Deciphering Pesticide Toxicity: Exploring Chemical Space, Structural Features, and Predictive Modeling through QSTR, q-RATAR, and Machine Learning
Lu Li
Co-Presenters: Individual Presentation
College: The Dorothy and George Hennings College of Science, Mathematics and Technology
Major: Biology
Faculty Research Mentor: Supratik Kar
Abstract:
Pesticides play a crucial role in agriculture but pose significant environmental and ecological risks. This study presents a comprehensive predictive modeling approach for assessing pesticide toxicity in rainbow trout using Quantitative Structure-Toxicity Relationship (QSTR), Quantitative Read-Across Structure-Toxicity Relationship (q-RASTR), and machine learning techniques. A curated dataset of 299 pesticides was analyzed after excluding 12 compounds with high residual errors. Molecular descriptors were computed using AlvaDesc, and feature selection was performed using genetic algorithms. Machine learning models, including Random Forest, were optimized through hyperparameter tuning, achieving an accuracy of 84.07%. The QSTR models demonstrated strong predictive power (R² = 0.68 for training, R² = 0.72 for testing), while q-RASTR models further improved prediction reliability by incorporating read-across descriptors (R² = 0.74 for training, R² = 0.81 for testing). Scaffold diversity analysis revealed high structural variability, with 81.8% of scaffolds being unique. External validation using the Pesticide Properties Database (PPDB) and PubChem databases confirmed the models' robustness, predicting toxicity levels with high reliability. These findings highlight the effectiveness of integrated computational approaches in environmental risk assessment, providing valuable insights into pesticide toxicity prediction.