GEMF: a novel geometry-enhanced mid-fusion network for PLA prediction

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning in bioinformatics

Background:

• Accurate prediction of protein-ligand binding affinity (PLA) is crucial for efficient drug discovery.
• Graph neural networks (GNNs) show promise for PLA prediction but often overlook critical geometric information like bond angles.
• Existing GNNs struggle to fully represent the complex binding process between proteins and ligands.

Purpose of the Study:

• To develop a novel geometry-enhanced network, GEMF, for comprehensive molecular geometry and interaction pattern learning.
• To address limitations in existing GNNs by integrating geometric properties and improving the representation of protein-ligand complexes.
• To enhance the accuracy of protein-ligand binding affinity prediction.

Main Methods:

• Proposed the Geometry-Enhanced Mid-Fusion (GEMF) network, incorporating a graph embedding layer, message passing, and multi-scale fusion.
• Represented protein-ligand complexes as graphs using physicochemical and geometric properties.
• Employed a dual-stream message passing framework to model covalent and non-covalent interactions, with an edge-update mechanism using line graphs to fuse distance and angle information.

Main Results:

• GEMF effectively learned comprehensive molecular geometry and interaction patterns.
• The dual-stream message passing and multi-scale fusion enabled detailed modeling of covalent, non-covalent, and heterogeneous interactions.
• Extensive experiments on benchmark datasets demonstrated GEMF's superior performance over state-of-the-art methods in PLA prediction.

Conclusions:

• GEMF significantly advances protein-ligand binding affinity prediction by integrating molecular geometry.
• The proposed network architecture offers a more robust representation of protein-ligand complexes and their interactions.
• GEMF provides a powerful new tool for accelerating drug discovery research.