Geometric Deep Learning for Protein-Ligand Affinity Prediction with Hybrid Message Passing Strategies.

Area of Science:

• Computational chemistry
• Structural biology
• Drug discovery

Background:

• Accurate protein-ligand affinity (PLA) prediction is crucial for accelerating drug discovery.
• Current deep learning methods often rely on 1D or 2D representations, neglecting critical 3D geometric information.
• 3D spatial features are hypothesized to play a significant role in molecular binding interactions.

Purpose of the Study:

• To develop a novel geometric deep learning approach, HybridGeo, for improved PLA prediction.
• To leverage 3D structural information and hybrid message-passing strategies for enhanced binding affinity modeling.
• To validate the generalizability and robustness of the proposed model on diverse datasets.

Main Methods:

• Developed HybridGeo, a geometric deep learning model utilizing dual-view graph learning for intra- and inter-molecular atomic interactions.
• Employed hybrid strategies to aggregate spatial information and a geometric graph transformer for residue-scale protein pocket analysis.
• Trained and evaluated the model on the PDBbind dataset and three external test sets.

Main Results:

• HybridGeo achieved state-of-the-art performance with a Root Mean Square Error (RMSE) of 1.172 on the PDBbind dataset.
• The model demonstrated superior performance across three external test sets, indicating strong generalizability and robustness.
• Ablation studies confirmed the effectiveness of individual modules, and case studies highlighted performance on macrocyclic compound complexes.

Conclusions:

• HybridGeo effectively integrates 3D geometric features for accurate protein-ligand affinity prediction.
• The model's superior performance and generalizability offer a promising advancement for computational drug discovery.
• The biological interpretability of predictions suggests potential for guiding rational drug design.