Single-cell mitochondrial morphomics reveals cellular heterogeneity and predicts complex I, III, and ATP synthase...

Mitochondrial heterogeneity drives diverse cellular responses in neurodegenerative diseases, complicating the evaluation of mitochondrial dysfunction. In this study, we describe a high-throughput imaging and analysis approach to investigate cell-to-cell mitochondrial variability. We applied known mitochondrial function inhibitors - rotenone, antimycin, and oligomycin to inhibit complexes I, III, and V (ATP synthase) function in human induced pluripotent stem cell-derived cortical neurons, a model commonly used in neurodegenerative disease research. We captured a large number of cell images and extracted a diverse range of mitochondrial morphological features related to shape, size, texture, and spatial distribution, for an unbiased and comprehensive analysis of mitochondrial morphology. Group-level cell analysis, which examines the collective responses of cells exposed to the same mitochondrial inhibitor, showed that cells treated with rotenone, antimycin, or oligomycin clustered together based on their shared morphological changes. Rotenone and antimycin, both targeting different complexes of the electron transport chain, formed sub-clusters within a larger cluster. In contrast, oligomycin, which inhibits ATP synthase, resulted in a distinct cluster likely due to its differing effect on ATP production. Single-cell analysis using dimensionality reduction techniques revealed distinct subpopulations of cells with varying degrees of sensitivity to each mitochondrial inhibitor, identifying the most affected cells. Mitochondrial feature differential expression analysis showed that neurite-related mitochondrial features, such as intensity and size, were more severely impacted than cell body-related mitochondrial features, particularly with rotenone and antimycin, which target the electron transport chain. In contrast, oligomycin which affects ATP synthesis by directly inhibiting ATP synthase showed relatively less severe alterations in neurite-related mitochondrial features, highlighting a distinct effect of the mode of action between inhibitors. By incorporating the most affected cells into machine learning models, we significantly improved the prediction accuracy of mitochondrial dysfunction outcomes - 81.97% for antimycin, 75.12% for rotenone, and 94.42% for oligomycin. This enhancement underscores the value of targeting highly responsive cell subpopulations, offering a more precise method for evaluating mitochondrial modulators and therapeutic interventions in neurodegenerative diseases.