Guiding antidepressant choice by applying deep learning approaches on individual patient-derived neurons: a window to drug efficacy in each patient

Depression treatment remains a significant clinical challenge due to the variability in patient responses to antidepressant medications. Outside of pharmacogenetic testing, established mostly for drug metabolism implications, current methods of selecting antidepressants are largely trial-and-error. Accordingly, the majority of patients do not respond to their first prescribed antidepressant, leading to prolonged periods of iterative drug-cycling and continued patient suffering. The advent of stem cell technology enables the generation of blood-derived neurons for each patient, and the interrogation of drug efficacy in-vitro on an individual patient basis. This study used computer vision and deep learning approaches on confocal microscopy images from patient-derived neurons to train, validate and test classifiers for patient antidepressant response. The classifiers were trained on thousands of microscopy image tiles from neurons derived from 6 patients from the STAR*D clinical trial after exposure to drug or vehicle. Balanced accuracy for drug response prediction ranged from 0.68 to 0.97 on the test set. Random labeling of tiles as vehicle or treatment led to a dramatically reduced balanced accuracy, ranging from 0.50 to 0.66, demonstrating signal specificity. This proof of concept study supports 1. The utility of patient-derived neurons as an in-vitro model in depression and 2. the use of unsupervised machine learning approaches to derive signals from these neurons for treatment optimization as well as novel biological insights. Ultimately, these methods can serve as a tool to match individual treatment regiments to each patient, as well as to support drug development efforts.