DiffBindFR: an SE(3) equivariant network for flexible protein-ligand docking.

Area of Science:

• Computational chemistry
• Structural biology
• Drug discovery

Background:

• Molecular docking is crucial for structure-based drug design, aiding in modeling protein-ligand interactions.
• Conventional docking methods struggle with flexible protein binding pockets and accurate prediction in real-world scenarios.
• Existing deep learning methods often neglect protein side chain flexibility and physical plausibility.

Purpose of the Study:

• To develop a novel, full-atom diffusion-based flexible docking model named DiffBindFR.
• To enhance the accuracy of predicting protein-ligand binding modes, considering both ligand and protein flexibility.
• To improve the applicability of docking for drug design using apo and AlphaFold2-predicted protein structures.

Main Methods:

• Developed DiffBindFR, a diffusion-based model operating on ligand movements and protein side chain torsions.
• Evaluated DiffBindFR's performance against existing docking methods for binding pose prediction.
• Tested the model's efficacy on both apo (ligand-free) and AlphaFold2-generated protein structures.

Main Results:

• DiffBindFR achieved higher accuracy in generating native-like protein-ligand binding structures compared to current methods.
• The model produced physically plausible and detailed atomic interactions.
• Superior performance was observed in predicting ligand binding poses and protein conformations for apo and AlphaFold2 structures.

Conclusions:

• DiffBindFR offers a powerful and accurate tool for flexible molecular docking.
• The model significantly improves structure-based drug design, particularly for challenging targets with flexible binding sites.
• DiffBindFR is well-suited for drug design applications utilizing apo and AlphaFold2-derived protein models.