Knowledge-Embedded Message-Passing Neural Networks: Improving Molecular Property Prediction with Human Knowledge.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Materials science

Background:

• Graph neural networks (GNNs) offer end-to-end molecular property prediction but require extensive data.
• High prediction accuracy often necessitates costly experiments for large datasets.
• Traditional descriptor-based quantitative structure-property relationships (QSPR) rely on manually designed features.

Purpose of the Study:

• To develop a novel GNN architecture, Knowledge-Embedded Message Passing Neural Network (KEMPNN), that integrates non-quantitative expert knowledge.
• To enhance molecular property prediction accuracy, especially in low-data regimes.
• To reduce the reliance on large experimental datasets for model training.

Main Methods:

• Extended a message-passing neural network (MPNN) to create KEMPNN.
• Incorporated non-quantitative knowledge annotations (e.g., substructure effects) into the GNN architecture.
• Evaluated KEMPNN on physical chemistry (ESOL, FreeSolv, Lipophilicity) and polymer (glass-transition temperature) datasets with limited training data.

Main Results:

• KEMPNN demonstrated improved prediction accuracy compared to standard MPNN when supervised with knowledge annotations.
• KEMPNN achieved accuracy comparable to or better than descriptor-based methods, even with small training datasets.
• The integration of expert knowledge effectively guided the model's learning process.

Conclusions:

• KEMPNN offers a powerful approach for accurate molecular property prediction with reduced data requirements.
• Integrating human expertise into GNNs is a viable strategy to overcome data scarcity in cheminformatics.
• This method holds promise for accelerating materials discovery and drug development by enabling reliable predictions from limited experimental data.