Enhancing Spectrometer\nPerformance with Unsupervised\nMachine Learning

Solid-state NMR spectroscopy (SSNMR)\nis a powerful technique to\nprobe structural and dynamic properties of biomolecules at an atomic\nlevel. Modern SSNMR methods employ multidimensional pulse sequences\nrequiring data collection over a period of days to weeks. Variations\nin signal intensity or frequency due to environmental fluctuation\nintroduce artifacts into the spectra. Therefore, it is critical to\nactively monitor instrumentation subject to fluctuations. Here, we\ndemonstrate a method rooted in the unsupervised machine learning algorithm\nprincipal component analysis (PCA) to evaluate the impact of environmental\nparameters that affect sensitivity, resolution and peak positions\n(chemical shifts) in multidimensional SSNMR protein spectra. PCA loading\nspectra illustrate the unique features associated with each drifting\nparameter, while the PCA scores quantify the magnitude of parameter\ndrift. This is demonstrated both for double (HC) and triple resonance\n(HCN) experiments. Furthermore, we apply this methodology to identify\nmagnetic field <i>B</i><sub>0</sub> drift, and leverage\nPCA to “denoise” multidimensional SSNMR spectra of the\nmembrane protein, EmrE, using several spectra collected over several\ndays. Finally, we utilize PCA to identify changes in <i>B</i><sub>1</sub> (CP and decoupling) and <i>B</i><sub>0</sub> fields in a manner that we envision could be automated in the future.\nOverall, these approaches enable improved objectivity in monitoring\nNMR spectrometers, and are also applicable to other forms of spectroscopy.