HITS-CLIP, PAR-CLIP, and iCLIP binding clusters/peaks data were retrieved from the gene expression omnibus and sequence read archive (SRA) (Barrett et al., 2013), the supplementary data of original references or directly from authors upon request. All binding sites coordinates were converted to hg19 and mm10 assemblies using the UCSC LiftOver Tool (Meyer et al., 2013).

Human gene annotations were acquired from GENCODE Version 17 (Harrow et al., 2012). Mouse gene annotations were extracted from Ensembl Gene Release 72 (Hubbard et al., 2009) and LiftOver to mm10 assembly. lncRNAs were further filtered to remove the transcripts overlapping with protein-coding genes. The aforementioned RBP CLIP clusters were used to intersect with the coordinates of all annotated transcripts to find their RBP binding sites, which were fed to Circos (Krzywinski et al., 2009) for visualization.

The Cancer Genome Atlas (TCGA) RNA-Seq expression datasets (level 3, IlluminaHiSeq_RNASeqV2) for 14 cancer types and gene annotation file (TCGA.hg19.June2011.gaf) were downloaded from TCGA Data Portal (Cancer Genome Atlas Research Network, 2008). Expression of 397 known lncRNAs can be measured in TCGA level 3 RNA-Seq data. Expression correlation (Pearson correlation coefficient) between lncRNAs and RBPs was estimated using co-expression program (the program is available from the authors upon request), which was written in C language and ALGLIB library, and p-value was adjusted with the false discovery rate (FDR) correction (Benjamini and Hochberg, 1995).

Disease/phenotype associated SNPs were curated from published GWAS data provided by the NHGRI GWAS Catalog (Welter et al., 2014), Johnson and O’Donnell (2009), dbGAP (Mailman et al., 2007), and GAD (Becker et al., 2004). Additional SNPs in linkage disequilibrium (LD) with reported disease-related loci were selected with the criteria requiring an r² value over 0.5 in at least one of the four populations (CEU, CHB, JPT, and YRI) genotype data of the HapMap project (release 28) (International HapMap 3 Consortium et al., 2010). For each SNP, rs ID were lifted to dbSNP build 141 based on the “RsMergeArch.bcp” and “SNPHistory.bcp” table from dbSNP, and genomic coordinates were lifted to the hg19 assembly using the UCSC LiftOver tool. All these disease-related SNPs or LD SNPs were mapped to exons and splicing sites (2 nt in the intron that is close to an exon) of the annotated lncRNA transcripts and further examined whether they were located in any RBP binding clusters.

RNA-binding protein–lncRNA interactions were deposited in our starBase V2.0 (Li et al., 2014) under the “Protein–RNA” section¹. For each interaction, we provided links to our enhanced deepView genome browser², which was written using a GD graphics library for PHP, to visualize RBP binding sites, lncRNAs, and other annotation tracks in an integrated display style similar to that of UCSC genome browser.

We curated 117 published CLIP-Seq datasets to profile the genome-wide binding maps of 65 RBP. Unique binding sites of distinct RBPs varied from thousands to millions, and the genomic context distributions of binding sites for different RBPs distinguished from each other (Figure 1; Table S1 in Supplementary Material). For example, PUM2, a translational repressor during embryonic development and cell differentiation (Huang et al., 2011), predominately bound to 3′UTR regions of protein genes, while another translation inhibitor FMRP (Napoli et al., 2008) tended to interact with CDS. The discrepancy in binding context preferences for RBPs could root from different amounts of available datasets, usages of various variants of CLIP-Seq, varying sequencing depth, and/or genuine distinctions in the underlying recognition mechanism of RBPs.

Despite that the majority of RBP binding sites were mapped to protein-coding genes, on average 1.1% of RBP binding sites lay within exons of human lncRNAs. In total, 21,073 and 1,662 RBP–lncRNA interactions were identified in human and mouse, respectively (Table 1). It is noteworthy that most well-studied lncRNAs interacted with chromatin modificators, acting as tethers or scaffolds (Khalil et al., 2009; Kung et al., 2013). Thus, we considered the binding features of Ezh2, a subunit of PRC2 complexes, by analyzing the CLIP-Seq data from mouse embryonic stem cells (Kaneko et al., 2013). Our results demonstrated that Ezh2 interacted with 35 lncRNAs including many imprinted RNAs, such as Tsix, Meg3, Rian, and Pvt1 (Figure 2), which was consistent with the epigenetic features of PRC2 (Zhao et al., 2010).

For the 12,255 human lncRNAs, 56.8% were found bound to at least 1 RBP. Surprisingly, 16 lncRNAs, including GAS5 and NEAT1, harbored binding sites of over 30 RBPs (Figure 3; Table S2 in Supplementary Material), indicating their diverse roles in biological processes when accompanied with different RBPs. Since one lncRNA could interact with multiple RBPs, it could be expected that some RBP binding sites were overlapped with each other. Therefore, we explored combinatorial effects among RBPs by employing integrated CLIP-Seq datasets. For example, we utilized PAR-CLIP data generated in HEK293 and intersect binding sites of three RNA destabilizer HuR, Ago2, and MOV10. The results showed that tens of lncRNAs, including cancer-related lncRNAs TUG1, DLEU2, and GAS5, were bound by at least two of the three RBPs at identical binding sites (Figure 4). This phenomenon suggested that the stabilities of these lncRNAs were likely under joint control of these three RBPs, which could be explained by their confirmed interplays in HEK293 (Chendrimada et al., 2007) and Hela cells (Kim et al., 2009).

To realize the roles of RBP–lncRNA interactions in cancer, we preformed co-expression analysis between RBPs and lncRNAs by virtue of 90 human CLIP-Seq datasets and expression data from more than 6,000 tumor samples in 14 types of cancer. Up to 583 pairs concerning 47 RBPs and 49 lncRNAs showed strong correlation at expression levels in at least 1 cancer type (Figure 5A). Marvelously, PUM2 and TUG1 involved with cell cycle regulation (Khalil et al., 2009; Huang et al., 2011) showed significant positive expression correlation (p < 0.05) in all 14 cancer types (Figure 5B). Two potential PUM2 binding sites on TUG1 have the consensus recognition motif UGURUAUA, which was highly conserved in mammals (Figure 5C).

Although GWAS over the years have revealed a significant number of genetic variants related to diseases or phenotypes, a considerable portion of these identified loci are not within protein-coding genes and therefore not functionally explained to date (Hindorff et al., 2009). Here, we tried to fill this gap by connecting RBP binding sites in lncRNAs and potential disease-related SNPs.

Altogether, 87,677 unique disease-related SNPs were collected from four public GWAS data source (Table S3 in Supplementary Material, detailed in Section “Materials and Methods”). Considering that additional SNPs in LD with reported disease-related loci may also map to RBP binding sites in lncRNAs, we perform LD analysis to extracted SNPs that had high LD relationship with disease-related SNPs using a threshold of r² > 0.5 in at least one population from the HapMap CEU, CHB, JPT, and YRI genotype data, which yielded a total of 895,968 disease-related or LD SNPs.

We found that 2431 of these SNPs were mapped to the exons of 2089 transcripts of 1489 lncRNA genes, among which 162 SNPs were also located in at least 1 binding sites of 29 RBPs (Table S4 in Supplementary Material). For example, three disease-related SNPs, namely, rs16902485, rs10283090, and rs2720659, resided in the exons of lncRNA PVT1. According to the GWAS annotations of Johnson and O’Donnell (2009), the latter two of the three SNPs were associated with “type II diabetes mellitus,” which was in good accordance with the recent reports showing that PVT1 may contribute to diabetic nephropathy (Hanson et al., 2007; Alvarez and DiStefano, 2011; Alwohhaib et al., 2014). These SNPs were also overlapped with binding sites of U2AF65, HuR, and eIF4AIII, respectively (Figure 6), suggesting variants in these sites might result in impaired binding of these RBPs to PVT1, which thereby might lead to the development of corresponding diseases.

Next, we checked whether disease-related SNPs might be located in the splicing sites of lncRNAs and affect the alternative splicing of lncRNAs. We defined a splicing site as the 2 nt within an intron close to the exon–intron junction. As a result, we found that only 24 SNPs lay within lncRNA splicing sites (Table S4 in Supplementary Material), among which only 1 SNP, rs17207481, was overlapped with binding sites of FUS and HuR. These results suggested that SNPs exerted limited effects on disease occurrence through the mechanism of disturbing alternative splicing of lncRNAs.

To facilitate comparative analysis of the CLIP-Seq datasets and exploration of RBP–lncRNA interactions, we developed the improved deepView Genome Browser² in starBase V2.0 (Li et al., 2014). In the query page of the browser, users can input one interested genomic region in the “search term” and select corresponding genome assembly to gain an integrated view of various genomic features. Information on binding sites of RBPs, predicted miRNA-target sites overlapped with CLIP-Seq data, as well as gene annotations from RefSeq and Ensembl were provided in toggleable tracks. The image of the browser will be updated immediately by clicking the “refresh tracks” button when users change track options. Figure 7 illustrated the visualization of FUS–MEG3 interactions with deepView. Users can click the “zoom in” or “zoom out” button at the top to shrink or extend on the center of the annotation tracks window by 1.5-, 3-, or 10-folds. Clicking the “View region at UCSC” button will redirect users to the UCSC page and exhibit the current region on the UCSC genome browser. To explore RBP binding sites on a particular gene, users can type its gene symbol in the position textbox and then click the “GO” button to update the display image to determine, which RBPs might participate in regulating the gene. Our visualization method allows a direct comparison of binding patterns of different RBPs, binding preferences of one particular RBP in different cell lines and tissues, and genomic contexts of RBP binding sites.