Deep Scoring Neural Network Replacing the Scoring Function Components to Improve the Performance of Structure-Based...

Area of Science:

• Computational chemistry
• Drug discovery
• Bioinformatics

Background:

• Accurate prediction of protein-ligand binding affinities is crucial for efficient drug development.
• Existing deep learning methods often overlook spatial relationships and interaction pairs between proteins and ligands.
• There is a need for advanced computational methods to improve virtual screening and drug candidate identification.

Purpose of the Study:

• To develop a deep learning-based virtual screening method, Deep Scoring, that incorporates relative spatial and atomic information.
• To enhance the prediction accuracy of protein-ligand binding affinities.
• To improve the identification of novel drug candidates.

Main Methods:

• Deep Scoring extracts relative positional and atomic attribute information from protein-ligand docking poses.
• Two ResNets are employed to extract features from ligand atoms and protein residues, creating an atom-residue interaction matrix.
• A dual attention network (DAN) is utilized to weigh the contributions of atoms and residues for binding affinity prediction.

Main Results:

• Deep Scoring demonstrated superior screening performance with an AUC of 0.901 on the DUD-E dataset.
• The method achieved high accuracy in pose prediction (AUC of 0.935 on PDBbind) and generalization (AUC of 0.803 on CHEMBL).
• Two novel compounds with potential ERK2 inhibitory activity were identified using Deep Scoring.

Conclusions:

• Deep Scoring offers a significant advancement over existing structure-based deep learning methods for protein-ligand interaction prediction.
• The method's ability to integrate spatial and interaction features improves virtual screening and drug discovery pipelines.
• Deep Scoring successfully identified promising novel inhibitors, validating its potential in drug development.