Machine Learning-Assisted High-Throughput SERS Classification of Cell Secretomes

Abstract

During the response to different stress conditions, damaged cells react in multiple ways, including the release of a diverse cocktail of metabolites. Moreover, secretomes from dying cells can contribute to the effectiveness of anticancer therapies and can be exploited as predictive biomarkers. The nature of the stress and the resulting intracellular responses are key determinants of the secretome composition, but monitoring such processes remains technically arduous. Hence, there is growing interest in developing tools for noninvasive secretome screening. In this regard, it has been previously shown that the relative concentrations of relevant metabolites can be traced by surface-enhanced Raman scattering (SERS), thereby allowing label-free biofluid interrogation. However, conventional SERS approaches are insufficient to tackle the requirements imposed by high-throughput modalities, namely fast data acquisition and automatized analysis. Therefore, machine learning methods were implemented to identify cell secretome variations while extracting standard features for cell death classification. To this end, ad hoc microfluidic chips were devised, to readily conduct SERS measurements through a prototype relying on capillary pumps made of filter paper, which eventually would function as the SERS substrates. The developed strategy may pave the way toward a faster implementation of SERS into cell secretome classification, which can be extended even to laboratories lacking highly specialized facilities.