MicroRNAs (miRNAs) are short, ~22 nt non-coding RNAs that control gene expression. miRNAs bind to the 3'UTRs of their regulated mRNA transcripts to facilitate mRNA degradation or translation inhibition [1, 2]. miRNAs are present in most metazoans and are thought to regulate a diverse range of biological processes [3, 4]. miRNA genes' evolution is unique since they mostly emerge through duplication events [5]; exhibit most often unidirectional evolution [6]; are generally gained and not lost [7]; show several cases of rapid evolution in primates [8, 9]; are rarely changed due to functional constraints [10]; and, show relatively rare evolutionary acquisition events (accounted by their relatively small numbers).

miRNA predictions based on computational methods, which followed initial extensive cloning efforts, are based on the secondary structure of the miRNA, its phylogenetic conservation and thermodynamic stability [11, 12]. miRNA gene repositories are constantly expanding giving rise to more than 3500 reported miRNAs in more than 30 animal species (Sanger miRbase database, Version 10.0 [13, 14]). However, it is apparent that even this comprehensive repository is far from completion, accounted by the very few miRNAs listed for dog (6) and chimpanzee (83), compared to human (533), to name two examples. Since these differences cannot be accounted merely by species-specific miRNAs, we saw the need for a computational tool for miRNA homologous searches.

We present miRNAminer, a tool for automatic identification of homolog miRNAs based on a given user defined query miRNA. The tool exploits numerous characteristics of miRNAs: high conservation of precursor sequences, very high conservation of mature sequences (particularly in the seed region, nt 2–8 [15]), and hairpin secondary structure with high folding energy and base pairing. miRNAminer first uses BLAST [16] to select candidate matches and ranks them according to their e-values. Then it employs a series of rigorous filters to improve specificity.

An input query consists of a precursor miRNA, mature miRNA, a set of filter threshold values and the number of best-fitted results requested in the output. We designed miRNAminer's algorithm to maximize specificity of matches. This is because the designed application of miRNAminer is to identify homolog matches after a miRNA has been experimentally confirmed. We estimated the default values presented below so that each filter by itself selects 95% of known miRNAs in training genomes (criteria was also based on [17]).

miRNAminer's algorithm follows these steps: (i) Use BLAST [16] to find matches in target genomes (the whole precursor miRNA from the query is used); (ii) Filter with e-value threshold (default 0.05 per chromosome); (iii) Extend the match by adding flanking nucleotides (default 50) up- and down-stream from the match (Ensembl genome database; [18]). Examine all possible extensions of the match within threshold length (default min 70 nt, max 180 nt); (iv) Filter with RNA secondary folding energy threshold (default -25 kcal/mole; RNAfold with options "-p -d2 -noLP" [19]); (v) Filter with minimal base-pairing threshold (default 55% pairing; with 20 gap penalty and 0.5 extension penalty); (vi) Filter with requirement for hairpin-shape secondary structure; (vii) Filter with alignment of precursor sequences (default 56% identity); (viii) Filter with alignment of mature miRNA sequences (default 80% identity); (ix) Filter with maximum number of mismatches in mature miRNA sequences (default 3 nt); (x) Filter with conservation of seed (2–8 nt, required 100% conservation [15]); (xi) Filter with position of mature miRNA on the hairpin (max 4 nt overlap of mature sequence and hairpin loop). miRNAminer's output includes detailed analysis of the identified genomic region(s) that passed the selected threshold criteria (Figure 1). Currently, miRNAminer supports searches in 10 metazoan genomes. We will regularly add additional genomes upon their release. After the query is issued, results are usually available within a minute (though this depends on the number of results requested) and can either be viewed on the screen or requested to be sent by email.

We compared our predicted miRNAs (in the human species; Table 2) to other prediction methods. We found that 18 and 36% of our miRNAs are contained within RNAmicro [23, 5] and Berezikov 2005 [24] databases, respectively, out of almost 3000 and 1000 miRNAs in each set, respectively. The overlap is not extensive however even when miRNAs derived from algorithms using very similar search parameters are compared only about 50% overlap is seen [25]. One of our identified miRNAs, which is not reported by any other study, was recently identified experimentally ([26] present in miRbase version 10.0 and not 9.0) increasing confidence in our unique miRNAs. Notably, even though miRNAs from all species were used to search for human homologs, the candidate miRNAs discovered are homologs to genes in two species only, M. musculus and R. norvegicus, indicating better miRbase coverage for mouse and rat than for other species. Two examples of the non-registered human homolog miRNA genes are presented in Figure 2. miRbase mouse miR-764 sequence, which has no known registered homologs was used as input for miRNAminer search (with default parameters). The output reported a homolog (presumably hsa-miR-764; see Table 2), which is located in the second intron of Human serotonin receptor 2C (HTR2C; NM 000868; Figure 2A). The mouse miRNA homolog is located in an intron of the same gene (HTR2C; NM 008312) suggesting an evolutionary conserved co-expression of miRNA and its host gene [27–29]. Second in the list of non-registered human homolog miRNA genes (sorted according to BLAST e-values) is miR-763. This miRNA spans the longest complementary sequence out of the list, has the lowest RNA folding energy (Table 2) and shows high conservation between many species (Figure 2C). Interestingly, human miR-763 is harboured in an intron of the high mobility group AT-hook 2 oncogene (HMGA2; for review see [30, 31]). Recent disrupted interplay between miRNAs and HMGA2 showed an increase in oncogenesis [32–34]. To regulate their targets, miRNAs bind to 'seed' regions in the 3'UTR, typically 6–7 nt long (nt 2 to nt 7 or 8 of the miRNA [15, 35, 36]; also see [37]). miR-763, possibly also co-expressed with its host gene [27–29], has a conserved binding site for its own harboured miRNA (nt 2 to 8 of the miRNA binds position 2192 in the 3'UTR which is conserved in human/mouse/rat). It is tempting to speculate a negative feedback regulatory role of newly identified human miR-763 and its oncogenic host when co-expressed in the same spatio-temporal context. To this end data from Expressed Sequence Tags (ESTs) supports this possibility (Figure 2C). Other identified miRNAs presented in Table 2 show high species conservation (for example, miR-670) or are located in exons (for example, miR-711) or exon-intron junctions (for example, miR-762). Interestingly, in a recent study involving deep sequencing, four of our human predicted miRNAs were confirmed (miR-760, 708, 543, and 665 [38], available in miRbase version 10.0). To conclusively confirm the presence of the identified candidates in the studied species, an experimental verification is required. However, the candidates identified by our method are close homologs to known miRNAs and as such are not required to meet as stringent criteria to be annotated as novel miRNAs [17]. In this study we looked at homolog genes which are genes related to each other by descent from a common ancestral DNA sequence. We do not segregate between orthologs, genes in different species that evolved from a common ancestral gene by speciation, and paralogs, genes separated by the event of genetic duplication. We cannot also rule out that similar miRNAs in different species have developed independently [39]. Our tool, which is based on evolutionary conservation, can only detect evolutionarily conserved miRNA genes. We are currently improving our algorithm to include multiple alignments of vertebrate miRNA sequences in order to better refine the boundaries of the miRNA precursor sequence.

For searches with relaxed parameters (reduced stringency) we suggest initially performing the following modifications: (i) do not 'Require seed conservation in mature miRNA (nt 2–8)' (uncheck box); (ii) increase 'maximal number of gaps in miRNA precursor alignment' from 10 (default) to 15; (iii) decrease 'minimal mature miRNA identity' from 0.8 (default) to 0.7; (iv) decrease 'minimal base pairing percentage in miRNA precursor' from 55 (default) to 40; and (v) change 'minimal/maximal length of precursor sequence (nt)' from 70/180 (default) to 50/250. In order to view miRNAs which are other than the top candidate we suggest increasing the 'number of results to report' from 1 (default) to 5. The parameters (i–v) above are listed in the order that would output an increasing total number of identified miRNAs. For example, reducing mature miRNA identity from 0.8 (default) to 0.7 increases miRNAs from 22 to 24 (9%) and 31 to 36 (16%) in human and mouse, respectively. On the other hand, we found that changing the length of the miRNA precursor from 70–180 nt (default) to 50–250 nt, added only 1 additional miRNA in human and none in mouse. This, however, might change when run in combination with other modified parameters. Altogether each of the modified parameters listed above will result, independently, in an average miRNA increase of 11% when tested on seven mammalian species.

Several approaches to identify miRNA homologs have been previously described, both in plants [40], and in animals [5, 41, 42]. However, the only tool that is available as a web service, microHARVESTER [40], is targeted for plants. miRNAminer is the first available miRNA gene homolog search tool for animal genomes.

Project name: miRNAminer; Project home page: http://pag.csail.mit.edu/mirnaminer; Operating system: Platform independent; Programming language: Java; License: Open source, see http://opensource.org/licenses/mit-licence.php; Code is available upon request. miRNAs identified using miRNAminer will be incorporated in next miRbase versions, see http://microrna.sanger.ac.uk.