MitoDecode: quantitative phenotypic decoding of mitochondrial stress states from fluorescence images

Abstract

The quantitative measurement of mitochondrial stress from fluorescence images of cells remains a challenging task. Existing methods rely on mitochondrial segmentation, manually predefined features, and discrete phenotype severity classes, which are prone to error and bias. Here, we present MitoDecode, a deep learning network that automatically assesses type and severity of mitochondrial stress from fluorescence microscopy images. MitoDecode eliminates the need for mitochondrial segmentation, is parameter-free, and operates across both diffraction-limited and super-resolution optical imaging modalities. It assigns a phenotypic probability profile for each individual cell, indicating both type and severity of phenotypic stress. The readout thus enables comparisons between different stressors and provides clues for underlying mechanisms. Furthermore, the method captures phenotypic shifts continuously across treatment doses, revealing subtle variations that would otherwise be missed. We demonstrate MitoDecode's adaptability across diverse cell types, imaging modalities, and stress conditions, and show that it can reliably detect mitochondrial phenotypes, even when stressors target non-mitochondrial pathways. We validate the reliability of the model with biochemical assays, showcase its utility for phenotypic stress profiling of drug treatments, and envisage its application for early-stage disease detection, drug discovery, and mechanistic lead generation.