GAABind: a geometry-aware attention-based network for accurate protein-ligand binding pose and binding affinity...

Area of Science:

• Computational Chemistry
• Drug Discovery
• Bioinformatics

Background:

• High-throughput profiling of protein-ligand interactions is crucial for drug discovery and optimization.
• Accurate prediction of binding pose and affinity remains a challenge due to computational costs and limitations in current molecular modeling techniques.

Purpose of the Study:

• To develop a geometry-aware attention-based deep learning model (GAABind) for predicting protein-ligand binding pose and affinity.
• To address limitations in molecular representation learning and intermolecular interaction modeling in existing methods.

Main Methods:

• GAABind utilizes a multi-task learning framework to predict binding pose and affinity.
• The model captures geometric and topological properties of proteins and ligands, employing expressive molecular representation learning.
• It models intermolecular many-body interactions and simulates ligand conformational adaptations using specialized networks.

Main Results:

• GAABind achieved state-of-the-art binding pose prediction (82.8% success rate) and comparable binding affinity prediction (Pearson correlation up to 0.803) on benchmark datasets (PDBbindv2020, CASF2016).
• On a SARS-CoV-2 main protease dataset, GAABind showed a 76.5% success rate in pose prediction and superior binding affinity prediction compared to baseline methods.

Conclusions:

• GAABind offers an effective deep learning approach for predicting protein-ligand interactions, enhancing computational drug discovery.
• The model's ability to integrate geometric awareness and advanced interaction modeling surpasses existing methods, paving the way for more efficient lead compound identification.