Predicting Pesticide Toxicity via QSAR and Ensembles

Chemical numbers are rising dramatically with limited hazard assessment testing completed. Vast majorities of compounds have not undergone even basic toxicity testing despite FIFRA (Federal Insecticide, Fungicide, & Rodenticide Act), TSCA (Toxic Substance Control Act), and increased safety standards via FQPA (Food Quality Protection Act) legislation. Manufacturer testing costs routinely reach $10 to $50 million for a single compound. There is simply not enough time or money to complete test batteries for even a portion of today's registered compounds. Necessary alternatives involve developing computational hazard assessment models, however, such modeling is restricted. Computational developments require methodologies for accurate quantitative predictions, utilization of only limited sparse experimental data, and applicability to a wide variety of compound classes & biological endpoints. This research tests feasibility for novel mathematical chemistry tools to meet computational and pesticide hazard assessment challenges via: 1) Software calculators that estimate non-empirical structural descriptors (QSAR, Quantitative Structure-Activity Relationships) for pesticides directly from structure to augment sparse experimental data sets and avoid measurement costs; 2) An innovative hierarchical hQSAR/TM/ strategy to select the descriptors most relevant for hazard endpoints; 3) Model building algorithms that use sophisticated machine-learning tools based on artificial neural networks (GEFS/TM/, Genetic Ensemble Feature Selection) to accurately predict toxicity.