Phenotype Classification of Intact Cells by NMR Spectroscopy through Machine Learning Approaches

NMR spectroscopy is a powerful, noninvasive tool to analyze complex biological samples. In vitro, high-resolution, 1D NMR spectra of biofluids and cell extracts make it possible to classify biological samples based on their metabolic fingerprint. However, such analysis is currently not possible with live cells or tissues, or by spectroscopic imaging in vivo, due to the line broadening arising from the intrinsic inhomogeneity of such samples, causing severe signal overlap. Here, we show that machine learning approaches applied to poorly resolved NMR spectra of live, intact cells recorded at high fields allow for the classification of different physiopathologically relevant cell types cultured in vitro. We demonstrate the successful classification of neural progenitor cells, neurons, and astrocytes, as well as the classification of mixed cell type samples, and show that a classifier trained on high-field NMR spectra can discriminate cells analyzed at lower fields, approaching those of current MRI instruments. In the future, this approach could be further developed for MRSI data analysis applications, potentially offering a noninvasive diagnostic tool for lesions of the central nervous system and reducing the need for biopsies.