Riboswitches are bimodular RNAs that are made up of a ligand-binding region (an aptamer) and a domain that controls gene expression. They are usually located in the 5' untranslated regions of bacterial mRNAs, where they control the expression of the gene by binding a low molecular weight metabolite that triggers a conformational change in the RNA 23242526. In recent years, many of these genetic control elements have been discovered, and it has become clear that they are structurally and functionally highly diverse 2728. Riboswitches control gene expression at both the transcriptional and translational levels, and can act as 'on' or 'off' switches. The majority of riboswitches are negative control elements, and among these, the first catalytic riboswitch discovered - glmS 14 - employs the ultimate method of switching off gene expression: when it binds its cognate ligand it cleaves itself, thus destroying the function of the mRNA of which it is a part.

The biological function of other recently discovered catalytic RNAs is less clear. Using an ingenious in vitro selection scheme, Szostak and co-workers 29 recently discovered an HDV-ribozyme-like element in an intron of a human mRNA and have demonstrated its biochemical activity. In this scheme, a library of uniformly sized, small circular DNAs was used as templates for rolling-circle transcription; self-cleaving RNAs can thus be identified by the appearance of unit-length RNA fragments. Cedergren and co-workers identified and biochemically characterized hammerhead ribozymes in the genomes of schistosomes 30 and cave crickets 31, and, using database searching, our group recently identified novel examples of hammerhead ribozymes 32 and found two hammerhead sequences encoded at distinct loci in the genome of Arabidopsis thaliana that we have characterized as catalytically active in vitro and in vivo 33.

To carry out RNA-based chemical catalysis, some parts of the ribozyme molecule must adopt very precise relative positions and orientations. In addition to specific recognition, there must be dynamic mechanisms for substrate binding and product release. With the notable exception of the ribosome, present-day ribozymes act on the phosphodiester backbone linking two consecutive nucleotides. Although the catalytic processes of such reactions are basically similar, they can be achieved in diverse ways and, in addition, as chemical convergence is pervasive, ribozymes display a rich repertoire of architectures that position the reactants appropriately. Furthermore, the number of conserved nucleotides and their dispersion throughout the molecule vary considerably from one ribozyme to the other: for example, the hammerhead ribozyme and the group I introns have about the same number of conserved residues - around seven - although the latter can be up to four times as large 3435. The positions and relative dispositions of the conserved structural elements with respect to the beginning and end of the ribozyme motif also vary (Figure 2). Most families of ribozymes can be subdivided into classes distinguished by their highly non-homologous peripheral elements 363738. However, the three-dimensional architectures of the ribozyme cores belonging to the same family are expected to be similar because they are maintained by tertiary constraints which, despite the conservation of short sequence segments, can form in diverse ways.

The hammerhead ribozyme well illustrates the difficulties of identifying new ribozymes either experimentally or by in silico approaches. Indeed, an incomplete catalytic RNA fold, which did not include tertiary contacts between elements away from the catalytic site of cleavage and sequence conservation, was accepted for a long time, until the full hammerhead ribozyme was (re)discovered 394041. A recent crystal structure 42 shows how the presence of tertiary contacts between loops far removed from the catalytically conserved region induces conformational changes in the core that promote the active state of the ribozyme. Importantly, all those contacts involve networks of non-Watson-Crick base pairing with patterns of evolution unlike those of Watson-Crick base pairs 4043. Fully biologically active hammerhead ribozymes possess structural complexity and strict sequence requirements (Figure 3b), but because of the non-Watson-Crick pairings, this is not immediately apparent from the sequence alone. In contrast, because of its convoluted pseudoknotted topology based on Watson-Crick pairs, the HDV ribozyme reveals most of its complexity immediately (Figure 3a). Incomplete hammerhead ribozymes without peripheral elements and with low sequence and structural complexity display reduced catalytic activities. Indeed, in vitro evolution, starting from random libraries, produced structurally diverse ribozymes with low activity, which contained some hammerhead variants 44. Another experiment selecting in vitro for self-cleaving motifs with hammerhead-type biochemical activity 45 led to the conclusion that the hammerhead motif makes the most common ribozyme fold and suggested that this motif has had multiple independent origins. The long-range interactions were not considered in those two in vitro selection schemes, as their importance had not been recognized at the time. The sequences collected during the second selection scheme would enable optimization of hammerhead ribozyme activity.

A ribozyme with a new branching activity, GIR1, has recently been experimentally identified in slime molds 46. On the basis of its secondary structure, this ribozyme belongs to the group I intron family. It carries out the first cleavage step of a group II intron, however, leading to the formation of a small lariat with a 2',5'-linkage at the 5' end of the endonuclease mRNA of which it forms a part, thereby protecting the message from exonuclease degradation. Thus, in this case, a similar secondary structure scaffold is the basis for two ribozymes catalyzing different chemical reactions: activation of an internal O2'-hydroxyl group in the case of the new ribozyme compared with activation of an O3'-hydroxyl group of an external cofactor for the rest of the group I intron family (Figure 1c). This is yet another example of the fact that similar RNA sequences can assume two different folds and catalyze two different chemical reactions, as shown by Schultes and Bartel 47. Minor variations could convert a starting sequence into either of these highly active ribozymes, demonstrating that the evolving paths of RNA sequence can easily cross in sequence space. Similarly, RNA folds recognizing different ligands may be very close in sequence space 48: for example, a small series of 'neutral' mutations (that is, mutations that have no effect on secondary structure) transformed a flavin-binding aptamer into a GMP-binding aptamer 49. Extensive networks of neutral variation in sequence space interconnect RNA regions with similar function and structure 5051, as confirmed by the recent elucidation of more three-dimensional RNA structures (see 4352 for reviews).

It is now recognized that the most common RNA-RNA binding contact is the so-called A-minor motif 53. This occurs between two contiguous adenines in one partner RNA and the shallow/minor groove side of two stacked Watson-Crick pairs in the other. An analysis of tertiary contacts shows that the contiguous adenines can originate from a variety of local environments (for example, bulging, apical or internal loops) and that the only molecular recognition requirement in the receptor RNA is the presence of two Watson-Crick base pairs 5455. In other words, coupled to the vast shape space accessible through mutations neutral for secondary structure, there are weak but crucial sequence constraints imposed by the tertiary contacts. In RNA architectures, the additional structural constraints originate from the topology of the secondary structure (junctions of helices, number of base pairs within helices, and so on). In short, RNA sequences (and thus their structure and function) are characterized by neutrality at all levels from molecular recognition between motifs to secondary structure and three-dimensional architecture.

This complex interplay between sequence conservation and neutral evolution on the one hand, and diversity in folds despite conservation in interaction protocols on the other, is central to the theoretical and experimental difficulties in identifying key regulatory RNA sequences from genomic sequence. For example, group I introns are characterized by an invariant core onto which is grafted a variety of peripheral elements 3656. Long-range contacts between those non-homologous peripheral elements are necessary for biological activity. All known group I introns contain a tertiary contact between two specific paired-segment regions (regions 5 and 9; Figure 4). However, the examples that have been crystallized (for a review, see 57) show that in each case, the contacts are achieved through different local topologies (Figure 4), each with different sequence constraints. Interestingly, in a first attempt at modeling the lariat-forming group I-like intron, GIR1, from slime molds, it was not possible to construct the usual intramolecular contacts between regions 5 and 9 58.

Are there more ribozymes that catalyze 2',5'-phosphodiester bond formation or cleavage to be discovered? Scattered evidence of the occurrence of 2',5'-bonds exists throughout the literature. A 2',5'-phosphodiester bond was observed in vitro 59 and in vivo 60 during circularization of the genome of the peach latent mosaic viroid, and the HDV ribozyme, unlike the hammerhead ribozyme, has been shown to cleave 2',5'-linkages efficiently 61.

Novel catalytic RNA entities can, in principle, be looked for either by database searches using defined consensus motifs from a given ribozyme or by experimentally testing candidate RNAs for biochemical activity. Both approaches have advantages and disadvantages. Database searches require RNA sequence alignments (as produced, for example, by Rfam 62) coupled with covariance analysis 6364656667. The quality of the sequence alignment is central to this process, however, and not many databases are as carefully hand-curated as the RNase P database 68. In database screening, the definition of what we consider to be the consensus motif of a given catalytic RNA is crucial. Even if a catalytic RNA motif is well defined, searches are complicated by the requirement to combine a complex assembly of structural (hairpin) and sequence information, which prevents simple solutions such as purely sequence-based homology searches. Generally, the tools available adequately identify isolated hairpins 69. Given a pattern description for a catalytic RNA motif, several programs, such as PatScan 70 or RNAMOT 71, can be used to screen the public databases. Hits from such searches require further analysis, and initially, a calculation of the secondary structure is necessary - although usually not sufficient. A secondary structure, calculated using a program such as RNAfold 72, is predictive if the required helical elements of the RNA motif under consideration will form in the hit sequence. Secondary-structure prediction programs have difficulty in accurately predicting large structures, however, and can also produce vast numbers of alternative structures when scanning whole genomes 7374.

For individual sequences found in a database search, a test of their particular biochemical activity (Figure 1) might be sufficient. However, functionally similar RNA molecules frequently exhibit numerous and highly divergent sequence insertions or deletions that interrupt the pattern of secondary-structure motifs and render the computer description of a given motif inadequate for finding sequences with similar activity. Furthermore, the use of pattern-description programs is incomplete if the complexity of the RNA structure - which goes way beyond the Watson-Crick base pairing 75 - is not taken into consideration. These issues, and whether the additional, essential tertiary interactions of a given RNA motif will form, can be addressed by a combination of comparative analysis of similar ribozymes with isostericity matrices, which give the geometrically equivalent base pairs for each particular type of base-base interaction 76. All pairwise base-base interactions present in nucleic acids have been classified into 12 families, where each family is a 4 × 4 matrix of the bases A, G, C, and U 75. This classification allows the deduction of all possible geometrically equivalent base pairs in a given family. The isostericity matrices have been verified for several RNA motifs using structural alignments anchored in crystal structures 77. Thus, for assumed structurally homologous positions in an RNA motif, one can compare the resulting pairwise interactions with the known isostericity matrices to assess the validity of an RNA motif assignment in an alignment 78. As this type of analysis is an iterative process, it is worth noting that it might also lead to refinement and extension of the pattern of the consensus motif that the search was started with. If applied to large assemblies of sequence information, as has been done for the kink-turn and C-loop RNA motifs 77, this approach allows a broader description (the comprehensiveness of which is currently unknown) and refinement of a given motif.

The analysis of co-variation of nucleotides in sequence alignments underlies most manual or automated secondary-structure determination. However, high sequence conservation (which is usually considered a marker for conservation of function) leads to serious ambiguities and difficulties in deriving secondary structures. The catalytic riboswitch glmS is a good example: the crystal structure 79 presents a different secondary structure from that deduced from sequences. The new helices involve pairings between segments, conserved at more than 95% in sequence, and thus giving no co-variation signal. The requirement for a well-defined RNA motif in database searches is also an intrinsic limitation of this approach.

While novel genomic locations of a known catalytic RNA can be identified from sequence similarities, novel activities cannot be so readily discovered. For this purpose, a recently introduced in vitro selection scheme 29 can be applied. Interestingly, from the human genome, a close variant of a known catalytic RNA motif was selected and characterized as a HDV-ribozyme-like sequence rather than a new catalytic RNA. It is intriguing that these sequences were discovered by their biochemical activity and not by in silico approaches (as an active HDV ribozyme can be made by both the 'genomic' sequence and its complementary sequence despite its intricate pseudoknotted secondary structure). Furthermore, additional, as-yet structurally uncharacterized, sequences were reported 29, and so this new selection scheme might actually have identified new self-cleaving RNA entities. The selection scheme described in 29 uses DNA minicircles that cover the genomic sequence of a given organism as the templates for rolling-circle RNA transcription (Figure 5). It can thus readily monitor RNA self-cleavage, but other activities will be missed. Thus, to assess the prevalence of a given catalytic RNA motif, the combined approach using sequence and three-dimensional structure information described above is most suitable, while novel in vitro selection schemes might be designed to discover activities other than RNA cleavage in a given organism.

As pointed out earlier, most of the reactions known to be naturally catalyzed by RNA (Figure 1) involve the breakage or formation of 3',5' (and occasionally 2',5') phosphodiester bonds. RNA has the potential to catalyze other chemical reactions, however. As well as peptide formation in the ribosome, Diels-Alder cycloaddition 80 and Michael addition 81 can be catalyzed by RNA, as shown by in vitro Darwinian evolution. Thus, reactions catalyzed by RNA in nature might be more diverse than currently known. The discovery of such activities is likely to be serendipitous and made by keen observers of RNA molecular behavior.

New small or large noncoding RNAs are regularly being discovered in both bacteria and mammals. Recent evidence shows that most of the mammalian genome is transcribed in complex patterns, producing tens of thousands of novel transcripts 8283. Novel RNAs are regularly predicted on the basis of their sequence conservation or secondary-structure elements 84858687. But these predictions do not utilize information on the non-Watson-Crick base pairing or tertiary structure so crucial to the activity of many ribozymes, and, as discussed above, these features are often not well conserved in the sequence. Nor do the predictive algorithms used give any indication of what the RNA function might be. Vertebrate genomes contain a large number of conserved noncoding elements (CNEs) or ultraconserved elements 8889, whose biological functions and mechanisms of action remain to be established. The evidence for transcription of most of these conserved elements is, however, still scanty 899091. In any case, the recent additions to the list of natural catalytic RNAs indicate that there are likely to be many more to come; new algorithms will be required that use all available information to identify and classify them.