Organelle-Aware Representation Learning Enables Label-Free Detection of Mitochondrial Dysfunction in Live Human Neurons

Mitochondrial dysfunction is a convergent hallmark of neurodegenerative diseases and represents a promising biomarker for early diagnosis and therapy. However, current in vitro assays rely on fluorescence or electron microscopy, which are invasive, low-throughput, and incompatible with longitudinal analysis. Here, we present a noninvasive framework by integrating label-free optical diffraction tomography (ODT) with organelle-aware representation learning to detect subtle mitochondrial dysfunction in live human induced pluripotent stem cell (hiPSC)-derived neurons. Through virtual staining of the nuclei, lysosomes, and mitochondria, we establish two complementary and interpretable classification pipelines: a deep learning model with organelle-aware encoder and a logistic regression model on morphometric descriptors. Both models achieve approximately 85% accuracy: the deep model provides end-to-end prediction with minimal feature engineering, whereas the logistic regression model offers a more interpretable, feature-based approach. To our knowledge, this is the first demonstration of ODT-based organelle-resolved virtual staining in live human neurons, establishing a scalable, non-invasive platform for mitochondrial disease modeling, drug discovery, and neurodegeneration research.