In vitro cardiac modeling of desmin-related myofibrillar myopathy and deep learning-based image analysis to develop a high-content screening assay

Mutations of DES, the gene encoding desmin, the main intermediate filament of muscle cells, are one of the major cause of Myofibrillar myopathy(MFM). This disease is mainly characterized by a muscular dystrophy but also, in many cases, by the progressive emergence of cardiac dysfunctions evolving frequently from dilated cardiomyopathy to heart failure. For the moment, no specific treatment can revert this disease. The following study relied on the development of an automated tool base on deep-learning to discriminate efficiently mutant and control cardiomyocytes in the perspective of drug screening. We use cardiomyocytes derived from human induced pluripotent stem cells (hiPSC-CMs) obtained from control or patients carrying a DES mutation (E439K). Phenotyping of hiPSC-CMs was performed by immunocytostaining of the sarcomeres or the desmin network. After imaging, the morphological perturbations were evaluated manually based either on the sarcomeric organization or on the aspect of the desmin network at single cell level. Then an automated classification model was generated by using machine-learning method trained with the previous images annotated with the results of the manual classifications. In parallel, another automated classification model based on cell profiling from sarcomeric, desmin and mitochondrial staining by using the deep learning EfficientNet B0 model was setting up in order to distinguish Control cell and DES mutated cell. Manual classification based on sarcomeric organization was established on 4 classes ranging from cells with disorganized sarcomeres to cells with a well-organized sarcomeric network on 1500 single cells. We observed that the distribution of DES mutated cells is shifted toward the class of disorganized sarcomere compared to control cells. Identically, we also manually classified 1200 single cells based on the organization of the desmin network (4 classes). We also observed a shift of DES mutant cells to be classified as cells with aggregated desmin network compared to control cells. Finally, the various validation tests of the two automated classification methods showed significant correlation with the manual classification methods. Here we demonstrate the efficacy of our automated classification software to be able to discriminates mutant and control cells. This analysis platform will then be used to perform high throughput drug screening assays to discover new therapeutic molecules.