The revolution in protein structure prediction by deep learning offers tremendous opportunities to accelerate biological discovery, provided that models can be validated with experimental data. NMR spectroscopy is a powerful biophysical technique to probe structure and dynamics of proteins at an atomic level. Traditional NMR methods require collection and interpretation of several large experimental data sets to calculate a structure. To reduce the amount of experimental data and accelerate analysis, we present a novel forward model for simulating NMR spectra from protein structures and use image analysis to score similarity between simulation and experiment. We have developed a software interface and analysis pipeline, BPHON (a ChimeraX extension), and tested its performance on standard protein data sets. We then apply BPHON to challenging, real-world experimental use cases, including a transporter membrane protein EmrE and alpha-synuclein fibrils observed in neurological disease. From the sets of candidate structures, BPHON quantitatively ranks the agreement with the experimental data, enabling high-resolution structure refinement for membrane proteins and classification of fibril samples from Parkinson disease subjects.