Enhancing Spectrometer Performance with Unsupervised Machine Learning

Abstract

Solid-state NMR spectroscopy (SSNMR) is a powerful technique to probe structural and dynamic properties of biomolecules at an atomic level. Modern SSNMR methods employ multidimensional pulse sequences requiring data collection over a period of days to weeks. Variations in signal intensity or frequency due to environmental fluctuation introduce artifacts into the spectra. Therefore, it is critical to actively monitor instrumentation subject to fluctuations. Here, we demonstrate a method rooted in the unsupervised machine learning algorithm principal component analysis (PCA) to evaluate the impact of environmental parameters that affect sensitivity, resolution and peak positions (chemical shifts) in multidimensional SSNMR protein spectra. PCA loading spectra illustrate the unique features associated with each drifting parameter, while the PCA scores quantify the magnitude of parameter drift. This is demonstrated both for double (HC) and triple resonance (HCN) experiments. Furthermore, we apply this methodology to identify magnetic field B₀ drift, and leverage PCA to "denoise" multidimensional SSNMR spectra of the membrane protein, EmrE, using several spectra collected over several days. Finally, we utilize PCA to identify changes in B₁ (CP and decoupling) and B₀ fields in a manner that we envision could be automated in the future. Overall, these approaches enable improved objectivity in monitoring NMR spectrometers, and are also applicable to other forms of spectroscopy.