The general approach is outlined in figure 1. Initially, a mRNA is compared with the genomic sequence to identify matching regions of 20–27 nucleotides with at most 2 mismatches (allowing 3 mismatches produced more than 10 000 matches per mRNA). These are called micromatches, and the genomic part is referred to as a genomic match. An average mRNA gives rise to about 1000 such micromatches, the vast majority (often all) of which we assume are spurious non-miRNA hits. However, it is possible, without comparing to other genomes, to filter the micromatches and achieve highly specific and fairly sensitive predictions of miRNA genes (Figure 1).

Six filters were used to identify a base set of genomic sequences as candidate miRNAs (with percentages of the initial micromatches that were remaining after each filter given in brackets): (1) they had high sequence complexity (26.9%); (2) they had no overlap with annotated exons on the same or the opposite strand (3.3%); (3) they had no overlap with repeat sequences defined by RepeatMasker (2.6%); (4) the putative miRNA:mRNA duplex should be relatively stable1721 with a calculated free energy of less than -34 kcal/mol (0.20%); (5) they had no more than identical 10 copies in the genome (0.19%), to eliminate repeated sequences not detected by standard repeat-masking; and (6) the miRNA was contained within a precursor structure that was similar to those observed in known Arabidopsis miRNA precursors, i.e. was predicted to be largely contained (at least 16 paired bases) within the stem of a double stranded stem-loop structure whose stem was predicted to have a free energy less than -60 kcal/mol, with at least 4 paired bases flanking the putative miRNA, and an intervening loop larger than 9 but less than 130 bases (0.0002%).

Although the base set predictions have a low number of false positives (see below), they can be even more refined to identify a subset of the predictions with extra confidence, because the probability of more than one mRNA matching a falsely predicted miRNA is minimal, unless the matching mRNA-targets are close homologs (in which case the multiple targets do not add much extra confidence). Most of the known miRNA in Arabidopsis are thought to have multiple targets often within the same family of homologous proteins22. If a known miRNA only has targets in a highly conserved protein family this filter can however be expected to falsely eliminate them.

In order to check the validity of our approach we took the mRNA targets of the known miRNAs and set out to see if using these as queries we would be able to correctly identify the known miRNA-genes. Of the 112 precursor sequences registered in RFAM (ver 5.1), we were able to map 92 perfectly to the current RefSeq assembly (TIGR ver 5.0) of the Arabidopsis genome; the remaining precursors were excluded from the positive control set. Likely targets for Arabidopsis miRNAs have previously been predicted allowing for up to 3 mismatches23. Repeating this procedure we find that our known miRNAs match 142 different annotated mRNA*. These are the positive control targets (refered to a 'known targets') and many have been experimentally confirmed2425. Initially, the 142 mRNAs in the positive control set yielded 359,976 micromatches after removal of low complexity sequences. However, the filtering procedure reduces this dramatically to 45 different loci (41 of which are already known) representing 16 different families (14 known). Assuming that the 'unknown' loci we find are false positives the procedure has 91% specificity and 45% sensitivity on the level of loci identified. Using the refinement step requiring more than one non-homologouos target only true positives are found, but at the expense of halving the sensitivity to 22%. The validity of the estimates of specificity and sensitivity is discussed below.

Applying the micromatcher procedure to all 28860 mRNAs annotated in Arabidopsis identifies 592 miRNA candidate loci (480 families) in the base set (Additional file 1). In the final step this is reduced to a set of 90 (70 new) when more than one non-homologouos target per miRNA is required. This is called the multi-target set and is a subset of the base set.

All miRNA gene predictions, their targets (with some basic annotation) and the predicted secondary structure of the precursor are available as supplementary data [Additional file 1], and at our website26

Using public databases we were able to acquire evidence for the expression of a small number of the predictions, 9 in the base set overlap with RNA molecules recently sequenced in a large scale cloning effort of Arabidopsis small RNA⁴, 109 have significant matches to Arabidopsis ESTs and 52 of the predicted precursors contain a 20-mer sequence tag from the Arabidopsis MPSS database27.

From an evolutionary point of view, it would seem to be a lot easier to adapt 20 bases in a miRNA for a new target than to evolve a protein for a specific regulatory task.

For mammals it has been suggested that the more targets a microRNA has the more likely it is to be conserved28 because of the additional constraints of having to match multiple targets.

Indeed also for plants: comparison of our predictions in Arabidopsis to two other plant species reveals that the more targets a miRNA is predicted to have, the more likely it is to be conserved (Figure 2). Although no Brassica species is yet completely sequenced and we had to use a conjunction of all single sequence Brassica entries from GenBank, significantly more of the predicted miRNAs are conserved in Brassica than in rice, indicating that many miRNA-genes have diverged beyond recognition since the divergence of monocots and dicots approximately 200 million years ago.

Thus, we speculate that the highly conserved miRNAs are likely to be central regulators, often of many target mRNAs (imposing the evolutionary constraint to stay conserved), and are more likely to be highly expressed. Whereas more recently evolved miRNA would have fewer targets, and a more localized spatiotemporal expression, making them less likely to be detected by cloning efforts.

Since evolutionary conservation is part of many of the previous discovery procedures, it is likely that the set of known miRNAs is biased towards those that are conserved, and our data suggest that in fact, miRNAs evolve fast and are less conserved than e.g. protein-coding genes.

It has been proposed that some miRNAs originate from inverted duplication of target sequences, exemplified by the single locus miRNAs miR-161 and miR-163, which have precursors that show extended homology to the target mRNAs also outside the mature miRNA sequence29. However, our structural filters require that the match between miRNA and target is in the range 20–25, effectively eliminating such miRNA with extended homology.

Of the predicted 592 precursors in the base set, 29 overlap with the 92 predictions made by Bonnet et al.30, and 4 of those by Wang et al.8. Thus, the different methods complement each other: The present method based on matching targets and miRNA is capable of finding non-conserved miRNAs, whereas the interspecies comparisons831 can find miRNAs without obvious targets.

The idea to use potential targets to find miRNA-genes has recently been employed in two other studies. Xie et al. 32 started by finding frequently occurring subsequences of human 3' UTR sequences conserved in other mammals and successfully searched the genome for new miRNA genes.

Moreover Adai and coworkers33 published results in Arabidopsis using potential targets to find new miRNA-genes. However, our approach differs significantly from theirs in the way the matches (that we term micromatches) are analysed and the kind of conclusions that can be drawn: Adai et al. looks for a 'cluster' of miRNA-genes that target the same sequence of a mRNA, and then aligns the candidates in such a cluster, scoring the alignment high if it shows a characteristic pattern where the miRNA and miRNA* are more conserved than the intertwining sequence. Thus, their method is limited to finding miRNAs that occur more than once in the genome, presumably as a result of duplication events. Moreover as a postfilter, Adai et al. require conservation in rice to generate their short-list used for experimental validation. Also, Adai et al. do not make any estimation of the specificity of their computational procedure and are consequently unable to speculate about the number of miRNAs.

In contrast our method is independent of whether a candidate has been duplicated in the genome or is conserved across species. Instead our aggressive filtering on the structural properties of the precursor enables us to make highly specific prediction (judging from the results using targets for known miRNAs as queries).

The multi-target miRNAs have a total of 528 different mRNA targets, which are involved in a variety of functions, but there is a notable over-representation of proteins with transcription factor activity and receptor binding activity as well as involvement in developmental processes (false discovery rate < 0.001, see Additional file 2). The predicted miRNA-genes are generally found scattered throughout the genome (Table 2). Unlike in mammals where 90 out of 232 miRNA-genes are within introns of protein coding genes 34, there is only one previously discovered Arabidopsis microRNA situated in an intron. This trend of plant microRNAs to be outside protein-coding genes also holds for our baseset predictions and even stronger for the multiple target predictions (Table 2).

Although estimating the sensitivity and specificity on the basis of the ability to correctly identify the small set of known miRNAs carries the danger of biasing, the presently most important concern must be not to massively overpredict new miRNA-genes. In constructing the filters we have therefore aim at high specificity at the expense of sensitivity. While false positives undoubtfully remain, the fact that the predictions share the properties of functional overrepresentation and bias of genomic location (properties not selected for in the filters) with known miRNAs provides independent indication that we indeed do not massively overpredict new miRNA-genes.

It is becoming evident that many regions between protein coding genes are transcribed (e.g. 3536). Indeed given the cases of miRNAs that have been suggested to regulate other miRNAs37 or RNAs that guide methylation DNA38, it would be interesting to extend our filtered intragenomic match approach to identify other possible miRNAs whose targets are not mRNAs.

Arabidopsis genome and annotation were the RefSeq sequences based on the 5.0 version released by TIGR. Known miRNAs were from the 5.1 release of the microRNA registry39.

BLASTN was used to search all Arabidopsis ESTs downloaded from GenBank on September 27, 2004. Hits longer than 70 nucleotides with more than 95% identity between a predicted precursor and an EST were considered positive. Sequences cloned and sequenced as part of the Arabidopsis Small RNA Project (ASRP)44, were downloaded from 45. All matches at least 15 long with at most one mismatch with our predicted mature miRNA-sequences were found using vmatch46.

To determine how many of our predictions were conserved in other plant genomes, we blasted the predicted Arabidopsis precursors against the rice-genome and brassica sequence downloaded from 47. A miRNA prediction was taken to be conserved if it had a significant (e-value < 0.01) blast hit containing the mature miRNA with no more than 2 mismatches and the homolog had flanking sequence capable of folding back on the mature miRNA with at least 15 base pairs between the miRNA and miRNA*.

For all candidate microRNAs in the baseset matching more than one mRNA, we found the number of different non-homologous targets by performing single linkage clustering on the aminoacid sequences of the corresponding mRNAs using the program 'blastclust' from NCBI. Two proteins were considered homologous if they had more than 70% identity across at least 50% of the length.

Micromatches with genomic start position within 4 nucleotides were logically grouped into the same locus.

We used the program vmatch48 to align and perform single linkage clustering of the predicted mature miRNA sequences. Candidate pairs aligning over at least 17 bases, allowing an edit distance of 1 were grouped in the same family.

We obtained gene ontology annotation (GOSLIM) from 49. From each GOSLIM category we constructed a 2 × 2 contingency table counting the number of targets vs non-targets with or without the GOSLIM annotation. We used R50 to calculate p-values with Fisher's Exact Test and employed the package 'qvalue'51 to correct for multiple testing setting a false discovery rate level at 0.001. The results are included as [Additional file 2], along with the R-code used.