Sihoon Lee (Sogang University)

Hanbin Oh (Sogang University)

Hanbin Oh (Sogang University)

Sihoon Lee (Sogang University)

 The relationship between a chemical's molecular structure and its physical or biological properties is modeled using the in silico approach known as the Quantitative Structure-Activity Relationship (QSAR). It is a viable substitute for using animals in method to determine the toxicity of substances. Reproductive and developmental toxicity is an essential regulatory criterion for assessing human health risks in the registration process of chemicals used for various purposes. Numerous in silico models have been developed to predict reproductive and developmental toxicity, but their performance and applicability domain are limited for practical regulatory use.

 In this study, a binary classification model for predicting reproductive and developmental toxicity was developed using the Graph Convolution Network algorithm. The toxicity data was curated from open-source government databases of four countries and relevant literature in accordance with GHS classification criteria. Structure-based Developmental and Reproductive Toxicity (DART) descriptors were designed, specific to the mechanisms of reproductive and developmental toxicity, and incorporated as global nodes. For performance evaluation, the model demonstrated a highest accuracy of 68.5% and an AUC of 79%. This study represents a departure from traditional QSAR modeling, which relies on molecular descriptors, by applying the Graph Convolution Network algorithm to the endpoint of reproductive and developmental toxicity.