HSQC2STRUC: A Machine Learning Model for Protein Secondary Structure Prediction using Unassigned NMR Spectra

Abstract

Dynamic changes in the secondary structure content of proteins can provide valuable insights into protein function or dysfunction. Predicting these dynamic changes is still a significant challenge but is of paramount importance for basic research as well as drug development. Here, we present a machine learning-based model that predicts the secondary structure content of proteins based on their un assigned1H,15N-HSQC NMR spectra with an RMSE of 0.11 for -helix, 0.08 for {beta}-sheet and 0.12 for random coil content. Our model has been implemented into an easy-to-use and publicly available web service that estimates secondary structure content based on a provided peak list. Furthermore, a Python version is provided, ready to be integrated into Brukers TopSpin software or own scripts.