miRmap web: comprehensive microRNA target prediction online

1MicroRNAs (miRNAs) are short (∼22 nt) noncoding RNAs that guide the RNA-induced silencing complex (RISC) to posttranscriptionally repress the expression of protein-coding genes by binding to targeted messenger RNAs (mRNAs). While the detailed mechanism of this guidance is not yet resolved, binding between the mRNA and the miRNA from position 2–7 (or 8) from the 5′ end usually triggers RISC-mediated repression. In the miRNA–mRNA binding map based on Ago HITS-CLIP (HIgh-Throughput Sequencing of RNA isolated by CrossLinking ImmunoPrecipitation), 73% of the studied target sites have a seed match. While searching for seed matches in 3′-UTR is a widely used feature for miRNA target prediction, it yields a very high number of potential targets. For example, human 3′-UTRs have 9.5 million potential targets defined as 7-mer seed matches of all human miRNAs, which means that on average a miRNA potentially targets ∼25% of the human genes. While this estimation is probably an upper bound, and as in a specific cell or tissue, not all miRNAs and all genes are expressed, thereby reducing the number of potential targets, prioritization of predictions is necessary to experimentally test the most probable targets first. We recently proposed a ranking function of miRNA-mediated repression strength parameterized from experimentally measured effects on mRNA or protein levels resulting in prediction of biologically meaningful potential targets.

2Implemented as an open-source Python software library, the miRmap library implements a comprehensive range of features using thermodynamic, probabilistic, evolutionary and sequence-based approaches to rank potential miRNA targets using information beyond the seed match. We compared the power of individual approaches to predict the repression strength of miRNA–mRNA pairs, assessed using data from transcriptomics, immunopurification, proteomics and polysome fractionation high-throughput experiments, and combined the features with a linear model to predict the miRNA repression strength as the ‘miRmap score’. This score doubles the performance of the TargetScan context score measured as the proportion of explained variance of a transcriptomics data set. Moreover, our comprehensive set of features covers features included in many other tools (Figure 1). For example, PITA ranks targets only with ‘ΔG total’, and PACMIT uses a combination of ‘ΔG open’ and ‘P.over binomial’. As the miRmap score is more predictive than these individual features, miRmap can be used as a reference tool. Here, we present ‘miRmap web’, a web application based on a public REST (REpresentational State Transfer) service. With miRmap web, biologists can easily, through a user-friendly and feature-rich graphical interface, browse precomputed miRmap predictions from model organisms, and also predict and rank miRNA targets on their own sequences.

3Thermodynamic features of miRmap, which we described previously, include ‘ΔG duplex’ and ‘ΔG binding’ that evaluate the energy of the miRNA–mRNA duplex, ‘ΔG open’ that evaluates the energy necessary to unfold the 3′-UTR and make the area accessible for RISC binding and ‘ΔG total’ [also named ‘ΔΔG’ by Kertesz et al.] that sum ‘ΔG duplex’ and ‘ΔG open’. Since the publication of the miRmap library, we have added new features, ‘ΔG seed duplex’ and ‘ΔG seed binding’, that specifically evaluate the seed region. Our model explains 12.7% (R2) of variance on the Grimson et al. data set, whereas the TargetScan context score (‘AU content’, ‘3′ pairing’ and ‘UTR position’) explains 7.49% with the same type of linear model. With the addition of the ‘ΔG seed duplex’ and ‘ΔG seed binding’ features, our model now explains 14.4% of the variance (Figure 1). It is worth nothing that TargetScan predictions also include an independent filter on conservation, named Pct. This score, based on the ‘Branch Length Score (BLS)’ feature, explains 4.88% of the variance. As described in our previous publication, fold-change distribution confirmed greater enrichment of stronger experimental repression with greater miRmap score (Supplementary Figure S1). This analysis was also applied specifically to 6-mer seeds and also showed increased performance (Supplementary Figure S2). The sequence-based features are, as described by the TargetScan context score, as follows: ‘AU content’ that is a weighted count of A and U nucleotides around the seed match, ‘UTR position’ that takes into account the enhanced repression for sites that are close to one end of the 3′-UTR and ‘3′ pairing’ that evaluates the contribution of the pairing on the opposite side to the miRNA seed. As only limited regions of 3′-UTRs have regulatory or structural roles, we introduced features in miRmap with probabilistic and evolutionary points of view. In the probabilistic approach, we modeled the 3′-UTR sequence composition with a Markov process (order 1) and determined the expected probability of finding at least n occurrences of the seed match, either with a binomial approximation (‘P.over binomial’) or with an exact solution (‘P.over exact’). The evolutionary approach assesses the conservation of the target sites compared with the rest of the 3′-UTR as measured by the ‘BLS’ and ‘PhyloP’ features. The ‘BLS’ evaluates the evolutionary time during which the seed match was present in a given set of species, while the ‘PhyloP’ feature detects negative selection pressure on the target sites within a statistical framework. MiRmap brings together common features of miRNA target prediction and extends these with novel features including ‘ΔG binding’, ‘ΔG seed binding’, ‘P.over exact’ and ‘phyloP’ to provide the first truly comprehensive open-source library for target predictions.

4As well as allowing predictions for user-submitted sequences, miRmap includes predictions for the annotated miRNAs [miRBase 19] and genes [Ensembl 69] of eight species: Human, Chimpanzee, Mouse, Rat, Cow, Chicken, Zebrafish and Opossum.

5As our service allows both interrogating a database and computing predictions, we developed a framework that ensures a reliable and responsive service. All computations on the server side are performed asynchronously: multiple users can search and compute predictions at the same time without impacting other users, with our web server computing capability as the only limit. This implementation was written using the Tornado framework (tornadoweb.org) and deployed using Nginx (nginx.org) reverse proxy.

6MiRmap web can be used as a web application or as a REST service. The web application is a single web page for ‘click&play’ usage integrating the form for the user query and the dynamically updated results table. Users can also easily retrieve prediction data sets or predict targets in a more programmatic way using the REST service. We used the widely used ExtJS Javascript toolkit (sencha.com) to build our web application and obtain a homogeneous look and feel as well as flexible configurability. A detailed documentation of the REST service, as well as examples, is included in help section of miRmap Web site.

7MiRmap web includes facilities for both browsing precomputed miRNA target predictions, and for online computing of miRNA targets on user-submitted sequences. The first step is to select a species (Figure 2). The user may then ‘Select’ or input the ‘Sequence’ of a miRNA and/or a protein-coding gene. It is important to note that for browsing precomputed predictions, users may select a miRNA or a gene, or both, with an auto-complete capability to facilitate rapid selections The gene selection and sequence input options may be combined such that the user may input a miRNA sequence and select a gene/transcript by name, or vice versa. When a gene name or identifier is entered, the canonical transcript of that gene is selected; it is also possible to select a specific transcript by directly entering the transcript identifier. When the form is valid, the user may then submit his/her query by clicking on the ‘Get targets’ button.

8Results from user-submitted queries return feature values for each miRNA–mRNA pair by default (the ‘Pair’ view), where each record is expandable to access additional details about each predicted target site, or for each target site by selecting the ‘Site’ view. By default, only a selection of the complete set of all miRmap features is displayed, and intuitive options from the toolbar allow users to display all features or customize their feature selections. Additional toolbar options enable (i) switching between ‘Percentile’ and ‘Raw’ score formats, (ii) multicolumn ordered sorting and (iii) flexible filtering of the results table to suit the user’s requirements. Raw miRmap scores for each feature, e.g. ‘ΔG duplex’ in kcal/mol and ‘P.over exact’ as a probability, can be converted to percentiles to simplify and standardize a scale from strong to weak repression. This is achieved by ranking targets for each species from the weakest to the strongest predicted repression for each individual feature and assigning scores to percentiles from 0 to 100%, with 100 representing the strongest repression. Users therefore do not have to be familiar with the ranges of different feature values, nor their directions, i.e. whether larger or smaller values correspond to higher or lower repression. Results are sorted by default according to their miRmap scores, but users may define their own sorting functions on each selected feature through a simple ‘drag&drop’ ordering of each column in the results table. In addition, filters may be applied to limit results to only canonical transcripts for each gene as well as by setting minimum or maximum scores for each feature using sliding selectors. These customized results tables may then be easily exported as CSV (comma-separated values) formatted files and saved for viewing in common spreadsheet applications such as Excel, or as BED files for viewing in the UCSC genome browser.

9With this integrated web interface, featuring easy filtering and sorting options, miRmap web aims to provide biologists interested in a particular miRNA or gene the means to prioritize their research with confident ranking of miRNA targets using the miRmap model. In contrast to most other miRNA target prediction web interfaces, miRmap web covers a wide range of possible usage by providing both precomputed and online predictions. Updated 3′-UTR annotations and newly discovered miRNAs can therefore be easily computationally tested with a state-of-the-art analytical tool.

10Supplementary Data are available at NAR Online: Supplementary Figures 1–2.