Work

ProtMS: Multi-level Structural Representation Learning for Protein Inverse Folding via Selective Space Graph Networks

Year: 2026
Type: conference-paper
Abstract: Protein inverse folding, facilitated by structure-driven sequence design, has emerged as a core technology for engineering customized functional proteins such as enzymes and therapeutic agents. However, existing models suffer from two key limitations: insufficient modeling of atomic-level feature correlations and ineffective capture of long-range amino acid interactions. To overcome these challenges, we propose a novel framework capable of capturing long-range amino acid interactions across multi-level structural representations while maintaining high inference speed. Specifically, our approach enhances atomic-level feature modeling by incorporating secondary structure information and comprehensively representing covalent bond angles. Furthermore, we employ a selective spatial graph model to deeply capture long-range interactions among all residues. Extensive experiments on multiple public datasets demonstrate that our model effectively extracts both structural features and global information without compromising inference efficiency.
Language: English
Open Access status: hybrid
OpenAlex ID: W7204010024
FWCI:
Cites: 23
Cited by:
Related to: 10