RNA functional analyses revealed the relevance of RNA secondary structure for their function. Including RNA secondary structure information, various bioinformatics approaches were developed to identify and analyze sRNAs. In silico screens were performed in several bacteria. In Escherichia coli, initially four comprehensive analyses of IGRs were conducted, based on comparative sequence- and secondary structure-analyses as well as promoter and terminator predictions. Several hundred sRNA candidates were identified and 36 validated experimentally [15–18]. Following these studies, dozens of sRNA candidates were predicted in other bacteria using similar approaches, e.g., in Helicobacter pylori [19], Pseudomonas aeruginosa [20], Nitrosomonas europaea [21]. Tools for de novo sRNA gene finding, such as RNAz [22] and EvoFold [23] use multiple genome alignments and focus on sequence and structure conservation. In contrast, the application of cmsearch [24] requires prior knowledge about a family of related sRNAs in different species. Scans with cmsearch base on a combination of HMMs and covariance models. Agreement between approaches is low, and a potentially large number of false positives is predicted [25]. As validation of candidates by experimental methods is usually required anyway, researchers have increasingly turned towards experimental screens.

High throughput studies based on the deep sequencing and tiling array technologies elevated the potential of sRNA identification enormously. Transcriptome studies of e.g., H. pylori, Caulobacter crescentus, and Synechocystis sp. PCC6803 revealed hundreds of new sRNAs [26–28]. In the order of Rhizobiales, transcriptome analyses, focused on sRNA identification, were reported for Rhizobium etli, Agrobacterium tumefaciens, and S. meliloti 1021. A tiling array study of the R. etli transcriptome resulted in identification of 17 putative trans-encoded sRNAs and 49 cis-encoded antisense sRNAs [29,30]. A deep sequencing approach in A. tumefaciens C58 identified 228 sRNA transcripts, 22 of which were experimentally confirmed via Northern blot experiments [31]. Beside individually detected and characterized sRNAs in S. meliloti, e.g., IncA, tmRNA, 4.5S RNA, and RNase P, bioinformatics based studies identified a set of sRNAs which were further validated by Northern blot experiments [32–37]. Recently, a comprehensive deep sequencing approach combined with microarray analyses extended the number of trans-encoded sRNAs to approximately 180 [38].

An experimental screen delivers bona fide sRNA transcripts, with no obvious hints towards a potential functional role. By necessity, it starts from a single species and does not by itself incorporate phylogenetic information. Hence, it calls for a subsequent in silico study where the transcripts obtained for one species are taken as pivot elements to study their conservation and distribution in larger phylogenetic units. One intrinsic limitation of this approach is clear: an sRNA widely distributed, e.g., in the Rhizobiales, but lacking in S. meliloti, cannot be found. Hence, a complete survey of the phylogenetic order or even class from a single pivot organism is not possible.

The present study starts from S. meliloti 1021 as the pivot organism and from 52 trans-encoded sRNA transcripts obtained in our aforementioned study [38]. For each transcript, we performed homology searches and constructed RNA family models (RFMs). Our goals are twofold:

We want to increase our knowledge about the distribution pattern of potential sRNAs conserved in the Rhizobiales;

The present article describes the RFM construction process, and discusses our observations made when applying these models to the Rhizobiales.

The endosymbiont S. meliloti exists in two different life forms, either in a free-living state as a soil bacterium or in a symbiotic relationship with its leguminous host plants, e.g., Medicago sativa. In response to flavonoids secreted by the host plant S. meliloti induces the formation of root nodules. These are colonized by the bacteria which inside the nodules differentiate to endosymbiotic bacteroids that are capable of nitrogen fixation. Bacteroids support the plant with ammonia and in turn receive C4-metabolites, e.g., succinate, from the host [39].

The genome of S. meliloti consists of three replicons, a single chromosome (3.65 Mbp) and two megaplasmids pSymA (1.35 Mbp) and pSymB (1.68 Mbp). The chromosome encodes 3,351 genes predominantly involved in housekeeping functions. The 1,293 genes on megaplasmid pSymA encode, among other functions, the symbiotic apparatus. pSymB carries 1,583 genes mainly involved in exopolysaccharide synthesis and transporter functions [40–42].

Within the order of Rhizobiales, sequenced plant symbionts include Mesorhizobium loti, Sinorhizobium fredii, R. etli, and Sinorhizobium medicae. The order of Rhizobiales also comprises completely sequenced human-, animal- as well as plant-pathogens. The animal pathogen B. melitensis, for example, generally infects sheep and goats, but can act as a human pathogen as well [43]. Bartonella henselae is responsible for the cat-scratch disease of humans [44]. A well studied plant-pathogen is A. tumefaciens, which infects several dicotyledons and acts as powerful tool in plant genetics [45].

We define our notion of RNA family models and give an informal overview of how they are constructed, before we proceed to report on the findings obtained with these models. Details of family model construction are presented in the Methods section.

The 52 analyzed trans-encoded sRNAs from S. meliloti and their relatives are collected in 39 RFMs (Figure 3, Table S1). At the first glance, they show a distribution of sRNAs in good accordance with phylogeny. In this subsection, we study their distribution in detail, moving from S. meliloti species to higher taxonomic levels. We will refer to Figure 3 and Table S1 throughout this discussion.

Among our 52 transcripts, 34 map to a single origin in S. meliloti 1021 and give rise to 34 RFMs. These 34 transcripts reveal a relative distribution of 50% (17), 29.4% (10) and 20.6% (7) on the chromosome, pSymA and pSymB, respectively. The remaining 18 transcripts reveal strong sequence similarity to other transcripts, originate from multiple loci and replicons, and were summarized to five RFMs RFM_SmelA003, RFM_SmelA075, RFM_SmelB044, RFM_SmelB053, and RFM_SmelB126.

Microsynteny means the preservation of the adjacent protein-coding gene upstream or downstream of a putative sRNA locus. Gene function is usually not affected by its location in relation to its genomic neighborhood. Consequentially, the degree of synteny is lost much faster than sequence similarity and represents a sensitive indicator for genome evolution [56–58].

To classify the dimension of microsynteny for the 39 RFMs, four categories were specified.

Complete microsynteny (type I) is determined for relatives of both neighboring genes;

According to this definition, microsynteny of type I, II, III, IV was observed for 9, 17, 1, and 11 RFMs, respectively (Figure 3, Table S1). Microsynteny analyses for the stand-alone sRNA SmelC749 was not performed.

Complete microsynteny was observed for RFMs SmelA022, SmelA054, SmelB095, SmelC032, SmelC055, SmelC165, SmelC549, SmelC775, and SmelC776, which are predominantly (8 out of 9) restricted to the Sinorhizobium/Ensifer group. An exception is given by RFM_SmelC165 with relatives in the Rhizobiaceae. Extensive microsynteny was observed for RFMs of SmelA001, SmelA020, SmelA056, SmelB003, SmelB008, SmelB009, SmelB075, SmelC289, SmelC416, SmelC500, SmelC507, SmelC601, SmelC671, SmelC023, SmelC434, SmelA033, and SmelC291 (Figure 3, Table S1). Similar to the aforementioned RFMs, with the exception of SmelA033, SmelC023, SmelC289, SmelC291, and SmelC671, this type of microsynteny was predominantly observed in the Sinorhizobium/Ensifer group as well. As expected, the degree of microsynteny is higher for RFMs that are restricted to closely related organisms and to RFMs with a predominant occurrence on descendants of the ancestral chromosome and megaplasmid [53,56].

All RNAs of RFM_SmelC023 are located adjacent to a DNA polymerase I encoding gene, except for RNA12 of A. tumefaciens str. C58. Furthermore, for the 14 RNAs identified in the Rhizobiaceae, a gene encoding a MarR-type transcriptional regulator is situated next to and in case of RNA4 of S. fredii overlaps the sRNA gene (Figure 5c, Table S1). Due to the aberrant length of the overlapping transcriptional regulator gene (compared to its homologous genes) and the presence of alternative start codons approximately 200 nt downstream of the predicted start we presume an annotation mistake. Except for RNA22, RNA25, and RNA28 of B. abortus S19, B. ovis ATCC25840 and B. melitensis M28, all RFM_SmelC023 members that occur in the Brucellaceae are located antisense to a predicted small peptide encoding region. Due to the fact that the sequence of the predicted ORFs is also present in other Brucellaceae which lack this annotation, most likely this ORF was missed during gene prediction.

Except for RNA3 of S. medicae WSM419, all relatives of SmelC289 are located next to a prolyl-tRNA synthetase gene (Figure 6c, Table S1). In case of RNA11, RNA13, RNA14, RNA15, RNA17, RNA19, RNA20, and RNA21 of the Brucellaceae, the prolyl-tRNA synthetase gene is indeed located adjacent to the corresponding sRNA genes, but these sRNA genes are overlapped in antisense by one or two presumably misannotated, small hypothetical genes. For 18 RFMs, overlapping genes were predicted of which the majority is annotated as hypothetical and thus their function and existence remain in question.

Partial microsynteny was only observed for RFM_SmelA014, while fragmented microsynteny is given for RFMs of SmelA003, SmelA018, SmelA019, SmelA075, SmelA099, SmelB033, SmelB044, SmelB053, SmelB064, SmelB126, and SmelC151 (Figure 3, Table S1).

In case of RFMs of SmelA003, SmelA075, SmelB053, SmelB044, and SmelB126, fragmented microsynteny is explained by their multiple copy numbers per genome. Members of these RFMs indicate a higher rate of intragenomic transfers. RFM_SmelA075 and RFM_SmelA099 occur with three and four hairpin loops, respectively, with similar loop motifs (see Section 2.7). Considering both RFMs, fragmented microsynteny is the dominant observation, but a closer look at specific taxonomy families, e.g., R. leguminosarum reveals “local” microsynteny. In detail, the R. leguminosarum relatives RNA9, RNA14, RNA16, RNA17, RNA18, RNA21, and RNA24 of RFM_SmelA099 and RNA12, RNA17, RNA19, RNA26, RNA35, RNA42, and RNA54 of RFM_SmelA075 occur in complete and extensive microsynteny, respectively RFM_SmelB126 contains 4, 3, 2, and 5 copies in S. meliloti 1021, BL225c, Ak83, and S. medicae WSM419, respectively An association to a potassium transporter encoding gene was observed for at least a single sRNA copy in each genome. Thus it is tempting to speculate that the potassium transporter associated sRNAs represent the ancestral version of this RFM.

Multiple copy numbers per genome as well as the scarce microsynteny of RFMs of SmelA003, SmelA075, SmelB044, SmelB053, and SmelB126 are in good agreement with the scattered occurrence of mobile genetic elements next to the sRNA loci. Mobile genetic elements probably contribute significantly to the genetic polymorphism in S. meliloti natural populations [59], since mobile genetic elements are able to copy and uncouple sRNA loci from their genomic context. Repeats and mobile genetic elements were also associated to members of RFM_SmelA014, RFM_SmelA054, RFM_SmelB003, RFM_SmelB008, RFM_SmelB009, RFM_SmelB064, RFM_SmelB075, and RFM_SmelC500.

Generally, transcripts have a varying number of sub-structural RNA-domains determined as stacked base pairs, internal loops, bulges and hairpin loops [60]. A number of sRNAs, e.g., Yfr1 of several cyanobacteria, reveal typical Rho-independent terminator-like features with a 3′-located, GC rich hairpin followed by a poly-U-tail [61,62]. Additional examples for sRNAs with typical terminator features are represented by RprA and Qrr1 of E. coli and Vibrio cholera, respectively [63]. RFMs of SmelB126, SmelB053, SmelC434, SmelC507, SmelC151, SmelC023, SmelC289, and SmelC671 include typical terminator structures. On the contrary, RFM_SmelB075, RFM_SmelA014, RFM_SmelA054, RFM_SmelC416, RFM_SmelC601, and RFM_SmelC165 contain stems with atypical hairpin loops, e.g., disrupted with internal loops, followed by poly-U-tails (Figure 5b and Figure 6b, Supplement S1). Otaka et al. [64] reported that besides transcription termination, terminator poly-U-tails of the noncoding transcripts SgrS and RyhB in E. coli are essential for Hfq interaction and riboregulation [64]. A similar pattern could be presumed in case of the aforementioned trans-encoded sRNA models. However, the remaining models reveal no terminator-like features (Supplement S1).

A complex situation is given for SmelB050, SmelB053, and SmelC691. The pSymB-located sRNA genes share typical trans-encoded sRNA gene features with a long distance to neighboring genes. Their transcripts have distinct 5′- and 3′-ends and form a triple stem loop structure [38]. SmelC691 has a similar pattern except for the first stem loop. The identified sRNA relatives, all collected in RFM_SmelB053, are mainly found in the Rhizobiaceae; only a single member occurs in Ochrobactrum anthropii in the Brucellaceae (Figure 3, Table S1). Comparison of all homologous sequences indicates the first stem loop as the most variable, sometimes completely missing domain, while the second stem loop shows strong conservation, at least in the loop motifs (GGAUGUA). The third stem loop has typical Rho-independent terminator-like features (Figure 4a, Supplement S1). In addition to the identified sRNA relatives of RFM_SmelB053, a number of putative 3′-UTRs were identified in the Rhizobiales, which occur in a sequence and structure pattern similar to the second and third stem loop of the SmelB053 relatives (Figure 4b). The majority of the identified 3′-UTRs (29 out of 33) are connected to genes coding for proteins involved in cold shock adaptation, e.g., SmelC521 in S. meliloti 1021 (Figure 4c, Table S2) [38]. Post-transcriptional regulation of cold shock genes via special 3′-UTR structures was reported in case of the 428 nt long cspA mRNA in E. coli. The mRNA has two stem loop structures at its 3′-end connected to regulation of degradation via binding of Hfq and Poly-(A)-polymerase I (PAP I) that prevents binding of polynucleotide phosphorylase and RNAseE [66,67]. Due to the structural similarity of the homologous sequences identified in our study compared to the cspA 3′-UTR of E. coli and the predominant connection of the homologous 3′-UTRs to cold shock genes in the Rhizobiales, we suggest similar functions of the 3′-UTRs in a Hfq dependent manner. The functional characteristics of the trans-encoded sRNA SmelB053 and its relatives remain unclear, but due to their conspicuous similarity to the conserved 3′-domains of cold shock genes, they might act as interceptor transcripts that sequester RNA degradation complexes and thus protect and stabilize mRNAs. Sequestration is a well characterized phenomenon in bacteria, e.g., exemplified by CsrB, 6S and GlmY RNA [14,63,68].

Strong sequence and thus structure conservation is a general feature between RFM members of S. meliloti-specific sRNAs (SmelA001, SmelA014, SmelA018, SmelA019, SmelA020, SmelA022, SmelA054, SmelA056, SmelB064, SmelC032). The lowest conservation is found in RFM_SmelB064 with a structure conservation index (SCI, see Methods section) of 0.91, while the remaining transcripts of each model reveal strong sequence and thus structure conservation, with SCI values of approximately 1 (Supplement S1). Functional characteristics of trans-encoded sRNAs are commonly provided by sub-sequences and structures, e.g., in case of DsrA and RyhB in E.coli [63]. Consequently in sRNA relatives, these crucial domains of sRNAs should be conserved in a more stringent manner than in non-functional domains. However, due to the kinship between and the high sequence similarities of the S. meliloti sRNAs, conclusions about conspicuous functional sub-structures of these transcripts remain impractical. As a matter of course, we see more sequence divergence for a subset of 22 RFMs with additional relatives in the Rhizobiaceae. RFM_SmelA003 and RFM_SmelB126 have several genome internal copies in each strain (Figure 3, Table S1). Presumably due to the duplication events, transcripts of these RFMs show strong sequence divergence. However, the structure as well as a motif section of the second hairpin loop of RFM_SmelA003 shows conservation (Supplement S1). RFM_SmelB126 reveals piecewise sequence conservations at more or less conserved positions, as well as a conserved 3′-located terminator hairpin. The best structural conservation for the 22 RFMs is represented by RFM_SmelC055 (SCI = 1.04), RFM_SmelC151 (SCI = 1.02), RFM_SmelB009 (SCI = 1.02), RFM_SmelC500 (SCI = 1.01), RFM_SmelC775 (SCI = 1.0), RFM_SmelB075 (SCI = 1.0), and RFM_SmelB008 (SCI = 1.0) (Supplement S1). The strongest nucleotide divergence is shown by RFM_SmelC549 with relatives in S. meliloti and S. fredii (SCI = 0.69) (Supplement S1).

In general, RFM_SmelC549 consists of four conserved stem loops but is disrupted by several internal loops and single stranded domains. These single stranded domains are the most deviating regions and thus they are responsible for the depressed SCI value. This high conservation implies a functional role of these four domains (Supplement S1).

The RFMs of SmelC023 (SCI = 1.11), SmelA075 (SCI = 0.99), SmelC289 (SCI = 0.9), SmelC291 (SCI = 0.55), and SmelC671 (SCI = 0.55) are composed of transcripts with variable sequence and structure conservations, presumably derived from a common ancestor (Supplement S1). In case of RFM_SmelC023 and RFM_SmelC289, four and five hairpin loops, respectively, are the main components of these transcripts. The 5′-located domain of RFM_SmelC023 is highly variable both in length and nucleotide composition. The 3′-located stem loop of RFM_SmelC023 is a Rho-independent terminator-like structure with high GC content (Supplement S1). This is supported by the poly-T-sequence following the annotated sRNA gene (data not shown). Presumably a degradation event of the initially annotated sRNA SmelC023 occurred and resulted in a processed 3′-end. The middle domain consists of two hairpin loop structures with highly conserved sequences within the loops, while the structural maintenance of these sub-domains is provided by stems with varying nucleotide compositions with evolutionary established base pair exchanges (Figure 5a,b, Supplement S1). This strongly suggests that the functional maintenance of this molecule is provided by both the hairpin structures and the loop sequences. sRNAs with several functional domains are a common feature in bacteria. Obvious examples are the trans-encoded sRNAs OxyS and DsrA in E. coli. The former binds to the fhlA mRNA in a Hfq-dependent manner. Two stem loops are presumed to be involved in mRNA binding, while the interaction site of Hfq is different to that of the mRNA binding site [3,69,70]. The latter exhibits different hairpins for different targets. In detail, DsrA consists of three stem loops of which the second is able to interact with hns mRNA, blocks the ribosome binding site (RBS) and thus inhibits mRNA translation. The third stem loop interacts with the rpoS mRNA and activates translation via remodeling of an inhibitory mRNA sub-structure [70]. Similar features could be presumed for RFM_SmelC291 (sra33) [54]. The 5′-domain of RFM_SmelC291 has a structurally conserved stem loop with a strongly conserved loop motif, UCCGCCGCAUCU, while the second stem loop shows extremely variable stem sequences but occurs with a dominant but different loop motif, UCCUCG as well (Supplement S1).

RFM_SmelC289 shows a 5′-located stem loop characterized by an AU-rich loop, stabilized by a variable stem of organism-specific nucleotide contents. Similar to RFM_SmelC023, the 3′-region contains a Rho-independent terminator-like structure and the middle domain consists of two conserved hairpin loops as well. The former hairpin shows complete conservation, while the latter reveals a highly conserved stem with a variable loop (Figure 6a,b, Supplement S1). However, the functional patterns of noncoding transcripts are not restricted to their presumed single stranded domains, e.g., the binding sites of RprA sRNA in E. coli are predominantly incorporated in stem structures [63]. Further hidden sRNA sequences that are essential for target interactions could be released due to a structural reformation of the sRNA, for example, the RNA binding protein Hfq is able to alter the secondary structure of the RydC sRNA in Enterobacteriaceae resulting in an active version of this transcript [71]. RFM_SmelC671 represents the most variable model of related sRNAs in this study. The transcript has a long single-stranded domain with a conserved CUCCCUGU motif, enclosed by down- and upstream located, highly variable hairpin loops of which the 3′-located domain acts as terminator (Supplement S1). A similar pattern was observed for Qrr1 in Vibrio cholera. This transcript reveals a long single stranded domain between its first and second loop, which was verified as the mRNA interaction site. This motif is highly conserved within Qrr2-4, which are paralogs of Qrr1 [4,63].

Large numbers of cis-encoded antisense RNAs were identified, e.g., via sequencing and tiling array studies in Synechocystis [28], H. pylori [26], S. meliloti [38], and R. etli [30]. Cis-encoded antisense sRNAs are located in antisense to their targets, act via perfect base pairing and mediate post-transcriptional regulation, e.g., stabilization or destabilization of target mRNAs [6]. Pairs of small noncoding transcripts that are located in antisense to each other remain rarely identified in bacteria. Georg and Hess [72] presumed that two small transcripts encoded by unlinked sRNA genes interact with each other due to a complementary section in both transcripts. However, the functional relevance of this feature needs to be further elucidated.

Here, we observe that SmelC776 relatives are located in antisense to RFM_SmelC775 sRNAs and thus most likely allow a mutual interference of both transcripts. This is in good agreement with the strong sequence conservation of SmelC776 (SCI = 0.95) within the overlapping domain (Table S1, Supplement S1). Approximately 90% (17 out of 19) of the nucleotide exchanges of RFM_SmelC776 in the S. fredii strain occur outside the overlapping part and thus suggest the antisense part as the functional transcript domain.

A remarkable situation was found in case of RFM_SmelA075 and RFM_SmelA099, which exhibit three and four hairpin loops, respectively. Each hairpin loop carries similar loop motifs, CCUCCUCCC, representing an anti Shine-Dalgarno (aSD) sequence, while the stems show more variability in their nucleotide content (Figure 7, Supplement S1). In several sRNAs, e.g., RNA III in Staphylococcus aureus, CyaR in Enterobacteria and ABcR1 in A. tumefaciens C58, loop sequences with at least partial aSD motifs were observed. Functional analyses demonstrated that the aSD motif is indispensable for sRNA binding to the RBS of the particular mRNA targets, as well as the resulting translation inhibition [73–79]. The sequencing profile as well as Northern blot analyses of SmelA075, which is a member of RFM_SmelA075, suggested SmelA075 as a stress-induced sRNA that occurs in several processed forms [38]. This is in good agreement with the trans-encoded sRNA RS0680a and its homologous transcripts identified in Rhodobacter sphaeroides. RS0680a represents a shorter version than the RFM members identified in this study. The transcript has two stem loops instead of three and four, and each comprises an aSD motif. It was suggested that RS0680a undergoes different maturation processes and is involved in the stress response in a more general pattern via binding to the RBS of several genes [80]. From a biological point of view, it was suggested to group all derivatives of RFM_SmelA075 and RFM_SmelA099 to a single sRNA family that consists of members with a varying number of hairpin modules. Implementing all facts, this RNA family consists of several copies per genome. It is somehow involved in stress adaption, presumably in post-transcriptional regulation via blocking the RBS of target mRNAs.

52 of approximately 180 trans-encoded sRNAs were selected and downloaded from GenDB [81], accessible via the RhizoGATE portal [1,2]. The choice of these 52 candidates was made in purely technical terms: Candidate transcripts should be well-covered by sequence reads, with clearly defined ends, and remote from any coding region. We plan to generate models for the remaining transcripts in the near future.

Complete genome sequences and annotations of all Rhizobiales available were obtained from the NCBI genomes FTP site [82]. For a complete list see Table S3. Additionally, whole genome and plasmid sequences were included that are not (yet) part of the above collection. Sequence data of A. sp. H13-3, B. melitensis M28, S. meliloti AK83, and S. meliloti BL225C, all members of the order of the Rhizobiales, were downloaded from the NCBI nucleotide database.

There is no standard and fully automated way to construct an RNA family model. The general difficulty of this process has been discussed, e.g., in [83]. Family model construction is supported by a variety of tools, but interspersed with modeling decisions and candidate screening by a human expert. In our case, we start with a trans-encoded sRNA from S. meliloti, say SmelXnnn, and construct an RNA family model RFM_SmelXnnn, which (1) comprises a set of orthologous RNA sequences from related organisms; and provides (2) a search function to find further family members in the Rhizobiales and beyond. In a few cases, we find that several sRNAs should be collected into the same family model, which is then named arbitrarily after one of them.

We constructed types of RFMs, Covariance models and Thermodynamic matchers. The use of these two complementary methods has already been motivated above (Section 2.1). We now add some details about both methods, and about the assessment step used with both.

Recall Figure 1, which gives an overview of our RFM construction pipeline. Phase 1 identifies putative homologous RNAs by iterative searches focusing on sequence homology. Phase 2 constructs an initial family model based on sequence and conserved structure, and uses this model to search all Rhizobiales for further homologs.

For specific RNA families, we found that a CM was inadequate to express their peculiarities. For example, SmelA075 has three hairpins, all of which exhibit a perfect loop motif (CCUCCUCCC) (cf. Section 2.7). It is widely distributed among the Rhizobiales, with the stems sequences highly diverged. As a consequence, the discriminatory power of the CM decreases. A TDM can be designed such that it gives no weight to the stem sequences, but enforces the loop motif. SmelB053 is a similar case with a strongly conserved structure and low sequence similarity, except for a prominent loop motif (GAUGUA).

Figure 9 shows a snapshot of TDM construction, where we depict a structure graphics with the help of the Locomotif editor [47], annotate it with conserved sequence information and compiled it into a search program at the Bielefeld Bioinformatics Server [90].

TDMs tend to run a bit faster than CMs (when their HMM filter is turned off). We use them to scan all Rhizobiales genomes. Resulting candidates are assessed like other candidates.

Design of a TDM also requires some iteration, as the motif description can be made more or less specific. Generally, it is a good strategy to begin with a rather restrictive motif and check that the known sRNAs are actually found, which verifies that the design is correct. Then, the motif is gradually relaxed. Search results may suggest adjustment to the original TDM design, such as relaxing the number of paired bases in a stem, or increasing the allowed size for a loop. This interplay of human design, matcher compilation and search continues as long as candidates pass the assessment step.

When an RNA family model is available, its search procedure can be used with more or less stringent cut-offs. In this study, however, we need to construct such a model in the first place. We start from a single family member, which may not even be a typical one, but has the virtue of being based on an experimental screen rather than being an in silico prediction. We want to collect a large number of homologs, and use further evidence to weed out unplausible candidates.

Some sequences show strong sequence and structure conservation and thus can be unambiguously identified in an automatic fashion. No further effort, human or computational, is spent on them. Divergent sequences without global conservation have to be curated integrating sequence and structure conservation with additional sources of information, such as genomic context and phylogenetic distribution. Figure 10 gives an overview of this assessment, which is described further below.

So far we completed 39 RFMs including 52 of 173 trans-encoded sRNAs published in [38]. There are 121 more to go. Further automation of the model construction process is highly desirable, but not easy to achieve. While some parts of our assessment step can be integrated into an automated workflow, there are also serious challenges arising from technical limitations of the available software. We discuss two such aspects in the sequel.