Development of Machine Learning and Chemical Language Model‐Based <scp>QSAR</scp> Models for Predicting Drug Residue Depletion Half‐Lives in Plasma and Tissues of Cattle Across Various Administration Routes

Accurate prediction of drug depletion half-lives plays a pivotal role in determining extralabel withdrawal intervals and ensuring the safety of food products derived from livestock. In this study, we employed machine learning (ML)-based quantitative structure-activity relationship (QSAR) models and an innovative chemical language model-based QSAR approach (ImprovedChemBERTa) to estimate plasma and tissue half-lives of drugs administered to cattle through different administration routes. Utilizing a dataset from the Food Animal Residue Avoidance Databank (FARAD) Comparative Pharmacokinetic Database, we developed one "descriptor-free" ImprovedChemBERTa model and 20 ML-QSAR models, integrating four different ML algorithms with five categories of molecular descriptors. Among ML-QSAR approaches, the deep neural network (DNN) method employing all descriptors achieved the highest predictive accuracy (test R²: 0.37). In contrast, the ImprovedChemBERTa model significantly outperformed traditional methods, reaching a test R² of 0.69, underscoring the superior capability and transfer learning potential of chemical language models. Our findings highlight the effectiveness of chemical language model-based QSAR strategies, which directly process raw chemical representations without requiring explicitly generated molecular descriptors. Overall, this work provides a robust foundation for advancing tissue-specific QSAR modeling in major food-animal species and supports global efforts toward enhanced food safety regulation.