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

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery
• Molecular modeling

Background:

• De novo drug design is crucial for early-stage drug discovery.
• Optimizing multiple properties of novel molecules simultaneously presents a significant challenge.

Purpose of the Study:

• To develop a de novo drug design model capable of efficiently optimizing multiple properties concurrently.
• To create an in silico framework for generating and prioritizing multi-property molecules.

Main Methods:

• Developed a teacher-student-interaction deep learning model fine-tuned by reinforcement learning (TSItransRL).
• Utilized a conditional transformer as the teacher model and a vanilla transformer as the student model.
• Employed interactive knowledge distillation and reinforcement learning for student model optimization.
• Evaluated performance using MOSES, conditional metrics, docking, Similarity Ensemble Approach (SEA), and scaffold novelty analysis.

Main Results:

• TSItransRL achieved high success rates: 98.36% for DRD2 targets and 98.90% for JNK3/GSK3β targets.
• The model demonstrated strong internal diversity (0.795) and outperformed most baseline methods.
• In silico assessments supported the preliminary prioritization of generated molecules based on predicted activity, drug-likeness, and synthetic accessibility.

Conclusions:

• TSItransRL offers a robust in silico framework for multi-property molecular generation and prioritization.
• The model effectively combines interactive knowledge distillation and reinforcement learning for efficient drug design.
• Generated molecules serve as computational candidates for subsequent medicinal chemistry assessment and experimental validation.