A flexible data-free framework for structure-based de novo drug design with reinforcement learning

Area of Science:

• Computational chemistry
• Drug discovery
• Molecular modeling

Background:

• Structure-based molecular generative methods are crucial for designing molecules with high binding affinity and target specificity.
• Current deep generative models face limitations due to atom-wise generation, impacting validity, synthetic accessibility, and adaptability across diverse targets.
• Existing methods often struggle with data dependency and chemical intuition, hindering broad applicability.

Purpose of the Study:

• To introduce a novel search-based framework, 3D-MCTS, for structure-based *de novo* drug design.
• To overcome the limitations of atom-centric generative models by employing a fragment-based molecular editing strategy.
• To enhance the efficiency, validity, and adaptability of drug design processes.

Main Methods:

• Developed 3D-MCTS, a framework utilizing a fragment-based molecular editing strategy with predefined retrosynthetic rules.
• Implemented multi-threaded parallel simulations and a real-time energy constraint-based pruning strategy for efficiency.
• Enabled incorporation of domain knowledge for generating molecules with desirable pharmacophores and enhanced binding affinity.

Main Results:

• 3D-MCTS outperforms state-of-the-art (SOTA) methods in producing molecules with enhanced binding affinity at a fixed computational cost.
• The fragment-based approach yields more dependable binding conformations with a 43.6% higher success rate than SOTA methods.
• Demonstrated a thirty-fold increase in high-affinity hits compared to traditional virtual screening, showcasing superior chemical space exploration.

Conclusions:

• 3D-MCTS offers a more chemically intuitive and synthetically accessible approach to *de novo* drug design compared to atom-centric methods.
• The framework's efficiency, adaptability, and ability to generate high-affinity molecules make it a powerful tool for diverse drug design scenarios, including metalloprotein applications.
• 3D-MCTS represents a significant advancement in structure-based drug design, addressing key limitations of current deep learning generative models.