An Ensemble Approach Based on Multi-Source Information to Predict Drug-MiRNA Associations via Convolutional Neural Networks

K Deepthi,A S Jereesh

doi:10.1109/access.2021.3063885

Abstract

Increasing evidence from recent research demonstrates that aberrant expressions of microRNAs (miRNAs) are linked to the development of chronic human diseases. Targeting miRNAs with bioactive small-molecules (or drugs) to regulate their activities provide an innovative insight into human disease treatment. Identifying the drugs that target particular miRNAs through the experimental study is complicated, time-consuming, and tremendously expensive. Therefore, computational researches by integrating information on drugs and miRNAs are essential for discovering potential drug-miRNA associations. Realizing the appropriate drugs that target the causal miRNAs behind diseases will contribute to miRNA mediated disease therapeutics and drug clinical applications. This study proposes an ensemble learning approach, ELDMA, that predicts novel drug-miRNA associations based on deep architecture-based classification. The method constructed features based on the integrated pairwise similarities of drugs and miRNAs and reduced the feature dimensions with principal component analysis (PCA). With the resulting features, the convolutional neural network is trained to extract intricate, high-level patterns. The deep retrieved features are given to the support vector machine classifier to infer potential drug-miRNA associations. We conducted global leave-one-out cross-validation (LOOCV), drug-fixed local LOOCV, miRNA-fixed local LOOCV, and 5-fold cross-validation to evaluate the model performance. ELMDA achieved corresponding AUCs of 0.9862, 0.7426, 0.9847 and 0.9928 for Dataset 1 and AUCs of 0.8643, 0.6742, 0.8671 and 0.8521 for Dataset 2, respectively. The results and case studies illustrate the effectiveness of ELDMA in identifying novel drug-miRNA candidates. The top predicted relationships are released for future wet-lab studies.

Highlights

MiRNAs are a novel class of non-coding RNAs involved in the process of RNA silencing and post-transcriptional regulation of gene expression [1], [2]
In this work, we proposed an architecture for identifying novel drug-miRNA relationships that consist of 3 main segments: principal component analysis (PCA), Convolutional neural network (CNN), and support vector machines (SVM)
In the cross-validation results, verified associations with predicted probabilities above the threshold were considered as True Positives (TP), and below the threshold were considered as False Negatives (FN)

Summary

Introduction

MiRNAs are a novel class of non-coding RNAs involved in the process of RNA silencing and post-transcriptional regulation of gene expression [1], [2]. They are involved in many basic functions for the growth and development of living organisms. MiRNAs act as regulators of various cellular pathways [3], and they achieve their function by binding with the complementary sequences of mRNA molecules. Many recent experiments disclose that miRNAs play essential roles as biomarkers and treatment. Abnormal miRNA expressions are correlated with complex diseases, including cancer, Parkinson’s, and immune-related diseases [7], [8]

Methods

Results

Conclusion