Receptor-Specific Diffusion Model: Towards Generating Protein-Protein Structures with Customized Perturbing and Sampling

Xin Juan, Kaixiong Zhou, Tianlong Chen, Ninghao Liu, Xin He, Rui Miao, Xin Wang

Keywords: protein structure prediction, diffusion model, graph neural network

TL;DR: We propose a novel receptor-specific diffusion model towards generating protein-protein structures with customized sampling

Abstract: Recent advancements in deep generative models have significantly facilitated protein-ligand structure design, which is crucial in protein engineering. However, recent generative approaches based on diffusion models in this field usually start sampling from a unified distribution, failing to capture the intricate biochemical differences between receptors. This may limits their capacity to generate reliable ligands for the corresponding receptors. Moreover, the current sampling process incurs a heavy computational burden and inefficiency, which further escalates the training demands on the model. To this end, we introduce a novel diffusion model with customized perturbing and sampling for the ligand design targeting the specific receptor, named as Receptor-Specific Diffusion Model (RSDM). In particular, the receptor-specific information is used to tailor fine-grained sampling distributions via changing the noise for customized perturbing. Meantime, we refine the sampling process using a predefined schedule to perform stepwise denoising and gradually decrease the influence of the receptor's guidence in the ligand generation for customized sampling. The experimental reaults indicate that RSDM is highly competitive with state-of-the-art learning-based models, including recent models like ElliDock and DiffDock-PP. Additionally, RSDM stands out for its faster inference speed compared with all baseline methods, highlighting its potential for generating dependable protein-ligand.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 7388