MGRNN: Structure Generation of Molecules Based on Graph Recurrent Neural Networks

Area of Science:

• Materials Science
• Computational Chemistry
• Drug Discovery

Background:

• Generating chemically valid molecular structures is a significant challenge in materials science and drug discovery.
• Deep generative models offer promising approaches for molecular structure generation.
• Existing methods often struggle with ensuring chemical validity.

Purpose of the Study:

• To propose a novel deep learning model, Molecular Graph Recurrent Neural Networks (MGRNN), for generating chemically valid molecular structures.
• To combine the strengths of iterative molecular generation algorithms with efficient deep learning training strategies.
• To develop a robust and efficient method for molecular structure generation in drug discovery.

Main Methods:

• Utilized a graph recurrent neural network architecture.
• Implemented an iterative sampling process for molecular generation.
• Integrated optional chemical domain expertise for valency checking.

Main Results:

• MGRNN demonstrated efficient computation and high model robustness.
• Generated 69% chemically valid molecules without explicit chemical knowledge.
• Achieved 100% chemically valid molecules when incorporating chemical rules.

Conclusions:

• MGRNN is an effective deep learning model for generating chemically valid molecular structures.
• The model offers computational efficiency and robustness for molecular design.
• MGRNN shows potential for accelerating drug discovery and materials science research.