An improved unsupervised learning approach for potential human microRNA–disease association inference using cluster knowledge

Abstract

Recently, microRNAs (miRNAs) have been shown to play significant roles in the progression of major human diseases. Identifying associations between human diseases and miRNAs using computational tools have attracted considerable attention. Experimental identification and validation of disease–miRNA associations are rare and time consuming. We propose a computational method for predicting associations to understand pathogenicity and uncover prognosis markers. Existing methods employed approaches based on network and machine learning to predict miRNA disease. However, these approaches do not consider the cluster information between miRNAs. Performance results of these methods’ predictions are not satisfactory due to more number of false positives and false negatives. The proposed methodology comprises of two phases: first, we use miRNA cluster information in addition to miRNA-functional similarity and disease semantic similarity. MiRNAs and diseases are placed in the same cluster by transforming the miRNA–disease association data matrix into a low-rank model using Principal Component Analysis (PCA). Second, the Adamic/Adar index was applied that computes the closeness of miRNAs and diseases based on shared neighbors, improving the prediction results. The problem of overestimation is resolved by incorporating similarity information about miRNA and disease. Case studies on Leukemia, Carcinoma, Glioma, Pancreatic Neoplasms, and Melanoma exhibit satisfactory performance with Area Under Curve (AUC) values ranging from 0.736 to 0.834. miRNAs predicted using the proposed method are ascertained to be matching the known associations found in the database of HMDD V3.2, dbDEMC V2.0, and PhenomiR V2.0.