BA-Pred and RMSD-Pred: Integrated Graph Neural Network Models for Accurate Protein-Ligand Binding Affinity and...

Area of Science:

• Computational chemistry and structural biology
• Artificial intelligence in drug discovery

Background:

• Accurate prediction of protein-ligand interactions is crucial for structure-based drug discovery.
• Existing methods often struggle to simultaneously optimize binding affinity estimation and pose evaluation.

Purpose of the Study:

• To develop an integrated deep learning framework to disentangle and improve protein-ligand binding affinity prediction and pose evaluation.
• To create complementary graph neural network models, BA-Pred for affinity and RMSD-Pred for pose assessment.

Main Methods:

• Utilized a Gated Graph Convolutional Network with Learnable Structural Positional Encoding (GatedGCN-LSPE) architecture for both models.
• Trained and validated models on established benchmarks like CASF-2016, Astex diverse set, and PoseBusters.
• Developed an integrated pipeline for virtual screening combining pose selection and affinity prediction.

Main Results:

• BA-Pred achieved state-of-the-art binding affinity prediction (1.10 pKd RMSE) on CASF-2016.
• RMSD-Pred demonstrated strong pose evaluation accuracy (96% top-1 success rate) and improved AutoDock-GPU pose selection by up to 33.1%.
• The integrated pipeline showed robust virtual screening performance, with an EF 1% of 21.1 on CASF-2016 and improved EF 1% on LIT-PCBA.

Conclusions:

• The graph neural network models offer balanced and accurate performance across diverse protein-ligand interaction prediction tasks.
• This framework shows significant potential to accelerate the drug discovery process.
• The integrated pipeline provides a promising tool for efficient virtual screening and lead optimization.