P18-100 A New High-Accuracy QSAR Predicting Thyroid receptor-related Endocrine Activity

In 2018, EFSA & ECHA regulatory agencies developed a framework involving testing levels foridentification of Estrogenic, Androgenic, Thyroid and Steroidogenesis (EATS) modality perturbation using in silico, in vitro and in vivo . While validated E, A and S receptor assays are available, thyroid receptor-specific assay development is still ongoing. Current regulatory methods for the thyroid system investigation rely on thyroid hormone concentration measurements from in vivo reproductive and developmental toxicity studies, which fail to reveal specific interactions with thyroid receptors (α and β). To address these limitations, a new high-accuracy quantitative structure-activity relationship (HA-QSAR) model named iVEMPS, was developed. The tool rapidly and cost-effectively assesses thyroid receptor endocrine modality disruption by chemical exposure, thus minimizing reliance on animal testing and in vitro studies requiring biological materials. The iVEMPS model was designed to predict thyroid receptor-related endocrine activity, underpinned by an innovative data curation methodology. A curated dataset, mainly sourced from EPA ToxCast, was refined through a semi-automated curation process guided by expert knowledge, ensuring the selection of high-quality data. The data splitting was performed using the Kennard-Stone algorithm. This ensures structural diversity and a balanced distribution of positive and negative responses across the training, validation, and test sets. The machine learning algorithm, namely a support vector machine (SVM) using circular molecular fingerprints, was trained and validated. A sophisticated applicability domain (AD) methodology was applied using a k-nearest neighbours’ model, based on the identification of structural analogues and the concordance of results with these analogues. Without including AD limitations, the model already demonstrated robust performance, with high sensitivity, specificity and accuracy, i.e. 99.34, 100 and 99.97% for the training set and 83.95%, 99.60%, and 98.90% for the external test set. Statistics of the external test set increased respectively to 92.06, 99.93 and 99.62% for sensitivity, specificity and accuracy following application of the AD model. This new methodology, which meets OECD principles for QSAR validation and ECHA regulatory standards, enhances prediction reliability, filling critical gaps in thyroid receptor research. Therefore, the iVEMPS model efficiently predicts Thyroid receptor-related Endocrine activity, supports regulatory and industrial assessments, enables prioritisation of low concern chemicals for EATS modalities, and promotes faster hazard assessments while reducing animal testing.