PFM: perturbed flow matching for structure-based drug design
Related Experiment Video
Updated: Nov 11, 2026
Multi-target Parallel Processing Approach for Gene-to-structure Determination of the Influenza Polymerase PB2 Subunit
Published on: June 28, 2013
Yankai Yu1, Guikun Xu2, Zhuyang Xie3
1School of Computing and Artificial Intelligence, Southwest Jiaotong University, Chengdu, Sichuan 611756, China.
Motivation:
Target-specific 3D molecular generation is a promising approach for structure-based drug design. However, existing diffusion-based methods are generally restricted to marginal distributions with closed-form solutions, limiting the flexibility of generative process design, especially for the forward noising process. As a result, these methods may be less effective in modeling the complex distributions underlying pocket-conditioned molecular generation.
Results:
We develop Perturbed Flow Matching (PFM), a flow-matching-based framework for protein pocket-conditioned 3D molecular generation. PFM employs a perturbed conditional probability path that incorporates both pocket binding-site information and coupled atom type-coordinate information, enabling more effective generation of molecules in protein binding environments. On CrossDocked2020, PFM achieves competitive structural quality and the best average Vina score among the evaluated unguided methods (-7.12); PFM-G further improves the score to -8.70 with tailored classifier-free guidance. Experiments on additional molecular datasets further confirm the effectiveness and generalizability of the proposed framework.
Availability And Implementation:
https://github.com/kurisu92725/PFM.