ASRmiRNA: Abiotic Stress-Responsive miRNA Prediction in Plants by Using Machine Learning Algorithms with Pseudo K-Tuple Nucleotide Compositional Features

Abstract

MicroRNAs (miRNAs) play a significant role in plant response to different abiotic stresses. Thus, identification of abiotic stress-responsive miRNAs holds immense importance in crop breeding programmes to develop cultivars resistant to abiotic stresses. In this study, we developed a machine learning-based computational method for prediction of miRNAs associated with abiotic stresses. Three types of datasets were used for prediction, i.e., miRNA, Pre-miRNA, and Pre-miRNA + miRNA. The pseudo K-tuple nucleotide compositional features were generated for each sequence to transform the sequence data into numeric feature vectors. Support vector machine (SVM) was employed for prediction. The area under receiver operating characteristics curve (auROC) of 70.21, 69.71, 77.94 and area under precision-recall curve (auPRC) of 69.96, 65.64, 77.32 percentages were obtained for miRNA, Pre-miRNA, and Pre-miRNA + miRNA datasets, respectively. Overall prediction accuracies for the independent test set were 62.33, 64.85, 69.21 percentages, respectively, for the three datasets. The SVM also achieved higher accuracy than other learning methods such as random forest, extreme gradient boosting, and adaptive boosting. To implement our method with ease, an online prediction server "ASRmiRNA" has been developed. The proposed approach is believed to supplement the existing effort for identification of abiotic stress-responsive miRNAs and Pre-miRNAs.

Keywords: abiotic stress; computational biology; machine learning; miRNAs; stress-responsive genes.

Figure 1

Feature selection for miRNA, Pre-miRNA, and Pre-miRNA + miRNA datasets. The optimal number of features were selected based on the higher accuracies in terms of auROC and auPRC. A total of 200, 250, and 500 features were selected for miRNA, Pre-miRNA, and Pre-miRNA + miRNA datasets, respectively.

Figure 2

Density graphs for the probabilities of prediction for different machine learning methods. It can be seen that most of the probabilities of prediction with SVM are higher than the random guess (0.5) as compared to random forest (RF), extreme gradient boosting (XGB), and adaptive boosting (ADB) methods. The XGB is observed to be the lowest performer among the considered methods. The variability in the prediction probabilities is lowest for the ADB and highest for the XGB methods.

Figure 3

Flow diagram showing the steps involved in the proposed approach for prediction of abiotic stress-responsive Pre-miRNAs and miRNAs.