Semi-supervised fuzzy-rough extreme learning machine for classification of cancer from microRNA.

The miRNA is a tiny, single-stranded RNA of nearly 22 nucleotides long that is transcribed from DNA and controls the genes in protein synthesis process. As expression levels of miRNAs vary significantly between cancerous and non-cancerous cells. Therefore, miRNAs can play a vital role in the development of cancer. Cancer classification from miRNA gene expression data is an area of interest to the bioinformatics and computational biology researches in recent past. However, traditional machine learning classifiers often do not produce the desired results due to lack of training patterns present in miRNA data. Additionally they are not often able to deal with overlapping, vague, uncertain, ambiguous and indiscernible cancer subtypes classes of the miRNA data. Motivated from the above said issues, we propose a novel semi-supervised fuzzy-rough based extreme learning machine (SSFRELM) method. Two stages are involved in the proposed method, where in the first stage, the soft class labels of the unlabeled samples are assigned using fuzzy-rough set theory, those are subsequently applied on the semi-supervised extreme learning machine in the second stage. The proposed SSFRELM method can improve classification accuracy since it utilizes unlabeled patterns together with limited labeled patterns in the learning process to develop the semi-supervised extreme learning machine and can handle the overlapping, vague, uncertainty, ambiguity, and indiscernibility usually present in miRNA gene expression data as it uses fuzzy-rough set theory. The proposed method is assessed using eight publicly available miRNA gene expression datasets from five different cancer types with respect to various classification evaluation measures in comparison to four other state-of-the-art methods. Experimental results justify that the proposed method outperformed the other counter-parts methods for cancer pattern classification. Paired t-test results also confirm the statistical significance of the better results achieved by the proposed SSFRELM over other methods for most of the cases. [ABSTRACT FROM AUTHOR]