pesticide data, 2) the algorithm accounts for predictor interactions and non-linear relationships between predictors and toxicity, and 3) the model produces species-specific predictions.
URLs/Downloads:
DOI: QSAR Evaluation and Development for the Prediction of Acute Responses in Fish to Exposure to Pesticides and their DegradatesPOSTER.PDF (PDF, NA pp, 1217.286 KB, about PDF)