Y.X. Wang, Fanhao Wang, Laiyi Feng, Changsheng Zhang, Luhua Lai
Accurate modeling of protein-peptide interactions is critical for elucidating peptide-mediated biological processes and advancing drug discovery. While traditional methods and recent deep learning-based approaches have shown promise, they often face limitations in accuracy, generalizability, computational efficiency, and the integration of prior binding knowledge. Here, we present DiffPepDock, an efficient protein-peptide docking tool based on SE(3)-equivariant diffusion models. DiffPepDock is pretrained on a carefully curated synthetic dataset of protein-fragment complexes and subsequently fine-tuned on high-quality experimental protein-peptide structures, combining generalization capability with task-specific accuracy. It also supports incorporation of user-specified binding priors including known binding motifs or reference ligands to facilitate pocket selection and enhance docking accuracy. Benchmarking on a non-redundant time-split test set demonstrates that DiffPepDock achieves accuracy comparable to state-of-the-art methods such as AlphaFold3, while substantially reducing inference time. Case studies further underscore its capability in accurately reconstructing native binding structures, particularly in scenarios where AlphaFold3 exhibits limitations. Moreover, DiffPepDock shows competitive in silico screening performance for identifying true peptide binders on AlphaFold-predicted targets, underscoring its practical utility in real-world applications. We anticipate that DiffPepDock offers a practical and reliable tool for protein-peptide docking, complementing existing biomolecular structure prediction methods and contributing to peptide therapeutic modeling efforts. The DiffPepDock tool is publicly available at https://github.com/YuzheWangPKU/DiffPepBuilder, with an interactive demonstration provided via Google Colab at https://colab.research.google.com/github/YuzheWangPKU/DiffPepBuilder/blob/main/examples/DiffPepDock_demo.ipynb.