Deep learning in human neurons predicts mechanistic subtypes of Parkinson's

Abstract

Parkinsons disease (PD) is a common, devastating, and incurable neurodegenerative disorder. Several molecular mechanisms have been proposed to drive PD, with genetic and pathological evidence pointing towards aberrant protein homeostasis and mitochondrial dysfunction. PD is clinically highly heterogeneous, it is likely that different mechanisms underlie the pathology in different individuals, each requiring a specific targeted treatment. Recent advances in stem cell technology and fluorescent live-cell imaging have enabled the generation of patient-derived neurons with different mechanistic subtypes of PD. Here, we performed multi-dimensional fluorescent labelling of organelles in iPSC-derived neurons, in healthy control cells, and in four different disease subclasses. We generated a machine learning-based model that can simultaneously predict the presence of disease, and its primary mechanistic subtype. We independently trained a series of classifiers using both quantitative single-cell fluorescence variables and images to build deep neural networks. Quantitative cellular profile-based classifiers achieve an accuracy of 82%, whilst image based deep neural networks predict control, and four distinct disease subtypes with an accuracy of 95%. The classifiers achieve their accuracy across all subtypes primarily utilizing the organellar features of the mitochondria, with additional contribution of the lysosomes, confirming their biological importance in PD. Taken together, we show that machine learning approaches applied to patient-derived cells are able to predict disease subtypes, demonstrating that this approach may be used to guide personalized treatment approaches in the future.