HMMER uses Hidden Markov Model (HMM) to create a position-specific score matrix based on primary sequence conservation (Krogh et al., 1994; Eddy, 1998; Mistry et al., 2013). Briefly, HMM is a statistical Markov model in which the modeled system is assumed to be a Markov chain with unobserved (hidden) states. nhmmer is a part of the HMMER and search nucleotide queries against a nucleotide sequence database (Wheeler and Eddy, 2013). The software can create a matrix with only one sequence, but its accuracy is improved when a larger set of reliable sequences is provided. A variety of input sequence file formats can be used as alignment file format (Stockholm, Aligned FASTA, Clustal, NCBI PSI-BLAST, PHYLIP, Selex, UCSC SAM A2M); and unalignment file (FASTA, EMBL, Genbank). The probabilistic HMM model searches for sequence homologs in an available sequence database, accounting for substitutions, insertions, and deletions. The program output ranks a list of the hits with the most significant matches to the query. Each hit represents a region of local similarity of the HMM to a subsequence of a full target database sequence. An alignment of the matched sequence to the model with the confidence value which each position is aligned is also shown. The current version of this software (Version 3.1) can be found at http://hmmer.org/.

Infernal (INFERence of RNA ALignment) was first created by Sean Eddy in 2002 (Eddy et al., 2002). The 1.1.2 version (July, 2016) (Nawrocki and Eddy, 2013b) is used by the Rfam database (Nawrocki et al., 2015) to infer a set of homologous sequences of an RNA family. Infernal's algorithm implements covariance models (CMs), a particular case of stochastic context-free grammar (SCFGs), to create a probabilistic model that accounts for RNA sequence and secondary structure conservation that can be used to search for a particular structural pattern in user-provided sequences (Eddy et al., 2002). Further reading about SCFG is available in Giegerich (2014). First, the software utilizes a set of reliable sequence alignments, along with a common secondary structure annotation (Stockholm format), to create the CM model specific for that target RNA family. Then, it uses a dynamic programming algorithm to find similar sequence and structural patterns in a set of target sequences. The alignment and a set of trustworthy RNA sequences can be found in the Rfam database (http://rfam.xfam.org/). Infernal has many functions that are based on analogous ones in HMMER, such as output formatting, which is very similar to the two software packages. Infernal is available at http://eddylab.org/infernal/.

Developed in 2004 by Bengert and Dandekar (2004), this web-based tool employs user-provided nucleotide sequence to infer putative riboswitches by searching for specific sequence motives and obtain its secondary structure. The algorithm was tested with a consensus set of 13 known Bacillus subtilis-like riboswitches sequences. To use the Riboswitch finder, merely give any RNA sequence up to three million base pairs of nucleotides. The tool provides information on the position of the putative riboswitch, the minimum free energy (MFE) and a putative secondary structure alignment. The secondary structure and MFE are calculated using the Vienna RNA package (Lorenz et al., 2011). Riboswitch finder is available at http://riboswitch.bioapps.biozentrum.uni-wuerzburg.de/server.html.

RibEx (Riboswitch Explore) (Abreu-Goodger and Merino, 2005) is a web server to detect riboswitches and riboswitch-like elements (RLEs). Among others, RibEx can detect the Gram-positive T-box and the PyrR protein binding site based on sequence motifs that are unique to each particular class. RibEx employs an algorithm capable of finding bacterial regulatory motifs, built exclusively on sequence conservation of regulatory regions associated with at least one cluster of orthologous groups of proteins that can be found in at least five non-redundant genomes. After submitting the target primary sequence of interest, up to 40,000 bases, the software provides a scheme of the open reading frame (ORF) with the regulatory element found. The sequence corresponding to this element can be addressed with the NCBI's BLAST tool. RibEx is currently available at http://132.248.32.45/cgi-bin/ribex.cgi.

The RiboSW is a webserver (Chang et al., 2009) able to identify up to 12 classes of riboswitches based on structural conformations and sequence conservation. The authors used the sequence and secondary structure information of 12 riboswitches annotated in Rfam to recognize fundamental structure components and to create HMM models. After a user sequence query, the software seeks for the combination of structural elements corresponding to one of the twelve riboswitch classes, and performs a functional local detection using HMMER (Mistry et al., 2013). This tool provides a sequence with secondary structure annotation and graph, the MFE, the HMM e-value and the RNA Logo graph, which compares the provided sequence with the corresponding Rfam riboswitch family.

The authors stated that the performance of their method is comparable to that of the CM employed by Rfam, except for the AdoCbl and TPP riboswitches, as RiboSW was not able to detect all the members of these riboswitch families. They suggested that this flaw may be due to the high structural variation found within these families. The webserver is hosted at http://ribosw.mbc.nctu.edu.tw/.

RNAConSLOpt is a program for predicting consensus stable local optimal structures for multiple sequence alignment of related RNAs (Li et al., 2012). The program also allows predicting alternate consensus structures for riboswitches elements in bacteria 5′ UTRs. De novo prediction, with no previous knowledge about sequences and structures of known riboswitches, is possible. To predict ncRNA, RNAConSLOpt incorporates information of free energies of structures, covariance and conservation signals into enumerating ConSLOpt stack configurations. The program output consisting of all probable consensus local optimal stack configurations and consensus stable local optimal stack configurations ranked according to both free energy and the associated minimal energy barrier. RNAConSLOpt is available for download at http://genome.ucf.edu/RNAConSLOpt/.

Using a dynamic programming algorithm that considers mismatches, the Denison Riboswitch Detector (DRD) (Havill et al., 2014) web server predicts up to 13 classes of riboswitches from DNA sequences. The applied algorithm breaks the query sequence into overlapping smaller sequences and searches for exclusive motifs belonging to each searched riboswitch class. Afterwards, the secondary structure is predicted using Mfold (Zuker, 2000) and is subsequently aligned with a consensus sequence of the riboswitch class. The query provides the position of the putative riboswitch along with its optimal and suboptimal secondary structure annotation and graph. The authors tested DRD using validated sequences annotated in Rfam. Overall, the software achieved 88–99% sensitivity and more than 99% specificity. The web server can be found at http://drd.denison.edu/.

Riboswitch Scanner (Mukherjee and Sengupta, 2016) is a web server capable of detecting 24 riboswitch classes and identifying novel riboswitches. Its algorithm utilizes 5-fold cross-validated HMM models created by HMMER 3 (Mistry et al., 2013) to determine putative riboswitches. Through the submission of the nucleotide sequences in FASTA format, the server provides the position of the riboswitch, MFE, HMM score and E-value, and secondary structure annotation obtained by RNAfold (Lorenz et al., 2011). It is available at http://service.iiserkol.ac.in/~riboscan/application.html.

Riboswitches control the expression of genes involved in the biosynthesis and transport of ligands, as well as transcription factors (Mandal and Breaker, 2004b). Given their significant regulatory role in bacteria and in a few eukaryotic organisms, it is crucial to develop tools for the accurate identification of different riboswitch classes. Several approaches have been used for the computational identification of riboswitch aptamers (Figure 3, Table 1). The current riboswitch search tools employ hidden Markov model algorithm, covariance model, and machine learning methods, which often use riboswitch aptamers identified from seed alignments performed with sequences retrieved from the Rfam database.

Most of the tools described here are web-based. These instruments often impose restraints on the input sequence length and number of riboswitches that can be detected at once. They also rely on sequence or structural conservation of the aptamer to perform that analysis. Therefore, the aptamer prediction affects the detection of more variable riboswitches, such as the TPP and the Cobalamin, or smaller ones, such as the guanine riboswitch.

Several computational methods have been created to identify novel riboswitches and to characterize those that are already known. Amongst the methods that use primary sequences, the HMMER and Infernal tools stand out due to their ability to run locally, with the advantage of not having upload limits. Both methods utilize similar approaches by applying probabilistic models to sequence datasets to infer patterns.

DRD group (Havill et al., 2014) compared their server with RiboSW. The advantage of DRD compared to the other server is the ability to scan genome-scale files for riboswitches. In analyses of overall sequences obtained higher sensitivity (0.95) than RiboSW (0.85). DRD server was able to detect 64 instances that RiboSW was not identified, and 12 instances in which the opposite was true.

In 2009, Singh and collaborators compared the performance of HMM with other two CM web-based tools (Riboswitch finder and RibEx) in the search for ten riboswitches families on Rfam or RefSeq databases (Singh et al., 2009). Their results showed that HMM models run faster than CM and were more accurate than Riboswitch Finder and RibEx. The recently released version 1.1 of Infernal (Nawrocki and Eddy, 2013b) is reported to be 100 times faster than earlier versions and has been used for the identification of functional RNA homologs in metagenomic data (Nawrocki and Eddy, 2013a).

Although we agree that HMMER is faster than Infernal, our experience in searching for riboswitches in genomes does not corroborate their report of similar searches. In our case, most of the candidate regions found while using HMMER were not identified by Infernal, and the ones that were in common were discarded according to the following threshold filters: positive values for HMMER and values above the Rfam gathering score for Infernal, or E-values greater than 0.01 in both cases (unpublished data).

In a recent review on the computational prediction of riboswitches (Clote, 2015), Infernal was considered the most valuable tool to predict riboswitch aptamers mainly because the relevant Rfam database relies on Infernal for maintenance and extension. Some studies have used Infernal to identify riboswitches and other ncRNAs in archaeal metagenomes (Nawrocki and Eddy, 2013a; Gupta and Swati, 2016), species in the phyla Actinobacteria (Kang et al., 2017) and Proteobacteria (Leyn et al., 2014), Methanobrevibacter ruminantium (Nawrocki, 2014), Neisseria gonorrheae (Remmele et al., 2014), and Brassica rapa (Pang et al., 2015). Based on our knowledge, we also recommend the use of the Infernal program because this tool takes into account secondary structure conservation.

Most functional ncRNAs have secondary structures that are strictly related to their functions and that have been conserved during evolution. RNA secondary structures are defined by the arrangement of a set of base pairs non-covalently bound through hydrogen bonds, and can be considered a substructure of the global 3D structure. As it is difficult to obtain the experimental elucidation of RNA 3D structures and these structures are hierarchically folded, the computational prediction of RNA secondary structures provides key information to clarify the potential functions of RNAs. So far, a large number of computational studies have been carried out in the field of RNA secondary structure prediction. Prediction methods can be classified into two groups: single sequence analysis and multiple sequence analysis. Single sequence analysis is a traditional approach that consists of finding the structure with minimum free energy (MFE) of a single RNA sequence. On the other hand, the multiple sequence analysis has the advantage of providing higher prediction accuracy compared to single sequence analysis because it incorporates evolutionary information. Nonetheless, this approach is not always suitable, given that knowledge of a set of homologous sequences is required. Below, we list a few programs that perform these analyses.

Similar to what is observed in proteins, the functions of an RNA molecule depend on its structure and dynamics, which are determined by its nucleotide sequence. The number of computational methods and algorithms to predict the 3D structure of proteins from its amino acid sequence is vast. Unfortunately, only a few are available for the prediction of RNA structure (Chojnowski et al., 2014).

To determine a 3D configuration with the best possible accuracy, knowledge-based approaches are the most suitable. Comparative (or homology-) modeling, for instance, which is based on sequence similarity, works properly when there is an experimentally elucidated structure to be used as a template (Baker and Sali, 2001). However, RNA templates are rarely available.

Physics-based approaches are successful for the prediction of relatively small molecules. These tools are comparatively more appropriate for building models of RNA molecules with less than ~40 nt and display reasonable reliability for molecules up to ~80 nt. The prediction of larger molecules is possible, but the reliability of the model decreases as the length of the sequence increases (Magnus et al., 2014).

The combination of knowledge- and physics-based approaches resulted in the development of the so-called de novo folding methods, which is the assembly of the target structure from small fragments derived from other known structures (Bujnicki, 2006). Here, we compiled some programs using different approaches to predict RNA 3D modeling.

Riboswitches undergo conformational changes upon ligand binding and act as a switch at the transcriptional or translational levels. Given that riboswitches are functional entities that can undergo conformational changes, knowing their structures is of essential importance to understand the molecular mechanisms associated with their regulatory functions. Hence, predicting the structure of riboswitches can provide useful insight into the mechanism through which small molecules bind to RNAs, as well as shed light on how this process induces conformational changes in riboswitches.

The application of energy minimization methods for secondary structure prediction of the riboswitch expression platform domain is still limited as it involves conformational change. However, the prediction of this domain may be useful to support experimental assays. Barash and Gabdank (2010) predicted a single point mutation positioned in the non-conserved TPP riboswitch region responsible for transforming the terminator to an anti-terminator state.

The recent developments in the secondary structure prediction allow to include probing data, like SHAPE and DMS, for restriction and prediction of a structure with high accuracy (reviewed in Sloma and Mathews, 2015). Among the programs listed by us, the RNAstructure includes an option of incorporating the probing data as restraints.

Prediction of a single RNA sequence is still limited, especially when long RNA sequences (reviewed in Hamada, 2015) are involved. Comparative approaches using homologous sequence information increase the accuracy of as secondary structure prediction. In many circumstances, homologous RNA sequences of the target RNA sequence can be obtained, and it would be of interest to know the common secondary structure to those sequences (Gardner and Giegerich, 2004).

The common secondary structure is a fundamental element in riboswitch aptamers prediction. Programs such as Infernal, Riboswitch Finder, RiboSW, DRD, and Riboswitch Scanner use structural conformations for homologous searching. Secondary structure information is also crucial for tertiary structure prediction. In template-based methods, it assists in modeling mutations or structural changes, whereas in de novo methods, it allows for base pair constraints when creating 3D models. For instance, the MC-sym tool was used to construct models of the SAM-I riboswitch RNA segment by incorporating elements of the expression platform, and allowing the formation of an antiterminator (AT) helix in the 3D structures (Huang et al., 2013).

RNAComposer uses 2D restraint to create models and has provided positive results regarding the structural prediction of riboswitches. The server has been tested using a set often riboswitches containing pseudoknots and extensive tertiary interaction (Purzycka et al., 2015). In this set, nine examples were characterized with high accuracy and acceptable recovery of canonical and non-canonical base pairing and stacking. Input and output files of tertiary structure tools are shown in Figure 4.

Prompted by the increasing number of 3D RNA prediction framework methods, the RNA-Puzzles was started in 2012 (Cruz et al., 2012). RNA-Puzzles is a CASP-like (Moult et al., 2014) event in which collective blind experiments for the evaluation of 3D RNA structure prediction are carried out (Cruz et al., 2012; Miao et al., 2015). In the three rounds of RNA-Puzzles, predictions based on homology models already attained a high-level precision, providing useful insight to understand the RNA structure (Miao et al., 2017). Moreover, the prediction of ligand binding and subsequent conformational changes can also be described, but cannot be reliably guaranteed. Moreover, the prediction of ligand binding and subsequent conformational changes can also be described, but cannot be reliably guaranteed.

Gong et al. (2017) demonstrate in their review other approaches that aid the investigation of folding kinetics of aptamers and co-transcriptional folding kinetics using coarse-grained SOP model and, BarMap and helix-based computational approach, respectively. A new method StreAM-Tg (Jager et al., 2017) also allows analyzing structural transitions. This method gain insights into RNA dynamics based on a coarse-grained representation of RNA MD simulations.

Current modeling methods for template-based predictions have consistently reached a high accuracy level, i.e., now it is possible to model nearly all the structural details, provided that a reliable homologous structure is identified. Also, the ligand binding sites were readily inferred via homology (Miao et al., 2017). Different classes of riboswitches can be found in the RCSB PDB (Table 2), facilitating the use of model-based approaches such as ModeRNA and MMB.

In the case of targets without sequence homology with previously experimentally resolved structures, modeling quality strongly depends on the size of the target. The third edition of RNA-Puzzles provided models for two small RNAs—the ZMP riboswitch (60-nt) and L-glutamine riboswitch (61-nt)—with approximately 6 Å of RMSD with the crystallographic structure. Although the tools are less accurate, they can correctly predict the overall global folding. Thus, the larger the targets without a template—ydaO riboswitch (108-nt)—, the less accurate will the predictions be (10 Å best-case RMSDs).

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.