Deep learning-based cell profiling based on neuronal morphology

Abstract

Treatment of neurons with {beta}-amyloid peptide (A{beta}1-42) has been widely used as a model to interrogate the cellular and molecular mechanisms underlying Alzheimers disease, and as an assay system to identify drugs that reverse or block disease phenotype. Prior studies have largely relied on high content imaging (HCI) to extract cellular features such as neurite length or branching, but these have not offered a robust/comprehensive means of relating readout to A{beta}1-42 concentrations. Here, we use a deep learning-based cell profiling technique to directly measure the impact of A{beta}1-42 on primary murine cortical neurons. The deep learning model achieved approximately 80% accuracy, compared to 54% for the cell phenotypic feature-based approach. The deep learning model could distinguish subtle neuronal morphological changes induced by a range of A{beta}1-42 concentration. When tested on a separate dataset, the accuracy remained comparable and dropped by only 2%. Our study demonstrates that deep learning-based cell profiling is superior to HCI-based feature extraction on neuronal morphology and it provides an alternative to a dose/response curve, where the modality of the response does not have to be pre-determined. Moreover, this approach could form the basis of a screening tool that can be applied to any cellular model where appropriate phenotypic markers based on genotypes and/or pathological insults are available.