Integration of a microfluidic multicellular coculture array with machine learning analysis to predict adverse cutaneous drug reactions

Abstract

Adverse cutaneous reactions are potentially life-threatening skin side effects caused by drugs administered into the human body. The availability of a human-specific <i>in vitro</i> platform that can prospectively screen drugs and predict this risk is therefore of great importance to drug safety. However, since adverse cutaneous drug reactions are mediated by at least 2 distinct mechanisms, both involving systemic interactions between liver, immune and dermal tissues, existing <i>in vitro</i> skin models have not been able to comprehensively recapitulate these complex, multi-cellular interactions to predict the skin-sensitization potential of drugs. Here, we report a novel <i>in vitro</i> drug screening platform, which comprises a microfluidic multicellular coculture array (MCA) to model different mechanisms-of-action using a collection of simplistic cellular assays. The resultant readouts are then integrated with a machine-learning algorithm to predict the skin sensitizing potential of systemic drugs. The MCA consists of 4 cell culture compartments connected by diffusion microchannels to enable crosstalk between hepatocytes that generate drug metabolites, antigen-presenting cells (APCs) that detect the immunogenicity of the drug metabolites, and keratinocytes and dermal fibroblasts, which collectively determine drug metabolite-induced FasL-mediated apoptosis. A single drug screen using the MCA can simultaneously generate 5 readouts, which are integrated using support vector machine (SVM) and principal component analysis (PCA) to classify and visualize the drugs as skin sensitizers or non-skin sensitizers. The predictive performance of the MCA and SVM classification algorithm is then validated through a pilot screen of 11 drugs labelled by the US Food and Drug Administration (FDA), including 7 skin-sensitizing and 4 non-skin sensitizing drugs, using stratified 4-fold cross-validation (CV) on SVM. The predictive performance of our <i>in vitro</i> model achieves an average of 87.5% accuracy (correct prediction rate), 75% specificity (prediction rate of true negative drugs), and 100% sensitivity (prediction rate of true positive drugs). We then employ the MCA and the SVM training algorithm to prospectively identify the skin-sensitizing likelihood and mechanism-of-action for obeticholic acid (OCA), a farnesoid X receptor (FXR) agonist which has undergone clinical trials for non-alcoholic steatohepatitis (NASH) with well-documented cutaneous side effects.