Overall, 42 482 RNAz hits corresponding to roughly 5 Mb in the D. melanogaster genome show evidence of evolutionarily conserved RNA secondary structure. About 20% of these overlap existing annotation. The 16 377 loci of the high confidence set covers approximately 2.1 Mb of DNA, see Tab. 1.

In total, 336 hits correspond to known non-coding RNAs according to at least one source of annotation (FlyBase: 316; BLAST against miRBase: 79; BLAST against Noncode: 44; BLAST against Rfam: 222; tRNAscan: 159). Tab. 2 summarizes the recall of the screen on several "classical" ncRNAs families. Note that some classes of ncRNAs were deliberately removed already in the Pecan alignments, notably the 5S rRNA sequences. We recovered 96% of the known D. melanogaster miRNAs.

The genomic distribution of structured RNA candidates in D. melanogaster is comparable to the observations in previous RNAz-based screens, see Fig. 1. As in the ENCODE data 22, the distribution of RNAz hits largely follows the patterns of sequence conservation. In the fly data, only 5'UTRs show a substantial enrichment relative to the input data. In contrast, the largest enrichment in the human ENCODE data was observed from 3'UTRs 22. The most striking difference between fly and human data is that the relative fraction of both intronic RNAz hits and intronic sequence conservation is twice as large in human.

In a recent article Manak and colleagues 5 describe widespread transcriptional activity in the D. melanogaster genome during 12 timepoints of early embryonic development detected by genomic tiling arrays. When comparing the RNAz hits to this data, we identify 4 236 (p > 0.5) and 1 713 (p > 0.9) hits that overlap a Transfrag in any of the 12 timepoints. A comparison of the fractions of RNAz hits from normal and control screen which overlap Transfrags in one, several or all timepoints yields, however, no significant enrichments (see Additional file 1 for details).

The distance distribution of intergenic RNAz hits reveals a striking difference between the situation in the human and the fly genome, Fig. 2. Since the D. melanogaster genome is much more compact than the human one, we need to compare the distribution of the distances between RNAz hits and the nearest coding sequence relative to the length distribution of the intergenic regions (IGR). In Fig. 2 we plot the relative frequency of IGR with a length exceeding a given distance D, and the relative frequency of RNAz hits with a distance larger than D from the nearest coding region. If intergenic RNAz hits are uniformly distributed within the IGR, the distribution of RNAz-CDS distances looks like the distribution of IGR distances, just shifted to the left by a factor of 4. Indeed, this is observed in the human data, albeit the shift is a factor between 3 and 4, indicating that the placement of intergenic RNAz hits in human is nearly uniform, with a small tendency of avoiding the proximity of coding genes.

In contrast, about 40% of the D. melanogaster RNAz hits in intergenic regions are located adjacent to coding sequences. This may indicate that current annotation of the fly genome lists boundaries of protein coding genes that systematically truncate the UTRs. If this is the case, however, then we would have to interpret more than 15% of the total RNAz hits as located in UTRs. Our data could be explained if a situation similar to the minifly gene is prevalent in the fly: For this gene a recent study 34 described several alternative poly-A sites and multiple small ncRNAs that are processed from the alternative 3'UTRs. At least one of these ncRNAs is structured: the snoRNA H1 was also detected in our screen. In any case, the structured RNAs by RNAz are on average much more closely linked to protein coding genes in flies than in human.

On the other hand, a small fraction (≈ 10%) of the intergenic RNAz hits, i.e., the tail in Fig. 2, is located much further away from CDS than expected for random placement. This suggests the existence of a distinct class of RNAz hits with a propensity for large IGRs. Most likely, these signals correspond to independently transcribed ncRNAs.

About 20% of the unannotated transcripts observed in D. melanogaster early development arise from stand-alone intergenic or intronic sources (relative to FlyBase annotation) 5. Only a relatively small fraction of the novel independent transcripts (5.1% of the total transcriptional output) had intergenic origin. In comparison, more than 13% [21.9% of 60%] of the transcriptional output recorded by comparable methods from the ENCODE regions has a distal intergenic source (in relation to annotated exons) 22. This difference is in agreement with closer association of most RNAz hits with protein coding genes in the fly.

In a recent study, Isogai et al. 32 identified TRF1 and BRF binding sites using high-resolution genome tiling microarrays and provided evidence that in Drosophila the alternative TRF1/BRF complex appears responsible for the initiation of all known classes of Pol III transcription. At the p > 0.9 significance level RNAz hits are about three-fold enriched in these regions. We have therefore analyzed the distribution of RNAz hits within the experimentally determined TRF1 and BRF binding regions. As reported in 32, most of the sites correspond to tRNAs, 7SL RNAs, and a subset of snoRNAs. In addition to these known ncRNAs, the loci contain 197 unannotated RNAz hits, which are prime candidates for novel Pol III transcripts.

In order to identify putative microRNAs, we screened all RNAz hits with RNAmicro 35. This results in 607 candidates, of which 541 are unannotated so far. 176 of these signals are located in annotated CDS and are therefore most likely false positives, leaving 365 plausible microRNA candidates. The recent discovery of hundreds of new human microRNAs that are not conserved beyond primates strongly suggests that "evolution of miRNAs is an ongoing process and that along with ancient, highly conserved miRNAs, there are a number of emerging miRNAs" 36. In the light of these data, a large number of drosophilid-specific microRNAs does not come unexpected.

Using SnoReport 58, RNAz hits are classified as putative box H/ACA snoRNAs, of which 4 intersect with previously annotated snoRNAs. Taking into account that for only 22 of the 250 annotated snoRNAs the annotation distinguishes between box H/ACA (3), box C/D (18), and scaRNAs (1), the small overlap with the existing annotation is not surprising. Again, recent experimental surveys in other species, including nematodes 3738 and mammals 39 have discovered a substantial number of previously unannotated snoRNAs in these species, suggesting that the current annotation of snoRNAs in D. melanogaster is also far from complete.

Finally, 1 700 RNAz hits have direct evidence for expression through ESTs that are not related to protein coding genes, i.e., through ESTs that do not intersect with the FlyBase, RefSeq, N-SCAN, Genscan, Human Proteins gene prediction and mRNA tracks of the UCSC Table Browser.

Since the 197 RNAz hits that overlap TRF1 or BRF binding regions 32 are good candidates without annotation for bona fide ncRNAs, we applied structure-based clustering to this small subset of our predictions to identify common secondary structures and, hence, putative novel functional RNAs. The complete clustering tree as well as a table of the most prominent clusters is given in the Additional file 1. Since all clusters have a mean pairwise identity less than 45%, structurally related candidates are typically highly diverged at sequence level.

In Fig. 3 an example cluster of complex structures is given. Clusters 22, 25 and 28 have a structure with two stem loops in common. All consensus structures show compensatory mutations.

Fig. 4 depicts a large cluster of simple hairpin structures. They show a relatively high structural conservation (high structure conservation index, SCI) whereas the sequence similarity (expressed as main pairwise identity, MPI) is small.

In order to study the phylogenetic distribution of the RNAz prediction we determine the last common ancestor for each RNAz hit that contains the corresponding sequence in the input alignment. Fig. 5 summarizes these results for both the true data and the "gentle" control screen.

More than 50% of the RNAz hits are found only within the melanogaster subgroup. To interpret this result we compute the ratio of newly appearing RNAz hits and the branch length for each branch in the tree leading to D. melanogaster. We observe little variation in the data, with the exception of a reduced rate of innovation along the most recent branch. This reduction is, however, most likely a methodological artefact, since the pairwise mutation distances between D. melanogaster, D. simulans, and D. sechellia are only about 0.1 and RNAz is known to be less sensitive for highly similar sequences.

Approximately 12% of the RNAz hits are conserved throughout all drosophilids. In comparison, a screen of vertebrate genomes 23 found about 3% of mammalian RNAz candidates (1 000 out of 36 000) to be conserved throughout vertebrates.

A comparison of true and shuffled screens furthermore indicates a small but significant decrease of the FDR with phylogenetic age of the RNAz hit.

For our analysis we used the Pecan 40 alignment of the 12 drosophilid genomes 3031 of the Comparable Analysis Freeze 1 (CAF1, Feb. 2006). The alignments were downloaded from 3141. We favored the Pecan alignments over two other sets of drosophilid alignments that are available at 3142. Visual inspection strongly suggested that Mavid alignments 4344 are more biased towards protein coding regions. We did not use the Multiz alignments, because they contain three additional genomes (insects), and removing those sequences would effectively require a complete realignment in order to obtain a fair comparison between screens performed on different input alignments. The Pecan alignments comprise the D. melanogaster chromosomes 2L (22.4 Mb), 2R (20.8 Mb), 3L (23.8 Mb), 3R (27.9 Mb), 4 (1.3 Mb), and X (22.2 Mb).

The current implementation of RNAz is restricted to input alignments containing at most 6 sequences and a maximum length of 400 nt due to the training of the underlying SVM 17. In addition, certain restrictions apply for the fraction of gaps and on the overall base composition as a consequence of the data sets that were used to train the SVM model. The original genomic alignments thus need to be re-processed. The protocol used in this contribution closely follows that of previous RNAz-based studies:

Alignments longer than 120 nt are cut into 120 nt slices in 40 nt steps, so that subsequent slices overlap in 80 nt. This default length is motivated by the fact that many structured RNAs are less than 100 nt long. Such short signals would "drown" in the noise of longer alignments that are then mostly unstructured. On the other hand, alignments that are too short do not yield reliable signals for secondary structure conservation. In a series of filtering steps, sequences were removed from the individual alignments or alignment slices if they are (a) shorter than 50 nt, or (b) contain more than 25% gap characters, or (c) have a base composition outside the definition range of RNAz (e.g. GC content > 0.75 or < 0.25).

Alignments were discarded completely if fewer than 3 sequences were left after the filtering steps, or they did not contain a D. melanogaster sequence, since this species serves as a reference and as the basis for subsequent annotation. All preprocessing steps were performed using the script rnazWindows.pl of the current release of the RNAz package 45.

For alignment slices with more than 6 sequences, rnazWindows.pl selects a representative subset consisting of the D. melanogaster sequence and five additional sequences in such a way that 6 sequences are as evenly distributed in the dataset as possible and approach an average pairwise sequence identity of 80%, the optimal working range of the RNAz program. In practice that means that only a single representative from nearly identical sequences is chosen, and highly divergent sequences are excluded provided there is sufficient sequence variation in the remaining alignment. For the technical details of the procedure we refer to the documentation of the RNAz package .45

Tab. 1 summarizes the initial filtering steps. Roughly 50% of the nucleotides in the Pecan alignments are still contained in the RNAz input data.

RNAz was applied to the filtered input alignment slices in both reading directions. Overlapping slices with a positive ncRNA classification probability of p > 0.5 were combined using rnazCluster.pl to a single annotation element, which we will refer to as "RNAz hit". From these data, we extract a subset of high confidence RNAz hits that contain at least one slice with a prediction confidence of p > 0.9.

In order to estimate the false discovery rate (FDR) of the screen we repeated the entire procedure with shuffled input alignments as described in 33. The alignments were (1) shuffled using the rnazRandomizeAln.pl script (part of the RNAz package). It wraps a conservative shuffling procedure that maintains local characteristics of an alignment, e.g. columns with the same gap and conservation pattern. All remaining RNAz hits of this control screen are then shuffled once again (2) using a more stringent shuffling method that explicitly shuffles all columns of a given alignment randomly (cp. Tab. 3).

The RNAz hits were annotated using the D. melanogaster sequence as reference. We performed the following annotation steps:

• Overlap with known D. melanogaster annotation

We used the coordinates of a set of D. melanogaster non-coding RNAs, publicly available as gff files at 46 and 3147 to identify already known D. melanogaster ncRNAs among our predictions. Furthermore, we computed the overlap of the RNAz hits with the CDS annotations from 3148 (file = dmel-all-r4.3.filtered.gff) and 49 (file = all_caf1_DGIL_TEX.gff).

• Overlap with public non-coding RNA databases

We furthermore performed BLAST 50 searches using rnazBlast.pl against the Rfam (version 7.0) 51, Noncode (version 1.0) 52, ncRNAdb 53, FlyBase (version 2006 00.2 Beta) 5455, and miRBase (version 9.0) 56.

• Tools for annotation of specific RNA families

We furthermore used tRNAscan 57 to annotate tRNAs and RNAmicro 35 to classify putative microRNAs. RNAmicro is an SVM based classification method that evaluates both thermodynamic stability and evolutionary conservation patterns.

We used SnoReport to recognize putative snoRNAs (for technical details see 58). Similar to RNAmicro, SnoReport is composed of a pre-filter, a secondary structure prediction step, and a subsequent SVM-based classificator. In brief, the prefilter searches for consecutive H (pattern: ANANNA) and ACA boxes 59. In the second step, the constraint folding option of RNAfold 60 is used to compute the secondary structure subject to the constraint that both boxes remain unpaired. If this results in an snoRNA-like secondary structure, several sequence and structure features are computed and passed to an SVM for classification. The model was trained on the set of snoRNAs that can be downloaded from the snoRNABase 61. C/D box and scaRNAs represented the negative and H/ACA box snoRNA sequences the positive samples. Estimated positive and negative prediction values for the model used here are 80% and 99.9%, respectively.

Structure-based clustering was performed as described in 62: The modified Sankoff algorithm implemented in the LocARNA program is used to compute local structural alignments and their consensus structure. The clustering tree is obtained by agglomerative clustering using LocARNA alignment scores as distance measures. To avoid that large scores influence the distance transformation we define distances by d(i, j) = max(0; q - score(i, j)), where q is here the 99% quantil of all pairwise scores. Since the procedure is computationally very demanding we have restricted this type of analysis here to a small subset of RNAz hits that are likely Pol III transcripts.

The phylogenetic relationships within drosophilids are taken from the AAA (Alignment/Analysis/Annotation of 12 related Drosophila species) web site 63. Branch lengths are genomic mutation distances computed from 4-fold degenerate sites in all coding regions corrected for base composition as in 3164. In order to determine the branch in the phylogenetic tree at which an RNAz hit first appears, we determine the last common ancestor (LCA) of the sequences in the corresponding input alignment and assign the RNAz hit to the branch in the tree leading to this internal node. Due to the fact that RNAz hits are a combination of single windows and each window represents a specific selection of sequences out of an n-way alignment, we only considered those sequences for the LCA analysis which are simultaneously present at all windows.