Interformer: an interaction-aware model for protein-ligand docking and affinity prediction.

Area of Science:

• Computational chemistry
• Structural biology
• Drug discovery

Background:

• Deep learning models are increasingly applied to protein-ligand docking and affinity prediction for structure-based drug design.
• Current models often fail to capture intricate atom-level interactions, limiting generalization and interpretability.

Purpose of the Study:

• To propose Interformer, a unified Graph-Transformer-based model for improved protein-ligand interaction modeling.
• To enhance generalization and interpretability in deep learning for drug design.

Main Methods:

• Utilizing a Graph-Transformer architecture to capture non-covalent interactions.
• Employing an interaction-aware mixture density network for detailed interaction modeling.
• Introducing a negative sampling strategy for affinity prediction correction.

Main Results:

• Demonstrated effectiveness and universality across benchmark and in-house datasets.
• Confirmed improved performance through accurate modeling of specific protein-ligand interactions.
• Achieved state-of-the-art (SOTA) performance on docking tasks.

Conclusions:

• Interformer effectively models protein-ligand interactions, advancing structure-based drug design.
• The model offers improved generalization and interpretability compared to existing methods.
• The proposed approach sets a new SOTA in protein-ligand docking and affinity prediction.