Identification of novel proteins binding the AU-rich element of α-prothymosin mRNA through the selection of open reading frames (RIDome)

Abstract

We describe here a platform for high-throughput protein expression and interaction analysis aimed at identifying the RNA-interacting domainome. This approach combines the selection of a phage library displaying “filtered” open reading frames with next-generation DNA sequencing. The method was validated using an RNA bait corresponding to the AU-rich element of α-prothymosin, an RNA motif that promotes mRNA stability and translation through its interaction with the RNA-binding protein ELAVL1. With this strategy, we not only confirmed known RNA-binding proteins that specifically interact with the target RNA (such as ELAVL1/HuR and RBM38) but also identified proteins not previously known to be ARE-binding (R3HDM2 and RALY). We propose this technology as a novel approach for studying the RNA-binding proteome.

List of Abbreviations

ARE

AU-rich element

BSA

bovine serum albumin

CFU

colony-forming unit

ELISA

enzyme-linked immunosorbent assay

MS

mass spectrometry

O/N

overnight

ORF

open reading frame

PBS

phosphate-buffered saline

RBD

RNA-binding domain

RBP

RNA-binding protein

RBPome

RNA-binding proteome

RT

room temperature

SPR

surface plasmon resonance

Introduction

In recent years, the RNA world has expanded tremendously as a huge array of new RNA molecules have been discovered through transcriptome analyses (reviewed in ^1,2). The vast majority of these RNAs lack the features that are typically considered to predict protein-coding capability ³ and are therefore globally classified as non-coding RNAs (ncRNAs). Although non-coding, many of them play fundamental roles in numerous biological processes, ranging from gene expression regulation and epigenetic modifications to RNA processing and translation (reviewed in ⁴). These findings are not surprising because it is now well known that many and diverse RNAs, including messenger (m)RNAs, can perform biological tasks by virtue of their ability to adopt complex secondary and tertiary structures. In the cell, these are targets of RNA-binding proteins (RBPs) ⁵ that mediate the assembly of functional ribonucleoprotein complexes (RNPs). Therefore, there have been intensive efforts to develop in vitro and in vivo methods to identify and profile the cohort of RBPs associated with RNAs of interest (reviewed in ⁶). These methods are based on a similar concept that uses specific RNA sequences to bait interacting RBPs that are then identified by specific reagents or by mass spectrometry (MS).

Recent updates to these approaches have extended the analysis to complex mixtures of transcripts, such as polyadenylated RNAs, allowing previously unprecedented views of the complexity and dynamics of RNA-protein interactions.^7,8 Nevertheless, the number of RBPs that have so far been associated with given transcripts is still a minor fraction of the entire predicted RNA-binding proteome (RBPome). This is likely due to some caveats of these approaches. The procedures consist of elaborate, multistep processes that require optimization and affinity reagents to purify the desired RNP complexes. Moreover, because of the sensitivity limitations of MS, RNPs composed of low-abundance RNAs and/or RBPs tend to be underrepresented,⁹ while proteins with low-complexity sequences are overrepresented.¹⁰

In principle, some of these limitations (e.g., poorly expressed RBPs) could be overcome by approaches based on the screening of protein libraries, with the advantage of coupling phenotype to genotype identification.

Several methods have been proposed to work in vivo ^11-13 or in vitro.^14-16 Generally, they use any format of protein library that can faithfully represent the full-length proteome of interest, either as an array of candidate proteins or as products of selected open reading frame (ORF) libraries. However, these methods are burdened by the effects of screening modes and contests on protein stability and folding, particularly when full-length proteins are tested.^17,18 These problems could be solved, or considerably attenuated, given the chance to select correctly folded and active proteins domains. Due to their smaller size and conserved structural folding,¹⁹ protein domains can be independently expressed while preserving their individual functions.²⁰ Thus, screening a library that faithfully represents most or all of the functional domains encoded by a genome (domainome), could provide a simple method to annotate gene products, including those encoding RBPs.

In previous works, we have demonstrated that applying β-lactamase-based filtering to randomly fragmented DNA or cDNA from diverse sources makes it possible to generate phage-display libraries enriched for genic ORFs.^21-23 We believe this is because only fragments of functional ORFs can form foldable domains that do not adversely affect the folding and activity of the fused β-lactamase reporter protein. Random ORFs do not fold into coherent domains and lead to the aggregation, misfolding and inactivation of the folding reporter. When combined with next-generation sequencing (NGS), as a way to analyze their complexity, these libraries are used as universal reagents that can be screened for several activities, including rapid interactome profiling.^23,24 In accordance with this assumption, we wanted to evaluate whether such an approach could add power to the field of RNA-protein interaction discovery.

Here, we show that by screening a human cDNA-derived ORF library with the AU-rich element (ARE) from the α-prothymosin (PTMA) mRNA, a well characterized RNA structure,²⁵ we succeeded in selecting a panel of RNA-binding domains (RBDs) ascribable to proteins known to bind ARE motifs, such as ELAVL1 and RBM38, and also including other proteins not previously known to be ARE-binding proteins (ARE-BPs), such as R3HDM2 and RALY. Overall, our method provides a simple and fast means to identify RBPs for a target RNA, independent of their relative abundance.

Results

This work stems from previous studies in which we succeeded in profiling the interactomes of single or complex mixtures of proteins ^23,24 using a platform that combines ORF phage display library selection with NGS analysis. Encouraged by these results, we decided to exploit the power of this platform for revealing RNA-protein interactions.

RIDome pipeline overview

The RNA Interacting Domainome (RIDome) pipeline features the following key steps, outlined in Figure 1A:

1. Selection of ORFs binding to a target RNA. ORF-displaying phages are challenged with a biotinylated RNA bait through 2 cycles of selection and amplification.

2. Massive sequencing of selected phage inserts and ranking of corresponding genes. ORF inserts are recovered from selected phages and sequenced by NGS, and the corresponding genes are ranked according to read frequency.

3. Recovery and validation of the top-ranking ORFs in vitro. High-scoring ORFs are recloned from the library by inverse PCR, and their interactions with the target are validated in vitro by ELISA- and Surface Plasmon Resonance (SPR)-based assays.

4. Validation of the interacting ORFs in vivo. RNA-binding ORFs are subcloned into a compatible mammalian expression vector and co-transfected into mammalian cells with a reporter DNA construct harboring the target sequence. The interactions are confirmed by RNA immunoprecipitation and functional assays.

Figure 1.

Overview of the RIDome technology platform. (A) The overall strategy consists of 4 key steps: (1) Selection of interacting ORFs on a target RNA. Displaying phages are challenged with a biotinylated RNA bait through 2 cycles of selection and amplification. (2) Massive sequencing of selected inserts and the ranking of reads. ORF inserts are recovered from the selected libraries, sequenced with 454 pyrosequencing, ranked and scored according to their frequency. (3) Recovery and validation in vitro of the top ranking ORFs. High-scoring ORFs are recloned from the library by inverse PCR, and their interactions are validated by ELISA-based assays. (4) In vivo validation of the interacting ORFs. RNA-binding ORFs are subcloned into a compatible mammalian expression vector and co-transfected in mammalian cells with a reporter DNA construct harboring the target sequence. The interactions are confirmed by RNA immunoprecipitation and functional assays. (B) The RNA baits used in the study and the prediction of their secondary structures with RNAfold.

RBPome coverage

As a proof of principle, we first assessed the extent of the predicted RBPome covered by our ORF library. For this purpose, we used library sequencing data that had been collected in previous and ongoing studies (Additional File 1) and compared those data to the largest dataset of RNA-binding proteins recently published by Gerstberger S. et al.²⁶ Although presumably not exhaustive, this data set includes 1542 experimentally and/or computationally predicted RBPs, and of these, we found that 1497 RBPs (> 97%) are represented by at least one read in our library (Fig. 2A). The same database also contains a census of 1704 transcription factors (TFs), a protein family numerically similar to RBPs. Similarly, we found that 1672 TFs are represented by at least one read in our library, corresponding to > 98% of TFs annotated in the database (Fig. 2A). These findings are consistent with data obtained by the thorough characterization of the library (unpublished results) and support the concept that coupling β-lactamase-based filtering to phage display enriches for genic rather than spurious ORFs, which is a prerequisite of selecting for functional domains.

Figure 2.

Enrichment of RNA-binding ORFs after phage selection. (A) RBPome coverage. To assess the extent of the predicted RBPome covered by our ORF library, we compared our data to the largest datasets of RNA-binding proteins ²⁶ and found that 1497 RBPs (> 97%) were covered by at least one read. The ORF library was also compared to a data set of 1704 transcription factors (TFs), and we found that 1672 TFs (> 98%) were represented. (B) Enrichment for RNA-binding proteins after phage selection. A shortened list of top-interacting clones was compared to the RBP and TF datasets. The selection was enriched specifically for RBPs. (C) Functional category enrichment of selected genes. Gene ontology (GO) terms analysis of the 36 top selected genes was performed using GOrilla. The y-axis shows significantly (p < 0.001) enriched GO terms relative to Process, Function and Component, respectively, and the x-axis shows the number of genes related to those terms.

Library screening and the identification of ARE-interacting domains

After assessing the RBPome coverage of the library, we decided to screen it using the AU-rich element (ARE), a well-characterized RNA sequence, as bait. ARE is a functional element found in the 3′-untranslated regions (UTRs) of many mRNAs, and it is known as the most common determinant of RNA stability in mammalian cells.²⁷ Although little sequence similarity is shared by different AREs, they generally contain an AUUUA motif embedded in U-rich sequences ²⁸ and are predicted to fold in stem-loop structures (Fig. 1B).^28,29 We chose the ARE from the α-prothymosin mRNA (ARE^PTMA) because it is a well-defined element ²⁵ predicted to fold in a stable, bulge-less stem-loop structure (Fig. 1B) that is known to be a target of RBPs in vitro and in vivo.^25,30 The sequence was synthesized as 3′-biotinylated RNA 24mers and used in the discovery pipeline. As a control, we synthesized a mutant (AREmut^PTMA) predicted to not fold in a stable stem-loop structure (Fig. 1B).

The library selection was performed as reported ^31,32 with minor changes of solutions composition to safeguard the RNA bait (see the Materials and Methods section). After two cycles of selection, phagemid DNAs were recovered, and ORF inserts were amplified and sequenced according to a 454 protocol. To limit the time-consuming analysis of large amounts of data, we decided to sequence a minimal fraction of the selection output, sufficient to contain an overview of the most likely repertoire of ARE-interacting ORFs. We obtained 71116 reads that were analyzed with the NGS Transcriptome profile explorer (NGS-Trex) system ³³ and mapped onto the human genome (NCBI build 36). Sequences matching annotated genes were then ranked as described,³¹ and the resulting list comprised 4971 genes (Additional File 1). Using a frequency of representation higher than 1/1000 reads as the cut off, only 36 genes were identified (Additional File 1). Redundantly mapped genes (DUXs, MTNDP4P12, MTNDP5P11, OVCA2, and STX16-NPEPL1) were removed from the list. The specificity of selection was validated by comparing the list of genes to the RBP and TF databases. Not surprisingly, we found that although no TFs were present, 15 of the 36 genes were RBPs (Fig. 2B).

Furthermore, we analyzed the identified genes for functional annotation using the Gene Ontology enRIchment anaLysis and visuaLizAtion tool (http://cbl-gorilla.cs.technion.ac.il).³⁴ We used the “2 lists of genes” running mode, setting the gene list from the (N)on (S)elected library as the “background” and the list from the selection as the “target” set. GOrilla associated only 32 genes with a GO term because 4 genes (DUX4L6, BSDC1, H19 and LOC400550) were not in the GO database. We found that the genes ranked in our list were enriched in 18 GO terms (Figure 2C and Additional File 2), most of which (14 of the 18) referred to RNA-related processes, a finding that confirms the utility of our selection as a simple and straightforward method for obtaining an overview of the potential ARE-RIDome.

Rescue and validation of ARE-interacting domains in vitro

To verify the binding properties of the enriched ORFs, we recovered the clones corresponding to the top 12 listed genes. To this end, each top gene was first individually analyzed with the NGS-Trex tool, and a typical result window is shown in Figure 3 (ELAVL1 and R3HDM2 are given as 2 examples). For each mapped gene, NGS-Trex shows its genomic context, with all supporting reads aligned to the gene. The blue bars at the top of each panel represent the genes, and the green boxes correspond to the exons (RefSeq mRNAs). From this analysis, it is possible to compare the supporting sequences obtained before (NS) and after selection. Particularly informative is the “focus index,” which represents the ratio between the depth of read coverage at the deepest site and the total number of reads per gene.³³ The closer this ratio is to one, the more ‘focused’ the reads are to a single site or domain within the gene, while a lower index indicates a wider distribution of the reads on each gene. As shown for ELAVL1, a substantial increase in the focus index (from 0.385 to 0.755) was obtained, with most reads centered on exons 3 and 4 (Fig. 3). Similarly, the focus index of R3HDM increased from 0.554 to 0.975, with most reads spanning exons 5 to 8 (Fig. 3).

Figure 3.

Analysis of ELAVL1 and R3HDM2 with the NGS-Trex. The picture shows the gene aligned to the supporting sequences obtained before (NS) and after selection (selected). The blue bars at the top of each panel represent the gene, and the green boxes correspond to the exons (RefSeq mRNAs) of the gene. An increase in the “focus index” following phage selection indicates that ORFs are enriched at a single site or domain within the gene. After analysis, the ORF clones are recovered from the selected library by inverse PCR, using a pair of back-to-back specific primers centered on the sequence shared by overlapping reads.

This analysis allowed us to exclude 3 top genes (highlighted in Additional File 1) because the corresponding fragments either were overrepresented in the NS library but not enriched by selection (H19) or did not correspond to annotated genic ORFs (MMP14 and DUX4L6). Candidate ORFs were rescued from the selected library by inverse PCR (see the Materials and Methods section). After transformation, 8 clones from each ligation (96 clones in total) were randomly picked and grown in a multiwell plate, sequenced to confirm gene/ORF correspondence and then analyzed by phage ELISA on ARE^PTMA, AREmut^PTMA and control targets. Among those tested, RBM38-, ELAVL1-, RALY- and R3HDM2-derived clones resulted in the most robust binding specificity (Fig. 4A). RBM38 and ELAVL1 are RBPs known to bind AU-rich elements in several mRNA 3′ UTRs and to regulate their stability, splicing and translation. Their RNA-binding properties are mediated by conserved RNA recognition motifs (RRMs). These RRMs are present as a single module in RBM38 ³⁵ or as a bipartite module in ELAVL1, where 2 tandem N-terminal domains (RRM1 and RRM2) selectively bind ARE, while RRM3 interacts with the poly(A) tail and other proteins.³⁶ Furthermore, when the ORF inserts of tested clones were sequenced, we observed that these inserts faithfully represented the predicted RNA-binding domains (Fig. 4B). For instance, all tested ELAVL1-derived clones encoded the RRM1-2 module, and no clones were found with different motif combinations (e.g.,, RRM1, RRM2 or RRM2-3), indicating that only the 2 N-terminal motifs can fold correctly and bind to the ARE sequence. This evidence is consistent with previous observations ³¹ and supports the power of our method for favoring the display of functional domains.

Figure 4.

In vitro validation of the selected RNA-binding proteins by ELISA-based assays. (A) Validation by phage ELISA. The reactivity of 12 top-ranking genes was tested on the ARE^PTMA RNA oligonucleotide. To test specificity, a mutated RNA (AREmut^PTMA), an ssDNA oligonucleotide and streptavidin served as controls. Values are indicated as the fold signal vs. the background (uncoated wells). (B) BLASTP analysis of ORF clones validated by GST ELISA. ELAVL1, RBM38, R3HDM2 and RALY contain at least one conserved RNA-binding domain. (C) Validation by GST ELISA. Selected ORFs with positive phage ELISA results were subcloned into a compatible pGEX vector and purified as GST fusion proteins. Assays were performed as in A.

Interestingly, RALY- and R3HDM2-derived clones were positive in phage ELISA (Fig. 4A). RALY is a member of the heterogeneous nuclear ribonucleoprotein (hnRNP) gene family and is involved in pre-mRNA splicing and tumor development.³⁷ RALY contains a single RRM at the N-terminus, with binding properties that have not yet been defined. R3HDM2 is a large protein containing a conserved motif consisting of an invariant arginine and a highly conserved histidine that are separated by 3 residues (R3H motif). Structural studies have indicated that the R3H domain might be involved in interactions with single-stranded nucleic acids.³⁸ None of these proteins had previously been shown to bind ARE sequences. Importantly, also in the case of RALY- and R3HDM2-derived phages, we observed that selected ORF sequences were limited to the corresponding binding domains (Fig. 4B), further evidence of the functional constraint exerted by filtering. The other 8 tested candidates either were negative by phage ELISA (RBMX2, MYH9, NPEPL1, PCBD2, MAP1A) or did not show binding specificity (TOP1, DNAJC7, EIF5B).

To further confirm the binding properties of positive clones, we chose a representative ORF for each gene and subcloned it into a modified pGEX vector to produce GST-fusion products with C-terminal FLAG-tags. Proteins were expressed in E. coli, purified by glutathione affinity chromatography, and then assayed by ELISA on ARE targets or controls. As shown in Figure 4C, the selected domains of ELAVL1, RBM38, RALY and R3HDM2, but not those of EIF5B, were confirmed to specifically bind the ARE^PTMA target, indicating that their folding and activity was not due to the phage context.

To directly compare the RNA-binding properties of ELAVL1-, RBM38-, RALY- and R3HDM2-derived GST-fusion products, we used surface plasmon resonance (SPR) to measure both the affinity and kinetics of binding to the ARE^PTMA sequence (see the Materials and Methods section). As shown in Figure 5, all fusion products bound the ARE^PTMA RNA with a K_D in the nanomolar range: RBM38 = 1 nM; ELAVL1 = 5 nM; R3HDM2 = 21 nM; and RALY = 55 nM. Because the binding properties of ELAVL1 to ARE targets have been characterized,^39,40 we wanted to compare these to the properties obtained with the GST-ORF_ELAVL1 product. Thus, we cloned the full-length coding sequence of ELAVL1 from HEK293 cells, expressed it in E. coli and performed SPR analyses as above. We also prepared full-length RBM38 (isoform b) as a recombinant GST fusion because no data are currently available on the affinity of this protein to target ARE sequences. As shown in Figure 5, full-length ELAVL1 and RBM38 proteins showed K_D values of 1 nM and 0.4 nM, respectively. The value we observed with ELAVL1 is in a good agreement with K_D values previously reported.³⁹ In contrast to previous reports, full-length ELAVL1 and GST-ORF_ELAVL1 did not show a significant difference in affinity for the ARE^PTMA target.³⁹ Similarly, we did not observe a significant difference in binding properties between full-length RBM38 and GST-ORF_RBM38 (0.4 nM vs. 1 nM, respectively).

Figure 5.

Kinetic analysis of RNA-protein interactions by SPR. Sensorgrams of ELAVL1, RBM38, R3HDM2 and RALY (ORFs) and ELAVL1 and RBM38 (full-length) binding to the ARE^PTMA RNA oligonucleotide are shown. Biotinylated RNA was captured on SA-coated sensor chips, and increasing concentrations of protein were injected over the surface. Injections were performed for 120 s (association phase), followed by a 300-s flow of running buffer to assess dissociation. The kinetic data were fitted to a 2-state binding model: 1/[(ka1/kd1) × (1+ka2/kd2)], where ka1 and kd1 are the association and dissociation rate constants, respectively, and ka2 and kd2 are the forward and reverse rate constants for conformational change.

Validation of ARE-Interacting Domains in vivo

To assess whether and to what extent the ORFs validated in vitro function in vivo, we made 4 pcDNA-based constructs (ELAVL1-, RBM38-, RALY- and R3HDM2-ORF) that allow the expression of V5-tagged products and used them to transfect HeLa cells. We also used a full-length ELAVL1 construct (ELAVL1-FL) as a reference because its localization and function are supported by published data.⁴¹

First, we analyzed the transfected cells by immunofluorescence to assess the cellular localization of ORF products. Representative results are shown in Supplemental Figure 1. As expected, the ELAVL1-ORF product localized mostly in the cytoplasm of transfected cells because it does not include the HNS sequence, which is responsible for the nuclear localization of ELAVL1.⁴¹ Consistent with reported data on the cellular localization of RBM38 isoforms,³⁵ the RBM38-ORF protein resulted in mostly cytosolic localization. Surprisingly, a prevalent nuclear/perinuclear localization of the RALY-ORF product (Supplemental Figure 1) was observed, although this protein does not seem to contain either a canonical nuclear localization signal (NLS) or the bipartite NLS element predicted by others.³⁷ We cannot exclude the possibility that the protein is carried to the nucleus by interacting cargos. Finally, the R3HDM2-ORF product showed mainly cytoplasmic localization.

Second, we tested whether ORF-encoded domains were able to bind target mRNAs in vivo. For this purpose, we generated a secreted luciferase-based ARE reporter construct (secNluc-ARE) containing the 3′ UTR from the human α-prothymosin (PTMA) gene and used it to co-transfect HEK293 cells with single-ORF or ELAV-FL constructs. RNP complexes were immunoprecipitated with an anti-V5 antibody. Immunocomplexes were split and analyzed by qRT-PCR to quantify the amount of bound mRNA (Fig. 6A) and by Western blotting (Fig. 6B) to assess the effective immunoprecipitation of V5-tagged ORFs. RNA quantification is expressed as the fold enrichment versus the negative control (cells transfected with the empty vector). RBM38-ORF and ELAVL1-ORF resulted in > 50- and 3.5-fold enrichment, respectively, while no RNA enrichment was observed in RALY-ORF and R3HDM2-ORF immunocomplexes. Notably, these results reflected the affinity differences shown by the ORF products in vitro (Fig. 5) and circumstantially support the rank differences of the selected ORFs.

Figure 6.

In vivo validation of RBM38, ELAVL1, RALY and R3HDM2 ORFs. (A, B) RNA immunoprecipitation. HEK293T cells were co-transfected with the ORF-V5 and the secNluc reporter constructs and subjected to RNA immunoprecipitation with an anti-V5 antibody. Following immunoprecipitation, both RNA and proteins were recovered and analyzed by A) qRT-PCR and B) Western blotting. (C) Luciferase reporter assay. The ARE of PTMA was cloned downstream of the secNluc in a luciferase reporter construct and transfected into HEK293T cells. Twenty-four hours following transfection, cell culture supernatants were collected and luciferase activity was measured. (D) RNA immunoprecipitation of endogenous PTMA mRNA. HEK293T cells were transfected with the V5-tagged ORFs. Twenty-four hour following transfection, cells were cross-linked with formaldehyde and RNP complexes were immunoprecipitated with an anti-V5 antibody as above. RNA was recovered by reverse cross-linking and analyzed by qRT-PCR to quantify the amount of bound mRNA. GAPDH was used as a negative control for background. The data are shown as the mean ± the standard error of the mean (n ≥ 3).

We then assessed whether ORF products can influence the metabolism of ARE-containing sequences in vivo by measuring luciferase activity directly from conditioned media at 24 hours post-transfection. As shown in Figure 6C, all constructs were able to enhance luciferase activity, and the ORF constructs were more effective than the ELAVL1-FL used as the positive control. We believe that these differences are due to a lower complexity of ORF-encoded RBDs in that the activity of corresponding RBPs – e.g., ELAVL1-FL – very likely reflects the interplay between different domains and/or other players.⁴² Intriguingly, RALY-ORF efficiently increased reporter protein expression (Fig. 6C), despite its prevalent nuclear localization (supplemental Figure 1).

Finally, we wanted to assess if the 2 novel ARE-BPs interact in vivo with endogenous PTMA mRNA. According to reference databases (http://www.proteinatlas.org and (http://www.hprd.org) PTMA is expressed in HEK293T cells. Thus, we first verified the presence of PTMA mRNA by RT-PCR and then we transfected HEK293 cells with RALY-ORF or R3HDM2-ORF or ELAV-FL (as positive control). Twenty-four hour following transfection, cells were cross-linked with formaldehyde and RNP complexes were immunoprecipitated with an anti-V5 antibody. RNA was recovered by reverse cross-linking and analyzed by qRT-PCR to quantify the amount of bound PTMA mRNA with the respect to GAPDH mRNA used as a negative control. As shown in Figure 6D, results confirm the ability of both RALY-ORF and R3HDM2-ORF products to bind PTMA mRNA in vivo, being R3HDM2-ORF more efficient than RALY-ORF and in accordance with respective affinities (Fig. 5C, D).

Discussion

Here, we present a novel in vitro method that allows the identification of RBPs bound to a target RNA via the selection of RNA-interacting domains. Essentially, this method consists of the screening of a cDNA-derived ORF-filtered phage display library with a target RNA, followed by the deep sequencing of the ORF inserts of selected phages to identify putative positive clones. The result is a ranked list of genes that directs analyses and validations to the best candidates.

Compared to other established RNA-centric methods for identifying RNA-protein interactions (reviewed in ^5,43,44), RIDome has certain practical advantage (summarized in the Supplementary Table 1).

First, RIDome is easy to implement and does not require mass spectrometry (MS) to identify RNA-interacting proteins. MS-based methods, particularly the in vivo formats, have proved to be powerful for revealing the complexity and dynamics of RNA-protein interactions. However, these methods require large amounts of starting material to purify enough protein to allow detection,^43,45 and despite significant improvements in MS technology, the identification of low-abundance RNA-protein complexes remains a major challenge. Moreover, the experimental setup for these approaches is generally complicated by the necessity to perform several parallel purifications on target or control RNAs to distinguish non-specific interactions and ensure the robustness of the results. In contrast, ORF-filtered display libraries are robust and stable reagents that can be repeatedly screened with single or multiple RNA targets. No laborious and costly work is required to grow cells and prepare extracts. If unavailable, libraries can be constructed by established protocols,³¹ starting from cDNA ^31,46 or even genomic DNA ⁴⁷ and using different formats, such as phages,⁴⁶ bacteria ⁴⁸ or yeast.⁴⁹ Library normalization and representativeness, which are crucial for facilitating the identification of less-expressed interactors, can be easily assessed by NGS analysis. When using a phage display library as demonstrated here, the procedure can be operationally very fast and straightforward. This method allows serial screenings on different target RNAs without the need to perform parallel experiments on negative controls because these can be used both as competitors during the selection process and as targets in the phage ELISA validation assays.

Second, as is typical for display technologies, the identification and cloning of the genotypes of selected phages are straightforward. Moreover, we show that the cloning of discrete functional domains is favored by the RIDome approach. We believe this is due to the ORF-filtering step introduced for the library construction; random ORFs, or those encoding partial domains, do not fold into coherent domains and lead to aggregation, misfolding and inactivation of the folding reporter ß-lactamase. We believe that this feature offers significant and practical value. For instance, validated domains can be readily manipulated (e.g., mutagenized) to obtain insights into the molecular basis of RNA-protein interaction specificities or used for structural studies or, conversely, used for a protein-centric identification of RNA targets.

Third, the RIDome output is a list of genes that bears some resemblance to the output from other RNA-centric methods used to identify RNA-protein interactions. The list can be analyzed using several bioinformatics tools to obtain inferences regarding cellular processes, localizations or the structural/functional constrains of candidate RBPs.^34,50 Nevertheless, the informative power of these analyses is strictly influenced by the robustness of experimental results and particularly by the background level of non-specific/false positives. In this respect, our approach provides a practical advantage. Phage ELISA is a simple and robust assay that can be used as a first validation step to discriminate between specific and non-specific binders (see Fig. 4A). Above all, it can be easily scaled up to hundreds or thousands of tests. For instance, as shown here (Fig. 4A), it has been as simple to assess the binding specificities of known (RBM38 and ELAVL1) or novel (RALY and R3HDM2) ARE-RBPs as to assess the non-specific binding of known (TOP1 and DNAJC7) or potential (MYH9 and EIF5B) RBPs. We think that it is important to note that the availability of a simple and high-throughput validation assay allows the reconsideration of the critical challenges encountered when purifying proteins from complex mixtures for MS analyses. Generally, great care must be taken to find the appropriate stringency conditions and negative controls to minimize non-specific interactions, particularly for abundant proteins. This is because validation can be labor- and time-consuming and can depend on the availability of specific reagents. Conversely, stringency might negatively affect the purification of weak but significant interactors, particularly if they are novel. We believe that these problems are overcome, at least in part, by the RIDome approach; despite the more or less permissive conditions used in the selection/purification step, any candidate gene listed in the NGS output can be validated without the need for specific reagents and with relatively limited effort. This can lead to the discovery of novel RBPs or to the assignment of binding properties to unpredicted RBPs.

The limits of the RIDome approach are mainly due to its in vitro format (summarized in the supplementary Table 1). As with other RNA-centric in vitro approaches, a synthetic RNA is employed to select/purify interacting phages. Consequently, it is possible to capture only the subset of proteins that bind static, correctly structured sequences within the RNA bait. The proteins that bind structures dynamically adopted by an RNA in vivo are likely missed. Compared to other approaches, a further limitation of RIDome is the type of starting material – cell extracts vs. library. Proteins that bind RNA because of the synergistic activities of other factors or because of post-translational modifications are likely not captured by the RIDome approach.

Although in this work, aimed at tuning and validating a novel approach, the bait was an extensively characterized short RNA sequence/structure that has previously been employed to identify ARE-RBPs, we anticipate that RIDome could be a helpful complementary approach for identifying the repertoire of RNA-binding proteins of diverse RNA elements, such as splicing sites,⁵¹ mRNA localization elements (LE),⁵² constitutive decay elements (CDEs),⁵³ RNA stability elements (RSEs)⁵⁴ and less characterized transcripts, such as long non-coding RNAs (lncRNAs).⁵⁵ Overall, the RIDome approach provides an RNA-centric global assessment of RNA-RBP interactions via the identifications of RNA-binding domains.

Materials and Methods

Biopanning procedures

The ORF phage library used in this study has been previously described ³¹; the production and rescue of phagemids were performed according to published protocols.³² For biopanning experiments, phage particles were suspended in PBS buffer at a concentration of 10¹¹ cfu/µl, and for each selection, 10¹² phages were used. The bait was a 3′ biotinylated RNA oligonucleotide corresponding to the AU-rich element of α-prothymosin (ARE ^PTMA, 5′-GGAAAUUUGUUUGUAUUUUUAGCU-biotin). Biopanning experiments were performed as follows: 20 µl of streptavidin-coated magnetic beads (New England Biolabs) were washed in TENT Buffer (10 mM Tris-Cl pH 8.0, 1 mM EDTA, 250 mM NaCl, 0.5% Triton X-100) and then incubated with 10¹² phages in 100 µl of TENT Buffer for 30 min at RT as a cleaning step. The ARE^PTMA oligonucleotide was diluted to 30 nM in TENT buffer containing 100 U/µl of the RNase inhibitor SUPERase-IN (Life Technologies), and then, 100 µl (3 pmoles) of the mixture was added to 20 µl of streptavidin magnetic beads and incubated for 20 min at RT. While fine-tuning the protocol, independent selections were performed using tRNA or ssDNA as competitors, and various strategies were attempted to recover bound phages (E. coli infection with or without RNase treatment), but we did not observe substantial differences among output ranks. Thus, as a standard procedure, pre-cleaned phages were added to the RNA-conjugated beads and incubated for 45 min at RT in the presence of 1 µg/µl tRNA or 1 µg/µl herring sperm ssDNA as competitors. The beads were then washed extensively in TENT buffer. The bound phages were eluted by infecting with 2 ml of E. coli DH5α (OD₆₀₀= 0.5) at 37°C for 45 min. The amplified phages were used for a second cycle of selection as reported.^32,56

Deep Sequencing of cDNA Inserts

After the second round of selection, colonies growing on agar plates were harvested, and the phagemid DNAs were isolated using a standard miniprep procedure. The cDNA inserts were processed for 454 sequencing according to the Titanium Rapid Library preparation method manual (Roche, Milano, Italy). The DNA library was quantified using a PicoGreen DNA Quantitation Kit (Life Technologies) and checked for quality by capillary electrophoresis (Agilent Bioanalyzer 2100 with the High Sensitivity DNA assay kit). The DNA library was then amplified in emulsion following the Titanium LIB-L emPCR protocol (Roche). The reaction was recovered by isopropanol emulsion breaking, and the beads carrying clonally amplified DNA fragments on their surface were enriched. The enriched sample was loaded onto one PicoTiterPlate (PTP) and was sequenced according to the 454 GS-FLX Titanium protocol on the GS-Junior sequencing platform.

Bioinformatics analysis

Sequences were processed with the NGS Transcriptome profile explorer (NGS-Trex) system (accessible at https://www.ngs-trex.disit.unipmn.it/Trex/cms/),³³ a custom analysis workflow procedure mainly based on PERL scripts. Both raw and analyzed data were stored in a relational database. Briefly, sequences were mapped onto the human genome (NCBI build 36) using GMAP software, and matching sequences were compared with annotated genes. Each gene was then ranked according to the number of supporting sequences (defined as ‘coverage’). The ‘depth index’ for each gene was defined as the maximum number of overlapping sequences (i.e., sequences supporting the same genic region). The ‘focus index,’ defined as (depth – 1)/rank, ranged between 0 (indicating a broad distribution of sequences over the gene) and 1 (indicating that all sequences were ‘focused’ on the same region).

Enrichment analysis

To implement functional enrichment, we used Gorilla (Gene Ontology enRIchment anaLysis and visuaLizAtion tool) for GO enrichment. A distinctive feature of GOrilla (http://cbl-gorilla.cs.technion.ac.il) is that it allows for the identification and visualization of enriched GO terms in ranked lists of genes.³⁴ It can be run in one of 2 modes: (1) searching for enriched GO terms that appear densely at the top of a ranked list of genes (“single-ranked list” mode) or (2) searching for enriched GO terms in a target list of genes compared to a background list of genes (“2 unranked lists” mode). Lists of ranked genes were retrieved from the NGS-Trex interface and analyzed in the “2 lists of genes” running mode, using the gene list from the NS library as the “background” set and the gene list from the selection as the “target” set. GO terms with p-values < 0.001 were chosen as statistically significant terms.

Rescue of phagemid clones by inverse PCR

A pair of specific back-to-back outward primers was designed for each of the tested genes, centering on the nucleotide region identified by the overlapping reads (oligonucleotide sequences are provided in Additional File 3). Here, 50 ng of the phagemidic DNA minipreps was used as the template, and inverse PCR reactions were performed with a Phusion High-Fidelity DNA Polymerase (Thermo Scientific). PCR products were gel purified, phosphorylated with T4 polynucleotide kinase, ligated by T4 DNA ligase and transformed into E. coli DH5αF′ competent cells. Colonies growing on ampicillin plates were randomly picked and inoculated into 1 ml 2x TY medium in a deep 96-well microtiter plate and grown under constant air supply using a custom-built air-well sparging minifermenter system.⁵⁷

Other cloning

For the bacterial expression of GST-fusion products, ORF fragments were excised with BssHII-NheI from the phagemid DNA and subcloned into a custom-designed pGEX-FLAG expression vector as previously described.²⁴ The vector harbors a FLAG-tag (DYKDDDDK) for the C-terminal tagging of expressed proteins. For the mammalian expression of the ORFs, we designed a vector (pcDNA-V5-AscI/NheI) from a pcDNA3.1 backbone (Invitrogen) by cloning a cassette Kozak-ATG-AscI-NheI-V5 tag-stop. ORFs were subcloned as described above. The full-length (FL) coding sequence of human ELAVL1 was obtained by RT-PCR from HEK293T cells and cloned into the pCI-neo vector (Promega). ELAV-FL was epitope-tagged by replacing the terminal stop codon with nucleotides encoding a V5-tag (GKPIPNPLLGLD). To generate a luciferase reporter under the control of the α-prothymosin AU-rich element, we first removed the EGFP cassette from the pEGFP-C2 vector (Clontech) and replaced it with the secreted NanoLuc® luciferase (Promega); then, 316 bp of the 3′UTR of α-prothymosin was amplified by RT-PCR from HEK293T cells and cloned at the 3′ end of the secNanoLuc® cassette, generating the pCMVsecNluc^ARE vector.

GST-fusion protein expression and purification

ORF fragments, subcloned in pGEX-FLAG, were transformed into E. coli BL21(DE3) cells. Bacterial cultures (100 ml) were grown at 28°C to an OD₆₀₀ = 0.5 AU, induced with 1 mM IPTG for 3 hours and centrifuged. Bacterial pellets were resuspended in 10 ml of lysis buffer (PBS containing 0.5 M NaCl, 1 mM imidazole, 200 µg/ml lysozyme, 20 µg/ml DNase, protease inhibitors), incubated for 30 min and sonicated for 2-3 minutes. Cell debris was removed by centrifugation, and the supernatants were incubated with glutathione-agarose beads (Sigma-Aldrich) for 1 h at 4°C with gentle rotation. After three washes with 0.1% PBS-Tween and 3 with PBS, GST-fusion proteins were eluted in 750 µl of elution buffer (50 mM reduced glutathione, 100 mM NaCl, pH 8.0). The proteins were dialyzed against PBS and checked for purity and concentration by SDS-PAGE. The quantitative densitometry of proteins stained with Coomassie brilliant blue was performed with ImageJ software,⁵⁸ using BSA as a reference for protein quantification. The integrity of GST-fusion proteins was also determined by Western blotting using 2 monoclonal antibodies targeting the GST and FLAG tags (Sigma-Aldrich).

ELISA

The expression and testing of selected clones in ELISA-based assays, either in the phage format or as soluble GST-fusion polypeptides, was performed according to previously described standard protocols,¹⁶ with some modifications. Briefly, phage ELISA was performed with Microlon plates (Greiner Bio-one) coated O/N at 4°C with 10 μg/ml streptavidin. Wells were blocked with TENT buffer and rinsed, and then, biotinylated RNA or control DNA oligonucleotides (5 pmol/well, diluted in 100 µl TENT buffer containing 5 units of SUPERase) were captured on the plates. The phage-containing supernatants of individual clones diluted 1:1 in TENT buffer (with 5 units of SUPERase) were added to the wells and incubated for 45 min, followed by incubation with HRP-conjugated anti-M13 monoclonal antibody (GE Healthcare). The complexes were revealed with TMB (3,3′,5,5′ tetramethylbenzidine), and the absorbance was read at 450 nm using a Victor™ X4 multilabel plate reader (Perkin Elmer).

ELISA on soluble GST-fusion polypeptides was performed as follows: biotinylated oligonucleotides were captured on Microlon plates as described above. Wells were then incubated for 1 hour at RT with the purified proteins (5 pmol/well, diluted in 100 µl TENT buffer containing 5 units SUPERase), washed extensively with TENT buffer, incubated for 1 hour with a mouse monoclonal anti-GST antibody diluted 1:5,000 in TENT buffer, followed by incubation with an HRP-conjugated secondary antibody (Sigma-Aldrich). The bound proteins were examined as described above.

Biacore experiments

The dynamics of RNA interactions were characterized by SPR using a BIACORE T100 instrument (Biacore) essentially as described.³⁹ The oligonucleotide used in the analyses was the ARE^PTMA. The chemical biotinylation of the oligonucleotide allowed the immobilization of the RNA onto streptavidin-coated sensor chips (Series S Sensor Chip SA, Biacore). RNA was diluted to a final concentration of 1 mM in HBS buffer (10 mM HEPES, pH 7.4, 150 mM NaCl), followed by heating at 80°C for 10 min and cooling to room temperature. The sample was then diluted 500-fold in running buffer (10 mM HEPES, pH 7.4, 150 mM NaCl, 1 mM DTT, 0.025% surfactant P20; Biacore) and injected over the sensor chip surface at 10 µl/min at 25°C to generate an approximately 500 response unit (RU). The proteins were serially diluted in running buffer to the concentrations indicated in Figure 5 and injected at 25°C at a flow rate of 30 µl/min for 2–3 min. Surfaces were regenerated by injecting 2 M NaCl at 50 µl/min for one minute. Analyses were performed in duplicate, and any background signal from a streptavidin-only reference flow cell was subtracted from every data set. To determine the kinetics (association/dissociation rate constants; ka/k_d) as well as the affinities (K_D) of the protein-RNA interactions, the data were analyzed using a 2-state binding model: 1/[(ka1/kd1) × (1+ka2/kd2)] (ka1 and kd1 are the association and dissociation rate constants, respectively, and ka2 and kd2 are the forward and reverse rate constants for conformational change).

Immunofluorescence microscopy

HeLa cells were seeded onto glass coverslips and transfected with the respective pcDNA-V5-ORF plasmids. At 24 hours after transfection, cells were fixed with 4% paraformaldehyde and 4% sucrose in PBS for 15 min at RT, permeabilized with 0.2% Triton X-100 for 10 minutes and blocked with 1% BSA in PBS for 30 min. The cells were incubated for 1 hour with a mouse monoclonal anti-V5 antibody (diluted 1:100) and then with a Cy5-conjugated secondary antibody (1:200, Jackson Immuno Research), and treated for 5 min with RNase A (100 μg/ml in PBS) followed by propidium iodide counterstaining. Images were captured with a Leica DMIRE2 confocal fluorescence microscope equipped with Leica Confocal Software v.2.61. (Leica Microsystems).

Luciferase assays and RNA immunoprecipitation (RNA-IP)

HEK293T cells were seeded in 10-cm dishes and transiently co-transfected with pCMVsecNluc^ARE and the respective pcDNA-V5-ORF plasmids. At 24 hours after transfection, culture supernatants were collected and diluted 1:20, and the luciferase activity was determined with the Nano-Glo® Luciferase Assay (Promega) using a Victor™ X4 multilabel plate reader (Perkin Elmer). Cells were then collected in 1 ml RIP lysis buffer (25 mM Tris-HCl pH 7.4, 150 mM KCl, 0.5% IGEPAL CA-630, 5 mM MgCl₂, 0.5 mM DTT, protease inhibitors and 10 mM ribonucleoside vanadyl complex as an RNase inhibitor) and sonicated with 2 short pulses (15 sec). The cell lysates were centrifuged, and the supernatants were pre-cleared with agarose beads. From each lysate, 50 µl was saved as “input,” and then, 3 µg of mouse monoclonal anti-V5 antibody was added to the lysates and incubated O/N on a rotary mixer. Protein A-agarose beads were added and incubated for 1 hour. The beads were washed 4 times with RIP wash buffer (same as lysis buffer but with KCl increased to 300 mM for higher stringency), and then, RNA and proteins were eluted from the beads with TriFast reagent (Euroclone), according to the manufacturer's instructions. The RNA was alcohol-precipitated using GlycoBlue (Life Technologies) as a co-precipitant, dissolved in 10 µl RNase-free water and reverse transcribed immediately (High-Capacity cDNA reverse transcription kit, Applied Biosystems). qPCR amplification was performed using a CFX96 Real-Time PCR Detection System (Bio-Rad) with the Platinum® SYBR® Green qPCR SuperMix-UDG kit (Invitrogen). Primers were annealed to the NanoLuc coding sequence (sequence reported in the Additional File 3). To construct a standard curve, 10-fold serial dilutions of the “input” cDNA from the negative control (cells transfected with pCMVsecNluc^ARE) were used and tested in triplicate. The standard curve was plotted as the mean Cq values versus the log cDNA dilution. Regression analysis, standard curve slopes and amplification efficiencies were calculated using CFX manager™ software (Bio-Rad). Proteins recovered from the organic phase were dissolved in Laemmli buffer and analyzed by Western blotting.

RNA immunoprecipitation of endogenous PTMA mRNA

HEK293T cells were transfected with the respective pcDNA-V5-ORF plasmids in 10 cm dishes. Twenty-four hours following transfection cells were collected in PBS and cross-linked with 1% formaldehyde for 15 min as described.⁵⁹ RNP complexes were immunoprecipitated with an anti-V5 antibody as above, RNA was recovered by reverse cross-linking and analyzed by qRT-PCR to quantify the amount of bound mRNA. GAPDH was used as a negative control to estimate background. RNA quantification is expressed as % of RNA recovery vs. input RNA.

Disclosure of Potential Conflicts of Interest

No potential conflicts of interest were disclosed.

Authors' Contributions

DC, LP, CS and DS conceived the study and designed the experiments. LP, DC and AC performed the experiments. CP performed DNA sequencing. FM and IB designed and implemented NGS-Trex and analyzed the data. FG performed the SPR experiments. SZ and SG critically discussed the results and direction of the project and contributed to drafting the manuscript. DC, CS and LP wrote the paper with assistance from the other authors. All authors have read and approved the manuscript for publication.

Additional Data Files

Additional File 1. Catalog of genes covered by the non-selected ORF library (NS_List); list of genes after selection (selected list); top ranking genes (top genes).

Additional File 2. Enrichment results for the inferred functions (Gene Ontology) of the top-ranking genes.

Additional File 3. Oligonucleotides used in the study.

Supplementary Material

Supplemental data for this article can be accessed on the publisher's website.

Funding

This work has been supported by grants from the Compagnia di San Paolo (Turin, Italy) and Fondazione Cariplo (Milan, Italy) to SZ and DS. LP is supported by a Compagnia di San Paolo PhD scholarship.