Development of Fluorescence Lifetime Imaging Techniques for Identifying and Predicting Cellular Metabolic Perturbations

Abstract

The detection of cellular metabolic perturbations is crucial for comprehending various physiological and pathological conditions. Autofluorescence lifetime imaging of two endogenous metabolic coenzymes reduced nicotinamide adenine dinucleotide (NADH) and oxidative flavin adenine dinucleotide (FAD) provides a label-free approach to monitor cellular metabolic changes. Despite the potential of autofluorescence lifetime imaging as a non-invasive method for quantifying cellular metabolism, it remains challenging to analyze and interpret autofluorescence lifetime imaging metrics in the context of cellular metabolism. The objective of this project is to develop analytical tools and prediction models to advance autofluorescence lifetime imaging as a platform technology for cellular-level detection of cellular metabolic pathway use. In this study, we investigated the measurement of metabolic variations in activated and quiescent T cells using different redox ratios. Through correlation analysis, we found that the fluorescence lifetime redox ratio (FLIRR) and various optical redox ratio formats were not correlated at the single-cell level. This work highlights that intensity and lifetime-computed redox ratios resolve metabolic perturbations but are influenced by different metabolic processes. Additionally, we conducted rigorous metabolic perturbation experiments and characterized the NADH and FAD lifetime signatures associated with glycolysis, oxidative phosphorylation, and glutaminolysis in cancer cells. Leveraging these findings, we developed two distinct models for predicting cancer cell metabolism. Firstly, a 1D conventional machine learning algorithm that uses cell-averaged autofluorescence endpoints was created and achieved over 90% accuracy in classifying cells as glycolytic or oxidative. Secondly, a 2D convolutional neural network using autofluorescence intensity and lifetime images was developed, yielding better accuracy than the 1D model, due to the inclusion of intra-cellular spatial patterns. These models were extensively validated across multiple datasets, including autofluorescence lifetime images of breast cancer cells, liver cancer cells, and T cells, data that was obtained from different microscopy systems. Furthermore, we introduced a novel approach to address metabolic states from autofluorescence lifetime data by developing a 3D convolutional neural network model trained on 3D NADH images that include two lateral spatial dimensions (X-Y) and one time dimension with the lifetime histogram of photon arrival times, or temporal point spread function. By incorporating both temporal and spatial patterns, this model successfully differentiated three major metabolic pathways (glycolysis, oxidative phosphorylation, glutaminolysis) in cancer cells with high accuracy. Notably, the model demonstrated applicability to macrophages with metabolism perturbed by mitochondrial gene mutations. In summary, our findings provide accessible and comprehensive approaches to analyze autofluorescence lifetime imaging data, facilitating a non-destructive understanding of cellular metabolic processes with single-cell resolution and high speed. Our results demonstrate the efficacy of machine learning and convolutional neural networks in predicting and comprehending metabolic phenotypes, bridging the gap between the autofluorescence lifetime variations and metabolic perturbations. Consequently, this research has significant implications for unraveling the underlying mechanisms of cellular metabolism and may contribute to advancements in disease diagnosis and treatment.