Identifying cancer cell metabolic states in autofluorescence lifetime images with machine learning

Cancer cells switch to glycolytic metabolic states even in aerobic environments to support enhanced growth and cellular functions. This phenomenon is known as the Warburg effect, and it inspires advancing interests in targeting metabolism for cancer therapy. Optical metabolic imaging (OMI) measures the fluorescence intensity and lifetime of the co-enzymes reduced nicotinamide adenine dinucleotide (NADH) and oxidized flavin adenine dinucleotide (FAD). OMI can quantitatively distinguish cellular metabolic activities in a label-free manner. The goal of this study is to identify key metabolic pathways of cancer cells using NADH and FAD fluorescence lifetime imaging and machine learning methods. MCF-7 breast cancer cells were exposed to different culture media and inhibitors to disturb their metabolic activities, and NADH and FAD fluorescence lifetime imaging were performed by a multi-photon microscope. Here, we proposed a potential method of training convolutional neural networks to predict cellular metabolic states. Adapting convolutional neural networks for the prediction of cancer cell metabolic conditions was anticipated to provide substantially better performance than traditional models with extracted features. In summary, this investigation offers a non-invasive, quantitative technology to detect metabolic perturbations at a cellular level, which improves the identification of different metabolic states of cancer cells.