De Novo Drug Design Using Reinforcement Learning with Graph-Based Deep Generative Models

Area of Science:

• Computational chemistry
• Artificial intelligence in drug discovery

Background:

• Deep generative models are powerful tools for chemical space exploration.
• Fine-tuning these models for specific molecular design tasks remains challenging.

Purpose of the Study:

• To develop a novel reinforcement learning scheme for fine-tuning graph-based deep generative models.
• To guide generative models toward producing molecules with targeted property profiles, including size, drug-likeness, and bioactivity.

Main Methods:

• A reinforcement learning framework was proposed to fine-tune pretrained graph-based deep generative models.
• The method was applied to tasks involving generating molecules with specific property profiles (size, drug-likeness, bioactivity).

Main Results:

• The computational framework successfully guided generative models to produce molecules with desired properties, even those absent in the training data.
• Generated molecules exhibited diverse structures and high predicted DRD2 activity (95% of sampled molecules).
• The proposed approach demonstrated superior performance compared to previously reported methods.

Conclusions:

• The developed reinforcement learning scheme effectively enhances deep generative models for de novo molecular design.
• This method enables the generation of novel molecules with specific, desirable properties, advancing drug discovery efforts.