Methods to study RNA-protein interactions

Abstract

Noncoding RNA sequences, including long noncoding RNAs, small nucleolar RNAs, and untranslated mRNA regions, accomplish many of their diverse functions through direct interactions with RNA-binding proteins (RBPs). Recent efforts have identified hundreds of new RBPs that lack known RNA-binding domains, thus underscoring the complexity and diversity of RNA-protein complexes. Recent progress has expanded the number of methods for studying RNA-protein interactions in two general categories: approaches that characterize proteins bound to an RNA of interest (RNA-centric), and those that examine RNAs bound to a protein of interest (protein-centric). Each method has unique strengths and limitations, which makes it important to select optimal approaches for the biological question being addressed. Here we review methods for the study of RNA-protein interactions, with a focus on their suitability for specific applications.

RNA and proteins are interconnected biomolecules that can promote each other’s life cycles and functions through physical interactions¹. The coding sequence of mRNA carries the instructions for protein synthesis and some regulatory sequences, and the untranslated regions of mRNA influence the fate of the encoded protein by regulating its protein translation, localization, and interactions with other proteins². Proteins, in turn, can bind and modulate RNA expression and function from RNA synthesis to degradation³. RNA-protein interactions are key to cellular homeostasis, and perturbations of RNA-RBP interactions can lead to cellular dysfunction and disease^4,5. Recent work has substantially expanded the number of putative RNA-protein associations in eukaryotes, underscoring the need for a versatile array of methods to identify and characterize their interactions^6,7.

Methods for studying the physical interactions between RNA and protein can be classified by the type of molecule they start with. RNA-centric methods start with an RNA of interest and are used to study proteins that associate with that RNA. Protein-centric methods, in contrast, start with a protein of interest and focus on the RNAs that bind it. Recent progress has expanded the number of both RNA-centric and protein-centric methods. Each currently available method has particular advantages and drawbacks, and thus method selection must be tailored to the relevant biological question. This review provides a selective overview of a subset of these methods and hopefully will assist scientists in their selection of optimal methods to address a particular research question. While there are several methods in each area, the review is focused on recent methods that have demonstrated substantial conceptual and technical advances.

Expanding the RNA-binding proteome

Canonical RBPs are defined by the presence of RNA-binding domains, such as the hnRNP K homology domain and the RNA-recognition motif⁸; however, recent efforts have identified novel RBPs with no annotated RNA-binding domains⁹. Thus, it is not possible to use protein sequence and structural information alone to predict whether an individual protein is indeed an RBP. Direct experimentation is required to generate a census of RBPs in the cell. Experimental methods to identify RBPs in cells use UV cross-linking to create a covalent bond between RNA and protein. Oligo(dT) capture has been used after cross-linking to isolate poly(A)-binding proteins for proteomic identification¹⁰. These methods are limited to identification of the RNA-binding proteome of poly(A) RNA. Recently, Chen and colleagues developed click-chemistry-assisted RNA interactome capture (CARIC), which uses metabolic labeling of RNAs with an alkynyl uridine analog to enable RNA capture independent of polyadenylation¹¹. UV-based methods for studying the RNA-binding proteome have been applied to several cell types in species ranging from Caenorhabditis elegans to humans^12,13. From these studies, a large number of RBPs have been discovered, suggesting that approximately 5% of the human proteome consists of RBPs¹⁰. The application of polyadenylation-agnostic methods such as CARIC in more cell types is likely to further expand the known repertoire of RBPs. Among the newly discovered RBPs are several metabolic enzymes such as adenylate kinase and fatty acid synthase¹⁴. It remains unknown how RNA binding affects the primary function of these metabolic enzymes. For example, how does the identified RNA-protein interaction affect the given RNA involved? How does it affect the metabolic function of the protein? These types of questions can be addressed with complementary RNA-centric and protein-centric methods.

RNA-centric methods: discovering RBPs bound to RNAs of interest

RNA is bound to protein throughout its life cycle. The changing medley of RNA-protein interactions is critical for cellular function, and is restructured on the basis of subcellular location and environmental stimuli³. These dynamic interactions are often transient, which makes it a challenge to identify the most important proteins bound to a given RNA. Broadly speaking, these methods can be classified into two categories (Fig. 1). In vitro methods commonly are used to study RNAs and proteins outside the context of an intact cell. In vivo approaches are used to investigate RNA-protein interactions in the cellular environment and are subdivided according to whether cross-linking is used. Each in vitro and in vivo RNA-centric method has particular strengths and drawbacks, which makes it important to select a method tailored to the biological question being addressed.

Fig. 1 |. Schematic representation of RNA-centric methods.

a, In vitro methods. Top, schematic of end-biotinylated-RNA pulldown. RNA is synthesized with biotin at the 5′ or 3′ end and combined with streptavidin. Recombinant or cellular-extract proteins bind to RNA. After being washed, the beads are boiled to elute and identify RNA-bound proteins. Middle, schematic of aptamer-tagged-RNA capture (S1, Cys4) methods. The RNA of interest is in vitro-transcribed with an RNA tag (blue). The RNA tag binds RNA (red) to a resin support. Proteins in the cellular extract bind to RNA. After washing steps, RNA complex is eluted with imidazole for Cys4 or biotin for the S1 aptamer method. Bottom, schematic of a protein microarray. RNA is in vitro-transcribed with Cy5. The RNA is then added to a human protein microarray spotted with −9,400 proteins. After washing steps, fluorescence is used to detect and quantitate RNA bound on spotted proteins on the microarray. b, Top, schematic of in vivo cross-linking methods. Cross-linking-based methods use either UV (RAP, PAIR, MS2-BioTRAP, TRIP) or formaldehyde cross-linking (CHART, ChIRP). Biotinylated oligonucleotide probes are hybridized to the RNA of interest, and the RNA and cross-linked proteins are purified for downstream analysis. Bottom, schematic of an in vivo non-cross-linking method (RaPID). BoxB RNA stem loops (blue) flank the RNA sequence of interest (red). RaPID (LN-HA-BirA*) fusion protein binding to BoxB sites leads to biotinylation of proteins associated with the inserted RNA sequence in living cells grown in biotin-containing media. Streptavidin beads are used to capture biotinylated protein.

In vitro methods.

In vitro methods commonly use in vitro-transcribed (IVT) RNA that contains a tag that binds to resin (Fig. 1a). After the IVT RNA is bound to resin, cellular extract is added, and subsequently washing steps are carried out to define the RBPs bound to the IVT RNA bait¹⁵.The speed and ease of these in vitro methods are balanced by several disadvantages. IVT RNA might not have the same modifications or structure that a given RNA has in cells. Likewise, if recombinant proteins are used, they may lack post-translational modifications that influence RNA association, and the use of high concentrations of protein may also promote nonspecific associations. The use of cellular extracts as a protein source may overcome some of these challenges, but it may also bias an experiment toward the detection of abundant proteins. A particular advantage of in vitro methods is the possibility of mutagenesis studies designed to identify the specific RNA nucleotides and protein amino acids involved in the binding of a given RNA-protein pair.

For in vitro studies, the simplest method of RNA tagging for pull-down is 5′ or 3′ RNA biotinylation. End-labeled biotinylated RNA is bound to streptavidin beads¹⁶, then cellular extract is added and beads are subsequently washed. With this method, there is no elution of RNA, and the beads are boiled in SDS for proteomic analysis¹⁶. An alternative approach is to add RNA aptamer tags to the RNA of interest^17,18. Numerous tags have been developed for in vitro methods, including several recently developed methods¹⁵. Stoecklin and colleagues used the S1 aptamer tag, which binds to streptavidin beads, thus enabling biotin to be used later to competitively bind streptavidin and elute the S1-aptamer-tagged RNA¹⁹. Doudna and colleagues leveraged Cys4 endoribonuclease, which binds to a Cys4 hairpin loop²⁰. Using imidazole, the nuclease can cleave off the hairpin loop and liberate the RBP-bound RNA (Fig. 1a)²⁰. Cellular proteins can bind nonspecifically to resin. Elution of RBP-bound RNA (rather than all protein) off resin reduces background by excluding nonspecific resin-bound proteins from downstream analysis.

An alternative in vitro approach is to use IVT RNA labeled with Cy5 dye and hybridize it to a protein microarray containing approximately 9,400 recombinant human proteins (Human ProtoArray)²¹. Proteins that capture the Cy5 RNA are detected via fluorescence reading (Fig. 1a). The protein microarray method does not require cellular extracts and enables the discovery of direct RNA-protein interactions²². However, the method is limited by the folding and post-translational modifications of spotted proteins and may be potentially distorted by artificial concentrations. As a category of methods, in vitro methods overall are particularly useful for characterizing the binding of specific known RNA-protein interactions.

In vivo methods with cross-linking.

Researchers can use protein-RNA cross-linking to identify in vivo interactions by purifying the RNA under denaturing conditions that remove noncovalent interactions, and subsequently extracting only the cross-linked proteins for identification. Formaldehyde is a small, bifunctional cross-linker that readily permeates cells and cross-links macromolecules within 2 A, including protein-protein, protein-DNA, and protein-RNA complexes, creating a reversible covalent linkage²³. UV light cross-links protein to nucleic acid at zero distance, and in an irreversible, covalent bond^24,25. Each method of cross-linking presents distinct advantages and disadvantages. Because UV light is a zero-distance cross-linker and because it does not cross-link protein-protein interactions with an efficiency anywhere near that of formaldehyde, it is a more specific cross-linker than formaldehyde. However, the efficiency of UV cross-linking is lower²⁴. Both UV and formaldehyde cross-linking have biases. UV cross-linking has a slight uridine preference²⁶, and double-stranded RNA is known to be poorly cross-linked²⁷. With respect to proteins, the efficiency of UV cross-linking varies by amino acid²⁸. The structure and surface area of the protein-RNA interaction are other factors that are likely to affect UV cross-linking efficiency, but the current understanding of cross-linking efficiency remains too incomplete to allow quantitative predictions. We expect further research to make progress on this essential problem. In contrast, with formaldehyde cross-linking, strongly nucleophilic lysine residues are preferentially cross-linked²⁹. Formaldehyde promotes cross-linking between proteins in addition to protein-RNA interactions, and thus it is difficult to distinguish proteins that contact RNA directly from those that are complexed with directly bound proteins. The biases and the low efficiency of both of these cross-linking modalities necessitate high input cell numbers, ranging from 10⁸ to 10⁹ cells, to maximize the capture of RNA-protein interactions^30–32.

Several in vivo methods use either UV or formaldehyde crosslinking (Fig. 1b). Methods that use UV cross-linking include RNA affinity purification (RAP)^33,34, peptide-nucleic-acid-assisted identification of RBPs (PAIR)³⁵, MS2 in vivo biotin-tagged RAP (MS2-BioTRAP)³⁶, and tandem RNA isolation procedure (TRIP)³⁷ (Table 1). Although they share a common UV cross-linking approach, these methods differ in experimental setup. RAP uses long, 120-oligo-nucleotide probes to pull down RNA-RBP complexes after crosslinking and has been used to study noncoding RNAs such as Xist and FIRRE^32,34. PAIR uses peptide nucleic acid probes with cell-penetrating peptides to gain entry into cells and hybridize to RNA, after which the RNA is purified, along with bound RBPs³⁵. MS2-BioTRAP uses the interaction between MS2 hairpin loop and MS2 coat protein to tether protein to RNA³⁸. Both MS2 hairpin RNA and MS2 coat protein are expressed in the same cell and form a stable complex, enabling the fusion MS2 coat protein to be used as a handle to purify the MS2-containing RNA after UV cross-linking³⁶. The ectopic expression of MS2-tagged RNA might not reflect physiological levels of RNA, which can potentially impair the accuracy of downstream proteomic analysis. TRIP is used to study polyadenylated RNA and uses a dual purification: poly(A) RNA is purified first, and then biotinylated antisense oligonucleotides (ASOs) are used to hybridize with the RNA of interest in the poly(A) mixture by base pairing, after which the RNA-ASO complex is purified with streptavidin beads³⁷. Chromatin isolation by RNA purification (ChIRP)^30,39 and capture hybridization analysis of RNA targets (CHART)³¹, in contrast, use formaldehyde to cross-link RNA to proteins. CHART requires an additional RNase H assay to identify accessible sites for probes, whereas ChIRP does not require prior knowledge of RNA accessibility and uses shorter, 20-mer oligonucleotide probes^30,40. These biotinylated probes hybridize with RNA in the cell and are captured with streptavidin beads for proteomic analysis. These UV and formaldehyde cross-linking methods allow the RNA-centric study of RNA-protein interactions in vivo.

Table 1 |.

Comparison of RNA-centric methods

Method	Application	Advantages	Disadvantages	References
Biotinylated RNA	SMN mRNA	Strong binding of biotin-end- labeled RNA with streptavidin beads	In vitro; potentially biased toward abundant proteins in cell extracts	16
S1 aptamer	ARE motif	Easy elution of RBP complex from streptavidin beads with biotin	In vitro; potentially biased toward abundant proteins	19
Cys4	Pre-miRNA	Elution of RBP complex with imidazole	In vitro; potentially biased toward abundant proteins	20
Protein microarray	TINCR, SNORD50	No cellular extract required; no MS required	In vitro; limited to direct interactions with proteins spotted on microarray	21,22
RAP	Xist, FIRRE noncoding RNA	In vivo; high specificity with UV cross-linking and long oligonucleotide probes (120 nt)	High input cell numbers	33
TRIP	p27 mRNA, CEP-1 mRNA	In vivo; high specificity with UV cross-linking	Two capture steps with poly(A) and biotinylated ASO capture decrease efficiency	37
PAIR	ANK mRNA	In vivo; high specificity with UV cross-linking	Cost and effort for production of peptide nucleic acid	35
MS2-BioTRAP	IRES	In vivo; high specificity with UV cross-linking	Requires MS2 conjugation to RNA, transfection/infection of RNA and labeler protein, and high input cell numbers	36
CHART	Xist, MALAT1, NEAT1	In vivo	Additional RNase H step to identify accessible sites for probes; high input cell numbers	31,52
ChIRP	TERC, Xist	In vivo; no prior knowledge of RNA accessibility required for probe design	Short probes may pull down similar sequence fragments; high input cell numbers	30,39
RaPID	ZIKV-host protein interactome; 3′ untranslated region motifs	In vivo; low number of cells required; direct labeling of protein	Requires BoxB link to RNA; short sequence limits; transfection/infection of RNA and labeler protein	44

In vivo methods without cross-linking.

Proximity proteomics has recently been applied for the RNA-centric study of RNA-protein interactions in living cells without the use of any form of crosslinking. ‘Promiscuous’ biotin ligases, predominantly used to study protein-protein interactions^41,42, convert biotin to a reactive bio-tin-5-AMP intermediate, which is released from the enzyme to covalently label any nearby exposed lysine amino acid residue⁴³. Because the intracellular environment is reducing, biotin-5-AMP is quenched and becomes unreactive within a distance of 20 nm of its point of release⁴¹. Hence, proteins that are within a distance of 20 nm will be labeled with biotin preferentially compared with other proteins in the cell. The RNA-protein interaction detection (RaPID) method allows one to use this spatial detection constraint to detect RBPs bound to RNA by tagging an RNA of interest with a BoxB aptamer to recruit a fusion protein of λ-N and a promiscuous biotin ligase⁴⁴. The biotin sprayer binds the BoxB motif through its λ-N domain and labels proteins proximal to its bound RNA (Fig. 1b). Because 20 nm corresponds to ~66 nt of linear RNA, placing BoxB aptamers both 5′ and 3′ of an RNA sequence of interest should enable the study of RNAs up to 132 nt long with this approach, and the structured nature of some RNAs might permit the study of substantially longer sequences. This approach relies on direct labeling of RBPs and does not require cross-linking or purification of the RNA⁴⁵. Although there are benefits to this approach, including speed, cost, the low number of cells needed, and ease of use, there are downsides as well. First, the BoxB site needs to be proximal to the RNA sequence of interest, and thus the ‘bait’ RNA has to be expressed artificially, such as by plasmid transfection. Additionally, RNA can fold into complex structures, so the positioning of the BoxB aptamer needs to be carefully considered for longer RNA species. Given these limitations, including the 20-nm limit for the ‘reach’ of this method, RaPID might be best for studies of shorter (≤132 nt) RNA motifs.

Proteomic analysis.

Most RNA-centric approaches use quantitative mass spectrometry (MS) to discover RBPs bound to RNA. At a practical level, MS techniques can be divided between those that use labeling and those that do not⁴⁶. Labeling methods can involve the differential use of stable isotope labels or chemical tagging of proteins in samples and controls. Hence, the ratio of labeled peptides can be used to obtain enrichment scores and identify true binding partners⁴⁶. Labeling MS methods, such as SILAC and iTRAQ, are particularly useful with formaldehyde cross-linking or an in vitro method, as these approaches have a lower specificity, and results could be contaminated by proteins bound nonspecifically⁴⁷. However, labeling methods require more technical expertise in MS data analysis and are more expensive. Label-free MS identifies proteins in both samples and controls; the challenge is in distinguishing true binding partners from nonspecific proteins. Analytical tools such as SAINT can be used with spectral count data from nonquantitative MS to effectively score the probability of a bona fide RBP-RNA interaction⁴⁸. The use of more than two replicates each for samples and controls is advisable to increase stringency and avoid false positives in nonquantitative MS analysis. For negative controls, we recommend an RNA derived by scrambling the sequence of the RNA of interest. Scrambled sequences have the same length and nucleotide composition but a different primary, and therefore secondary, structure of RNA. Positive controls where the RNA has known binding partners can be included in experiments designed to validate both the experimental setup and proteomic analysis. In methods that require the purification of cellular RNA (CHART, RAP, etc.), verifying the isolated RNA by sequencing will increase confidence that proteomic analysis has indeed found interactions with the RNA of the interest. Proteomic analysis requires multiple steps from raw data acquisition to mapping and differential analysis⁴⁹. There are a variety of open access and proprietary software tools for this, with attendant advantages and disadvantages^50,51. It is critical to keep track of and report the tools and settings used for analysis.

RNA-centric method selection

Given the diversity of methods available, which one is optimal for a given application? Three factors can help to guide the selection of an appropriate RNA-centric method.

The first relates to the choice between an in vitro or in vivo approach. In vitro approaches are particularly powerful for determining which nucleotides and amino acids contribute to known RNA-protein interactions. In vivo methods are helpful in studies that rely on the cellular environment, where the localization of RNAs and proteins may vary from organelle to organelle (nucleus, cytoplasm, endoplasmic reticulum, etc.). Thus, in vivo methods may be best for discovering and studying RNA-protein interactions when subcellular localization, RNA and protein modifications, or a dynamic range of local protein concentrations are a factor.

The second consideration relates to RNA abundance. The copy number of the RNA of interest is critical for the detection of RNA-protein interactions. The higher the RNA copy number, the fewer cells are required in order for RNA-protein interactions to be detected in vivo. For example, the noncoding RNAs MALAT1 and NEAT1 have high expression, and thus fewer cells are needed to obtain their RBP interactome⁵². Many in vivo methods that use cross-linking to detect RBP interactions, however, require 100 million to 1 billion cells^30,33,52. Thus, the effort and cost required to scale up experiments to study low-copy-number RNA can be over-whelming. In such a setting, the better alternative might be to use in vitro methods to study RNA-protein interactions.

The third factor involves the strength of the studied RNA- protein interaction, which especially influences the choice of cross-linking approach. Experimentally measured RNA-protein dissociation constants vary widely, ranging from high nanomolar to picomolar concentrations⁵³. Cross-linking modalities differ in efficiency, and in general, the weaker the interaction between RNA and protein, the less likely it will be captured with UV cross-linking as opposed to formaldehyde-cross-linking-based methods. Thus, more cells might be needed to capture RNA-protein interactions in methods that use UV light versus formaldehyde for cross-linking. Furthermore, with cross-linking methods, the RNA needs to be purified with the use of oligonucleotide capture probes before bound RBPs can be isolated. The efficiency of the oligonucleotide capture further decreases the efficiency of RNA-protein interaction capture, and this necessitates higher input cell numbers. Taken together (Fig. 1), these considerations can help to guide the selection of an appropriate RNA-centric methodology.

Protein-centric methods: characterizing RNAs that bind a protein of interest

Protein-centric methods start with a protein of interest and characterize its interaction with RNA. Commonly, these approaches either directly purify the protein to find associated RNAs or use selective chemical modification of RNA in a way that relies on its association with the protein of interest. The overwhelming majority of studies that identify RNAs bound to a given protein do so by purifying the protein of interest. The most common approach in this case is to make use of the long-known fact that protein will chemically cross-link to nucleic acid in vivo when hit by UV light at approximately 254 nm (refs. ^54,55). The use of 254-nm induced cross-linking played an important role in the initial identification of RBPs⁵⁶. Almost all amino acids cross-link (D, E, N, and Q being the exceptions)⁵⁷. Although RNA-protein cross-linking is generally thought to proceed through an initial excitation of the nucleobase, it should be noted that UVC light also induces some protein-protein cross-linking^58–60. Methods that involve UV cross-linking followed by purification of the protein of interest and identification of bound RNAs are broadly termed cross-linking immunoprecipitation (CLIP)⁶¹ methods, with those that use high-throughput sequencing (HITS) forming the CLIP-seq family of methods⁶². Methods that use a similar protocol but an alternative cross-linker are also sometimes referred to as CLIP methods²⁷.

A common difficulty with CLIP-seq is immunopurifying enough cross-linked RNA, which can become a problem with poor crosslinking efficiency, low RNA-ribonucleoprotein complex abundance, poor antibodies, inefficient library preparation, or combinations thereof. Unfortunately, there is no universally accepted answer to the question of how much purified cross-linked RNA is enough for CLIP. Practically, a common approach includes CLIP followed by an evaluation of whether cross-linked RNA can be visualized by dye or ³²P labeling. Another common tack is to see whether a given CLIP-seq effort generates satisfactory libraries. CLIP-seq is covered in greater detail below.

If sufficient UV cross-linked complexes cannot be purified, then options diverge: if (1) indirect interactions are tolerable and (2) binding sites within RNAs do not need to be determined, then the standard method is RIP-seq^63,64. RIP-seq may be conceptualized as RNA-seq after protein purification, or CLIP-seq without the removal of non-cross-linked RNAs. Essentially, immunopurification is carried out under nondenaturing conditions that are intended to preserve cellular complexes, thereby removing the need for cross-linking. RIP-seq can also provide RNA binding sequence locations if the RNase digestion is optimized⁶³. The conventional wisdom is that CLIP has a higher signal-to-noise ratio than RIP, as might be expected from the removal of noncovalent interactions, and as supported by the often vanishingly small noepitope CLIP-seq control datasets. However, it is conceivable that RIP might have a higher signal-to-noise ratio for proteins with very poor cross-linking.

In the case that either indirect interactions are not tolerable or RIP-seq is not satisfactory, there are alternative cross-linking reagents. PAR-CLIP⁶⁵, which uses 4-thiouridine and/or 5-thioguanine as a nucleotide analog, is advantageous in cases where UV light is not penetrating deep enough into the sample. However, we note that PAR-CLIP results have generally been similar to those of regular cross-linking^66,67. Investigators choosing a method for their protein of interest may consult ref. ¹⁰, which identifies proteins that can be cross-linked to RNA either by 254-nm UV light or by nucleotide analog, and compares their relative efficiencies.

Recently, formaldehyde was used as a cross-linking reagent for CLIP on a double-stranded RNA-binding protein²⁷, which are generally thought to UV cross-link poorly. Methylene blue has been used to cross-link double-stranded RNA to RBPs⁶⁸, but this method has not yet been widely applied in CLIP. Many compounds known to crosslink RNA to protein, such as diepoxybutane⁶⁹, 2-iminothiolane⁷⁰, and DTT⁷¹, have not been studied as reagents for CLIP, and it is likely that many more uncharacterized cross-linking compounds exist. Regardless of whether standard UV cross-linking or alternative methods are used, methods that rely on protein purification for protein-centric RNA studies constitute a mainstay of the field, foremost among which are the quickly expanding varieties of CLIP.

CLIP-seq

A vast number of alternative CLIP-seq protocols have been published. We present a subway-map view of how these protocols progress from immunoprecipitation to PCR amplification in Fig. 2 (also see Supplementary Table 1). Some steps, such as initial dephosphorylation of RNA, ligation to the 3′ end of RNA, and reverse transcription, are universal, whereas others are method specific. In the original CLIP-seq protocol, 5’ and 3’ adaptors are ligated to purified RNA, and reverse transcriptase has to proceed through the crosslinked nucleotide⁶². Reverse transcription through the cross-linked nucleotide is 10–25% efficient for SuperScript IV⁷², but manganese appears to increase this rate⁷³. CLIP variants 1 (ref. ⁶⁷) and 2 (refs. ^74,75) streamline the original protocol by doing both ligations on-bead. The RNA processing steps in the CLIP protocol CRAC (UV crosslinking and analysis of cDNA) are the same as in variant 1, but are preceded by a denaturing purification⁷⁶. iCLIP removes the 5′ adaptor ligation to RNA, replacing it with a circularization of the cDNA⁷⁷. eCLIP replaces the 5′ adaptor ligation with a 3′ cDNA ligation⁷⁸, and monitored eCLIP uses both a 5′ ligation and a 3′ cDNA ligation⁷². irCLIP is similar to iCLIP but makes use of a biotinylated, fluorescent 3′ DNA adaptor⁷⁹. BrdU-CLIP uses a nucleotide analog in reverse transcription to separate cDNA from unreacted reverse-transcription primer⁸⁰. The recent GoldCLIP method is a shortened iCLIP protocol that removes the protein gel step and, like CRAC, includes an on-bead denaturation⁸¹. To date, eCLIP has probably produced the largest number of datasets, owing to its use it in the ENCODE project.

Fig. 2 |. Subway map of CLIP protocols, from immunoprecipitation to PCR.

The chart highlights steps of various representative CLIP-seq protocols. Not all steps are included. XL, UV cross-link; IP, immunopurification; phosphatase, removal of 3′ phosphate; kinase, addition of 5′ phosphate; RT, reverse transcription; L3, 3′ adaptor ligation to RNA or DNA; L5, 5′ adaptor ligation; PK extraction, proteinase K extraction from nitrocellulose membrane; Ppt/column, alcohol precipitation or column cleanup of nucleic acid; TBE, Tris-borate-EDTA; SA, streptavidin.

There are currently no published data that would allow an estimation of the overall efficiency of any CLIP-seq method, which would require quantification of the number of input cellular complexes and the total library complexity at the end. In fact, it is rare for CLIP-seq methods to quantify the efficiency of any step. Investigators looking to begin CLIP with an optimal method face some confusion, as there is little in the way of true quantitative comparisons. Some advantages to each method are noted in Table 2. Fluorescent DNA adaptors (as in irCLIP⁷⁹) may be particularly helpful for investigators initiating CLIP studies, as they do not interfere with any method and provide a way to track sample throughout library preparation, which can reveal where sample may be lost. The technical end point of CLIP-seq method evolution would be the sequencing of bound RNAs directly, which would bypass nearly every step in Fig. 2, but direct RNA sequencing⁸² has not yet been coupled to CLIP-seq. An intrinsic challenge to all CLIP-seq methods that rely on antibody immunoprecipitation is the requirement for high-quality antibodies. Although this can be overcome by the insertion of epitope tags, concerns about the physiological expression of tagged proteins and the possibility of the tag itself affecting RNA binding are limitations. CLIP-seq methods, however, continue to evolve, and represent an active area of protein-centric method innovation.

Table 2 |.

Features of selected CLIP methods

Method	Advantages	Disadvantages	References
HITS-CLIP/PAR-CLIP	Well-established method; can identify interactions at the 3′ end of RNA	Long protocol; UV cross-linking may be poor; reverse transcriptase must bypass the cross-linked nucleotide	62,65
iCLIP	Well-established method; reverse transcriptase does not need to bypass the cross-linked nucleotide	Long protocol; UV cross-linking may be poor; interactions near the 3′ end of an RNA may be unidentifiable because reverse transcriptase stops at the cross-linked nucleotide	77
eCLIP	Avoids the circularization step of iCLIP, which can be an unreliable reaction	Long protocol; UV cross-linking may be poor; must ligate a single-stranded DNA adaptor to single-stranded cDNA	78
irCLIP	Uses a fluorescent adaptor to visualize cross-linked RNA at each step; efficient	Long protocol; UV cross-linking may be poor; interactions near the 3′ end of an RNA may be unidentifiable because reverse transcriptase stops at the cross-linked nucleotide	79
HITS-CLIP variants 1 and 2	Short protocol; increased reliability because both ligations are carried out on-bead	UV cross-linking may be poor; reverse transcriptase must bypass the cross-linked nucleotide	67,74–76
GoldCLIP	No gel purification step; short protocol	UV cross-linking may be poor; must express a fusion protein; interactions near the 3′ end of an RNA may be unidentifiable because reverse transcriptase stops at the cross-linked nucleotide	81
fCLIP	Replaces UV cross-linking with formaldehyde, which may have a higher efficiency for proteins binding double-stranded RNA	Long protocol; formaldehyde cross-linking for CLIP is not as well understood as standard UV cross-linking	27
BrdU-CLIP	BrdU in reverse transcription removes ‘empty adaptor’ reads that can clutter HITS data	Long protocol; UV cross-linking may be poor; interactions near the 3′ end of an RNA may be unidentifiable because reverse transcriptase stops at the cross-linked nucleotide	80
TRIBE	No need to purify the protein of interest; no dependence on UV cross-linking; RBP interaction may occur anywhere in RNA	Less extensive examples of effective use of the method	90
RNA tagging	No need to purify the protein of interest; no dependence on UV cross-linking; straightforward protocol	Has not been demonstrated outside Saccharomyces cerevisiae; might not work on RBPs distant from the 3′ end of RNA	91

Advantages and disadvantages of different methods of identifying RNAs bound to a protein of interest. Only a selected subset of CLIP methods is included, and many other excellent protocols exist.

CLIP-seq analysis.

There is no universal standard for CLIP-seq analysis, and new pipelines continue to emerge. This might be due to variation in study aims and in how background is defined. Table 3 summarizes some of the different analytic approaches. Several features of CLIP data may influence the analytic path taken. In chromatin immunoprecipitation (ChIP)-seq, the amount of nucleic acid (DNA) is fixed at two copies, whereas the abundance of transcription factors varies. In CLIP-seq, the abundance of both RNAs and RNA-binding proteins varies over orders of magnitude. At some frequency, all RNAs contact all proteins, and high abundances of RNA or protein may make low-affinity interactions common enough to be easily detected by CLIP As a result, the identification of an interaction between a specific RNA and protein as having occurred by, say, the clustering of cross-link-induced mutations does not by itself necessarily provide robust insight into its frequency or physiological relevance.

Table 3 |.

Approaches used to interpret CLIP-seq data

Study	Approach	References
Nova	CLIP data are evaluated on the basis of enrichment relative to a control immunopurification.	62
Ago	Calculates the expected number of reads per RNA from RNA abundance and randomizes their positions to determine the odds of a peak as high as that observed.	94
eCLIP	CLIP data are evaluated on the basis of enrichment relative to RNA from the input lysate run on an SDS-PAGE gel and transferred to a nitrocellulose membrane.	78
FOX2	Reads are scrambled randomly in transcript to calculate the odds of the observed peak height. This differs from the method used for Ago above94 in that the reads are drawn from the CLIP data, not RNA abundance. As a result, this approach identifies a locational bias within a given RNA, not affinity for the RNA molecule of interest, although the latter is indirectly tested by the requirement for having enough reads to identify a locational bias.	95
Puf2p	Raw peak height cutoff, enrichment relative to local CLIP signal, and enrichment over expectation from RNA abundance. This method is a combination of a local signal pileup analysis like that of FOX295, enrichment relative to RNA abundance like that of Ago94, and a third raw peak height requirement.	67
FBF	Compares each of the different approaches available at the time of its development: local pileup analysis (similar to FOX2 above95), enrichment over RNA abundance (similar to Ago94), enrichment over a control immunopurification (similar to Nova62), CIMS96, and CITS80. Enrichment over a control immunopurification was found to most accurately reflect known biology.	97
Nova, Ago	Frequency of cross-link-induced mutation pileup versus random permutations. Termed CIMS. The notes on FOX2 apply to this approach as well.	96
hnRNPC, TIA1, Rbfox	Frequency of reverse-transcription termination positions versus random permutations. Termed CITS80. The notes on FOX2 apply to this approach as well.	80, 98,99
Many	Frequency of PAR-CLIP-specific mutations. We refer the reader to a more specialized review of CLIP-seq analysis tools100.	75, 101, 102

A non-exhaustive list of selected approaches is noted. This table presents examples of the different concepts of CLIP-seq analysis, rather than comparing the numerous implementations of those approaches.

Quantitation of CLIP-seq data is one current area of analytic challenge. If PCR duplicates are removed, each read in a CLIP library should, in principle, correspond to a single cross-linked protein-RNA complex isolated from a cell. That is, barring technical artifacts, each sequenced DNA fragment must have been transcribed from a single cross-linked RNA fragment purified from the cell. If the CLIP library-preparation protocol has approximately the same efficiency for nearly all RNA fragments, as is commonly assumed, then each stack of reads in a region per all reads corresponds to that region’s proportion of all cross-linked complexes. As the reads-per-million value represents the read count per all reads multiplied by 1 million, reads-per-million in CLIP-seq data represents the frequency of cross-links at a given RNA out of all crosslinks. Because this frequency includes a cross-linking efficiency factor (the proportion of RNA-protein complexes that react with UV light to form a covalent bond) that is dependent on the RNA and may vary considerably, it might not always be accurate to treat cross-link frequencies as true RNA-protein interaction frequencies. The currently unpredictable variability of cross-linking efficiency for different RNA sequences introduces a second complexity in CLIP-seq compared with ChIP-seq. The number of reads at a region in a sequenced CLIP library is equal to the number of cross-links at that region multiplied by a long string of multiplicative efficiency factors, one for each step in library construction. That is, there is some number of cross-linked RNA-protein complexes containing a given RNA present in the lysate input to a CLIP experiment, and at each subsequent step of library construction only a fraction of molecules is correctly processed. As a result, CLIP signal tends to vary on a log scale between replicates, and, as with RNA-seq data⁸³, a conversion to a log scale is frequently helpful.

A large number of CLIP analysis programs have been published (Table 3); however, none has as yet been universally adopted as a standard, and it is not uncommon for a method to be used in only a handful of studies. This is likely due to the fact that an experienced bioinformatician can write a script to implement any of the approaches in Table 3 with a similar amount of labor as it would take to adapt an existing tool, thereby avoiding giving up both control of and detailed understanding of the analytic pipeline. CLIP analysis programs are likely to become more user-friendly with time, but meanwhile, multipurpose software libraries to deal with HITS data have also become more extensively functional and easier to adapt to situations like CLIP. For example, the combination of Python libraries HTSeq⁸⁴, NumPy⁸⁵, SciPy⁸⁶, statsmodels⁸⁷, and pandas⁸⁸ (to name a few) provides an extremely powerful framework for CLIP analysis, and Jupyter⁸⁹ notebooks create an interactive environment for programming analysis suites. It is possible that no complete CLIP analysis package will become standard, although a consensus may form regarding well-written, especially fast algorithms for specialized CLIP-related tasks. At present, it probably remains optimal for CLIP analysis to involve mostly bespoke code that uses multiple forms of read count and mutation analysis, although programs such as those referenced in Table 3 may be satisfactory for many experiments.

Methods not requiring protein purification

Methods to identify the RNA targets of an RBP without purifying the RBP are relatively new, and currently rely on two different chemical modifications of RNA. In the first, TRIBE⁹⁰, the RBP of interest is fused to the enzyme ADAR, which deaminates nearby adenosines, after which deaminated bases are subsequently identified by sequencing. In the second, RNA tagging⁹¹, the RBP of interest is fused to the enzyme poly(U) polymerase, which adds poly(U) tails to bound RNAs. Tails are subsequently identified by sequencing of the 3′ ends of RNA. The coupling of RBPs to peroxidase tags has also been used to identify RNAs in specific subcellular compartments; this method has not yet been used to identify direct RNA targets⁹². Far more enzymatic modifications of RNA are possible, and we expect this field to see rapid growth as more and better methods are found. Especially exciting are the possibility of multiple distinct chemical marks being made by separate RBPs for studies of combinatorial regulation, and the combination of chemical modification with RBP purifications to study the locations of RBP-RNA interactions.

Conclusion

The technical and conceptual advances in methods for studying RNA-protein interactions have shed light on complex and critical RNA-protein interactions in cells. There are areas where further innovation could spur the accessibility of these methods to researchers. Cross-linking is the crux of several methods used to identify RBPs and to define RNAs bound to protein. Current UV and formaldehyde cross-linking approaches are inefficient, and better cross-linking methods could capture RNA-protein interactions efficiently with fewer cells. Advances in orthogonal areas could also generate novel methods to study RNA-protein interactions. The discovery of RNA-specific Cas proteins could be adapted to probe RNA-protein interactions in cells⁹³. Fusion of these Cas proteins with enzymes could label either RNA or protein at specific spots along an RNA. With potentially exciting new tools on the horizon, it is important for researchers to be aware of the strengths and limitations of different methods. In virtually all cases, orthogonal methods are essential to validate results based on a single method. For example, RBPs discovered via an RNA-centric method should be validated by the complementary protein-centric method (CLIP, etc.) for confident identification of bona fide RNA-protein interactions. Taken together, the expanding arsenal of both RNA-centric and protein-centric methods for the study of RNA-protein interactions will accelerate progress in this expanding area of biology.