Modeling to predict the aquatic toxicity of neutral compounds using toxic mode of action classification and linear free energy relationships

Di Toro, Dominic M.With the multitude of new chemicals developed each year, understanding their potentially toxic effects on aquatic life is of current concern. The goal of this work is to provide a methodology for estimating the aquatic toxicity of new compounds using modeling techniques and chemical descriptors. For this analysis, a database of 2049 chemicals with 47 associated modes of action (MoA) is compiled from literature. The specific mode of action refers to the molecular initiating event that results in a toxic effect in the organism. Narcosis, pyrethroid sodium channel modulation, and acetylcholinesterase inhibition are key groups considered. The database includes alkanes, polycyclic aromatic hydrocarbons (PAH), pesticides, inorganic, and polar compounds. MoA information from 14 sources were assigned using a variety of reliable experimental and modeling techniques. Conflicting MoA assignments from two or more sources were noted and removed from model development. The database was expanded to include CAS numbers, chemical names, SMILES strings, Log Kow, Abraham parameters, and molecular weight information. Aquatic toxicity and water solubility data was collected and curated for the compounds used in model development. ☐ Understanding the mode of action associated with a compound results in improved toxicity predictions.1-4 The proposed methodology involves the use of a feature weighted k-nearest neighbors (k-NN) classification technique to assign an MoA label to a test compound. Abraham solute parameters are employed to determine group separation as they describe the interactions between the compound and the target site in the organism. The model utilizes a binary discrimination technique with a training/test split of 80/20 to characterize compounds as positive (within the specific mode of action group) or negative (not in the group). This method ensures that if membership in every MoA is associated with a low probability, the compound is left as unassigned across all tested groups. Thorough investigation of linear and nonlinear discrimination methods resulted in a feature-weighted k-NN methodology. Optimizing the influence of each solute parameter in the model with k=3 resulted in the most successful MoA assignment results. ☐ Once grouped into specific MOA classes, linear free energy relationships (LFER) are used to quantify acute toxicity levels (e.g. LC50 values) for the test compound. The poly-parameter target site model (TSM) was developed for each specific MoA using multilinear regression analysis. With this procedure, critical target site concentrations (CTSC) are determined for each genus. The resulting CTSC were used to determine Final Acute Values that are appropriate for establishing criteria guidelines for 95% species protection.University of Delaware, Department of Civil and Environmental EngineeringPh.D