A Hierarchical Graph Neural Network Framework for Predicting Protein-Protein Interaction Modulators With Functional...

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Predicting small molecule modulators of protein-protein interactions (PPIMs) is crucial but challenging.
• Current machine learning models require extensive manual feature engineering.
• Deep learning, particularly graph neural networks, shows promise but often overlooks molecular hierarchy and domain knowledge.

Purpose of the Study:

• To develop a novel hierarchical graph neural network framework (HiGPPIM) for improved PPIM prediction.
• To integrate atom-level and functional group-level molecular features guided by chemical knowledge.
• To enhance molecular representation learning for PPIM identification and potency prediction.

Main Methods:

• Constructed atom-level and functional group-level graphs based on chemical principles.
• Employed graph attention networks for learning representations from these graphs.
• Utilized a hypergraph attention network to aggregate and transform the two-level graph information.

Main Results:

• HiGPPIM achieved state-of-the-art performance on PPIM identification and potency prediction tasks.
• Evaluated on eight protein-protein interaction (PPI) families, demonstrating robust predictive capabilities.
• Confirmed the effectiveness of incorporating functional group information for guiding PPIM prediction.

Conclusions:

• The proposed HiGPPIM framework effectively leverages hierarchical molecular structures for PPIM prediction.
• Integrating functional group information significantly enhances the accuracy of PPIM identification and potency prediction.
• HiGPPIM offers a powerful, knowledge-guided deep learning approach for accelerating drug discovery efforts.