Identifying microRNAs involved in cancer pathway using support vector machines

Highlights

• Construction of a two-step SVM classifier for identifying miRNA associated with cancer.

• Features are extracted from sequence, thermodynamics and miRNA–mRNA hybridization interactions based on experimentally data.

• For miRSEQ – Positions 1, 6, 10, 19, GG and CC repeat in the miRNA sequence form the optimal feature subset.

• Optimal features vary significantly based on the number of seed formed by hybrid for miRINT.

• Final classifier obtained a good performance with cv-rate ranging from 92 to 87.

Abstract

Since Ambros’ discovery of small non-protein coding RNAs in the early 1990s, the past two decades have seen an upsurge in the number of reports of predicted microRNAs (miR), which have been implicated in various functions. The correlation of miRs with cancer has spurred the usage of this class of non-coding RNAs in various cancer therapies, although most of them are at trial stages. However, the experimental identification of a miR to be associated with cancer is still an elaborate, time-consuming process. To aid this process of miR association, we undertook an in-silico study involving the identification of global signatures in experimentally validated microRNAs associated with cancer. Subsequently, a support vector machine based two-step binary classifier system has been trained and modeled from the features extracted from the above study. A total of 60 distinguishing features were selected and ranked to form the feature set for classification – 26 of these extracted from the miR sequence itself, and the remainder from the thermodynamics of folding and the hybridized miRNA–mRNA structure. The two step classifier model – miRSEQ and miRINT had reasonably good performance measures with fairly high values of Matthew’s correlation coefficient (MCC) values ranging from 0.72 to 0.82 (availability: https://sites.google.com/site/sumitslab/tools).

Introduction

miRNA (miR) are small non-coding, single stranded RNAs (about 22 nucleotides in length) involved in several regulatory pathways in the cell cycle. They bind to the untranslated regions (UTRs) of mRNA, (particularly the 3'UTR) and play an important role in the post-transcriptional regulation of gene expression (Bartel, 2004, Filipowicz et al., 2008). Recent studies suggest that these noncoding RNAs can bind to 5'UTRs (Ragan et al., 2009) and coding regions (Hausser et al., 2013) of mRNA as well, but little is known about the mechanism of binding and their regulation. Binding of a miR to a specific target in an UTR with complete complementarity either leads to degradation of the mRNA itself or induce translational repression (Esquela-Kerscher and Slack, 2006). In tissues associated with various tumors, it has been observed that the expression pattern of miRs is altered considerably (Cummins et al., 2006, Zhang et al., 2006). Additionally, gene mapping reveals that most of the human miRs are located in chromosomal positions which are susceptible to rearrangements (Calin and Croce, 2007). Hence, it can be asserted that miRs in humans play a major role in the cancer pathway.

Previous studies by several authors have investigated the involvement of different types of base pairing in miR–mRNA interactions and target prediction algorithms have been formulated based on these precincts. These algorithms predominantly considered Watson Crick base pairing between the miR and its respective mRNA – especially with the 2nd to the 8th nucleotide positions of miR – as the potential target sites. However, in later studies, it was found that animal miRs do not bind to mRNA with perfect complementarity (unlike in plants); rather their binding leaves several imperfections like loops, mismatches or bulges and often involves GU(non-Watson Crick) base pairing as well (Axtell et al., 2011, Didiano and Hobert, 2008). Other than these determinants, AU richness around the seed regions and folding of mRNA play a vital role in target binding (Grimson et al., 2007, Robins et al., 2005). All these factors need to be considered, not in isolation but together to hypothesize miR:mRNA interactions.

Some of the computational methods used in the functional annotation of miRs involved in cancer mainly rely on the expression profile of various cancer cell types and statistical analysis for further classification (Jayaswal et al., 2011). These methods utilize the expression profile but they fail to consider the fact that a single miR can bind to several mRNA target sites and regulate the cell differently. Our aim at feature selection was, therefore, to embrace all these redundancy checks. Other attempts to classify miRs into oncogenes and tumor suppressor genes (TSGs) were based on functional and evolutionary features (Wang et al., 2010) like conservation, expression levels, chromosome distribution, etc.

The present study involved a search and analysis of features involved in the interaction of a miR:mRNA associated with cancer. These features encompassed sequential, hybridization and thermodynamics of validated miR:mRNA interactions only. Based on the curated and prioritized features, we developed a two-step machine based classifier model – miRSEQ and miRINT, which will identify a miR to be associated with cancer and also classify the type of its association, i.e., either with an oncogene or a tumor suppressor. Prioritization of the features and a diversification of the models according to the number of seed regions drastically improved the performance of the classifier, as compared to generalized features and holistic hybridization. The incorporation of seed based classification in the determination of features is a novel approach in our algorithm. The final classifier thus developed had good performance with experimentally validated datasets giving good prediction accuracy (cross validation (cv-rate) ranging from 92% to 87%).