Predictive genetic panel for adult asthma using machine learning methods

Abstract

Background: Asthma is a chronic inflammatory disease of the airways that is heterogeneous and multifactorial, making its accurate characterization a complex process. Therefore, identifying the genetic variations associated with asthma and discovering the molecular interactions between the omics that confer risk of developing this disease will help us to unravel the biological pathways involved in its pathogenesis.

Objective: We sought to develop a predictive genetic panel for asthma using machine learning methods.

Methods: We tested 3 variable selection methods: Boruta's algorithm, the top 200 genome-wide association study markers according to their respective P values, and an elastic net regression. Ten different algorithms were chosen for the classification tests. A predictive panel was built on the basis of joint scores between the classification algorithms.

Results: Two variable selection methods, Boruta and genome-wide association studies, were statistically similar in terms of the average accuracies generated, whereas elastic net had the worst overall performance. The predictive genetic panel was completed with 155 single-nucleotide variants, with 91.18% accuracy, 92.75% sensitivity, and 89.55% specificity using the support vector machine algorithm. The markers used range from known single-nucleotide variants to those not previously described in the literature. Our study shows potential in creating genetic prediction panels with tailored penalties per marker, aiding in the identification of optimal machine learning methods for intricate results.

Conclusions: This method is able to classify asthma and nonasthma effectively, proving its potential utility in clinical prediction and diagnosis.

Keywords: Asthma; genetics; machine learning; prediction; single-nucleotide variants.

Fig 1

Cumulative sequential analysis, adding one feature at a time from highest to lowest importance score. Even for algorithms with great potential, dozens of markers are necessary for predictive accuracy to be powerful in asthma outcomes. Penalized and nonpenalized systems are similar, but in some cases, such as K-nearest neighbor and C5.0, prediction becomes satisfactory with fewer markers. The vertical lines in each graph represent the first appearance of 80% predictive accuracy (black lines = original data; red lines = penalized scores). Horizontal lines: dotted = 80%, dashed = 85%, filled = 90%.

Fig 2

ROC curves of the genetic panel predictions. Compared with other algorithms, the ROC curve and AUC generated by the SVM model were superior. AUC, Area under the curve; KNN, K-nearest neighbor; ROC, receiver-operating characteristic.