Identification of RNA–protein interaction networks using PAR-CLIP

The Application of Photoactivatable Ribonucleosides for RNA–Protein Crosslinking

As an alternative to the crosslinking of natural nucleic acids with short-wavelength UV light, modification of RNA with photoactivatable nucleoside analogs has been shown to increase the efficiency of nucleic acid crosslinking. The best characterized photoactivatable ribonucleoside analogs (Figure 3(a)) are halogenated pyrimidines, such as 5-iodouridine and 5-iodocytosine,⁶⁶ and thione-containing ribonucleosides, such as 4-thiouridine (4SU) and 6-thioguanosine.⁶⁷ Photoactivatable ribonucleosides are excited with long-wavelength UV (>310 nm), where natural nucleotides no longer crosslink, and yield photoadducts of distinct structure (Figure 3(b)) depending on their reaction mechanisms involving either shortlived free radicals as intermediates or direct photoaddition without reaction intermediates. Amino acid side chains within proteins show different reactivity toward the different nucleoside analogs with the reactions involving predominantly aromatic amino acids such as phenylalanine, tyrosine, and tryptophan but also lysines and cysteine.⁶⁶

Initially, photoactivatable nucleosides were introduced into RNA by in vitro transcription from DNA templates or by chemical synthesis and used to probe RNA–RNA and RNA–protein interactions.^68–71 RNPs are biochemically reconstituted from cell extracts or with recombinant proteins and modified RNA. After long-wavelength UV irradiation, the crosslink sites within RNAs were detected by monitoring accumulation of prematurely terminated cDNA products from stalling of reverse transcriptase in primer extension assays. The reconstitution of RNPs was also accomplished in living cells by transfection of modified RNAs followed by crosslinking and identification of the crosslinked protein partners.⁷²

Some photoactivatable nucleosides are readily bioavailable by providing it in tissue culture medium for uptake.^73,74 4SU is imported into metazoan cells by at least one member of equilibrative nucleoside transporters (ENT1 and SLC29A1).⁷⁵ Presumably, 6SG is taken up by the same or a related mechanism. 4SU-labeling of RNA can also be achieved in mammalian cells as well as tissue- and developmental stage-specific fashion using transgenic animals by conditionally expressing Toxoplasma uracil phosphoribosyltransferase (UPRT) and feeding of 4-thouracil.^76–78 UPRT catalyzes the conversion of uracil analogs and 5-phosphoribosyl-1-R-diphosphate to uridine analog monophosphates. Thioribonucleosides are randomly incorporated into nascent RNA, substituting a fraction of the corresponding natural nucleosides. The incorporation rate of modified nucleosides depends on various factors including the identity of the nucleoside, nucleoside concentration, incubation time, and cell type. Consequently, modified nucleoside incorporation rates should be determined for a given experimental condition by recovering RNA and quantifying the nucleotide composition (Figure 3(c)). Ideally, the fraction of incorporation is adjusted to accomplish one crosslink for every interacting RNA segment, but bioavailability and toxicity are limiting. Incorporation rates of up to 4%4SU relative to uridine were obtained in cultured human cells (i.e., 1 substitution for every 25 Us) without measurable change in gene expression or toxicity.^59,78 The direct application of 4SU in whole animal models like mice has not been thoroughly tested. We suspect that achieving 4SU incorporation in the tissue(s) of choice must balance optimal delivery with toxicity. Thus the use of photoactivatable nucleosides remains a tractable approach for cell culture. One way to circumvent this limitation is to work with cultured primary cells or cell lines that closely approach or match the transcriptome of interest.

Photoactivatable ribonucleoside-enhanced CLIP (PAR-CLIP)⁵⁹ takes advantage of the random transcriptome-wide incorporation of photoactivatable ribonucleoside analogs into newly synthesized RNA in living cells and their subsequent crosslinking to interacting proteins. The current versions of PAR-CLIP use 4SU and 6SG (Figure 4(a)). The photocrosslinking efficiency is dependent on the ribonucleoside analog, the sequence composition of the RRE and its flanking sequence, and the amino acid composition of the protein surface. In general, crosslinking efficiency of modified RNA exceeds the efficiency of UV-254 crosslinking using the same energy dose.^59,63 However, only a small number of proteins were analyzed in a comparative manner. PAR-CLIP of IGF2BP1, with a 4SU substitution ratio between 2 and 4%, yielded several hundred folds more crosslinked RNA than UV-254 irradiated cells grown without 4SU.⁵⁹ Under similar conditions, another study reported AGO2 crosslinked twofold better while ELAVL1 (HuR) crosslinked with similar efficiency as the UV-254 method.⁶³ The crosslinking efficiencies were estimated based on the intensity of the radioactive RNA signals obtained upon resolving the crosslinked and IPed RBPs on SDS-denaturing protein gels. Ultimately, however, the yield of reverse transcribed full-length cDNA from UV-254 or −365 crosslinked RNA needs to be determined to better judge which of the crosslinking methods is more efficient at binding site determination.

Use of 4SU and 6SG leads to cDNA Sequence Changes that Define the Position of Crosslinking

Sequence analysis following a 4SU PAR-CLIP experiment unexpectedly revealed frequent, non-random mutations of T to C in clusters of overlapping sequence reads that aligned to the reference genome. Typically, more than 50–70% of the sequence reads corresponding to a crosslinked RBP-binding site show T to C changes at U positions within or in close proximity of the RRE (Table 3). Sequence reads without the characteristic T to C change likely represent reads from non-crosslinked RNA background. An analysis using synthetic 100% 4SU-substituted oligoribonucleotides demonstrated that RT of non-crosslinked 4SU leads to about 20% misincorporation of dG across from 4SU, likely due to tautomerism between the thiol and thione form. The dG misincorporation ratio increased to over 90% incorporation upon crosslinking of 4SU to protein, presumably due to changed hydrogen donor–acceptor and base stacking properties (Figure 4(b)).⁵⁹ Non-crosslinked RNA recovered from 4SU-treated cells is therefore expected to yield less than1% T to C changes assuming equal nucleotide composition and a 4% 4SU incorporation ratio.

When using 6SG, the crosslinking sites were discerned as a G to A change, albeit the frequency of these changes was much reduced (26% G to A frequency for a 6SG IGF2BP1 PAR-CLIP).⁵⁹ The crosslinked 6SG residue represents a structurally different adduct for RT as compared to 4SU. Therefore, crosslinked 6SG either shows an intrinsically altered deoxynucleotide misincorporation frequency, or it is the result of a reduced probability to read through the crosslinking lesion and a dominance of non-crosslinked background signal.

The T to C (or G to A) transition frequency of clusters of sequence reads is used to separate clusters derived from crosslinked RNA segments from those originating from non-crosslinked background. The background is typically composed of error-free reads matching abundant cellular RNAs, e.g., rRNA, miRNA, tRNA, or bacterial RNAs present in preparations of recombinant RNAligases used in the protocol. For typical mRNA-binding RBPs, such as IGF2BPs, PUM2, or QKI, the fraction of reads mapping to mRNAs (intronic and exonic) varied between 1.6 and 7.6% relative to all reads mapping to the human genome. However, whereas the reads corresponding to each category of background RNAs are basically error-free, clusters of reads corresponding to mRNAs show the 50–70% T to C transition frequency characteristic of crosslinked RNA and prototypical mRBPs.

The frequency of deletions induced by UV-254 in regular CLIP was reported to be as high as 20% in clusters of sequence reads.⁶⁴ However, this number was determined by counting the non-redundant sequences per cluster of reads and cannot be directly compared to those reported for PAR-CLIP—since PAR-CLIP considers read frequency in its accounting. Given its high rate of T to C changes, PAR-CLIP requires less sequence reads to capture crosslink evidence compared to UV-254 CLIP approaches. It readily identifies thousands to hundreds of thousands of binding sites⁵⁷ and is well suited for characterization of RBPs binding with lower sequence specificity and to many sites. Since UV-254 CLIP also generates a signal predominantly based on nucleotide deletions from crosslinked uridines, the sequence bias intrinsic to the use of 4SU in PAR-CLIP is the same. The use of 6SG by PAR-CLIP already represents one approach on how to overcome a uridine bias and it is conceivable that other modified nucleotides may reduce the current nucleotide biases in defining crosslinking sites.

Computational Analysis

Solexa sequencing of a cDNA library generated by the CLIP procedure currently yields over 20 Mio sequence reads and several bioinformatic pipelines for data analysis have been described.^79,80 Sequence reads are categorized by alignment to genome and reference databases annotating mRNAs, rRNAs, tRNAs, and so forth. Mapping of sequence reads has to allow for at least one error (substitution, insertion, or deletion) to capture the reads arising from crosslinked RNA and to calculate the crosslink-induced mutation ratios (T to C and G to A). Allowing for more than one error can increase the percentage of mapped sequences, but can also increase the incidence of assigning sequence reads to multiple genomic locations, only one of which represents its origin. Sequence reads of >20 nt length are preferable to minimize mapping to multiple locations. Aligning the sequences to both transcriptome and genome reference sequences allows the annotation of intronic, exonic, exon–intron, and exon–exon sequences, the ratios of which can provide insight into the potential roles and subcellular localizations of RBPs. PAR-CLIP experiments using mRBPs typically yield thousands to hundreds of thousands of clusters of overlapping sequence reads representing RBP-binding sites (Figure 5).

Occasional single-nucleotide T to C (or G to A) polymorphisms present within the background of non-crosslinked RNAs may be misinterpreted as crosslinking sites, and clusters of reads with precisely 50 or 100% mutation frequency at a single site in diploid cell lines should be considered cautiously. RNAseq analysis of the transcriptome of the corresponding cell line or tissue can assist in identifying these polymorphic variations and in masking these clusters. Relating RNAseq transcript abundance to read numbers from crosslinked sites also provides a measure for the affinity of a RBP to its target sites and it is also useful to determine the level of transcript abundance required for a saturating discovery of RNA-binding sites.⁵⁸

To facilitate the discovery of underlying RREs by motif-finding programs, it may be useful to rank clusters and only use a few hundred of the top clusters. PAR-CLIP clusters can be ranked according to the number of different positions that show the characteristic sequence change upon crosslinking, the fraction of reads with the crosslink-induced mutation frequency, and the total number of reads defining a cluster. When multiple binding sites occur in close proximity, the corresponding sequence reads will overlap and generate large clusters. These clusters can be resolved into individual binding sites by crosslink peak-finding approaches. One software tool that accomplishes this task is PARalyzer, developed by Öhler and colleagues (http://www.genome.duke.edu/labs/ohler/research/PARalyzer/). The program was first applied to isolate short oligoU-rich RREs found in AU-rich regions of 3′ UTRs and intronic sequences bound by ELAVL1.⁵⁷ This approach is also useful for studying RBPs that contain more than one RBD, and/or repetitions of the same domain (Figure 1(b)).

There are several algorithms that can search for common sequence motifs, among them (Phylo)Gibbs, MEME, MDScan, Consensus, Weeder, cERMIT, or Gimsan.^81–88 Many RBPs contain RBDs that typically recognizes short RREs, e.g., 4 nt in KH, 4–5 nt in RRM.³ For example, NOVA proteins, which contain 3 KH domains recognize the degenerate HYCAY (where H denotes A, C, or U and Y denotes C or U),²⁸ the IGF2BPs with 4 KH and 2 RRM domains recognize CAUH, and QKI with its single KH domain binds to AYUAAY-containing sequences⁵⁹ (Figure 5).

Motif-finding based on primary sequence may not always be successful because certain RBPs, such as those containing dsrm RBDs predominantly recognize structural elements without primary sequence conservation. Secondary structure prediction algorithms such as Mfold⁸⁹ may be used in conjunction with an analysis of the effects of sequence mutation on the RBP–RNA interaction. In binding sites that exhibit sufficient evolutionary conservation, it may be possible to determine if base-pair covariation⁹⁰ reliably predicts RNA secondary structure elements within crosslinked sequences.

It may ultimately be necessary to confirm predicted RREs and the contribution of flanking sequences to protein binding. For these validation experiments, recombinant protein is required such that gel shift or similar binding studies can be performed. Analyses using di- and tri-nucleotide repeat sequences corresponding to identified RREs, as well as the testing of natural binding sites and its sequence variants will ideally corroborate the results of the PAR-CLIP.

As the various CLIP methods gain broader use, they will likely identify a wider variety of sequences that are recognized by a single RBP, which has been underestimated since only in vitro characterized sequences with highest affinity have been considered. The fact that CLIP methods deeply capture RNA–protein interactions in vivo changes the relevance of motif discovery from target prediction to characterization of RRE usage and occupancy, affinity differences between RBDs, and consequences to PTGR.

Crosslinking Efficiency and the Influence of RNA and Protein Sequence and Structure

Irradiation of 4SU- and 6SG-modified RNAs by long-wavelength UV triggers photoaddition reactions, which require electronic orbital overlap between reactive amino acid side chains and the excited nucleotide. Therefore not only nucleotide and protein composition are critical, but also the relative orientation of the reaction partners. This has been well documented from analysis of PAR-CLIP crosslinking patterns.⁵⁹ QKI protein with its RRE AYUAAY preferably crosslinked to U at position 2 (U2), but rarely to the U at position 3 (U3). QKI-binding sites having cytidine at position 2 (C2) were identified from crosslinking to non-conserved Us immediately adjacent to the RRE, but not by crosslinking to U3. Therefore, U3 is situated in a structural environment not optimal for crosslinking. Similarly, a preference for crosslinking to U7 but not U1 or U3 of the PUM2 RRE UGUANAUA was observed with sequence variants at position 7 showing crosslinking predominantly outside of the RRE. These examples illustrate that 4SU residues flanking the RRE can participate in crosslinking and it is not necessary to have U-rich RREs for identification of binding sites. It is interesting to note that only 0.4% of the human transcriptome is composed of 32-nt windows devoid of any U residue. Although it was previously argued that 4SU PAR-CLIP is a U-biased method in comparison to UV-254 crosslinking, the predominant signal in the UV-254 protocol also involves crosslinked uridines that characteristically lead to deletions. The use of 6SG provides some remedy to this shortcoming, but as discussed before, the G to A transition events in cDNA are less frequent and may require deeper sequencing. It will therefore be interesting to explore other photoactivatable nucleotides and their respective signature on crosslinking and RT.

RNase Treatment

The choice of RNase and the extent of RNase treatment affect the RNA background as well as the recovery and recognition of crosslinked sites. RNase digestion of cell lysates is required for fragmentation of large mRNPs to smaller subunits facilitating IP and reducing the risk of detection of RNA-dependent interactions with other RBPs. Furthermore, reduction of the RNA fragments close to the borders protected by the footprint of the RBP facilitates the identification of the binding sites. We find that RNA of 20–35 nt length is optimal for transcriptome-wide definition of binding sites for multiple reasons. (1) RNA of this size window avoids spreading of the RBP–RNA band on the denaturing SDS gels. (2) Shorter RNAs are more efficiently converted to small RNA cDNA libraries due to better adapter ligation. (3) Sequencing reads of up to 50 nt are obtained using the cost- and time-effective Solexa short-read sequencing approach. Full-length sequencing is necessary to capture the mutations characteristic for crosslinking. (4) Sequence reads of 20 or more nucleotides tend to map with little ambiguity to single genomic loci at the required error distance of 1. (5) Short RNA segments also reduce the chances of copurifying other RPBs independently bound in vicinity to RNA segments occupied by the RBP of interest.

RNases are highly processive enzymes and it is non-trivial to adjust reaction conditions to reproducibly generate a desired 20–35 nt crosslinked RNA peak. The RNases must be sufficiently active given that these enzymes will be used in cell lysates containing abundant non-coding RNAs, such as rRNA and tRNA that will compete with mRNAs and pre-mRNAs. However, to minimize the chances for overdigesting the input RNAs, RNases with sequence specificity are preferable, such as RNase A, which cleaves at the 3′ end of pyrimidines or RNase T1 that cleaves at the 3′ end of guanosines; both enzymes lead to the formation of products with 3′ phosphates via a 2′,3′-cyclic phosphate intermediate and 5′ hydroxyls. Our laboratory prefers the use of RNase T1 because of its high activity and good specificity cleaving only after G. Even when RNase digestion is allowed to proceed to completion, crosslinked RNA segments, which are free of Gs around the borders of the RRE are recovered, and allow for the initial identification of the unknown RRE. Subsequently, the RNase digestion conditions may be altered to capture all possible target sites. The initial definition of the RRE under stringent RNase T1 digestion conditions also facilitates the annotation of sites obtained under less stringent conditions and the realization of the presence of other sites due to contaminating crosslinked RBPs.

Other laboratories developed a strategy of using a variety of other endoribonucleases in place of, or in conjunction with, RNase T1, such as micrococcal nuclease,⁵⁰ RNase I,⁵⁴ and RNase A.⁴⁵ To date, only one study has compared the potential biases in the recovery of RNA-binding sites introduced by various RNases or digestion conditions using the RBP ELAVL1,⁶³ which recognizes AU-rich sequence elements. As anticipated, overdigestion by RNase T1 led to marked depletion of G-rich clusters as compared to conditions using minimal RNase T1 digestion alone or in combination with micrococcal nuclease.