Transformer-Based Multi-Task QSAR for Organ-Specific Toxicity and Drug Efficacy Prediction

Drug discovery is a complex and time-consuming process with a high failure rate, mainly due to unexpected toxicity and poor therapeutic effectiveness of candidate molecules. Early and reliable prediction of molecular toxicity and bioactivity is therefore essential to improve drug safety and reduce development costs. Traditional QSAR models rely on handcrafted molecular descriptors and single-task learning, which often struggle to capture complex structure–property relationships and offer limited interpretability. Recent transformer-based deep learning methods address these challenges by treating SMILES strings as a chemical language and learning rich molecular representations directly from sequence data. In this work, we present a transformer-based multi-task QSAR framework that simultaneously predicts organ-specific toxicity and drug efficacy across multiple biological targets. The model uses a ChemBERTa-style encoder with shared and task-specific components, enabling effective knowledge sharing across related endpoints. Attention-based explainability provides insight into important chemical substructures influencing predictions. Experiments on large curated datasets show improved accuracy and efficiency, demonstrating the effectiveness of the proposed approach for early-stage drug discovery.