Tightly coupled equivariant flow matching for molecular docking with multimodal physical constraints

Area of Science:

• Computational chemistry
• Drug discovery
• Artificial intelligence in pharmacology

Background:

• Deep learning molecular docking methods often fail to produce physically valid protein-ligand conformations.
• Post-processing with force fields improves plausibility but introduces computational costs and weakens interactions.
• Existing diffusion and internal coordinate-based methods face challenges with inference efficiency and per-step computational cost.

Purpose of the Study:

• To develop a novel molecular docking method that generates physically realistic and accurate conformations efficiently.
• To address the limitations of existing deep learning docking techniques, including physical validity and computational cost.
• To propose MPFDock, a tightly coupled equivariant flow matching approach guided by multimodal physical constraints.

Main Methods:

• MPFDock utilizes a tightly coupled equivariant flow matching framework in Cartesian space.
• The method incorporates geometry-grounded ligand-protein physical interactions during the training phase.
• Force field-guided flow matching is employed during inference to enhance physical realism and accuracy.

Main Results:

• MPFDock generates physically realistic and high-accuracy docking conformations, validated by PoseBusters metrics.
• The method achieves superior performance compared to existing approaches on the PoseBusters and DeepDockingDare benchmarks.
• MPFDock significantly reduces the number of inference steps required while maintaining high accuracy and physical realism.

Conclusions:

• MPFDock offers a significant advancement in molecular docking for drug discovery by enhancing both accuracy and physical validity.
• The proposed method overcomes the trade-offs between accuracy, physical realism, and computational efficiency seen in previous techniques.
• MPFDock represents a promising, tightly coupled approach for generating reliable molecular docking poses.