Scaffold-based molecular design with a graph generative model.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Drug discovery often relies on modifying known molecular structures.
• Designing new molecules around a specific core structure (scaffold) is crucial.
• Existing methods require strategies for scaffold-based molecular design.

Purpose of the Study:

• To propose a graph generative model for scaffold-based molecular design.
• To ensure generated molecules contain a specific molecular scaffold.
• To enable conditional generation of molecules with desired properties.

Main Methods:

• Developed a graph generative model that takes a molecular scaffold as input.
• The model sequentially adds atoms and bonds to extend the scaffold.
• Incorporated conditional generation to control molecular properties.

Main Results:

• Generated molecules are guaranteed to contain the input scaffold.
• The model demonstrates generalizability to unseen scaffolds.
• Simultaneous control over multiple chemical properties was achieved.
• Successfully applied to design epidermal growth factor receptor inhibitors.

Conclusions:

• The proposed model effectively supports scaffold-based molecular design.
• The model's ability to control properties and generalize is valuable for drug discovery.
• A semi-supervised extension enhances applicability with limited data.