NMR Spectroscopy for the Validation of AlphaFold2 Structures

The introduction of AlphaFold has fundamentally changed our ability to predict the structure of proteins from their primary sequence of amino acids. As machine learning (ML) and artificial intelligence (AI) based protein prediction continues to advance, we examine the potential of hybrid techniques that combine experiment and computation that may yield more accurate structures than AI alone with significantly reduced experimental burden. We have developed heuristics comparing N-edited NOESY spectra and AlphaFold predicted structures that seek to determine whether the predicted structure reasonably describes the structure of the protein which generated the NOESY. These heuristics are fundamentally different than those used before in that they concentrate on interresidue contacts and not specific HH pairs. We present a large collection of data connecting entries across the BMRB, PDB and AlphaFold Database that includes experimentally derived structures and corresponding spectra, establishing it as a means to develop and test hybrid methods utilizing AlphaFold and NMR spectra to perform structure determination. These data test the new heuristics ability to identify inaccurate AlphaFold structures. A support vector machine was developed to test the consistency of NMR data with predicted structure and we show its application to the structure of an unsolved engineered protein, LoTOP.