Multi-Property De Novo Drug Design Using Deep Learning-Based Knowledge Distillation and Reinforcement Learning

Computer-aided de novo drug design has been widely explored for early-stage drug discovery, yet the multi-property optimization of novel molecules remains challenging. We aimed to develop a de novo drug design model to efficiently optimize multiple properties simultaneously. We developed a teacher-student-interaction deep learning model fine-tuned by reinforcement learning (TSItransRL) using bioactivity datasets (DRD2 and JNK3/GSK3β targets). A conditional transformer was pretrained as the teacher model to incorporate multi-property information. A vanilla transformer served as the student model and was subsequently optimized through interactive knowledge distillation and reinforcement learning. An evaluation was conducted using MOSES and conditional metrics on two tasks, specifically generating molecules with DRD2-targeting activity and generating molecules with dual JNK3/GSK3β-targeting activity, with the analyses including docking, the similarity ensemble approach (SEA), and scaffold novelty. TSItransRL achieved success rates of 98.36% and 98.90% for the DRD2 and JNK3/GSK3β tasks, respectively, with an internal diversity of 0.795, outperforming most baselines. The docking, SEA, scaffold, and ADMET analyses were used as exploratory in silico assessments to support the preliminary prioritization of selected generated molecules. TSItransRL provides an in silico framework for benchmark-level multi-property molecular generation and prioritization, combining interactive knowledge distillation with reinforcement learning to explore molecules that satisfy predefined predicted-activity, drug-likeness, and synthetic-accessibility criteria. The generated molecules should be regarded as computational candidates for a further medicinal-chemistry assessment, independent validation, and experimental testing rather than experimentally validated leads.