IgFlow-LM: De Novo Antibody Design via Joint Flow Matching on SE(3) and Protein Language Models Probability Flows

Zhenjie Luo, Zheyu Zhou, Yuanhao Li, Aoyun Geng, Yajie Meng, Feifei Cui, Shuting Jin, Chao Zha, Xin Gao, Zilong Zhang

Keywords: antibody, artificial intelligent, de novo design, protein language model

Abstract: In this work, we present IgFlow-LM, a multi-modal deep generative model for de novo antibody design based on a flow-matching framework that integrates protein language models (PLMs). By learning the joint distribution over SE(3)-equivariant structural flows and PLM-derived probabilistic flows, IgFlow-LM enables the coordinated generation of antibody 3D structures and latent embeddings in the PLM space. Experimental results demonstrate that, in unconditional design, IgFlow-LM generates antibody structures that closely resemble naturally occurring antibodies. IgFlow-LM generates antibodies closely resembling naturally observed ones, with backbone dihedral angles exhibiting strong agreement with reference antibody distributions and overall backbone conformations adhering more closely to physical constraints. Furthermore, we benchmark IgFlow-LM against baseline models on two commonly studied conditional CDR design tasks. IgFlow-LM demonstrates superior overall performance compared to baselines and generates CDR sequences with higher diversity.

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

Submission Number: 15513