DeepBindGCN: Integrating Molecular Vector Representation with Graph Convolutional Neural Networks for Protein-Ligand...

Area of Science:

• Computational chemistry
• Drug discovery
• Bioinformatics

Background:

• Large-scale drug virtual screening requires accurate identification of high-affinity binders from vast molecular libraries.
• Protein pocket, ligand spatial information, and residue/atom types are critical factors influencing binding affinity.

Purpose of the Study:

• To develop a novel computational model for accurate and efficient prediction of protein-ligand binding affinity.
• To create a screening pipeline integrating the new model for identifying potent drug candidates.

Main Methods:

• Utilized pocket residues or ligand atoms as nodes, constructing edges based on neighboring information to represent molecular structures.
• Employed a graph convolutional network (GCN) model, DeepBindGCN, incorporating pre-trained molecular vectors for enhanced representation.
• Developed a screening pipeline integrating DeepBindGCN with other methods, validated using TIPE3 and PD-L1 dimer examples.

Main Results:

• DeepBindGCN achieved a root mean square error (RMSE) of 1.4190 and a Pearson r value of 0.7584 on the PDBbind v.2016 core set.
• Demonstrated comparable prediction power to state-of-the-art models that rely on 3D complex structures.
• The model is independent of docking conformation, preserving spatial and physicochemical features.

Conclusions:

• DeepBindGCN offers a powerful, non-complex-dependent tool for predicting protein-ligand interactions.
• The model significantly enhances the efficiency and accuracy of large-scale virtual drug screening.
• This approach holds promise for various applications in drug discovery and development.