Variable-Length Generative Protein Design via Generalized Poisson Flow
Published in arXiv preprint, 2026
Recommended citation: Chaoran Cheng, Zhanghan Ni, Yanru Qu, Yuxin Chen, Ruihan Guo, Jiajun Fan, Ge Liu. "Variable-Length Generative Protein Design via Generalized Poisson Flow." arXiv preprint arXiv:2607.09039, 2026. https://arxiv.org/abs/2607.09039
GPFlow generates proteins without fixing their length in advance: it learns the rate function of an inhomogeneous generalized Poisson process by maximum likelihood, with guarantees for recovering the joint distribution and a bound on the KL divergence to the data. It is evaluated on structure and sequence design, motif scaffolding and peptide co-design.