Combining Group-Contribution Concept and Graph Neural Networks Toward Interpretable Molecular Property Models

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in chemistry

Background:

• Quantitative structure-property relationships (QSPRs) are crucial for accelerating compound discovery but often lack generalizability and accuracy.
• Existing machine learning and deep learning models for QSPRs frequently suffer from a lack of transparency and interpretability.
• Traditional group contribution (GC) methods provide interpretability but may have limitations in predictive power.

Purpose of the Study:

• To develop novel, interpretable graph neural network (GNN) models for predicting physicochemical properties of organic compounds.
• To integrate the fundamental concept of group contributions (GC) into GNN architectures for enhanced interpretability.
• To improve the accuracy and generalizability of QSPR models while maintaining transparency.

Main Methods:

• Development of two interpretable GNN models: attentive group-contribution (AGC) and group-contribution-based graph attention (GroupGAT).
• Integration of group contribution (GC) principles within the GNN framework.
• Utilization of attention mechanisms to highlight substructures with the highest attention weights in molecular representations.

Main Results:

• The proposed AGC and GroupGAT models demonstrated superior performance compared to classical GC models.
• The developed GNN models outperformed other existing GNN approaches for predicting aqueous solubility, melting point, and various enthalpies (formation, combustion, fusion).
• The interpretability feature successfully highlighted key molecular substructures relevant to specific properties, consistent with semiempirical GC models.

Conclusions:

• The developed interpretable GNN models offer a powerful and transparent approach for QSPR modeling.
• Integrating GC concepts into GNNs enhances model interpretability by identifying critical molecular substructures.
• These models represent a significant advancement in predicting chemical properties accurately and understandably, aiding in rational compound design.