Mammalian MicroRNA Prediction through a Support Vector Machine Model of Sequence and Structure

Ying Sheng, Pär G. Engström, Boris Lenhard · 2007

Read Mammalian MicroRNA Prediction through a Support Vector Machine Model of Sequence and Structure by Ying Sheng, Pär G. Engström, Boris Lenhard (2007) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/plos-pone-0000946.txt.

Introduction

Ying Sheng, Par Engstrom and Boris Lenhard built a computer model to predict microRNA genes in mammals, published in PLOS ONE in 2007. MicroRNAs are tiny RNA pieces, about twenty-two letters long, that tune gene activity. They can order a message destroyed or block its use, so finding them matters.

Hundreds were already known in mouse and human, but more were surely missed. The authors train a support vector machine, a pattern learning program, on both sequence and folded shape. Shape matters because microRNAs form a distinct hairpin. A better predictor hands biologists a fuller parts list for gene control.

Catalog record

CREATORYing Sheng, Pär G. Engström, Boris Lenhard
YEAR2007
COLLECTIONPAPERS
FORMATTXT
FILE SIZE65 KB
LICENSECreative Commons Attribution License
RIGHTSCC BY
SOURCEPLOS ONE

Source: https://journals.plos.org/plosone/article?id=10.1371/journal.pone.0000946

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Mammalian MicroRNA Prediction through a Support Vector Machine Model of Sequence and Structure Authors: Ying Sheng, Pär G. Engström, Boris Lenhard Published: 2007, PLOS ONE DOI: 10.1371/journal.pone.0000946 ABSTRACT MicroRNAs (miRNAs) are endogenous small noncoding RNA gene products, on average 22 nt long, found in a wide variety of organisms. They play important regulatory roles by targeting mRNAs for degradation or translational repression. There are 377 known mouse miRNAs and 475 known human miRNAs in the May 2007 release of the miRBase database, the majority of which are conserved between the two species. A number of recent reports imply that it is likely that many mammalian miRNAs remain to be discovered. The possibility that there are more of them expressed at lower levels or in more specialized expression contexts calls for the exploitation of genome sequence information to accelerate their discovery. In this article, we describe a computational method-mirCoS-that uses three support vector machine models sequentially to discover new miRNA candidates in mammalian genomes based on sequence, secondary structure, and conservation. mirCoS can efficiently detect the majority of known miRNAs and predicts an extensive set of hairpin structures based on human-mouse comparisons. In total, 3476 mouse candidates and 3441 human candidates were found. These hairpins are more similar to known miRNAs than to negative controls in several aspects not considered by the prediction algorithm. A significant fraction of predictions is supported by existing expression evidence. Using a novel approach, mirCoS performs comparably to or better than existing miRNA prediction methods, and contributes a significant number of new candidate miRNAs for experimental verification. INTRODUCTION MicroRNAs (miRNAs) are an abundant class of ∼22 nt long endogenous non-protein-coding RNAs that function by binding to target sites on 3′-UTRs of messenger RNAs (mRNAs) to repress translation or mediate mRNA degradation (reviewed in [1] ). Mature miRNAs are synthesized from longer 70–100 nt precursors (pre-miRNAs), each of which forms a hairpin structure that contains one or two mature miRNAs in either or both of its arms. Thus far, a total of 4584 miRNAs have been identified and reported to miRBase (release 9.2, May 2007), a database that stores experimentally validated miRNAs and their homologs [2] , [3] . Several recent observations indicate that miRNAs, in general, may be essential for organisms to differentiate into multiple cell- and tissue types and/or to keep cells in a particular differentiation state [4] . miRNA target prediction in mammals indicates that ∼10–30% of protein-coding genes may be under control of currently known miRNAs [5] , [6] . This number may still increase because many additional miRNAs have been predicted. There are currently 475 human miRNAs and 377 mouse miRNAs in miRBase, but recent studies have suggested that the number of miRNAs in a vertebrate genome can be as many as 800–1000 [7] , [8] . The high number of miRNA genes, their diverse expression patterns [9] – [12] and the abundance of potential miRNA targets suggest that miRNAs are likely to be involved in a broad spectrum of human diseases. Indeed, components required for miRNA procession and/or function have been implicated in fragile X mental retardation [13] , DiGeorge syndrome [14] , and cancer [15] . Lu et al. [16] demonstrated recently that miRNAs can indeed be developed into potent cancer markers. As miRNAs are likely to play a central role in development and also in disease, it is important to understand their function. An important step towards this would be to assemble a complete catalogue of miRNA genes. Experimental cloning efforts have successfully identified highly expressed miRNAs from various tissues [9] , [11] , [17] – [32] . However, cloning methods are highly biased towards miRNAs that are abundantly and/or ubiquitously expressed. On the other hand, computational prediction of miRNAs could become a powerful aid for finding tissue-specific or lowly expressed miRNAs. A number of computational methods for miRNA prediction have been described and appear to complement each other because they take different approaches to miRNA prediction (reviewed in [33] ). Support vector machines (SVMs) are machine learning algorithms widely used to solve classification problems. A SVM assigns an object to one of several classes based on a set of input features associated with the object. In bioinformatics, superiority of SVMs over other classification methods has been shown for prediction of DNA-binding proteins [34] , gene function [35] and protein subcellular localization [36] . Here we describe a method-mirCoS-to predict conserved miRNAs in mammalian genomes. Being fundamental functional RNA elements, many pre-miRNAs have maintained sequence and secondary structure conservation across large evolutionary distances. Conservation can be described by several features which can serve as inputs for a SVM model. Currently, five other SVM-based miRNA prediction algorithms are available. Four of them are aimed at non-conserved miRNA prediction [37] – [40] , and therefore they can not achieve sufficient specificity when applied to entire large genomes. The fifth is aimed at annotating the results from non-coding RNA prediction [41] , and has better specificity, predicting about 5000 miRNAs in the human genome. However, this number is still about five times higher than recent estimates [7] , [8] and may include many false positives. To improve on this, we have added a number of previously unused features, and built a composite SVM model consisting of three sequentially applied SVMs and applied it to the human and mouse genomes. We predict about 3400 human-mouse conserved candidate miRNA genes, many of which show evidence of sequence and secondary structure conservation across all vertebrates (including fish) and are supported by independent evidence not used by our prediction method. Finally, we show that mirCoS performs better than or comparably to other recent methods. Because the great majority of our predictions are novel, the method described here constitutes a worthy addition to the arsenal of computational methods aimed at completing the mammalian miRNA collection. RESULTS AND DISCUSSION SVMs classify objects based on a set of features for each object. With the goal of predicting mature miRNAs, we selected features describing three aspects of precursor and mature miRNAs: (1) sequence conservation of pre-miRNA, (2) secondary structure of pre-miRNA and its conservation, and (3) placement and secondary structure of mature miRNA within its pre-miRNA ( Table 1 ). We chose to train one SVM (SVM1) for the first of these feature sets, another (SVM2) for the second, and two SVMs (one for human and one for mouse data, collectively termed SVM3) for the third feature set. We applied the SVMs sequentially, so that only candidates that were classified as positive by SVM1 were passed on to SVM2, and similarly for SVM2 and SVM3 ( Figure 1 ). The main reason for dividing the prediction task over three SVMs was to reduce running time: the second and third feature sets require secondary structure predictions that are expensive to compute for whole genomes, and the number of objects to classify increases almost 26-fold for the third feature set, because there are many putative miRNA positions within each pre-miRNA hairpin. Additionally, the sieve effect of sequential application of three SVMs aided in increasing the specificity of predictions, which we consider more important than sensitivity at this stage of the search for unknown candidate miRNAs and their selection for experimental validation. Two

Read the complete text