Interpreting Raman spectra using machine learning: towards a non-invasive method of characterizing single cells

Abstract

Raman microscopy has the potential to non-destructively measure the biomolecular changes in single cells in a label-free manner. However, the extent to which Raman spectra can effectively infer biologically-relevant information is not well understood. In this thesis, we use machine learning methods to explore the ability of Raman microscopy data to infer cell states in microbes and the gene expression values of ten genes in mouse embryonic fibroblasts (MEFs) undergoing a dynamic cellular reprogramming process. Using a multi-modal, supervised learning approach, we provide evidence that Raman spectra can accurately resolve microbial cell types. This thesis also presents a robust computational pipeline to preprocess Raman spectra, calibrate multi-modal data, and segment nuclei; an analysis of methods to increase the signal-to-noise ratio of Raman spectra; and an analysis of Raman spectral features important for predicting microbial cell-type. Together, the results suggest Raman microscopy be considered as a useful modality for distinguishing cell-types and potentially tracking cellular dynamics, a common goal of many consortia including the Human Cell Atlas.