MGCVAE: Multi-Objective Inverse Design via Molecular Graph Conditional Variational Autoencoder.

Area of Science:

• Computational chemistry
• Materials science
• Drug discovery

Background:

• Generating molecules with specific properties is crucial for drug development and new organic materials.
• Existing methods for de novo molecular design require improvement in efficiency and property control.

Purpose of the Study:

• To propose and evaluate a molecular graph generative model for de novo molecular design.
• To compare the performance of a molecular graph conditional variational autoencoder (MGCVAE) against a standard molecular graph variational autoencoder (MGVAE).
• To apply multi-objective optimization to MGCVAE for simultaneous property satisfaction.

Main Methods:

• Development of a molecular graph generative model based on an autoencoder architecture.
• Utilizing a molecular graph conditional variational autoencoder (MGCVAE) for targeted molecule generation.
• Implementing multi-objective optimization with calculated logP and molar refractivity as targets.

Main Results:

• MGCVAE generated 25.89% optimized molecules, a significant improvement over MGVAE's 0.66%.
• The model successfully produced drug-like molecules possessing two specific physical properties simultaneously.
• Demonstrated the efficacy of graph-based, data-driven models for designing molecules with tailored properties.

Conclusions:

• MGCVAE is a highly effective tool for the de novo design of molecules with desired properties.
• The proposed graph-based generative model offers a powerful data-driven approach for molecular discovery.
• This methodology advances the design of functional organic molecules for various applications.