Learning Toxicity Classifications From Semantic Knowledge-Base Data

In late 2017, the Interagency Coordinating Committee on the Validation of Alternative Methods (ICCVAM) proposed a project to develop in silico models of toxicity titled "Predictive Models for Acute Oral Systemic Toxicity". The goal was to predict five specific endpoints needed by regulatory agencies:

• “Very toxic” chemicals (LD50 less than 50 mg/kg)

• “Nontoxic” chemicals (LD50 greater than or equal to 2000 mg/kg)

• Point estimates for rodent LD50s

• Categorization of toxicity hazard using the U.S. Environmental Protection Agency (EPA) codes

• Categorization of toxicity hazard using the United Nations Globally Harmonized System of Classification and Labelling (GHS) codes

In collaboration with EPA's National Center for Computational Toxicology (NCCT), they provided training and validation datasets for a sizeable body of oral acute toxicity data from rat models. We did not participate directly in this shared task, but used the published datasets for a proof of concept to determine the accuracy of a semantic knowledge-based classifier. We focused our efforts on the categorical classification tasks (all endpoints but the LD50 point estimations).