An effective sequence structure representation for long non-coding RNA identification and cancer association using machine learning methods

Abstract

The invent of high-throughput technologies and consequent developments in Bioinformatics research unveiled many important non-coding transcript molecules such as Long non-coding RNAs (lncRNAs). The available studies confirmed that lncRNAs play important genetic and epigenetic roles in higher-order species like the human and their differential expressions leads to complex diseases like cancer. Even though there are arrays of studies and related tools for the analysis, less conserved patterns in the sequences and intractable structural properties challenge the understanding of varying functionalities of lncRNAs. For the better approximation of these characteristics, higher quality feature representation is required. This paper proposes an extended hybrid sequence-structure feature set for machine learning based lncRNA analysis. Here, the sequence features are derived from various frequencies of k -mer patterns, GC content and molecular weight. The structure representations consider the context of different secondary structure elements which include stems, interior loops, multi-loops and hairpin loops. These features are used for the classification of lncRNA/mRNA and cancerous/non-cancerous lncRNAs. The classifications use machine learning algorithms such as LDA based topic model, Random Forest, SVM and Naïve Bayes. The results show that the proposed feature set is effective in classifying lncRNAs and provide a direction towards the analysis of the role of secondary structure elements in cancer-related lncRNAs.