A deep learning convolutional neural network distinguishes neuronal models of Parkinson's disease from matched controls

Abstract

Parkinsons disease (PD) is a neurodegenerative disorder that results in the loss of dopaminergic neurons in the substantia nigra pars compacta. Despite advances in understanding PD, there is a critical need for novel therapeutics that can slow or halt its progression. Induced pluripotent stem cell (iPSC)-derived dopaminergic neurons have been used to model PD but measuring differences between PD and control cells in a robust, reproducible, and scalable manner remains a challenge. In this study, we developed a binary classifier convolutional neural network (CNN) to accurately classify microscopy images of PD models and matched control cells. We acquired images of iPSC-derived neural precursor cells (NPCs) and dopaminergic (DANs) and trained multiple CNN models comparing control cells to genetic and chemical models of PD. Our CNN accurately predicted whether control NPC cells were treated with the PD-inducing pesticide rotenone with 97.60% accuracy. We also compared control to a genetic model of PD (deletion of the Parkin gene) and found a predictive accuracy of 86.77% and 95.47% for NPC and DAN CNNs, respectively. Our cells were stained for nuclei, mitochondria, and plasma membrane, and we compared the contribution of each to the CNNs accuracy. Using all three features together produced the best accuracy, but nuclear staining alone produced a highly predictive CNN. Our study demonstrates the power of deep learning and computer vision for analyzing complex PD-related phenotypes in DANs and suggests that these tools hold promise for identifying new targets for therapy and improving our understanding of PD.