Unified and explainable molecular representation learning for imperfectly annotated data from the hypergraph view.

Area of Science:

• Computational Chemistry
• Drug Discovery
• Machine Learning

Background:

• Molecular representation learning (MRL) shows potential in drug development.
• Dataset imperfections and lack of explainability hinder MRL model design.

Purpose of the Study:

• To develop a unified and explainable multi-task MRL framework.
• To address challenges in MRL for drug discovery.

Main Methods:

• Formulated molecules and properties as a hypergraph, capturing property-property, molecule-property, and molecule-molecule relationships.
• Developed OmniMol, integrating a meta-information encoder and task-routed mixture of experts (t-MoE).
• Implemented an SE(3)-encoder with equilibrium conformation supervision for physical symmetry and conformational relaxation.

Main Results:

• Achieved state-of-the-art performance in chemical property prediction.
• Demonstrated top performance in chirality-aware tasks.
• Provided explainability for all three key molecular relationships.

Conclusions:

• OmniMol offers a unified, explainable, and high-performing framework for molecular representation learning.
• The framework effectively captures correlations among properties and physical principles among molecules.
• OmniMol shows promise for practical applications in accelerating drug development.