VFDiff: SE(3)-Equivariant Vector Field Guided Diffusion Model for Target-Aware Molecule Generation in 3D

ICLR 2025 Conference Submission1453 Authors

Keywords: Diffusion Model, Molecule Generation, Structure-Based Drug Design

TL;DR: We proposed a molecular generation diffusion model guided by docking energy, which surpassed previous methods in important indicators such as docking score, high affinity ,and QED, becoming a new SOTA model.

Abstract: Structure-based drug design (SBDD) is a key challenge in drug discovery that aims to generate small molecules capable of binding tightly to specific protein pockets. However, current diffusion models have focused on the complementarity of ligand molecules and protein pockets in physical space while ignoring the docking energy requirements, resulting in only generating suboptimal docking postures. In this paper, we present VFDiff, a novel SE(3)-equivariant diffusion model for 3D molecular generation, guided by vector fields derived from protein-ligand binding energy. In contrast to current diffusion models, VFDiff incorporates energy-based guidance in both forward and reverse processes to ensure ligand molecules are spatially complementary and energetically matched to their target pockets. Our approach includes three fundamental mechanisms: energy-planning, which adjusts diffusion trajectories based on energy gradients; force-guiding, which refines molecular generation; and position-tuning, which improves sampling accuracy. Extensive experiments on the CrossDocked2020 dataset demonstrate that VFDiff outperforms state-of-the-art methods, achieving superior binding binding affinity with an impressive Avg. Vina Score of up to -7.37, while maintaining competitive molecular properties, and diversity. This work introduces a new framework for generating target-specific molecules with improved structural and functional fidelity, offering a significant advancement in SBDD.

Primary Area: generative models

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.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

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: 1453