Although the distribution of MRS appeared to correlate broadly with several other genome features, the specific nucleotide composition of each sequence window will influence the number of MRS. By randomising the genome sequence whilst maintaining nucleotide composition (mononucleotide randomisation), we estimated the number of MRS expected in the sequence due to nucleotide composition alone. Additional randomisation models were used in order to account for relationships between adjacent bases. The mononucleotide randomisation model generated sequence in which the frequency of each of the four nucleotides matched that observed in the chromosomal sequence. More complex first, second and third order Markov chain randomisation processes reflected the di-, tri- and tetra-nucleotide content of the chromosomal sequence. For each 2 Mb non-overlapping window used in Figure 1, the nucleotide sequence was randomised 1000 times, and MRSfinder was used to map and count the number of MRS in each randomised sequence. A comparison of MRS counts for chromosome I under each randomisation process is shown in Figure 2 (results for second order Markov chain randomisation of the other chromosomes can be found in Additional file 3). The observed number of MRS in mononucleotide randomised sequence was similar to that found in real sequence, while the first, second and third order Markov chain randomised sequence yielded far fewer MRS. As MRS occurrence was best modelled by the mononucleotide randomisation process, subsequent analyses focussed on this method of randomisation.

Figure 3 shows the difference in observed MRS count for each 2 Mb window from the mean count in the mononucleotide randomised sequences, in terms of standard deviations from the mean. Throughout the length of each chromosome, the number of MRS in real sequence was generally lower than in the mononucleotide randomised sequence. The arms were particularly poor in MRS and the chromosome centres were at most only slightly enriched for MRS. In contrast to the autosomes, the distribution of MRS along chromosome X (Figure 3, broken line) was much more even and similar to that found in mononucleotide randomised chromosome X sequence.

One effect of randomising the genome sequence in relatively large sections of 2 Mb is that nucleotide content (or nucleotide local pattern) becomes more uniform across each section, eliminating, for example, local peaks of very high AT%. To identify the effects of local areas of extreme nucleotide composition, mononucleotide randomisation was applied to smaller sections of sequence (10 bp, 100 bp, 1 kb, 50 kb, 2 Mb and the whole chromosome length) to C. elegans chromosome I. The number of MRS found in the whole chromosome under each mononucleotide randomisation regime, averaged over 1000 iterations, is shown in Figure 4. The numbers of MRS found when the chromosome was randomised along its entire length in one section and in 50 kb sections were very similar to the 2 Mb randomised sequence (about 10% higher than in the actual sequence). However, at randomisation sections of less than 50 kb the total number of MRS found rose dramatically. A similar effect was observed in the second order Markov chain process randomised sequence (data not shown). Compared to actual genomic sequence, the average number of MRS observed in mononucleotide randomised sequence doubled when the chromosome was randomised in sections of 10 bp.

The above results show that the number and distribution of MRS in the C. elegans genome is distinct from that found in random sequence. To investigate how this distribution is related to other genome features, the degree of overlap between MRS and different functional parts of the genome was assessed. The number of MRS occupying the same genome space as exons, introns, 3' untranslated regions (UTR), 5' UTR, genes and intergenic regions, is given in Table 1. The expected score indicates how many MRS would be expected to lie in a feature, based on the total size of the feature and assuming a uniform distribution of MRS across the genome.

The ratios of actual and expected MRS numbers showed large differences in MRS abundance in each of the genome features. MRS were particularly rare in exons, which contained less than half the MRS expected. As a result, the number of MRS in genes was also lower than expected, despite enrichment for MRS in introns and untranslated regions. Intergenic regions had slightly more MRS than expected. However, the 5' UTR and 3' UTR were by far the most MRS-enriched parts of the genome, by factors of 4.2 and 2.8 respectively. The relative enrichment of introns, 5' UTR and 3' UTR for MRS provides an explanation for the spatial relationship between genes and MRS described in Figure 1.

The MRS is AT rich and so is more likely to occur in AT rich sequence (see Additional File 4). To control for this bias, an AT-correction factor was used to adjust the expected number of MRS. The correction factor was based on the number of MRS found in mononucleotide random sequence with AT content equivalent to that of each feature, as a proportion of the number of MRS found in random sequence with AT content equivalent to that of the whole genome. When this correction is applied, the AT-poor exons appeared enriched for MRS, while the AT-rich introns had fewer than expected. Both genes and intergenic regions had approximately the number of MRS expected.

However, even with AT correction, the untranslated regions, particularly the 5' UTR, showed strong enrichment for MRS. Alternative overlap scoring systems that take into account partial MRS-feature overlaps did not affect these results. Although UTR form only a small part of the genome and contain only a small proportion of the total MRS, the degree of MRS enrichment and their proximity to genes points to a functional role for MRS.

To clarify the relationship between genes, especially their 5' and 3' UTRs and MRS, the frequency of MRS in the regions surrounding gene boundaries was investigated. Using the data from MRSfinder, MRS locations were plotted on a section of sequence extending 1000 bp upstream of the translation start site (ATG codon) through the first 400 bp of the coding sequence (CDS) from each C. elegans gene. The same analysis was carried out on sequence from the last 400 bp of the CDS through to 1000 bp downstream of the stop codon (Figure 5A).

As expected from the overlap of MRS with genes and intergenic regions reported in Table 1, the frequency of MRS in regions outside the CDS was higher than in the CDS itself. The enrichment of MRS in the 5' and 3' UTRs shown in Table 1 correlates with striking increases in MRS frequency in the regions immediately flanking genes. The MRS frequency sharply rose and fell over a span of 350 bp, peaking 50–100 bp upstream of the CDS start. At the 3' end of the CDS the MRS frequency spike had an even greater amplitude, increasing by more than 3 fold in 200 bp.

One explanation for the MRS spikes bounding CDS is that they are related to AT content of these areas. For example, in the case of 3' UTR the apparent over-representation of MRS was reduced when AT content was taken into account (Table 1). Plotting AT content in the region surrounding CDS revealed a pattern of sharp spikes similar to that observed for MRS frequency (Figure 5). However, on closer inspection there were subtle differences between the MRS frequency and AT content variation. Firstly, the upstream AT peak occurred in the 50 bp immediately preceding the start codon, 50–100 bp after the MRS peak. Similarly at the downstream end, the AT peak occurred in the 50 bp immediately following the stop codon, again 50–100 bp separate from the MRS peak.

Another difference was that the AT content dropped to 58% in the first 50 bp of the CDS, then rose to about 62% for the middle part of the CDS. The pattern was similar at the end of the CDS, where the AT dropped to near 58% in the last 50–100 bp. In both locations this AT dip was not matched by a dip in the MRS frequency. The variation in AT content in the vicinity of gene boundaries is an intriguing observation. A similar pattern was described previously by Zhang et al. 29 but further discussion of this phenomenon is beyond the scope of this paper.

An analysis of the MRS frequency surrounding gene boundaries was also performed on a related nematode, Caenorhabditis briggsae (Figure 5B). As in C. elegans, the frequency of MRS was higher in C. briggsae intergenic regions near genes than in CDS. However, from 1 kb upstream to 1 kb downstream of the CDS, the frequency of MRS was generally lower in C. briggsae than in C. elegans. The main difference in the pattern of MRS frequency between the species was that while C. briggsae displayed the same striking increase in average MRS frequency at the 3' end of the CDS, it lacked any increase in frequency at the 5' end. The possibility that less robust gene annotation in C. briggsae could have lead to this discrepancy was addressed by filtering the dataset to ensure all CDS started with ATG and ended with a stop codon, and that the selected sequence was complete and of high quality (i.e. no Ns). However, the possibility that the C. briggsae gene set is systematically lacking upstream exons cannot be excluded.

The difference between MRS frequency and AT content is even more marked in C. briggsae than in C. elegans. Although C. briggsae lacked an upstream MRS peak, an increase in AT content from about 63% to 66% was evident in the 50 bp immediately preceding the CDS start. In common with C. elegans, the downstream AT peak occurred 50 bp before the MRS peak and the AT dip at the start and end of the CDS was not matched by a dip in MRS frequency.

The distinctive increase in MRS frequency at the downstream end of both C. elegans and C. briggsae CDS could be due to conservation of MRS in specific genes, or simply a reflection of a general tendency. To investigate this, the occurrence of MRS within 200 bp of the CDS stop codon in C. elegans genes was compared to MRS occurrence in the same region of the corresponding C. briggsae ortholog (Table 2). Surprisingly, of the 224 C. briggsae genes annotated as orthologs of C. elegans genes with an MRS within 200 bp of the CDS stop codon, only 18 had an MRS in a similar position. Nonetheless, a small but significant degree of correlation between C. elegans genes and their C. briggsae orthologs for the presence or absence of MRS was detected (log odds ratio = 0.641, p value = 0.006). Therefore, the peak of average MRS frequency at the downstream end of C. elegans and C. briggsae CDS was due partly to apparent conservation of MRS in specific genes.

If the MRS is related to a cis regulatory function then the presence of an MRS near a gene may be associated with a particular functional group of genes. This possibility was examined by identifying the set of C. elegans genes with an MRS within 200 bp of the CDS stop codon, and searching for over-represented Gene Ontology (GO) terms within this set. The top most over-represented GO slim terms are shown in Figure 6. The most over-represented term was the molecular function "receptor activity": 89/509 genes in the MRS set had this annotation (17.5%) compared to 1122/9102 genes in the reference set (12.3%). None of the other terms were significantly over-represented after correction for multiple testing. Analyses were conducted to detect correlation of MRS-associated genes with other genomic and functional genomic features, including expression pattern (as determined by Serial Analysis of Gene Expression data) and position in operons, but no significant associations were obtained (data not shown).

The identification of MRS on a genome-wide scale was automated through the use of a custom perl program, MRSfinder. Using the description of the MRS given by van Drunen et al. 5, MRSfinder locates all occurrences of the MRS in a given sequence in either orientation and reports their start and stop positions. The program is freely available 30.

Version WS150 of the C. elegans genome was downloaded from the WormBase ftp site 31. The associated gene annotation for WS150 was downloaded using WormMart 32 and additional annotation was downloaded from the WormBase genome browser 33.

Version cb25 of the C. briggsae genome was downloaded from WormBase ftp site 34. This version is assembled into 578 contigs. The associated annotation was downloaded using WormMart 32.

Each chromosome was divided into consecutive, non-overlapping 2 Mb windows, with the first window starting at chromosome base position 1. Where the final window did not contain 2 Mb, the counts for that window were scaled proportionally. For each window, the number of MRS (from MRSfinder) and gene start positions (from WormBase) were assessed. Where a gene was annotated as having more than one transcript or gene model, one transcript and model was randomly selected.

For randomisation of sequences >= 32,000 bp, a roulette wheel selection algorithm was used where a nucleotide's chance of selection was based on its frequency in the original sequence. Due to the stochastic nature of this randomisation method the nucleotide frequency was verified to ensure it fell within 0.2% of that found in the original sequence. For sequences <32,000 bp, the sequence was randomised using a Fisher-Yates shuffle. Each sequence was randomised 1000 times. Each chromosomal sequence was split into consecutive, non-overlapping windows of the appropriate length with correction for shorter end windows as above. Following randomisation, MRSfinder was used to identify all the MRS in the randomised sequence. The mean and standard deviation of the MRS counts for each randomised version of the sequence were calculated.

First, second and third order Markov chain processes were used to randomise the genome sequence following the algorithm of Workman and Krogh 35. In a first order Markov chain process, the first nucleotide is chosen by sampling from the mono-nucleotide frequency. Subsequent nucleotides are added by sampling the probability distribution derived from the frequency of the four di-nucleotides that start with the previous nucleotide. Higher order Markov chain processes are used to generate randomised sequence in a similar fashion.

Genes, introns, exons, 3' UTR and 5' UTR were identified based on the GFF file for the appropriate C. elegans chromosome. Intergenic regions were defined as all sections of DNA not annotated as belonging to a gene. Where two or more incidents of a single feature type overlap, they were joined to form a single incident of that feature. The genomic coordinates of each feature were used to identify MRS that lay wholly within and partially overlapping a unit of that feature.

The number of MRS expected to lie wholly within each feature type (i.e. complete overlap) was calculated using the formula:

M(F((f-m)+1))/c

The expected number of MRS expected to partially overlap a feature:

M(F(2(m-w)))/c

When the average size of the MRS exceeds that of the feature, a complete overlap is defined as a feature lying wholly within an MRS. The expected number was calculated using the formula:

M(F(m-f)+1))/c

The expected number of partial overlaps when the average size of the MRS exceeds that of the feature:

M(F(2(f-w)))/c

where M = number of MRS, F = number features of specific type, f = average length of feature, m = average length of MRS, w = minimum number of nucleotides required for a partial overlap and c = chromosome length.

Three different scoring methods were used to combine the number of partial and complete overlaps to give an overall score. In method 1 complete overlaps = 1 point, partial overlaps = 0 points, method 2 complete overlaps = 1 point, partial overlaps = 1 point method 3 complete overlaps = 1 point, partial overlaps = 1/2 point. In all scoring methods, the minimum number of nucleotides required for a partial overlap was 12. An AT content correction factor was calculated based on the ratio of the number of MRS found in random sequence with the same AT content as each feature to the number of MRS found in random sequence with the same AT content as the genome. The number of MRS found in random sequence of specific AT content is shown in Additional File 4.

In this analysis, one CDS per gene was used: where a gene was annotated with multiple transcripts and/or gene models, a single transcript/model was randomly selected to represent the gene. The CDS were then subjected to quality filters to remove poor quality sequence (containing Ns), CDS with insufficient sequence upstream or downstream and CDS that did not start with ATG or end with a stop codon. Of the 20,052 C. elegans CDS originally identified, 20,032 passed these filters. The 19528 C. briggsae CDS were reduced to 12954 after filtering. Each successfully filtered CDS was then split into consecutive, non-overlapping 50 bp windows, starting 1000 bp upstream of the CDS start site and continuing to 1000 bp downstream of the CDS stop site. The total number of MRS mid-points occurring in each window across all CDS was divided by the number of CDS used to produce a frequency of MRS occurrence in that window.

CDS were selected, filtered and split into consecutive, non-overlapping 50 and 10 bp windows as described above. For each window the AT content was calculated as a percentage of the window length. The mean AT% for each position across all CDS was calculated.

The cb25 version of the C. briggsae genome sequence and annotated orthologs to C. elegans were downloaded from WormBase. After subjecting the 11,953 orthologs to filtering for length (i.e. sufficient sequence upstream and downstream for further analysis), poor quality (sequence containing Ns), and CDS not starting with ATG or ending in a stop codon, 4132 genes remained. MRSfinder was used to detect MRS within 200 bp of the CDS stop for each of these filtered genes in C. elegans and C. briggsae. To test for association between a C. elegans gene having an MRS and the C. briggsae ortholog having an MRS we calculated the log odds ratio (a × d)/(b × c) where a is the number of orthologs with an MRS within 200 bp of the CDS stop codon in C. elegans and C. briggsae, b is the number of orthologs where an MRS is only found within 200 bp of the CDS stop codon in C. briggsae, c is the number of orthologs where an MRS is only found within 200 bp of the CDS stop codon in C. elegans and d is the number of orthologs where neither organism has an MRS within 200 bp of the CDS stop codon.

A set of C. elegans genes with an MRS within 200 bp of the CDS stop codon was analysed to identify over or under-represented Gene Ontology (GO terms). The Gene Ontology annotation file for C. elegans was downloaded from the Gene Ontology website 36. Following Vavouri et al. 37, only GO terms inferred from electronic annotation (evidence code IEA) were used due to the bias of RNAi phenotypes on the GO annotations of C. elegans genes. The Perl script map2slim and version 1.2 of the generic GO slim ontology (both available from the GO website 36) were used to obtain GO slim term association counts for the C. elegans gene set. Of the 1057 genes in the set, 509 were associated with a GO slim term. The GO slim term counts for this gene set were compared to a reference set, containing all the remaining C. elegans genes. For each GO slim term the log odds ratio (a × d)/(b × c) was calculated, where a is the number of genes in the MRS set associated with the term, b is the number of genes in the reference set associated with the term, c is the number of genes in the MRS set not associated with the term and d is the number of genes in the reference set not associated with the term. To account for multiple testing, the Benjamini and Hochberg method was used to calculate a p value threshold for a 5% false discovery rate 38.