Science China Information Sciences Компьютерные науки (общие)
OPUS-Design: a 3D convolutional approach for protein sequence design targeting complex spatial microenvironments
Аннотация
Complex three-dimensional (3D) protein environments, including monomers, oligomers, and ligand-bound small-molecule complexes, form a canonical class of biological systems. Protein sequence design based on backbone structure, commonly referred to as protein inverse folding, is central to modeling these intricate systems, as it seeks to uncover the intrinsic link between 3D structural topology and target amino acid sequences within complex spatial microenvironments. State-of-the-art methods for this task are dominated by graph-based models such as ProteinMPNN and ESM-IF, which deliver strong performance. However, 3D convolutional neural networks (3DCNNs) offer a promising alternative. These networks feature inherent rotation invariance and excel at capturing dense local spatial contexts in complex protein microenvironments, yet their full potential for protein inverse folding remains underexplored. Here, we introduce OPUS-Design, a two-stage 3DCNN-based framework for protein sequence design. Tested on CASP15 and CAMEO targets spanning monomers, oligomers, and ligand-bound complexes, OPUS-Design matches or outperforms top graph-based methods in sequence recovery. These findings validate the efficacy of 3DCNN for protein inverse folding and suggest its potential for broader applications in related structural prediction tasks.