Circulating microRNAs and Both Association With Methacholine PC20 and Prediction of Asthma Exacerbation in the Childhood Asthma Management Program (CAMP) Cohort

Abstract

Background: Circulating microRNAs have shown promise both as a non-invasive biomarker and a predictor of disease activity. Prior asthma studies with clinical, biochemical, or genomic tests have failed to demonstrate excellent prediction of asthma exacerbation. This thesis hypothesizes that circulating miRNA would reveal: 1) candidate biomarkers related to airway hyperresponsiveness (AHR) and provide biologic insights into asthma epigenetic influences, 2) a panel of circulating miRNA in a pediatric asthma cohort or a combined microRNA-clinical asthma exacerbation score that may have superior predictive capability compared to the clinical asthma exacerbation score alone, and 3) the power of machine learning techniques such as backpropagation neural networks and XGBoost that may provide both improved prediction of exacerbations or perhaps yield insight into complex non-linear behavior. Methods: Serum samples were obtained at randomization for 160 children in the Childhood Asthma Management Program and were profiled using the TaqMan miRNA array set containing 754 miRNAs. For Aims 2 and 3, only 153 subjects had complete information on steroid bursts in the first year after randomization. Aim 1: The association of the isolated miRNA with methacholine PC20 was assessed. Network and pathway analyses were performed. Functional validation of two significant miRNAs was performed in human airway smooth muscle cells (HASMs). Aim 2: Dichotomized data for asthma exacerbation from the first year after randomization to the inhaled corticosteroid arm were used for binary logistic regression with microRNA cycle threshold and clinical exacerbation score. Ontology and pathway analyses were performed for significant miRNAs Aim 3: Dichotomized data for asthma exacerbation as in Aim 2 was used for both backpropagation and resilient backpropagation neural networks and XGBoost with miRNA features selected by the Kruskal Test for binary and multiclass classification. Randomized hyperparameter optimization was used for all 3 algorithms. The neural network metric for classification was mean misclassification error. The XGBoost metric for classification was AUROC. Results: Aim 1: Of 155 well-detected circulating miRNAs, eight were significantly associated with PC20 with the strongest association with miR-296-5p. Pathway analysis revealed miR-16-5p as a network hub, and involvement of multiple miRNAs interacting with genes in the FoxO and Hippo signaling pathways by KEGG analysis. Functional validation of two miRNA in HASM showed effects on cell growth and diameter. Aim 2: Of the 125 well-detected circulating miRNA, 12 had significant odd ratios for exacerbation with the most significant being miR-206. Each doubling of expression of the 12 miRNA resulted in between a 25-67% increase in risk of exacerbation. Stepwise logistic regression resulted in a three miRNA model that, when combined with the clinical score, demonstrated an AUROC of 0.81, which was superior to either the clinical model alone (AUROC 0.67) or miRNA model (AUROC 0.71). The three microRNAs also had biological relevance with involvement in NK-kB signaling and inactivation of Gsk3 by AKT pathways. Aim 3: For the neural network analysis, the backpropagation algorithm with all miRNA features had an improved mean misclassification error (0.248) compared to the resilient backprogation algorithm with all features and the top 10 features. The XGBoost model had an AUROC of 0.69, which was similar to the AUROC of the 3 miRNA panel found in the logistic regression analysis (Aim 2). Conclusion: Aim 1: Reduced circulatory miRNA expression at baseline is associated with an increase in PC20. These miRNA provide biologic insights into, and may serve as biomarkers of, asthma severity. miR-16-5p and -30d-5p regulate airway smooth muscle phenotypes critically involved in asthma pathogenesis, supporting a mechanistic link to these findings. Functional ASM phenotypes may be directly relevant to AHR. Aim 2: Circulating microRNAs combined with a clinical model of asthma exacerbations demonstrate prediction of exacerbation status in subjects taking inhaled corticosteroids with use of logistic regression. Aim 3: These preliminary analyses with machine learning algorithms including neural network and XGBoost did not result in improved prediction of asthma exacerbation compared to logistic regression. Neural networks were computationally complex and did not result in improved accuracy compared to the no information rate. XGBoost and logistic regression results used different feature selection (non-parametric vs. parametric) and data pre-processing steps (dense matrix input vs. median imputation of missing values) yielding similar predictive capability for asthma exacerbation with circulating miRNA. While further analyses are needed, these preliminary results support robust prediction with circulating miRNA especially given miR-206 was a key feature among models despite different feature selection and analytic techniques.