Detection and monitoring of the autophagy response in mammalian cells using Raman spectroscopy and machine learning

Abstract

Autophagy is a highly conserved biological process in eukaryotic cells with an essential role in cellular homeostasis, where dysfunction in this process is a key factor in a multitude of diseases. To this end, accuracy and reliability in the detection and monitoring of autophagy is critical for medicine and research. Current methods for autophagy detection can be costly, time consuming and labour intensive, which hampers their use outside of research laboratories. Raman spectroscopy is a spectrographic technique which allows visualisation of the chemical environment within a sample in a non-destructive and non-invasive manner. The use of Raman spectroscopy has never been fully realised for the detection of autophagy and can offer key benefits to the process. By using Raman spectroscopy to assess the chemical changes in cells undergoing autophagy key changes can be mapped out and used to create a predictive model via the use of machine learning. Raman collection was optimised via the use of drying cells onto a gold coated mirror, to facilitate high Raman scattering intensities with relatively short collection times. The processing of the spectra was optimised with a focus on creating a more generalisable process. Testing different baseline correction algorithms it was found that the asymmetric least squares algorithm performed the best across multiple distinct Raman spectra from different cell lines and different experiments. Raman spectra of cells after 4 hours complete amino acid deprivation show little difference in their Raman spectra, whereas the spectra of cells after 1, 2 and 5 days without l-glutamine show significant changes. These changes are often cell line specific but a commonly seen change is a drop in the 1675 cm-1 shoulder region associated with the protein secondary structure β-pleated sheets. Many of these changes however are also seen in ATG5(-/-) cells suggesting the changes are not autophagy specific. Using an artificial neural network, trained on the first 20 principal components of the Raman spectrum and the cell type, could distinguish between control cells, autophagy induced ATG5(+/+) cells and autophagy induced ATG5(-/-) cells with a 98.2% accuracy. This research lays the groundwork for the use of Raman spectroscopy as a method of detection for autophagy. Larger and more comprehensive data sets would need to be collected to further the applicability of this method.