Machine learning-based image analysis of Parkinson's disease iPS-derived neurons predicts genotype and reveals mitochondria-lysosome abnormalities

Abstract

Mitochondrial and lysosomal dysfunction are central features of Parkinsons disease (PD) across major genetic forms including PRKN, SNCA, and LRRK2. We applied cell morphomics, a machine-learning-based framework combining high-content imaging with quantitative feature extraction, to analyse mitochondrial and lysosomal morphology at single-cell resolution in iPS cell-derived cortical neurons from PD patients and healthy controls (13 lines total). Supervised machine-learning models distinguished PD neurons from controls with high accuracy (AUC = 0.87) and reliably separated individual genotypes. Feature importance and attribution analysis revealed genotype-specific organelle biases, with mitochondrial features dominating classification in PRKN neurons, balanced mitochondrial and lysosomal contributions in SNCA neurons, and a greater lysosomal contribution in LRRK2 neurons. Multi-class models retained strong performance, and findings were reproduced across two independent laboratories using different dyes and imaging conditions. These results demonstrate that morphomics provides a robust and scalable framework to quantify genotype-specific organelle abnormalities in PD neurons and supports its application for cellular stratification and biomarker discovery.