Gargoyles: An Open Source Graph-Based Molecular Optimization Method Based on Deep Reinforcement Learning

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Optimizing chemical compounds is crucial for drug discovery and material design.
• Existing machine learning models often generate compounds de novo, limiting optimization of user-defined molecules.
• A gap exists for methods that refine existing compounds efficiently.

Purpose of the Study:

• To develop a novel compound optimization method using deep reinforcement learning.
• To enable exploration and optimization of user-defined compounds.
• To enhance specific molecular properties like Quantum ElectroDynamics (QED).

Main Methods:

• A deep reinforcement learning framework was employed for molecular graph-based optimization.
• The method generates compounds fragment-by-fragment, adding atoms iteratively.
• Optimization was guided by Quantum ElectroDynamics (QED) as a target metric.

Main Results:

• The method successfully enhanced the Quantum ElectroDynamics (QED) of compounds.
• Optimization was demonstrated by improving the activity of a compound targeting dopamine receptor D2 (DRD2).
• Generated compounds maintained structural similarity to the starting molecules while improving activity.

Conclusions:

• The developed method is suitable for optimizing molecules from a given starting compound.
• This approach offers a powerful tool for targeted drug discovery and molecular design.
• The fragment-by-fragment, atom-by-atom generation strategy allows for high-density exploration around existing structures.