Molecular Design Method Using a Reversible Tree Representation of Chemical Compounds and Deep Reinforcement Learning

Area of Science:

• Computational chemistry
• Material informatics
• Drug discovery

Background:

• Automated molecular design is crucial for material informatics and drug discovery.
• Existing methods may lack efficient representations for complex molecular structures.

Purpose of the Study:

• To introduce a novel molecular representation, the Reversible Junction Tree (RJT).
• To formulate molecular design as a tree-structure construction problem using deep reinforcement learning (RJT-RL).
• To demonstrate the applicability of RJT-RL for molecular optimization and fine-tuning in drug discovery.

Main Methods:

• Developed a coarse-grained tree representation of molecules (Reversible Junction Tree; RJT).
• Formulated molecular design and optimization as a tree-structure construction problem using deep reinforcement learning (RJT-RL).
• Ensured all intermediate and final states in RL are convertible to valid molecules.

Main Results:

• The RJT representation is reversely convertible to the original molecule without external information.
• RJT-RL efficiently guided the optimization process in simple benchmark tasks.
• Demonstrated applicability to multiobjective optimization and fine-tuning for realistic drug discovery scenarios.

Conclusions:

• The novel RJT representation and RJT-RL method offer an efficient approach for automated molecular design.
• This method facilitates the optimization and fine-tuning of molecular properties for drug discovery.
• RJT-RL shows promise for advancing computational drug discovery and material informatics.