Molecular property prediction based on graph structure learning

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in drug discovery

Background:

• Molecular property prediction (MPP) is crucial for computer-aided drug discovery.
• Graph-based models have advanced MPP, but often overlook inter-molecule relationships.
• Integrating molecular relationships can potentially enhance prediction accuracy.

Purpose of the Study:

• To propose a novel graph structure learning (GSL) based approach for MPP.
• To effectively incorporate relationships between molecules into the prediction process.
• To improve the performance of molecular property prediction models.

Main Methods:

• Utilized graph neural networks (GNNs) to extract molecular representations from molecular graphs.
• Constructed a molecule similarity graph (MSG) using molecular fingerprints.
• Applied molecule-level GSL on the MSG to fuse intra-molecule and inter-molecule information for enhanced molecular embeddings.

Main Results:

• The proposed GSL-MPP method achieved state-of-the-art performance on most benchmark datasets.
• The approach demonstrated particular effectiveness in classification tasks.
• Visualization studies confirmed the generation of high-quality molecular representations.

Conclusions:

• The GSL-MPP method successfully integrates inter-molecule relationships for improved MPP.
• The approach offers a promising direction for advancing molecular property prediction in drug discovery.
• The developed method provides superior molecular embeddings by considering both internal molecular structure and external molecular similarities.