Knowledge-aware contrastive heterogeneous molecular graph learning.

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Molecular representation learning is crucial for predicting properties and aiding drug design.
• Current methods using homogeneous graphs struggle with integrating external knowledge and multi-granularity representations.

Purpose of the Study:

• To propose a novel framework, Knowledge-aware Contrastive Heterogeneous Molecular Graph Learning (KCHML), for enhanced molecular representation.
• To overcome limitations of traditional homogeneous graph encoding in integrating external knowledge and handling multi-level molecular structures.

Main Methods:

• Developed KCHML, a framework encoding molecular graphs into heterogeneous structures.
• Utilized contrastive learning to enrich molecular representations with external knowledge.
• Conceptualized molecules using three distinct graph views (molecular, elemental, pharmacological) with heterogeneous graphs and a dual message-passing mechanism.

Main Results:

• KCHML demonstrated superior performance in molecular property prediction compared to state-of-the-art models.
• The framework effectively captures intricate molecular features.
• Showcased capability in downstream tasks like drug-drug interaction prediction.

Conclusions:

• KCHML offers a paradigm shift in molecular graph encoding, moving from homogeneous to heterogeneous structures.
• The proposed method significantly improves molecular representation learning for property prediction and drug discovery.
• KCHML's ability to integrate external knowledge and multi-granularity views enhances its applicability in complex chemoinformatics tasks.