A new high-accuracy QSAR model based on ultra-curated data for predicting thyroid receptor-related endocrine activity

Endocrine active substances interfere with hormone signalling and represent a major regulatory concern. The thyroid system is particularly vulnerable, with disruption linked to neurodevelopmental, metabolic, and cardiovascular effects. Current in vitro assays may improve mechanistic understanding but are resource-intensive while screening databases (e.g., Tox21 assays) is insufficient for large-scale chemical coverage or for assessing new materials. In silico tools such as QSAR models offer scalable alternatives, yet existing thyroid receptor (TR) models suffer from poor data curation, class imbalance, and limited regulatory acceptance.

This study aimed to develop a high-accuracy QSAR model, aligned with the OECD QSAR Assessment Framework, to accurately predict TR-related endocrine activity, thereby supporting the reduction of animal testing and regulatory decision-making. In this work, data from Tox21 TR luciferase agonist and antagonist assays were combined with curated BindingDB data, followed by expert review of activity-cytotoxicity relationships and strict filtering of low-affinity interactions (termed here as "ultra-curation"). Agonists and antagonists were grouped together as "actives". The final dataset comprised 291 active and 6,049 inactive compounds. Each substance was associated with a standardised SMILES, to locate duplicates (subsequently removed), and the dataset split into training, internal validation, and external validation sets with chemical diversity balancing. A Support Vector Machine (SVM) classifier using circular fingerprints (FCFP6) was optimised via grid search. Model performance was evaluated with cross-validation and external validation, complemented by an applicability domain (AD) framework based on analogues.

The SVM achieved strong performance in training (balanced accuracy = 96.03%; AUC-ROC = 99.91%; AUC-PR = 99.21%) and robust performance on the external validation set (balanced accuracy = 92.46%; sensitivity = 87.65%; specificity = 99.42%). Incorporation of the AD framework improved minority class (active) detection, raising balanced accuracy to 96.15%, sensitivity to 92.30%, and specificity to 100%, while excluding ~14% of compounds with uncertain predictions.

The QSAR model developed here demonstrates good predictive accuracy for thyroid receptor-related endocrine activity and addresses key regulatory concerns through rigorous data ultra-curation and applicability domain assessment. This high-accuracy QSAR provides a promising in silico tool to support integrated approaches to testing and assessment (IATA) and regulatory evaluation of endocrine disruption potential.