Mammalian cells contain various macromolecules that can be investigated non-invasively with Raman spectroscopy. The particular mixture of major macromolecules present in a cell being probed are reflected in the measured Raman spectra. Determining macromolecular identities and estimating their concentrations from these mixture Raman spectra can distinguish cell types and otherwise enable biological research. However, the application of canonical multivariate methods, such as principal component analysis (PCA), to perform spectral unmixing yields mathematical solutions that can be difficult to interpret. Non-negative matrix factorization (NNMF) improves the interpretability of unmixed macromolecular components, but can be difficult to apply because ambiguities produced by overlapping Raman bands permit multiple solutions. Furthermore, theoretically sound methods can be difficult to implement in practice. Here we examined the effects of a number of empirical approaches on the quality of NNMF results. These approaches were evaluated on simulated mammalian cell Raman hyperspectra and the results were used to develop an enhanced procedure for implementing NNMF. We demonstrated the utility of this procedure using a Raman hyperspectral data set measured from human islet cells to recover the spectra of insulin and glucagon. This was compared to the relatively inferior PCA of these data.
© 2017 The Author(s)PDF Article