Assisting and accelerating NMR assignment with restrained structure prediction

Abstract

Accurate dynamic protein structures are essential for drug design. NMR experiments can detect protein structures and potential dynamics, but the spectrum assignment and structure determination requires expertise and is time-consuming, while deep-learning-based structure predictions may be inconsistent with experimental observations. A symbiosis between experiments and AI methods is therefore essential for solving such problems. Here, we developed a Restraint Assisted Structure Predictor (RASP) model and an iterative Folding Assisted peak ASsignmenT (FAAST) pipeline directly leveraging experimental information to improve the AI-assisted structure prediction and facilitate experimental data analysis in an integrative way. The RASP model improves structure prediction, especially for multi-domain and few-MSA proteins. The FAAST pipeline for NMR NOESY analysis reduces the time consumption to hours and yields high quality structure ensemble. Both methods show high consistency between predicted structures and restraints, provided or iteratively assigned. This strategy can be expanded to other types of sparse experimental information in structure prediction.