Srnascanner: a Computational Tool for Intergenic Small Rna Detection in Bacterial Genomes

Background: Bacterial non-coding small RNAs (sRNAs) have attracted considerable attention due to their ubiquitous nature and contribution to numerous cellular processes including survival, adaptation and pathogenesis. Existing computational approaches for identifying bacterial sRNAs demonstrate varying levels of success and there remains considerable room for improvement.


Introduction
Systematic experimental and computational approaches have led to the identification of ,92 small RNAs (sRNAs) in Escherichia coli K12 MG1655 alone [1]. Many sRNAs have been assigned regulatory roles in the survival and physiology of the organism [2]. Prokaryotic sRNAs are known to play roles in regulation of sporulation [3], sugar metabolism [4], iron homeostasis [5], survival under oxidative stress [6], DNA damage repair, maintenance of cell surface components [7] and regulation of pathogenicity [8]. Though sRNAs do not code for peptides they exert their function through antisense modes by RNA-RNA base pairing [9,10] or by antagonizing target proteins through RNAprotein interactions [11]. Genomic screens for sRNAs have been most extensively conducted in the model organisms E. coli K-12 [12,13] and Bacillus subtilis [3]. More recently, significant numbers of sRNAs in pathogens such as Staphylococcus aureus [14], Pseudomonas aeruginosa [15] and Listeria monocytogenes [16] have been identified, though functional roles of the majority remain to be determined.
Most computational methods, such as QRNA [17] and Intergenic Sequence Inspector [18], use intergenic sequence conservation among related genomes to identify sRNAs. By contrast, the RNAz [19] and sRNAPredict [15,20] programs utilize estimated thermodynamic stability of conserved RNA structures and existing 'orphan' promoter and terminator annotations for sRNA predictions, respectively. Previous studies by Argaman et al. [12], Chen et al. [21], Pfeiffer et al. [22] and Valverde et al. [23] had used promoter and terminator signals to predict sRNAs but did not provide computational scripts for general use. This study implements a generic transcriptional signal detection strategy and applies it systematically to obtain reproducible computational results and matching 'prediction scores'. Furthermore, sRNAPredict [15,20] and SIPHT [24] require available promoter information and databases of rho-independent terminators predicted by TransTermHP [25] to identify sRNAs. Moreover, sRNAPredict2 requires as inputs sequence and structure conservation data as identified by Blast and QRNA, respectively, markedly hampering detection of sRNAs mapping to non-conserved intergenic sequences. The proposed tool overcomes these limitations by searching genome sequences for orphan transcriptional signals and integrating signal co-ordinates to identify candidate intergenic sRNAs without any pre-requirements.
Comparative genomic approaches are restricted to identifying sRNA candidates located within conserved genomic backbone regions common to closely related bacteria [26]. However, most bacterial species have significant cumulative spans of multiple strain-specific sequences or islands, dispersed along the genome, many of which play key adaptive and/or pathogenesis-related roles [27,28]. Indeed, genomic island-borne sRNAs have been identified in S. aureus [14] and Salmonella enterica serovar Typhimurium [22,29]. Furthermore, sRNAs transcribed from strain-specific regions of S. Typhimurium were reported to partake in complex networks for stress adaptation and virulence regulation [8,22,28,29] leading Toledo-Arana et al. [8] to emphasize the need for identification of strain-specific sRNAs in pathogens. S. Typhimurium is an important food-borne pathogen that causes a substantial burden of diarrhoeal disease globally. Life-threatening systemic infections can also occur in those with severe comorbidities, at extremes of age and/or with impaired immune systems.
We have constructed a position weight matrix (PWM) based tool named sRNAscanner, using E. coli K-12 MG1655 sRNAspecific transcriptional signals as positive training data, for the identification of intergenic sRNAs. Experimentally characterized E. coli sRNA promoters appear to vary slightly in base distribution frequencies when compared to E. coli mRNA promoters (Table S1a), though it remains possible that observed differences may be statistically insignificant. sRNAscanner cut-off thresholds were identified using the known E. coli K-12 MG1655 sRNAs as a positive dataset [30]. The predictive abilities of sRNAscanner and sRNAPredict2 [20] were then compared by analysing 13 bacterial genomes representative of diverse species. As a specific case study, we analyzed a S. Typhimurium complete genome sequence and experimentally validated a small set of previously uncharacterized predictions. Our results strongly support the accuracy and utility of sRNAscanner as a tool for the discovery of novel sRNA genes within intergenic regions of bacterial genomes and hint at the broader power of customized PWMs as a generic strategy for detection of defined genomic features in diverse bacterial genomes.

Methods
Summary of the sRNAscanner program sRNAscanner uses as inputs matching complete bacterial genome sequence and protein coding table files in standard FASTA and tab-delimited text formats, respectively, to identify sRNA genes in intergenic regions. The sRNAscanner suite consists of algorithms to perform the following functions: (a) construct PWMs from sRNA-specific transcriptional signals, (b) search complete genome sequences using constructed PWMs to identify 'orphan' intergenic promoter and terminator locations, (c) perform coordinate based integration of promoter/terminator signals to define putative intergenic transcriptional units (TU) and (d) select predicted TUs based on cumulative sum of scores (CSS) values above a nominated threshold. The CSS value is determined by summating three individual matrix-specific sum of scores (SS) values for each candidate TU (see below for calculation of SS value). sRNAscanner uses pre-computed PWM and the following pre-defined parameters to predict intergenic sRNAs: promoter box 1 SS value ($2), promoter box 2 SS value ($2), terminator SS value ($3), spacer 1 range (defines distance between promoter boxes 1 and 2; 12-18), spacer 2 range (defines distance between promoter box 2 and terminator signal; 40-350), Unique Hit value (200) and CSS ($14). The Unique Hit value identifies potential TU from a set of overlapping hits based on the presence of closely located start coordinates mapped within a defined window size which by default is set at 200 bp. sRNAscanner selects the TU with the maximum CSS value from each overlapping set as a unique representative hit for the set. Note: all parameters can be altered by users as required. Predicted TUs are examined for the presence of a putative ribosome binding site and initiation codon; if both signals are identified the TUs are classified as coding for putative mini-proteins [28]. Remaining TUs are considered to code for candidate sRNA molecules. A flowchart summarizing the sRNAscanner algorithm is shown in Figure 1.

Construction of PWMs from training data
sRNAscanner computes a PWM of four rows and x columns for N input sequences each having x residues; N and x can be any positive integer. The program uses multiple sequences of sRNAspecific transcriptional signals in fasta format as input for the construction of alignment matrices. The alignment matrix captures the number of occurrences, n i,j , of letter i at position j across the set of aligned sequences. Subsequently, actual occurrence values were converted into log-odd scores; values that reflect the positional weights of each of the four bases (A, T, G, C) at each position. Frequency calculations and scoring schemes were adopted from previous algorithms and the positional weights were derived from the alignment matrix itself. A PWM was then derived from the above alignment matrix using the following formula (see Hertz and Stormo, 1999 [31] for details): In this formula N is the total number of input sequences and p i is the a priori probability of the letter i occurring at position j of an input sequence; by definition for a four component system (A, C, G & T) this expected frequency is 0.25 for each of the four nucleotides, f i,j = n i,j /N is the frequency of the letter i in position j. Importantly, the precise genomic base frequency of the training or test genomes do not have a role in the construction of PWM. The log-odd scores are used for the construction of PWM; the algorithm was implemented using the PWM_create module of the sRNAscanner program. We have used ten promoter boxes and twenty one rho-independent terminators [21] of experimentally-verified E. coli K-12 sRNA genes as training data to construct PWM1 (promoter box1), PWM2 (promoter box2) and PWM3 (rho-independent terminator) (Table S1 and Figure S1).
Identification of intergenic sRNA specific transcriptional units PWM1, PWM2 and PWM3 matrices were used individually to scan entire genome sequences, one nucleotide at a time, by a sliding window method as described previously [31]. The width of each sliding window was equal to the length of its matching input PWM. The matrix-specific SS value of each DNA sequence window was calculated by adding the PWM-determined scores corresponding to each of the respective bases within the window as described previously [31]. Each successive sliding window was assigned a SS value and it was compared against a selected threshold SS value obtained by analysis of the 92 known E. coli K-12 sRNA genes from the sRNAMap and Rfam datasets (http:// srnamap.mbc.nctu.edu.tw/). sRNAscanner was run with an arbitrary minimum SS value of 1 for each of the three matrices to identify potential intergenic TUs which were then compared manually with the known K-12 sRNA genes to identify concordant pairs. Using these criteria and no imposed CSS cutoff, 66 of the 92 known sRNAs were identified as possessing sRNAscanner-detectable potential transcriptional signals (Table  S2). Re-iterative empirical analyses using progressively higher matrix-specific SS values were performed to identify matrixspecific default SS thresholds that sought to maximize sensitivity whilst minimizing false-positive hits; SS cut-offs determined were as mentioned previously. Sequences having PWM1-, PWM2-and PWM3-specific SS values above the threshold scores were selected as potential promoter box 1, promoter box 2 and terminator signal hits, respectively. Next, the orientation, relative position and spacing of PWM-detected hits were examined against pre-defined allowable ranges for spacer 1 and spacer 2 to identify potential TUs. Spacer parameters used were based on analysis of the length and transcriptional signal spacing features of known E. coli and other Enterobacteriaceae sRNAs. Sequences satisfying both spacer checks and a selected CSS cut-off value were identified as likely TUs. The PWM3 SS value was expected to contribute most to the CSS score as for the known E. coli K-12 TUs detected by the program, PWM3 scores varied from 4.54-11.19, whilst the top values for PWM1 and PWM2 were 4.98 and 6.03, respectively. Importantly, higher SS values on one or both of the other matrices would not have compensated for a single below-threshold score. Identified TUs were compared with protein coding annotation files. Non-redundant, intact, non-overlapping TUs identified within intergenic regions alone and lacking putative ribosome binding sites and start codons were reported as probable sRNAspecific intergenic TUs. Abbreviations shown are to describe the eleven growth conditions. Salmonella pathogenicity island 1 (SPI-1) induced cultures  were grown with high salt-containing LB broth (0.3 M NaCl) for 12 hours at 37uC/220 rpm in tightly closed tubes. Salmonella pathogenicity island 2 (SPI-2) induced cultures  were prepared by inoculating 70 ml of SPI-2 medium [32] in 250 ml flasks, with 1/100 inoculums grown in SPI-2 medium overnight, and incubated at 37uC/220 RPM until reaching an OD 600 = 0.3. The above cultures were spun down and the cell pellets mixed with stop mixture [95% ethanol (v/v), 5% phenol (v/v)] and immediately frozen in liquid nitrogen.

RNA isolation and Northern blot analysis
Total RNA was prepared from frozen cells using the TRizol (Invitrogen) method and treated with DNase I (Fermentas) as described previously [32]. Approximately 10 mg of RNA for each growth condition was added to 26 RPA buffer and run on 6% polyacrylamide/7 M urea gels, along with a pUC8 DNA ladder (Fermentas). After separation RNA was transferred to Hybond-XL nylon membranes (GE Healthcare) and UV cross-linked. Potential sRNA transcripts were detected using c-ATP end-labeled oligonucleotide probes (Table S3).

RACE mapping of RNA transcripts
59RACE experiments were performed as described by Vogel and Wagner [33]. In summary, primary transcripts were treated with tobacco acid pyrophosphatase (TAP), ligated to A4 RNA adapters (500 pmol) at the 59ends and reverse transcribed into cDNA with random hexamers (400 ng) using Superscript II Reverse Transcriptase (Invitrogen). Next, the first strand of the cDNA molecule was PCR amplified using an adapter-specific primer (JVO-0367) and matching sRNA-specific primer (Table  S3). Amplified 59 RACE products were cloned into TOPO pCR2.1 and sequenced from both ends with M13 primers.

Results and Discussion
Optimization of sRNAscanner with known E. coli K-12 MG1655 (NC_000913) sRNA data We analysed the E. coli K-12 MG1655 (NC_000913) genome using pre-defined parameters (see User Guide) and matrices trained with data from ten promoter boxes and twenty one rhoindependent terminators [21] of experimentally verified E. coli K-12 sRNA genes. To maximize sensitivity at the expense of specificity, we ran this analysis without application of a CSS cutoff. Predicted intergenic sRNA-specific transcriptional units were compared with the 92 reported E. coli K-12 sRNAs available in sRNAmap [1] and/or Rfam [34]. Physical locations of 66 of the 92 experimentally-validated sRNAs fully or partially overlapped with sRNAscanner-identified putative TUs. However, application of the program without a CSS cut-off led to extremely low specificity with .2,500 putative intergenic TU identified. Subsets of known MG1655 sRNA predicted by sRNAscanner and other computational and experimental methods are shown as a Venn diagram ( Figure 2). The mean and standard deviation of the CSS of experimentally verified MG1655 sRNA transcriptional units detected by sRNAscanner were used to define a stringent CSS cutoff value of 14 (mean + standard deviation = 13.87). Nevertheless, the substantial overlap between whisker plots of CSS values for the known sRNAs and the uncharacterized sRNAscanner hits ( Figure 3A) and the fact that these two sets remained unresolved even when CSS score distributions were plotted as a histogram ( Figure 3B), suggested that many genuine E. coli K-12 intergenic TUs remained to be experimentally defined or that the matrices and/or the sRNAscanner algorithm lacked specificity. Interestingly, the single uncharacterized hit outlier with a CSS = 19.56 has also been predicted by SIPHT ( Figure 3A). Lists of sRNAscanner-predicted (CSS.14) known and novel candidate sRNA TUs in MG1655 are as shown (Table S2 and Table S4).

Analysis of sRNAscanner performance characteristics
sRNAscanner was run with the training set derived matrices and pre-defined parameters. Excluding the 10 sRNAs used to inform the PWM1 and PWM2 matrices, sRNAscanner (CSS.14) detected 24% of the known E. coli K-12 sRNA genes [1]. Assessment of the specificity of sRNA prediction tools remains extremely challenging as there are no gold standards and known bacterial sRNAs are likely to represent no more than the tip of a vast 'RNome' iceberg. Even experimental validation is problematic as individual sRNA may only be expressed under highly specific conditions and/or at extremely low levels. We have attempted to examine the specificity of sRNAscanner through three bioinformatics approaches. sRNA genes used to inform the training dataset were included in these subsequent analyses. Firstly, we have generated a conventional Receiver Operating Characteristic (ROC) plot [35] based on analysis of the E. coli K-12 genome ( Figure 4A). The set of known K-12 sRNAs predicted by sRNAscanner were defined as the 'True positive' set and the impact of the full range of CSS cut-off values was assessed. The ROC plot and related normalized frequency distribution graph ( Figure 4B) suggested a major sensitivity-specificity sacrifice with there being no classical optimum point; favoring either led to a marked deterioration of the other. However, even by these criteria the sensitivity (Sn) -specificity (Sp) performance of sRNAscanner at CSS.14 (Sn = 32%; Sp = 95%) was comparable to that of sRNAPredict2 (Sn = 20%; Sp = 96%). Secondly, we compared the performance of the pre-computed training-set-derived PWMs with those of randomly generated 'equivalent' matrices and used both sets of matrices to analyse the E. coli K-12 genome sequence. Equivalent random matrices were generated by randomly shuffling entire columns within each matrix (R1 random matrices) ( Figure S2), the numbers within individual columns (R2 random matrices) ( Figure S3), and a combination of these two shuffling strategies (R3 random matrices) ( Figure S4). This approach preserved the precise SS characteristics for matching genuine and random matrices and allowed the same SS and CSS thresholds to be used. However, only the R1 random matrices represented the same combination of nucleotide preferences, though present in distinct permutations as compared to the original matrices. The training and random PWM sets were used to search the E. coli K-12 genome to identify occurrences of each motif and, through integration of these data, TU-like arrangements. The occurrence frequencies (OF) of individual motifs were defined as the number of predictions per nucleotide of the genome. The ratios of OF obtained with the random and rationally-derived original matrices were expected to be inversely proportional to the ratios of matrix specificities [36]. However with the exception of the comparison between the genuine and R1 versions of PWM2, all three training PWM had higher OF than matching random matrices when applied to the K-12 genome sequence ( Figure 4C). This was most marked for PWM3 with its three random versions exhibiting less than 20% of the hits observed with the training set-derived matrix. These data strongly argued against the random nature of bacterial intergenic DNA and demonstrated the relative abundance of terminator-like motifs Figure 2. Venn diagram showing the set of known E. coli K-12 MG1655 sRNA genes detected or missed by sRNAscanner. The program was run using the training set-derived PWMs and parameters described in the text. The pale green elipse shown in dotted outline highlights the set of 66 known sRNA genes detected when the program was run without a CSS cut-off threshold. The darker green vertical oval indicates the set of 22 known sRNAs and a further 170 potentially novel intergenic sRNA detected using a CSS.14 cut-off. The sets of known E. coli K-12 MG1655 sRNA genes predicted bioinformatically by Wassarman et al. [13], Argaman et al. [12] and Chen et al. [21] are shown in blue-, red-and green-outline ovals, respectively. A further 61 sRNA genes identified through diverse experimental and bioinformatic means are shown in the yellow-outline oval. doi:10.1371/journal.pone.0011970.g002 in intergenic regions. Hits identified by the random matrices were compared with known sRNA regions to identify the number of known sRNA TUs detected. The stringent requirement for the correctly ordered, orientated and appropriately spaced occurrence of each of the three independently detected transcriptional signals was expected to filter out much of the noise. Indeed, use of the training dataset-derived PWMs resulted in identification of 66 known sRNA TUs (CSS scores [mean, range]: 12.87, 8.65-17.57), while use of the R1 random PWM, the best performing of the random versions, yielded only 14 known sRNA TUs with lower CSS scores (11.42, 9.77-14.09). The R2 and R3 shuffled matrices identified 5 and 9 potential sRNA TUs, respectively. Hence, the training matrices detected more than four times as many known sRNA TUs but only approximately twice as many total 'TU' hits as the R1 matrices ( Figures 4D and 4E). Nevertheless, as the random matrices yielded up to 68% as many total 'TU' hits as the training set-derived PWMs it would appear that even with a stringent CSS.14 cut-off, that at best only about 40% of positive calls were valid. As a third approach, we hypothesized that the ratio of the numbers of hits obtained with the full complement of concatenated genuine intergenic DNA to those found on randomly shuffled intergenic sequences would provide a qualitative measure of specificity. The concatenated sequence comprising all K-12 intergenic sequences fused end-to-end (VIGS) was subjected to random nucleotide shuffling to generate ten random variants (RIGS-1 -RIGS-10). A length distribution histogram of the 'sRNA' hits in the VIGS and RIGS sequences is shown in Figure 4F. Consistent with a moderate level of specificity, the concatenated native intergenic sequence yielded approximately three times as many hits as those identified on the 'average' random intergenic sequence (435 vs 152) ( Table S5). Use of future additional filters and/or genus-adapted PWMs may lead to incremental increases in specificity, perhaps with minimal loss of sensitivity. For example, TransTermHP-2.07-predicted rho-independent terminators in E. coli K-12 and S. Typhimurium LT2 typically exhibited PWM3 scores of $6 as opposed to the PWM3 minimum score criterion of .3, suggesting a possible route to specificity gain.

Head to head comparison of sRNAscanner and sRNAPredict2
A diverse group of bacterial genome sequences representative of Enterobacteriaceae, Vibrionaceae, Pseudomonadaceae, Bacillaceae, Clostridiaceae, Chlamydiaceae and Lactobacillaceae were analyzed using sRNAscanner. Intergenic transcriptional unit data derived from sRNAscanner analyses were compared with previously reported sRNAPredict2 results [20]. Manual curation of these predictions identified partial or complete overlaps with known sRNAs. sRNAscanner (CSS.14) and sRNAPredict2 detected a total of 180 (Sn = 31.3%) and 184 (Sn = 32%) known sRNA genes, respectively, across all 13 bacterial genomes investigated (Table 1). However, across the genomes analyzed 0 to 23 known sRNAs per genome, comprising a total of 88 known sRNAs, were predicted uniquely by sRNAscanner. By comparison, 92 known sRNAs were predicted uniquely by sRNAPredict2. However, sRNAPredict2 yielded appreciably more uncharacterized hits than sRNAscanner (2953 vs 2344), suggesting a higher signal-to-noise ratio for the latter. Similarly, large numbers of novel hits missed by sRNAPredict2 were predicted by sRNAscanner, and vice versa. Indeed, combined use of the two tools may potentially offer a degree of cross-validation. However, sRNAscanner as optimized presently appeared to be more appropriate for the analysis of genomes of Enterobacteriaceae and other medium/low G+C organisms. sRNAscanner sensitivity versus known sRNAs ranged from 51% for Clostridium tetani E88 (28.6% G+C) to 24% for Salmonella Typhi CT18 (51.9% G+C) to 0% for Mycobacterium tuberculosis CDC1551 (65.6% G+C). Detailed lists of known and putative sRNA regions predicted by sRNAscanner in the above genomes are provided as supplementary data files (see Table S4 and File S1).

Typhimurium SL1344
Analysis of the S. Typhimurium LT2 genome using sRNAscanner under default conditions yielded a total of 38 known and 118 novel candidate sRNAs ( Figure 5, Table S4). The genomic locations of the 118 novel sRNA candidates were compared with putative intergenic transcripts detected in deep sequencing libraries derived from Hfq-co-immunoprecipitated RNA obtained from S. Typhimurium SL1344 grown under multiple conditions [32,37,38] [unpublished data, J. Vogel]. S. Typhimurium SL1344 was used for all subsequent experimental validation as no comparable RNA deep sequencing dataset was available for S. Typhimurium LT2. Sixteen novel sRNA candidates were detected by both sRNAscanner and deep sequencing analysis (Table 2).

Northern and 59 RACE based verification of novel sRNAs predicted by both sRNAscanner and deep sequencing
Northern blot experiments using oligonucleotide probes targeting the 16 novel sRNA candidates mentioned above were performed (Table S3). RNA samples were harvested from cells grown and/or subjected to eleven different growth conditions. Six of the candidates (sRNA1, sRNA3, sRNA6, sRNA8, sRNA10 and sRNA12) yielded distinct Northern-detectable transcripts of broadly similar sizes to the sRNAscanner-predicted entities ( Figure 6). The additional non-specific bands seen with sRNA3-, sRNA6-and sRNA8-specific probes may comprise degraded and/ or processed forms of the matching sRNAs or overlapping mRNA  [39] and Rfam [34], Padalon-Brauch et al. [29] and Sittka et al. [32,38]. The circles shown in red dotted outline and green solid outline, excluding the central pale green curve-sided triangular area, indicate the numbers of known sRNAs predicted by sRNAscanner without and with the use of a CSS cut-off (CSS.14), respectively. The central pale green curve-sided triangular area, including the innermost circle outlined in purple, represents the 118 novel, intergenic, non-overlapping candidate sRNAs predicted in this study; the innermost circle outlined in purple represents the 16-member subset comprising sRNA candidates found to have likely mRNA transcripts by comparison with RNA deep sequencing datasets [32,38]. The $ superscript symbol indicates the five candidates belonging to both the Pfeiffer et al. [22] and Sittka et al. [32,38] sets; the asterisk symbol denotes the one sRNA candidate mapping to the Padalon-Brauch et al. [29], Papenfort et al. [39] and Sittka et al. [32,38] [37,38] were chosen for experimental validation by Northern and 59RACE analyses; five of these sixteen deep sequencing-supported hits, shown underlined, were also identified by TargetRNA. The remaining 17 sRNA candidates listed were associated with TargetRNA-identified putative mRNA targets.
transcripts. Given the above assumption, sRNA1 and sRNA12 were expressed under all growth conditions tested; sRNA8 and sRNA10 were detected in late stationary phase samples only, whilst sRNA3 appeared to be induced specifically under cold shock conditions. The sRNAscanner-predicted sRNA6 overlapped with a previously proposed processed 59UTR fragment of the yhiI transcript [38] that was likely to match the transcript we detected under ESP-2.0 conditions. However, in this study the sRNA6 locus was also found to express a distinct ,70 nt transcript found under LSP and SPI-1/SPI-2 inducing conditions only.
The 59ends of six candidate sRNA transcripts corresponding to the same Northern-supported candidates were successfully mapped by 59RACE analysis. The 59 RNA termini identified for sRNA1, sRNA6 and sRNA10 were coherent with computationally predicted transcriptional start sites but start-sites of the remaining three candidates varied significantly from those predicted by sRNAscanner ( Table 2). The extents of overlap between sRNA predicted entities, deep sequencing identified sequences and 59RACE mapped start-sites are shown schematically in Figure 6; Northern-detected transcripts were excluded as their precise locations could not be conclusively inferred on the basis of available data.

Potential biological significance of sRNAscanner predictions for Salmonella Typhimurium
Recent discoveries of three sRNAscanner identified hits that had originally been classified as novel provide further biological validation of this algorithm; sRNA17, sRNA20 and sRNA29 are now known as isrM [29], STnc410 [22] and rseX [39,40], respectively. As many functionally characterized sRNAs are antisense regulators of cognate mRNA targets [41], we hypothesized that the presence of a matching TargetRNA hit may allow for more reliable identification of genuine sRNAs. However, we emphasize that bioinformatically-derived predictions of sRNA-mRNA interactions remain fraught with problems. Consequently, pending experimental validation by gel-shift assays or other methodologies TargetRNA data need to be treated as truly putative. We identified 22 sRNAscanner hits with TargetRNAidentified potential mRNA targets ( Figure S5); five had also been detected in the deep sequencing dataset (Table 2). Several TargetRNA-identified genes play roles in pathogenesis. sRNA18 putatively targets STM1403 that codes for SscB, a type III secretion system (T3SS) chaperone encoded by Salmonella pathogenicity island 2 (SPI-2). SscB is needed for normal secretion and function of the SseF T3SS effector, which in turn is required for Salmonella-induced epithelial cell filamentation and bacterial proliferation in macrophages [42]. sRNA33 is believed to regulate ssaP, which is postulated to code for part of the SPI-2 T3SS translocon apparatus itself [43]. sRNA23 is predicted to regulate RcsF which has been proposed as one of two proximal membranelocated sensors for the Rcs phosphorelay signal transduction system that coordinately regulates expression of SPI-1/SPI-2, flagellar, fimbrial and capsule-related colonic acid synthesis genes [44]. sRNA28 is hypothesized to target stiB, a fimbrial chaperone gene, potentially allowing for sRNA28-based fine-tuning of Sti fimbriae expression [45]. sRNAs have also been shown to regulate S. Typhimurium outer membrane protein (OMP) profiles in response to envelope stress [46] or nutrient availability [39]. Similarly, sRNA29 and sRNA7 are predicted to interact with OMP-encoding genes (Table 2). Clearly, data supported solely by sRNAscanner and TargetRNA bioinformatics predictions remain speculative and robust experimentation would be required to validate these prior to drawing firm conclusions.

Conclusions
We have developed and implemented a simple PWM-based strategy for the discovery of intergenic sRNA genes. Despite use of a small, single species-derived training set, we have demonstrated the major utility of sRNAscanner to predict large numbers of potential sRNA genes in diverse bacterial species. Undoubtedly, it is vital to further experimentally validate the predictive accuracy of sRNAscanner and other sRNA prediction programmes using Northern blot analysis, ultra-high-density cDNA sequencing [37,38] and other emerging tools. Nevertheless, caution is advisable in interpretation of results as each experimental method has its own strengths and weaknesses. Furthermore, transcriptional signals would be expected to vary considerably between phylogenetically distant organisms. Consistent with this idea, we found that the E. coli-derived PWMs used in this study performed well with medium and low GC genomes but not with high GC genomes. Consequently, we propose that an organism-targeted approach is likely to lead to significantly enhanced performance characteristics. Importantly the tool developed and the strategy proposed would allow users to generate individualized PWMs based on species-, genus-or family-derived training sets to better identify sRNA genes in selected bacterial organisms. In addition, a reiterative process of PWM optimization and selection of rationally informed cut-offs based on newly discovered and validated sRNAs may allow for progressively higher levels of specificity without excessive loss of sensitivity. Finally, we propose that PWM-based scanning strategies may in time prove to be a powerful way of revealing other cryptic codes not only in DNA but in protein molecules as well.