A Unified Framework for Molecular Property Prediction based on Hierarchical Multi-Granularity Molecular...

Area of Science:

• Computational Chemistry
• Drug Discovery
• Machine Learning

Background:

• Molecular property prediction is crucial for accelerating drug discovery.
• Current methods often underutilize functional motifs and global attributes, focusing mainly on atom-bond information.
• Existing approaches lack effective mechanisms for integrating information across different structural levels, leading to incomplete chemical semantic representations.

Purpose of the Study:

• To develop a unified framework for hierarchical multi-granularity molecular representation learning (HMG-MRL).
• To systematically integrate domain knowledge from atom-bond, functional motif, and molecular attribute levels.
• To improve the accuracy and interpretability of molecular property prediction models.

Main Methods:

• Proposed HMG-MRL, a framework integrating fine-grained atom-bond, medium-grained functional motifs, and coarse-grained molecular attributes.
• Introduced an atom-bond bipartite graph approach for explicit modeling of atom and bond interactions.
• Developed a Motif Transformer for capturing global motif interactions and a cross-granularity communication module.

Main Results:

• HMG-MRL achieved competitive predictive performance across nine benchmark datasets.
• The framework effectively integrates information from multiple structural granularities.
• Case studies demonstrated the model's ability to identify key molecular components and complementary structural patterns.

Conclusions:

• HMG-MRL offers a robust approach to molecular representation learning by incorporating hierarchical information.
• The framework enhances molecular property prediction accuracy and provides insights into chemical semantics.
• This method advances drug discovery by enabling more effective identification of promising drug candidates.