<p>Molecular docking is a fundamental task in drug discovery, yet many deep learning–based docking methods still struggle to generate physically valid conformations. Although post-processing with empirical force fields can partially improve physical plausibility, it breaks tightly coupled modeling by introducing a decoupled refinement stage, increases computational cost, and often weakens ligand–protein interactions and leads to pose deviations. In recent years, diffusion-based molecular docking methods have improved both accuracy and physical validity, but they typically require many sampling steps, resulting in inefficient inference. Internal coordinate–based methods can reduce the number of sampling steps, but their complex pipelines significantly increase per-step computational cost, limiting their practical applicability. To address these challenges, we propose MPFDock, a tightly coupled equivariant flow matching method for molecular docking guided by multimodal physical constraints in Cartesian space. By incorporating geometry-grounded ligand–protein physical interactions during training and employing force field–guided flow matching during inference, MPFDock generates physically realistic and high-accuracy docking conformations according to PoseBusters metrics with significantly fewer inference steps. Experiments on the PoseBusters and DeepDockingDare benchmarks show that MPFDock consistently outperforms existing methods in terms of docking accuracy and physical realism on the evaluated benchmarks.</p>