Databases and resources for human small non-coding RNAs

In 2001, three groups published independent reports on the discovery of a new class of small non-coding RNAs (sRNAs), which were named micro-RNAs (miRNAs) [1–3]. These comprise a large family of small, ~22 nucleotide-long, non-coding RNAs that have emerged as key players in post-transcriptional gene regulation [4]. Subsequent years have witnessed the discovery of many new types of sRNAs. In humans, apart from the hundreds of miRNAs detected so far, there are also many endogenous small interfering RNAs (endo-siRNAs)[5] and piwi-interacting RNAs (piRNAs)[6, 7]. These and other short non-coding RNA molecules collectively are called 'sRNAs'. They are generally short (~18-30 nucleotides [nt]); do not code for proteins; exert their function as RNA molecules generally combined with protein factors; and represent a substantial portion of the RNA output of cells. Moreover, sRNAs encompass a diverse, widespread and basal regulatory system: they are known to regulate genes and genomes at different levels, including chromatin structure, transcription, RNA stability and translation [8–10]. Furthermore, they can act as activators or inhibitors and their disruption has been linked to disease [11]. The explosion of information on sRNAs makes necessary its organisation--in terms of their biogenesis, expression properties and functional characteristics--into public databases.

Although it is generally assumed that a single precursor miRNA molecule leads to a single functional miRNA, there is evidence that precursors can be processed with heterogeneous ends, giving rise to isomiRs [33]. A recently published database collects isomiR sequences from high-throughput sequencing of miRNAs from human 293T cells [34] (Table 1). This database allows one to retrieve all reads and isomiRs assigned to a specific miRNA, with results linked to miRBase and Ensembl [35].

The endo-siRNAs, which were first observed in plants, are also a very abundant class of sRNAs [10]. They share some properties regarding biogenesis and function with miRNAs [36, 37]. More recently, other sources of endo-siRNAs have been identified, such as convergent mRNA transcripts and sense-antisense pairs [10]. To date, there is only one database specific for siRNAs, siRNAdb,[38] which contains the collections of endogenous and exogenous siRNA molecules from the literature that have been experimentally verified. Moreover, siRNAdb also includes predicted siRNAs based on a combination of computational prediction methods [39–41]. Additionally, the database includes information about targets and experimental sources. A set of target predictions is also available for the non-experimentally verified siRNAs.

In 2006, a new abundant sRNA species of approximately 30 nt was observed in extracts of total RNA from mouse testes [42]. These somewhat larger types of sRNA, called piRNAs, were found to exert their function most clearly in the germline [43] and possibly in cancer cell lines [44, 45]. In contrast to miRNAs and endo-siRNAs, they derive from single-stranded precursors [10]. The only database exclusively dedicated to piRNAs is piRNAbank,[46] which contains piRNA sequences collected from Genbank [47] and published data, indexed through unique identifiers linked to NCBI and with additional information on gene name and genomic position.

Small nucleolar RNAs (snoRNAs) are longer types of non-coding RNA (ncRNA; 60-300 nt in length), but are also regarded as sRNAs. These are a highly evolutionarily conserved class of RNAs which function mostly in the nucleolus and participate in the chemical modification of other RNAs, mainly ribosomal RNAs (rRNAs) [48]. The snoRNA-LBME-db database [49] was created to collect the available information on human snoRNAs. Besides sequence information, it also includes predicted RNA targets and potential base-pairing interactions with these targets. The data have been collected from the literature and can be accessed by the name of the snoRNA or can be downloaded in their entirety. This database includes the Cajal body-specific RNAs (scaRNAs). This is a class of nuclear sRNAs similar to snoRNAs which accumulate specifically in Cajal bodies and guide modifications of small nuclear RNAs (snRNAs). The latter are also considered to be sRNAs and are part of the spliceosome [50]. All of these nuclear sRNAs and other structural RNAs are present in the Rfam database (Table 1), which is a general collection of structured RNA families [51] represented by their sequences and a structural model. Rfam also includes models for miRNAs.

Recent advances in deep-sequencing technologies have yielded a large number of short RNA sequences,[16, 37, 52, 53] leading to an exhaustive detection of both known and novel sRNAs. In order to facilitate access to deep-sequencing data, the Short Read Archive (SRA) of NCBI [54] and the European Nucleotide Archive [55] (ENA) http://www.ebi.ac.uk/ena/ have been created. They provide centralised access to published sequencing data, including sRNAs, and, more notably, all the data are accessible through powerful search tools. Many of these sRNAs, however, still await classification. They vary significantly in origin and structure and their function is often unknown, which makes it difficult to develop databases for their storage and analysis.

A growing number of public databases are now available for sRNAs obtained from deep-sequencing experiments that classify them in terms of the experimental origin, function (if known) and genomic localisation. These databases try to integrate novel heterogeneous data and sometimes include curated data. The NONCODE database [56] (Table 1) provides organised information for snoRNAs, piRNAs, miRNAs, scaRNAs and other ncRNA classes present in GenBank, classified according to cellular process. Similarly, RNAdb [57] (Table 1) provides access to data from snoRNA-LBME-db, miRBase, the FANTOM project [58] and the H-invitational project [59]. All the datasets are divided according to the main sRNA classes. Additionally, RNAdb contains predictions based on comparative methods. All RNAdb data are available for downloading.

The deepBase [60] database (Table 1) contains the most heterogeneous collection of sRNAs from deep-sequencing data from different libraries (Table 2) [19, 25, 27]. The deep-sequencing data have been mapped to the human genome assembly and annotated according to various sRNA classes, which are defined by the location of the mapped sRNAs: non-coding RNA-associated small RNAs (nasRNAs); sRNAs that overlap promoter regions (pasRNAs), which overlap with the transcription start site (TSS) of genes; exon-associated small RNAs (easRNAs),[60] which overlap with exons from RefSeq [61] genes; and repeat-associated small RNAs (rasRNAs),[60] which overlap with repeat elements from the University of California, Santa Cruz (UCSC) genome browser [62]. Additionally, deepBase includes: all known miRNAs from miRBase and snoRNAs from snoRNA-LBME; novel predicted miRNAs and snoRNAs; and RNA clusters built from the sRNAs that are proximal in the genomic sequence.

Finally, fRNAdb [63] (Table 1) is a searchable database, with functional and genomic annotation for ncRNAs of various lengths from snoRNA-LBME-db, miRBase, the FANTOM and the H-invitational projects, NONCODE and RNAdb. At the same URL, the authors provide visualisation of the annotated and predicted ncRNAs with a mirror of the UCSC browser.

The sRNA world is expanding: it includes many new sRNA sequences that share many features but also differ in others, which makes it difficult to integrate them into the same resource. Despite the current wealth of public databases, there are many recently discovered sRNA classes that are not present in any specific or general database. Below, we enumerate some of the datasets that we cannot find in the previously mentioned databases, and which are only available directly from the publications.

Deep sequencing has led to the detection of at least three new classes of sRNAs linked to the region proximal to the promoter and TSS of genes: promoter-associated sRNAs, which are hypothesised to result from the transcription of independent capped short transcripts or as cleavage products of longer RNAs and which were specifically called 'PASRs';[20] TSS-associated RNAs (TSSa RNAs),[19] included in deepBase and similar to PASRs (Table 2); and transcription initiation RNAs (tiRNAs)[21] (Table 2), which are predominantly 18 nt in length and originate mostly from the region downstream of the TSS, possibly from the backtracking of RNA polymerase II (RNAPII) at the start of transcription [64]. Interestingly, a new class of nuclear sRNAs, also of about 18 nt in length, recently has been found to be associated with the 5' splice sites of genes (spliRNAs)[22] (Table 2). PASRs, TSSaRNAs and tiRNAs datasets are available at NCBI GEO, and spliRNAs at the FANTOM web page [65]. A different class of sRNAs, also with positional biases, are the termini-associated short RNAs (TASRs)[23] (Table 2) and their antisense counterparts (aTASRs),[24] obtained via a change of the sequencing protocol. These are found to be located antisense of the 3'-untranslated regions of genes. They are only available from the SRA at NCBI (Table 2).

Interestingly, some authors have proposed that some miRNAs and other sRNAs could stem from the processing of snoRNAs by a still unknown mechanism [66–68]. Moreover, there is also some evidence from deep-sequencing experiments that other sRNAs also could originate from transfer RNAs (tRNAs) [28–30]. The sRNAs derived from snoRNAs [31, 68] and tRNAs [28–30] (see Table 2) are not available in any of the databases mentioned above.

Additionally, there are datasets of sRNAs expressed in specific cell lines, like novel miRNAs obtained from human leucocytes [26] (Table 2), which have not yet been integrated into any database. These examples of new classes of sRNAs display diverse positional and length biases and their function is mostly uncharacterised. They are therefore difficult to integrate with each other and within the existing sRNA databases.

The sRNA data currently available--especially those obtained from deep-sequencing experiments--are very heterogeneous. Ideally, these sRNA datasets should be compiled and integrated together, thereby facilitating functional and computational downstream analyses. One of the main problems with the integration of sRNA data is the lack of functional information. While miRNAs and siRNAs are well characterised in terms of how they interact with their targets,[10, 36, 69] nothing is known so far about the mode of action of many of the new classes of sRNAs. This is further complicated by the fact that sRNAs may have multiple functions. For instance, exogenous siRNAs may affect gene expression [70] and also splicing [71] by inducing chromatin changes. Likewise, miRNAs can also affect gene expression through a similar pathway affecting chromatin,[72] and other sRNAs originating from sense-antisense transcription can also trigger similar mechanisms [73].

For many of the new sRNA classes, then, we still lack a function and a target definition. Interestingly, the Argonaute (AGO) family of proteins is known to be essential for the function of miRNAs, piRNAs and endo-siRNAs [74]. It may be possible, therefore, that some of the new classes of sRNAs exert their function through one or more members of the AGO family as well, and may interact with DNA or RNA in a similar way. Further evidence on the proteins that interact with the new classes of sRNAs will help in defining their function and possible targets.

Another important issue for the classification of sRNAs is the characterisation of their biogenesis. Interestingly, some of the new sRNAs have similar positional biases. The sRNAs localised near the TSS of genes have been linked to transcription activity [21]. Likewise, TSSa RNAs and PASRs have been associated with transcription and processing of other RNAs [20]. These similarities may be indicative of a common biogenesis mechanism. Furthermore, many snRNAs are found to be independent of the processing machinery responsible for the biogenesis of miRNAs [75]. Thus, there must be other factors involved in the generation of these sRNAs. Recent experiments have shown that there is considerable endonucleolytic activity associated with still undetermined proteins [76]. These proteins may be the key factors responsible for the biogenesis of some of the novel classes of sRNAs.

In summary, high-throughput methods--and especially deep-sequencing technologies--provide a unique opportunity to explore the wealth of RNA species in diverse cellular contexts. Different types of RNA have been characterised using these technologies, giving rise to new sRNA categories. Further experiments will be necessary to determine the function and biogenesis of these new molecules. Considering the large amount of data that are being generated, we are just touching the tip of the iceberg regarding the sRNA world. The coming years will witness new and exciting developments in molecular and computational biology in the area of sRNAs.