Label-free metabolic clustering through unsupervised pixel classification of multiparametric fluorescent images

Abstract

Autofluorescence microscopy is a promising label-free approach to characterize NADH and FAD metabolites in live cells, with potential applications in clinical practice. Although spectrally resolved lifetime imaging techniques can acquire multiparametric information about the biophysical and biochemical state of the metabolites, these data are evaluated at the whole-cell level, thus providing only limited insights in the activation of metabolic networks at the microscale. To overcome this issue, here we introduce an artificial intelligence-based analysis that, leveraging the multiparametric content of spectrally resolved lifetime images, allows to detect and classify, through an unsupervised learning approach, metabolic clusters, which are regions having almost uniform metabolic properties. This method contextually detects the cellular mitochondrial turnover and the metabolic activation state of intracellular compartments at the pixel level, described by two functions: the cytosolic activation state (CAF) and the mitochondrial activation state (MAF). This method was applied to investigate metabolic changes elicited in the breast cancer cell line MCF-7 by specific inhibitors of glycolysis and electron transport chain, and by the deregulation of a specific mitochondrial enzyme (ACO2) leading to defective aerobic metabolism associated with tumor growth. In this model, mitochondrial fraction undergoes to a 13% increase upon ACO2 overexpression and the MAF function changes abruptly by altering the metabolic state of about the 25% of the mitochondrial pixels.

Keywords: Artificial intelligence; Fluorescence lifetime imaging microscopy; Live cell metabolic imaging; Machine learning; Metabolic clustering; NAD(P)H FLIM.