Comprehensive Drug-Likeness Prediction Using a Pretrained Transformer Model and Multitask Learning

Drug-likeness is essential in drug discovery, indicating the potential of a compound to become a successful therapeutic. However, existing rule-based and machine learning methods are limited by their reliance on hand-crafted features, poor generalizability across chemical spaces, and insufficient adaptability to the diverse contexts of drug development. To overcome these limitations, we introduce an innovative framework that integrates molecular pretrained transformer models with multitask learning. This approach enables the simultaneous capture of complex chemical features and facilitates knowledge sharing across related prediction tasks. Our framework features two models: SpecDL, tailored for specialized drug-likeness assessments, and GeneralDL, designed for comprehensive, cross-data set evaluation. SpecDL achieved an average ROC-AUC of 0.836 across four tasks, while GeneralDL reached an average ROC-AUC of 0.781 on six internal and external test sets, both surpassing the leading existing methods. Furthermore, GeneralDL demonstrated robust generalization to toxicity and biological activity predictions and provided interpretable outputs via attention weight analysis. These results establish our framework as a powerful, generalizable tool for drug-likeness prediction with significant potential to enhance early-stage drug discovery.