The Landscape of long non-coding RNA classification

Abstract

Advances in the depth and quality of transcriptome sequencing have revealed many new classes of long non-coding RNAs (lncRNAs). lncRNA classification has mushroomed to accommodate these new findings, even though the real dimensions and complexity of the non-coding transcriptome remain unknown. Although evidence of functionality of specific lncRNAs continues to accumulate, conflicting, confusing, and overlapping terminology has fostered ambiguity and lack of clarity in the field in general. The lack of fundamental conceptual un-ambiguous classification framework results in a number of challenges in the annotation and interpretation of non-coding transcriptome data. It also might undermine integration of the new genomic methods and datasets in an effort to unravel function of lncRNA. Here, we review existing lncRNA classifications, nomenclature, and terminology. Then we describe the conceptual guidelines that have emerged for their classification and functional annotation based on expanding and more comprehensive use of large systems biology-based datasets.

The non-coding RNA universe

The classic view of the transcriptome landscape and its mRNA-centric paradigm for transcript annotation has undergone a fundamental change [1, 2]. The ENCODE project estimates that (mostly non-coding) transcripts cover 62–75% of our genome [3], and contribute greatly to the overall estimate of 80% potentially functional sequence in our DNA [4]. Similarly, RNAseq studies show that transcripts from these non-coding regions dominate the population of non-ribosomal non-mitochondrial RNAs in a human cell [5]. Non-coding RNAs (ncRNAs) have emerged as a major source of biomarkers [6–10], targets for therapeutics [8, 11], and potential explanations for the function of non-coding GWAS variants [12]. Constant expansion of the evidence for broad functionality of ncRNAs [13], in processes ranging from heritable epigenetic change [14] to species-specific changes in cognition [15], may finally answer the long standing question of the role of non-coding DNA in eukaryotic biology [16, 17].

Although there has been an emphasis on the annotation and classification of ncRNAs with properties similar to those of protein-coding mRNAs, the vast majority of genomic space used for RNA production remains under-explored. Moreover, although some ncRNAs share properties with coding mRNAs, such as splicing and polyadenylation [18, 19], sequence conservation [18, 19], and export to the cytosol [20], many others do not, highlighting the differences in the functionalities of coding versus ncRNAs [5, 11, 13, 21–24].

The state of non-coding annotation is still in its early days, but the field now has sufficient perspective to establish a logical conceptual framework for the classification of the universe of transcripts that emanate from non-coding genomic regions. New methodologies for integrated classification have accompanied a massive expansion of global transcriptome datasets, particularly from genomics consortiums such as FANTOM [25], ENCODE [3], and GTEx [26]. Methods for grouping RNA sequencing (RNA-seq) (Box 1) reads into a single transcribed region have improved rapidly [27–30]. Progress has also been made in machine learning approaches aimed at the integration and biological interpretation of diverse datasets [29, 31, 32]. All this has culminated in widely used sets of annotations of lncRNAs, such as the one provided by the GENCODE consortium [33], and others [31, 34–37].

Box 1. Overview of high-throughput technologies used to detect and quantify ncRNAs.

RNA sequencing (RNA-seq): currently, one of the most commonly used procedures in transcriptome profiling. Typically, RNA is converted into cDNA using random hexamers followed by massive random sequencing of the resulting cDNAs using next-generation sequencing technologies. As a result, millions of short sequence tags can be generated per experiment. Subsequent mapping of the tags reveals the genomic position encoding the RNA and its relative mass in the cell. The procedure is suitable for various aspects of transcriptome research: RNA mapping, quantitation, alternative splicing analysis etc.

Messenger RNA sequencing (mRNA-seq): RNA-seq on polyA+ fraction of RNA, often synonymous with RNA-seq.

Direct RNA sequencing (DRS): sequencing of native RNA, without library preparation including cDNA conversion step [143], has been successfully used to sequence native polyA+ and identify alternative polyadenylation sites. DRS is particularly useful in applications where artifacts of reverse transcription are undesirable, such as precise strand of origin determination, and in applications that deal with minute amounts of nucleic acids such as single-cell applications. Theoretically, it can provide multiple tags per molecule, however so far it has been used in applications that provide a single tag per molecule at the polyadenylation site.

Cap assisted gene expression (CAGE): a transcriptome profiling procedure that targets RNAs with a 5′-cap [144]. CAGE generates short (typically 27 nt) sequence tags from 5′ ends of such RNAs, with one tag per RNA molecule. It enables accurate mapping of 5′ ends this subset of RNAs.

Serial analysis of gene expression (SAGE): targets polyadenylated messages and generates a single internal (typically close to the 3′ end) tag per RNA molecule [145].

Paired-end tag (PET): also targets polyadenylated RNAs and generates a tag that combines information on 5′ and 3′ ends of the same RNA molecule [146].

Rapid amplification of cDNA ends (RACE): an ‘outward’ PCR-based method designed to identify sequences connected to a given region, which can be used in conjunction with NGS or microarrays, for deep transcriptome profiling of a specific locus [44].

Targeted RNA sequencing: selection of RNAs from a locus of interest using tiling microarrays followed by RNA-seq to achieve the same goal [45].

GRO-seq: A typical RNA profiling experiment measures steady-state levels of RNA. On the other hand, GRO-seq [134] combines nuclear run-on experiments and NGS analysis to provide information on transcription competent RNA polymerase complexes.

Here, we review existing lncRNA classes and then describe the conceptual guidelines that have emerged for their classification and functional annotation based on the expanding and more efficient use of large systems biology-based datasets. This framework endeavours to guide researchers in the classification of ncRNAs and interpretation of next-generation sequencing (NGS) data, especially in non-coding portions of the genome.

Criteria and features of existing classes and categories of lncRNAs

The classification of the great majority of lncRNAs relies on the empirical attributes originally used to detect them (Table 1, Figure 1). This reflects their short history relative to protein-coding genes, and provides a convenient basis by which to classify these un-characterized RNA species.

Table 1.

Different known classes of lncRNAs

Category	Abbreviation	Reference	Specific Examples
Classification based on transcript length
Long non-coding RNA	lncRNA	[38, 39]
Long-intergenic non-coding RNA; large Intervening Non-Coding RNA, long-intervening non-coding RNA;	lincRNA	[18]	ANRIL [117], H19 [149], HOTAIR [18], HOTTIP [150], lincRNA-p21 [151], XIST [152], Paupar [153]
Very long intergenic non-coding RNA	vlincRNA	[29]	HELLP transcript [42], Vlinc_21, vlinc_185, vlinc_377, vlinc_500 [29]
macroRNA		[28, 154]	Airn, Gtl2lt, KCNQOT1, Lncat, Nespas (reviewed in Ref [154]), STAiR1 [28]
Promoter-associated long RNA	PALR	[38]
Classification based on association with annotated protein-coding genes
Intronic ncRNA; Stable intronic sequence RNA; totally intronic RNA, partially intronic RNA	sisRNA, TIN, PIN	[49, 50, 54] additional references in the text
Circular intronic RNAs	ciRNAs	[55]
Sense ncRNA		[44]
Natural antisense ncRNA	asRNA, NAT	[57]	BACE1-AS [155], aHIF [156], Tsix [157]
Mirror antisense		[44, 67]	Globin antisense [67]
Exonic circular RNAs	ecircRNAs	[62]	cANRIL [118]
Chimeric RNAs, trans-spliced RNAs, exon juxtaposition		[44, 63–65]
Standalone ncRNAs made from 3′UTRs	uaRNA	[60]
Chromatin-interlinking RNA	ciRNA	[68]
Transcription start site-associated RNAs	TSSa-RNAs	[158]
Classification based on association with other DNA elements of known function
Enhancer-associated RNA	eRNA	[159]
Promoter-associated long RNA	PALR	[38]
Upstream antisense RNA	uaRNA	[160]
PROMoter uPstream Transcript	PROMPT	[89]
Telomeric repeat-containing RNA	TERRA	[161]
Classification based on protein-coding RNA resemblance
mRNA-like noncoding RNAs	mlncRNAs	[18]
Long-intergenic non-coding RNA; large Intervening Non-Coding RNA, long-intervening non-coding RNA;	lincRNA	[18]	ANRIL [117], H19 [149], HOTAIR [18], HOTTIP [150], lincRNA-p21 [151], XIST [152]
Classification based on association with repeats
C0T-1 repeat RNA		[162]
Long interspersed nuclear element	LINE1/2	[163]
Transcribed endogenous retroviruses		[81]
Expressed Satellite Repeats		[164]
Non-coding RNA driven by promoters within repeats	vlincRNAs, NASTs	[29, 76]	Vlinc_21, vlinc_185, vlinc_377, vlinc_500 [29]
Polypurine-repeat-containing RNA	GRC-RNA	[165]
transcribed pseudogenes		[83]	PTENP1 and KRASP1 [86]
Classification based on association with a biochemical pathway or stability
Nrd1-unterminated transcript	NUT	[166]
miRNA primary transcripts		[167]	H19 [168]
piRNA primary transcripts		[169]
Cryptic unstable transcript	CUT	[88]
PROMoter uPstream Transcript	PROMPT	[89]
Xrn1-sensitive unstable transcript	XUT	[91]
Stable Uncharacterized Transcript, Stable Unannotated Transcript	SUT	[92]
Classification based on sequence and structure conservation
Transcribed-Ultraconserved Regions	T-UCR	[95]	UCR106 [95]
Hypoxia-induced noncoding ultraconserved transcript	HINCUT	[100]
Long-intergenic non-coding RNA; large Intervening Non-Coding RNA, long-intervening non-coding RNA;	lincRNA	[18]	HOTAIR [18], HOTTIP [150]
RNA-Z regions		[97]
EvoFold regions		[98]
Classification based on expression in different biological states
Long stress-induced non-coding transcript	LSINCT	[101]
Hypoxia-induced noncoding ultraconserved transcript	HINCUT	[100]
Non-Annotated Stem Transcript	NAST	[76]
Classification based on association with subcellular structures
Chromatin-associated RNA	CAR	[102]
Chromatin-interlinking RNA	ciRNA	[68]
Nuclear bodies associated RNAs		[170]
PRC2 associated RNAs		[19, 103]
Classification based on function
Long noncoding RNAs with enhancer- like function; ncRNA-activating	ncRNA-a	[108]	ncRNA-a7 [108]
miRNA primary transcripts		[167]	H19 [168]
piRNA primary transcripts		[169]
Competing endogenous RNA	ceRNA	[109]	PTENP1 and KRASP1 [86]

Figure 1. Schematic diagram illustrating various classes of ncRNAs.

Three hypothetical loci are shown. Protein coding exons are shown as green (locus 1) or yellow boxes (locus 3). Locus 2 signifies a pseudogene of locus 1. Regulatory regions of locus 1 are shown in purple (promoter) and magenta (enhancer). Repeats are denoted by brown boxes. Lines with arrows represent ncRNAs. CAR: chromatin-associated RNA. ceRNA: Competing endogenous. RNA ciRNA: chromatin-interlinking RNA (grey) or circular intronic RNA (green). ecircRNA: exonic circular RNA. eRNA: enhancer-associated RNA. lincRNA: long intervening non-coding RNA. ncRNA-a: activating non-coding RNAs. PALR: promoter-associated long RNA. PIN: partially intronic RNA. TIN: totally intronic RNA. TSSa-RNA: transcription start site-associated RNA. T-UCR: Transcribed Ultraconserved Regions. uaRNA: 3′UTR-derived RNAs. vlincRNA: very long intergenic non-coding RNA. The role depicted here for CARs and ciRNAs in stabilizing a chromatin loop is hypothetical.

Classification based on transcript length

The length estimate of ncRNAs serves as the most commonly used attribute for their classification. Typically, a threshold of 200 bases separates long from short ncRNAs [38, 39] (Table 1). Often, our knowledge is limited to sequence reads mapped to a ‘region of transcription’, and even with improvements in NGS read length [40], this will probably continue for the foreseeable future.

Building transcribed regions based on RNA-seq profiling of total RNA (rather than the polyA+ fraction, see below) led to the discovery that intergenic space encodes thousands of very long intergenic non-coding RNAs (vlincRNAs), whose primary transcripts can range in length from 50 kb to 1MB [28, 29, 41]. Spanning at least 10% of human genome [5, 29], vlincRNAs have been implicated in important biological processes such as pluripotency [29], cancer [28, 29], apoptosis [29], cell-cycle progression [28, 42], and cellular senescence [41].

Classification based on association with annotated protein-coding genes

This commonly used attribute (Table 1, Figure 1) serves as the foundation of the GENCODE classification of lncRNAs [33]. It underlies the logical challenge of overlapping non-coding and coding transcripts at a given locus - called “transcriptional forests” by the FANTOM consortium [43]. Targeted methods (Box 1) based on Rapid Amplification of cDNA Ends (RACE) experiments [44] and RNA-seq [45] indicate that transcriptional forests constitute a general feature of the human genome. A prominent category of ncRNAs has emerged from these transcriptional forests composed of sense ncRNAs that overlap coding mRNAs on the same strand and share some sequence with the latter, yet do not encode proteins [44, 46–48]. This category includes unspliced sense partially intronic RNAs (PINs) [49], and spliced transcripts that combine exons from coding and non-coding regions of a gene [47, 48]. GENCODE recognizes the existence of such spliced lncRNAs in their “sense overlapping” biotype [33].

The PIN and “sense overlapping” categories allow for overlap between lncRNAs and exons of a protein-coding gene. However, a protein-coding gene can produce lncRNAs found exclusively in its introns, known as totally intronic RNAs (TINs) [49] (Table 1, Figure 1). TINs make up the majority (~70%) of all non-coding (non-rRNA) nuclear-encoded RNA and 40–50% of all cellular (non-rRNA) RNAs by mass, as established by single-molecule sequencing [50]. Evidence that large numbers of introns encode standalone RNAs originally came from microarray expression profiling [49, 51] and in silico analysis of expressed sequence tag (EST) databases [49, 52]. Even genomes as compact as those of human viruses can encode functional intronic RNAs [53]. Large numbers of standalone intronic RNAs recently found in Xenopus oocytes [54] and mice [50] support the conclusion that introns encode functional ncRNAs on a global scale. Some of these transcripts likely represent circular intronic ncRNAs (ciRNAs) (produced from introns that escape debranching) that can accumulate in cells and regulate expression of their parent genes [55] (Figure 1). Overlap on the opposite DNA strand from their associated protein-coding gene represents another frequently used attribute of lncRNAs. These natural antisense transcripts (NATs) (Figure 1) occur in 50–70% of all protein-coding genes [56, 57].

ncRNAs can also be composed solely of sequences of exons of protein-coding mRNAs (Figure 1, Table 1). For example, transcript cleavage followed by post-transcriptional 5′-cap addition [58, 59] can result in production of stand-alone ncRNAs from various parts of mRNAs [58], notably 3′ UTRs [60]. In fact, the type 0 variant of the cap structure may associate with the post-transcriptionally capped 5′ ends [61]. Additional cellular processes could produce this type of ncRNA, such as the reverse-splicing implicated in the production of circular exonic RNAs [62], trans-splicing leading to production of chimeric RNAs [63, 64], exon juxtaposition [65], and presumed RNA copying, leading to production of “mirror antisense” transcripts [44, 66, 67]. Finally, RNAs whose sequences have features of bona fide coding transcripts may have other roles as revealed by the class of chromatin-interlinking RNAs (ciRNAs). These RNAs participate in maintaining interphase chromatin configuration and mostly include spliced transcripts with long 3′UTRs [68].

Classification based on association with other DNA elements of known function

Notable classes of such RNAs include enhancer- and promoter-associated long RNAs (Table 1, Figure 1). These long RNAs are involved in linking the dynamics of nuclear architecture, chromatin signalling plasticity, and transcriptional regulation [69]. Interestingly, enhancers that give rise to RNA species have greater likelihood of functionality in reporter assays than those that do not [70], arguing for a functional, rather than spurious, link between RNA and this type of genomic element.

Classification based on mRNA resemblance

As mentioned previously, research has focussed on ncRNAs with a spliced structure, conserved sequence, and a polyA tail [18, 19, 34, 35, 71] (Figure 1). In fact, lncRNAs annotated by GENCODE - even those solely confined to intronic sequences - represent primarily spliced transcripts [33]. These features were used to identify thousands of transcripts in mouse [18] and human [19], called long intervening ncRNAs (lincRNAs) [18]. This approach has revealed many important functional lncRNAs, such as HOTAIR, which mediates gene silencing by facilitating localization of the epigenetic repressor Polycomb repressive complex 2 (PRC2) to its target sequence [72]. Expression analysis of ~10,000 human lincRNAs across 1,300 tumour samples using microarrays revealed hundreds non-coding transcripts potentially driving four different cancer types [73]. Numerous other studies have implicated lincRNAs in human development and disease [74]. As an indication of their functionality, expression analysis of 11 tetrapod species found 2,508 lincRNAs expressed in at least three species and originating more than 90 million years ago [71].

Classification based on association with repeats

About half of the human genome consists of repeats of various categories, and many ncRNA-encoding genomic regions overlap these elements (Figure 1). Promoters within repeats drive expression of many ncRNAs [75], especially in pluripotent [29, 76, 77] and cancerous cells [29]. Promoters within the long tandem repeats (LTRs) of endogenous retroviruses specifically associate with ncRNAs from several classes of non-annotated stem transcripts (NASTs) [76] in pluripotent stem cells, including lincRNAs [77] and vlincRNAs [29] (Table 1). In addition, LTR-driven vlincRNAs identify common regulatory architectures between stem and cancer cells [29], an interesting reminder of ideas from the stem cell theory of cancer [78].

Individual copies of repeats are expressed from their own promoters and contribute to the ncRNA transcriptome. For example, RNA polymerase III (Pol III) transcribes non-coding repeated elements, such as Alu, B1, and B2, that can bind to RNA Pol II and affect its activity in response to stress [79]. Long interspersed nuclear elements (LINEs) comprise 20% of the genome and express mostly non-coding transcripts due to 3′ truncation and accumulated mutations [80]. Similarly, expression of non-coding ERVs is a well-documented phenomenon, [81]. Finally, examination of transcripts containing repeat copies continues to reveal additional regulatory functions for these molecules, as exemplified by Alu-mediated inter-molecular interaction of coding and ncRNAs in trans described recently [82].

Transcripts from a specific subset of repeated sequences – non-coding copies of protein-coding genes or mRNAs (pseudogenes) [83] have gained prominence upon realization that they can function in various ways [84], including binding and titrating regulatory molecules that normally interact with the functional copies [85, 86]. Moreover, pseudogenes can be transcribed from the opposite strand thus producing transcripts capable of inter-molecular interaction with the productive copy [87] or its promoter [85].

Classification based on a biochemical pathway or stability

Classification of ncRNA based on their association with substrate pools of different RNA degradation pathways and enzymes has recently gained popularity. Inhibition of components of the exosome (RRP6, RRP40, and RRP44) or nonsense mediated decay (XRN1) has revealed populations of ncRNAs not easily observed in wild-type cells [88–92] (Table 1). This approach also provides information about the pathways of their metabolism. The latter is another attribute used for classification of ncRNAs, as exemplified by XUTs (Xrn1-sensitive unstable transcripts) [91] (Table 1). Most of the pathways analysed so far in this classification involve RNA degradation, and these lncRNAs overlap with classes of NATs [91] and promoter-associated RNAs [89, 90, 92].

Classification based on sequence or structure conservation

Sequence conservation, though highly informative in predicting functional protein-coding mRNAs, remains a metric with controversial merit in the non-coding space. Its absence - typical for lncRNAs [93] - does not universally imply lack of functionality [22, 24]. Still, many ultra-conserved regions (UCRs) – sequences of DNA 100% conserved in human, rat, and mouse – map to the non-coding space of the genome [94]. A large number of UCRs are transcribed as ncRNAs, and some are associated with malignant states [95]. As secondary RNA structure plays a crucial role in ncRNA function [24, 96], a number of bioinformatics approaches such as RNA-Z [97] and EvoFold [98] leverage structure rather than sequence conservation to predict ncRNA-encoding regions (Table 1) [99].

Classification based on biological states

A number of cancer-associated transcribed UCRs (T-UCRs) encoding ncRNAs were induced by hypoxia and thus further sub-classified as “hypoxia-induced noncoding ultraconserved transcripts” (HINCUTs) [100]. They serve as an example of another attribute used for ncRNA classification: induction after treatment with a stimulus or association with a certain biological state. Another example is “long stress-induced non-coding transcripts” (LSINCTs) [101].

Classification based on subcellular localization

RNA localization can provide important clues to its function. ncRNAs tend to be enriched in the nucleus [38, 56], which suggests their involvement in the temporal-spatial regulation of nuclear architecture. For example, chromatin-associated RNAs (CARs) – both intronic and intergenic – form an integral component of chromatin, with the potential to regulate the expression of nearby genes [102] (Figure 1). The ENCODE consortium performed extensive profiling of three sub-nuclear compartments (chromatin, nucleolus, and nucleoplasm) revealing their RNA compositions [3]. Within the nucleus, ncRNA association with, and targeting of, the gene-silencing PRC2 complex led to the identification of thousands of PRC2-associated ncRNAs in mouse embryonic stem cells [103] and human cell lines [19]. ncRNAs form components of other nuclear sub-compartments such as paraspeckles, the nucleolus, and the nuclear matrix [104]. Presumably, additional classes of ncRNAs associated with these and other compartments likely await discovery. Interestingly, some lncRNAs localize to the cytosol [105] and actually associate with ribosomes. Even the small mitochondrial genome encodes lncRNAs [106], underscoring the variety of different processes in which these transcripts could participate (see below).

Classification based on function

lncRNAs can participate in a plethora of different cellular processes: chromatin remodelling, regulation of transcription and translation, RNA stability, scaffolding, and innate immunity just to name a few. We discuss here only examples of functions used for classification, and direct the reader to other reviews focusing on lncRNA molecular mechanisms [6, 7, 13, 24, 39, 107].

Activating ncRNAs (ncRNA-a), which have enhancer-like properties, represents an example of classification based on function (Table 1). This class is distinguished from enhancer RNAs (eRNAs) [108] by positively regulating nearby genes (Figure 1). One notable member of this class, designated ncRNA-a7, regulates the Snai1 transcription factor. Depleting ncRNA-a7 leads to major phenotypic changes at both cellular and molecular levels [108]. The category of ncRNA-a will probably continue to grow, as the accumulation of high-quality expression datasets identify more lncRNAs that positively correlate with nearby genes (St. Laurent et al, submitted).

Another example is Competing endogenous RNAs (ceRNAs) [109]. They share sequence similarity with protein-coding transcripts and function by competing for regulatory molecules [109]. Any ncRNA sharing a sequence with another (coding or non-coding) RNA could potentially be a ceRNA, such as transcribed pseudogenes, which represent important ceRNAs [86] (Figure 1). Conceivably, the ceRNAs could form part of a complex regulatory matrix driven by differential affinity among many contextually associated RNA molecules [110].

Some lncRNAs serve as precursors for shorter functional RNAs as exemplified by primary transcripts for mi- and piRNAs (Table 1). In fact, the long and short cleavage products could have distinct functions, as evidenced by short non-coding tRNA-like molecule produced during maturation of MALAT1 lncRNA [111]. The ENCODE consortium estimates that approximately 6% of all annotated coding and non-coding transcripts overlap with short RNAs [3]. A recent report suggests that an 18 nt short RNA produced from a coding mRNA regulates translation [112]. Also, ncRNAs derived from 3′ ends of mRNAs associate with Argonaute proteins, suggesting they represent novel regulatory molecules [90]. Short RNAs derived from protein-coding transcripts can also mediate trans-generational silencing of the parent gene [14]. Conceivably, cleavage could also generate functional lncRNAs from a longer precursor ncRNA, where both the precursor and the product could have different functions.

Finally, we note that not every lncRNA transcript functions solely as a non-coding element. Peptide sequencing data revealed presence of 250 novel mouse peptides encoded by presumed lncRNAs [113]. The full extent of the novel mammalian proteome encoded by lncRNAs is not yet clear however. Although many lncRNAs appear to associate with ribosomes [105, 114], this frequently does not result in protein synthesis [115, 116], but instead could reveal lncRNA regulation of translation [105]. Nonetheless, the protein coding potential is currently used as one metric in lncRNA definition [37].

Challenges of current lncRNA classification

As described above, the existing classifications of lncRNAs rest on their descriptive and distinctive properties: from their size, to their localization, to their function. For example, the GENCODE system, as one of the few practical and up-to-date classifications available, also classifies lncRNAs into “Antisense RNA” or “lincRNA”, in addition to intron-associated biotypes [33]. Although logical principles guide these classifications, they have inherited a number of unavoidable shortcomings. First, the existing classes capture a small fraction of lncRNAs present in the cell as illustrated in Figure 2. The various lists of lncRNAs annotated based on resemblance to protein-coding mRNAs account for only 0.05–1.12% of cellular RNA (Figure 2) while functional intronic RNAs could constitute as much as 16% [50]. Second, the overlap between multiple existing annotations of lncRNAs derived by different groups is small [33]. Third, the descriptions of the classes in the current annotation schemes can be vague. For example, a lncRNA could initiate at an enhancer element or initiate a large distance away and merely overlap it, yet currently they would both be classified as eRNAs. Fourth, the existing classes are not mutually exclusive. Thus, a lincRNA could theoretically also classify as an eRNA, and an LSINCT, and a CAR, and a T-UCF, and so on. For example, ANRIL is a lncRNA [117], a NAT [117], and a circular RNA [118]. This point is particularly problematic, as few datasets cover all of these characteristics and therefore many ncRNAs are not comprehensively assessed. Fifth, they lack systematization: following the current schemes, in the future one might expect hundreds of overlapping classes of ncRNAs as new knowledge is incorporated into the classification. There are already at least 50 associations with a multitude of biological or biochemical processes (Table 1). Sixth, an attribute used in the current classifications may decline over time in relevance or utility. Considering the growing role of trans regulation by ncRNAs via intermolecular interactions [42, 82, 119], the fact that a lncRNA associates with an enhancer, promoter, or intron, or is antisense to a known gene may not reflect the actual function of that ncRNA. Instead, the latter could function by interacting with transcripts derived from elsewhere in a genome.

Figure 2. Properties of different published lists of human transcripts representing various classes of ncRNAs.

Sequence conservation was defined by the conserved elements from the Vertebrate Multiz Alignment & Conservation (100 Species) from the UCSC Browser [147]. Relative conservation represents the fraction of conserved bases relative to the total lengths for each list of ncRNAs. Relative mass and expression levels represent averages of several malignant and normal tissues profiled using single-molecule RNA-seq analysis [5, 29]. Only uniquely aligning non-rRNA and non-chrM reads were considered. Relative mass represents proportion of reads mapping to a particular genomic element relative to all reads. The relative expression is the relative mass divided by the total length of each list and normalized to the relative expression of coding exons (defined by UCSC Genes). Promoter-associated RNAs were defined by the regions 3 kb upstream of annotated start sites of UCSC Genes. Given the lack of a comprehensive list of standalone human intronic RNAs, we extrapolated the relative mass of those based on mouse data [50]. The GENCODE annotations [33] are based on v19.

The consolidated conceptual framework of lncRNA classification

The concepts driving lncRNA classification have begun to benefit from recent developments in the annotation of non-coding transcripts and the dramatically improved techniques for measuring them. Below, we review the conceptual components that provide the basis for these ongoing improvements (illustrated in Figure 3).

Figure 3. Outline of the consolidated conceptual framework of ncRNA classification.

Highly accurate empirical RNA-seq data drives both annotation and quantification of the longest ncRNA (Tier 1) and of processed ncRNA species (Tier 2) across the entire genome. The quantitation data serves as the basis for the combined global matrix of knowledge of expression of each (coding and non-coding) RNA gene and transcript across multiple biological sources (Tier 3). This information provides the input for the functional annotation of non-coding transcripts using systems biology approaches. Mapping of RNA modifications provides the final layer of knowledge in this scheme.

Tier 1: mapping the longest unprocessed transcript

The fact that the existing lists of lncRNAs miss a large fraction of ncRNA mass (Figure 2) argues that the annotation process has to start at a higher level. The first logical step in this effort is mapping the longest non-coding transcript (Figure 3).

Subdividing the intergenic space into standalone ncRNA loci (genes) has obvious benefits. First, it allows for consolidation of disparate and often incomplete ncRNAs represented by ESTs, lincRNAs, and mRNAs, into a single locus. As an illustration, the clinically important 8q24 region upstream of the MYC gene contains a number of different lncRNA elements [12] (Figure 4). Given the distances that separate them, it may not be obvious that these annotations are part of the same transcript, yet the RNA-seq signal clearly groups them together into one locus associated with a specific regulatory region (Figure 4). Second, such a grouping would allow experiments to focus on the locus rather than its many different genomic elements, allowing for seamless integration of the data from independent experiments. Third, it would clarify the issue of RNA association with different genomic features, for example enhancers, by showing whether the transcripts originate from these DNA elements, or merely overlap them. Overall, the longest transcripts would serve as scaffolds to bring together all the disparate annotations into gene-like structures with their own specific transcription regulatory regions, helping to resolve the problem of overlap. In this case, promoter information and CAGE tag data (Box 1) [25] would help in both assessing the quality of the map and understanding the regulation of such genes.

Figure 4. A genomic view of the 8q24 region upstream of the human MYC gene.

This clinically-important locus containing many GWAS hits associated with several cancers represents an example of a genomic region that could clearly benefit from the new annotation scheme. The RNAseq analysis reveals fairly strong signal on both strands covering most of this >1Mbp region. Yet, the known lncRNA annotations represent only a small fraction of this locus and judging by the distribution of the RNAseq signal and known promoters, are likely part of much larger transcript units (for example vlincRNAs shown on the figure). Transcriptome RNA-seq data is represented by the polyA- nuclear RNA from normal epidermal keratinocytes (NHEK) and embryonic stem cells (H1) generated by the ENCODE consortium [3]. In addition, vlincRNAs [29], promoters [32], and disease-associated variants from genome-wide association studies [148] (GWAS) are shown. Reproduced with permission from [12].

A lncRNA may not always be produced from its own dedicated promoter, as exemplified by circular intronic RNAs [55]. Such standalone functional intronic RNAs would however have certain features, such as low correlation with other exons or introns of the same gene, relatively high expression levels with low variance, and occasionally, differential expression in a biological time course [50]. These properties can now be measured by highly quantitative analysis of RNA levels across multiple diverse samples. Thus, defining standalone transcripts would require an additional dimension – quantitative measurement to allow for analysis co-expression with multiple neighbouring transcripts.

Tier 2: defining processed transcripts

The transcriptional forest concept implies that multiple RNA species share the same genomic space, either transcribed independently or derived by processing of longer precursors [38, 43, 120]. Mapping sites of polyadenylation [121, 122] provides additional information on completeness of such maps. Application of highly-sensitive methods targeted to specific regions using RACE [47] or capture-sequencing [45] would increase the discovery of processed species. Multiple levels of processing exist, such as A to I editing [123, 124] and others [125] each under their own regulatory control.

Tier 3: the additional dimension of expression levels

In the past, genomic coordinates alone determined genomic annotation. However, in the case of overlapping transcripts it cannot predict which isoforms likely function in a particular tissue. Thus, progress of our understanding of the complexities of the transcriptome (Box 1) argues for an additional dimension – expression of each RNA produced at a given locus (Figure 3). The pioneering efforts by the FANTOM Consortium [25] make this undertaking possible.

Tier 4: RNA modifications

A map of all (>100) RNA modifications [125] constitutes the final layer of annotation (Figure 3). These patterns could represent an information rich source for distinguishing RNA molecules, thus assisting in their classification. So far technological limitations prevent us from efficient genome-wide mapping of most RNA modifications, and assaying technically accessible modifications is fraught with pitfalls [124] such as false discovery due to technological noise [126]. Hopefully, existing [127] and emerging technologies [128] will enable progress in cataloguing RNA modifications.

From consolidated conceptual framework to function

The first key of the new framework is consolidation, achieved by grouping disparate lncRNA transcripts into genes or standalone Tier 1 transcripts. The second aspect moves from phenomenological description of lncRNAs to their genomic coordinates, which parallels the evolution of the concept of the gene [129]. The third aspect uses empirical data to unravel the layers of overlapping transcripts, as exemplified by the study of intronic RNAs in mice [50].

The fourth aspect assigns functional weight to a lncRNA by integrating information from diverse high throughput methods [130]. Among others, these methods include CLIP-seq [131] for detection and measurement of RNA – protein interactions, SHAPE-seq [132] for analysis of RNA secondary structure, ChIRP [133] for measurement of RNA–chromatin interactions, and GRO-seq (Box 1) to measure transcription [134]. This multi-faceted approach combines independent sources of evidence for the functionality of an RNA molecule, underscoring the complexities of lncRNA involvement in the flow of biological information [24, 135]. Fortunately new machine learning methods have emerged to identify and decipher complex patterns in the data, yielding probabilistic evaluation of ncRNA function across large populations of transcripts [71, 136].

The evolution of the conceptual framework of ncRNA classification described above provides a roadmap for the analysis of an RNA-seq experiment and its integration into a broader knowledge base of high-throughput multi-dimensional information. The availability of a common set of genomic coordinates for various stages of ncRNA processing (Figure 3, Tiers 1 and 2) provides the key resource that enables the integration of data from multiple experiments. Tiers 3 and 4 assist in refinement of the classification by separating overlapping transcripts that have different patterns of gene expression and modification.

For effective data integration, systems biology approaches require expression datasets that cover large numbers of biological sources with extraordinary precision [137]. Small yet biologically important effects [138] can be lost in technological noise [139]. Similarly, loss of ncRNAs and transcriptome complexity can occur during library preparation [140] and RNA isolation [141]. Processing of NGS data also presents a number of challenges. For example, algorithms building transcripts (Tier 1) should account for regions of the genome that have poor alignability due to repetitive regions [142]. NGS reads unassigned after these steps can then serve as input into ab initio algorithms such as the genomic binning approach [50, 139] to detect differentially expressed regions [27–30]. Without a doubt, cycles of iterations consisting of annotation, expression measurement, and addition of new transcripts and transcribed regions from global RNA measurement experiments will illuminate the puzzle of pervasive transcription.

Concluding remarks

Assigning functions to the mass of lncRNAs produced in the cell requires novel thinking and approaches. Many of the classic reductionist methods that worked well for coding genes have proven less useful to the challenges of deciphering the elaborate populations of transcripts generated by pervasive ncRNA transcription. Instead, global, systems-biology and genomics-driven approaches have emerged, which rely on an integrative framework of annotation and classification. This framework increases emphasis on the quality of genome-wide RNA measurements to allow for the ready integration of data from multiple types of experiments. It facilitates the development of improved tools for the integration of the highly multi-dimensional data from these experiments into the classification framework, thereby revealing associations between both coding and non-coding transcripts. Finally, it supports the rational and structured selection of subsets of these predictions for biological follow-up using reductionist methods.

Highlights.

• Non-coding RNAs constitute the majority of transcriptional output of the human genome.

• Lack of functional knowledge of most of ncRNAs has led to classification with a number of issues.

• A new classification is emerging rooted in unbiased whole-genome surveys of RNA.

• The new classification should allow for easy data integration across multiple experiments.

Glossary

5′-cap

an altered nucleotide present at 5′ ends of a eukaryotic RNA and vital for its functioning

ENCODE project

the Encyclopedia Of DNA Elements, a public research consortium launched in September 2003 by the National Human Genome Research Institute. The goal of the project is to identify all functional elements in the human genome sequence

Endogenous retrovirus (ERV)

a genomic element that was traced back to a retrovirus integrated into an ancestral genome and since retained. ERV sequences comprise ~8% of the human genome

Expressed Sequence Tag (EST)

a relatively short and typically partial sequence of a longer RNA molecule

FANTOM consortium

an international research consortium established by scientists at RIKEN, Japan in 2000, initially to assign functional annotations to the full-length cDNAs collected during the Mouse Encyclopedia Project. FANTOM has since developed and expanded over time to encompass different fields of transcriptome analysis

Genomic bin approach

an approach designed to detect differentially-expressed regions of the genome in the regions where no annotation is available

Long tandem repeat (LTR)

identical pieces of DNA found at the ends of retroviruses and critical for viral life cycle. LTRs contain elements required for viral gene expression. LTRs of ERVs often retain these elements and thus can initiate or control expression of host transcripts

Paraspeckle

a subcellular compartment that could be identified in nuclear interchromatin space

Polycomb repressive complex 2 (PRC2)

a multi-protein complex that reversibly modifies chromatin structure and silences target genes

Tiling microarray

a microarray design (typically oligonucleotide-based) where probes interrogate an entire genomic region of interest at regular intervals agnostic of genomic annotations. This design differs from other microarrays that target only specific genomic features of interest, like exons of known genes