The conformational landscape of proteins is crucial to understanding their<br>functionality in complex biological processes. Traditional physics-based<br>computational methods, such as molecular dynamics (MD) simulations, suffer from<br>rare event sampling and long equilibration time problems, hindering their<br>applications in general protein systems. Recently, deep generative modeling<br>techniques, especially diffusion models, have been employed to generate novel<br>protein conformations. However, existing score-based diffusion methods cannot<br>properly incorporate important physical prior knowledge to guide the generation<br>process, causing large deviations in the sampled protein conformations from the<br>equilibrium distribution. In this paper, to overcome these limitations, we<br>propose a force-guided SE(3) diffusion model, ConfDiff, for protein<br>conformation generation. By incorporating a force-guided network with a mixture<br>of data-based score models, ConfDiff can can generate protein conformations<br>with rich diversity while preserving high fidelity. Experiments on a variety of<br>protein conformation prediction tasks, including 12 fast-folding proteins and<br>the Bovine Pancreatic Trypsin Inhibitor (BPTI), demonstrate that our method<br>surpasses the state-of-the-art method.