A hierarchical interaction message net for accurate molecular property prediction

Area of Science:

• Computational chemistry
• Drug discovery informatics
• Machine learning in pharmacology

Background:

• Predicting molecular properties like ADMET is crucial for efficient drug discovery.
• Current deep learning models struggle with effective multi-level feature interaction.
• Limitations exist in balancing global and local chemical information for property prediction.

Purpose of the Study:

• To introduce the Hierarchical Interaction Message Net (HimNet) for improved molecular property prediction.
• To enable interaction-aware representation learning across atomic, motif, and molecular levels.
• To enhance feature extraction for predicting drug activity and ADMET profiles.

Main Methods:

• Developed a Hierarchical Interaction Message Passing Mechanism as the core of HimNet.
• Utilized hierarchical attention-guided message passing for feature integration.
• Evaluated HimNet on eleven diverse datasets, including MoleculeNet benchmarks and specialized ADMET datasets.

Main Results:

• HimNet demonstrated superior or near-superior performance across most molecular property prediction tasks.
• The model effectively balanced global and local information for feature extraction.
• Achieved high accuracy in predicting metabolic stability, malaria activity, and liver microsomal clearance.

Conclusions:

• HimNet provides an accurate and efficient solution for predicting molecular activity and ADMET properties.
• The model facilitates advanced decision-making in early-stage drug discovery.
• Hierarchical interaction learning significantly advances deep learning applications in pharmacology.