The mRNA Binding Proteome of Proliferating and Differentiated Muscle Cells

Muscle formation is a coordinated process driven by extensive gene expression changes where single cells fuse together to form multinucleated muscle fibers. Newly synthesized mRNAs are then regulated by RNA binding proteins (RBPs), affecting post-transcriptional transcript metabolism. Here, we determined how large-scale gene expression changes affect the catalog of RBPs by studying proliferating and differentiated muscle cells in healthy and dystrophic conditions. Transcriptomic analysis showed that the expression of more than 7000 genes was affected during myogenesis. We identified 769 RBPs, of which 294 were muscle-specific and 49 were uniquely shared with cardiomyocytes. A subset of 32 RBPs (half of which were muscle-specific) was found to be preferentially associated with target mRNAs in either myoblasts (MBs) or myotubes (MTs). A large proportion of catalytic proteins were bound to mRNAs even though they lack classical RNA binding domains. Finally, we showed how the identification of cell-specific RBPs enabled the identification of biomarkers that can separate healthy individuals from dystrophic patients. Our data show how interactome data can shed light on new basic RNA biology as well as provide cell-specific data that can be used for diagnostic purposes.


Introduction
Skeletal muscle development and regeneration require a coordinated series of events during which single proliferating muscle stem cells fuse to form multinucleated muscle fibers. This unique process is driven by a cascade of transcriptional events that determine whether cells are committed to become contracting muscle or remain in a quiescent state. While transcriptional studies have elucidated the gene expression changes required for successful muscle formation in in vitro and in vivo models [1][2][3], studies focusing on how the observed gene expression changes are regulated by RNA binding proteins (RBPs) are lacking. RBPs are necessary for mRNAs to exert their functions within the cell environment by forming ribonucleoprotein complexes. RBPs have been described to affect the way mRNAs and miRNAs act by coordinating mRNA turnover, localization, and translation [4][5][6][7][8]. The fundamental role of RBPs is further underlined by their role in the development of genetic conditions such as neurodegenerative diseases [9] and neuromuscular diseases, for example in myotonic dystrophy type 1, which is caused by the sequestration of specific RBPs [10,11].
In the last five years, there has been great progress in the identification of RBPs. The development of improved enabling technologies has allowed high-throughput capture and identification of RBPs bound to polyadenylated RNAs. One of these methods is the interactome capture method, which has been used to identify the repertoire of RBPs in a variety of human and murine cell lines [12][13][14][15][16][17]. These studies have not only expanded our knowledge of the hundreds of proteins able to bind RNA, but they have also shown how several proteins bind their RNA targets through unconventional RNA binding domains [18]. Interestingly, different cell types have shown different repertoires of RBPs, suggesting that particular proteins may act as RNA binders in certain cell types but not in others.
In this work we aimed to uncover the catalog of RBPs present in proliferating and differentiated muscle cells and to study whether differences exist in cells obtained from patients affected by Duchenne muscular dystrophy (DMD), where the gene expression signature is highly altered.

Interactome capture enables identification of RBPs in muscle cells
To identify muscle-specific RBPs, proliferating muscle cells (myoblasts, MBs) were cultured and differentiated into multinucleated myotubes (MTs), mimicking the generation of muscle fibers. Both MB and MT cells were then irradiated with UV light to induce crosslinking via covalent bond formation between RNA and proteins. RBPs were purified after capturing polyadenylated mRNAs with an oligo-dT bound to magnetic beads. Samples were then split to enable both protein and RNA analyses ( Figure 1A). Silver staining images showed strong enrichment in protein content in crosslinked (CL) samples compared to uncrosslinked (noCL) samples ( Figure 1B, Figure S1A). To test whether the crosslinking protocol enabled protein enrichment after poly-A purification, we compared two replicates of healthy control (HC)-derived MB CL to two replicates of HC-derived MB noCL. Protein identification for both MB CL and MB noCL showed an increased number of identified proteins in CL conditions that was reproducible ( Figure 1C and D). In proliferating MBs, we identified 160 RBPs characterized by a positive fold change (FC) in CL samples compared to noCL samples and a significant difference according to differential representation analysis as tested by DESeq2. DESeq2 analysis provides a formal demonstration of the enrichment; nevertheless, in this case the enrichment is also clearly visible in the gel and evident in the mass spectrometry count data ( Figure 1B, Figure S1A and B). GO annotation of the identified proteins for both molecular function and biological process showed strong enrichment for RBPs and proteins involved in RNA biology ( Figure 1E and F). An extended analysis involving both MB and MT samples obtained from HCs and DMD patients (a total of eight samples with four MB and four MT samples from CL or noCL conditions) was performed (data available in Table S1). This analysis again showed a significant enrichment of 124 proteins out of the identified 655 proteins in the CL samples compared to the noCL samples ( Figure S1B). As expected, we were able to pull down proteins known to bind RNA such as CELF2, HNRNPM, and HDLBP, as well as RBPs involved in muscle diseases such as MBNL1, which plays a key role in the pathogenic mechanism of myotonic dystrophy type 1 ( Figure S1C). Therefore, we have shown that crosslinking is required to enrich for RBPs and to study the interactome of muscle cells. Further analyses of different experimental conditions were performed by comparing CL samples only.

Comparison of proliferating and differentiated cells
We compared proliferating single nucleated MB cells with differentiated multinucleated MT cultures from HCs or DMD patients (HC MB, HC MT, DMD MB, and DMD MT, respectively; Figure 2A). HC samples consisted of two replicates of KM155 cells and one replicate of AB117 cells, and DMD patient samples consisted of two replicates of 8036 cells and one replicate of 6311 cells (see Materials and methods for details).

RNA-seq analysis shows differences between cell types and between HC and DMD cells
RNA purification showed that the quantities of RNA recovered were reduced after crosslinking ( Figure 2B), and the total number of reads was similar across samples (51-64 million per sample). Sequencing of the purified mRNAs was performed to assess whether the transcriptional reprogramming of cells occurred in both HC-and DMD-derived cells and to test whether differences existed between the conditions. As expected, a strong signature marked the two cell stages with over 7000 genes differentially expressed between MB and MT samples ( Figure 2C and D). Overrepresentation analysis of the differentially expressed genes (DEGs) showed terms such as contractile fiber, myofibril, and sarcomere among the top GO terms for cellular component, indicating cellular programming towards formation of multinucleated muscle fibers (Table S2). As expected, we identified genes involved in muscle function such as those encoding dystrophin and myosin binding protein C. We also showed an increase in ribosome coverage, as previously described [16]. Pathway analysis with Reactome showed enrichment of processes involved with mRNA translation, which is compatible with the synthesis of myofibrillar proteins needed to form muscle fibers (Table S3). Analysis using WikiPathways was congruent with Reactome analysis with highlighted pathways such as EGF/EGFR signaling pathway, which induces growth, differentiation, and adhesion. About 700 genes were found to be differentially expressed between HC-and DMD-derived cells, with enrichment of GO terms such as skeletal muscle organ development and genes such as DMD (the affected gene, presumably undergoing nonsense mediated decay; Figure 2E and F). Pathway analysis using WikiPathways also highlighted pathways known to be affected in DMD such as oxidative stress, focal adhesion, NRF2 pathway, Wnt signaling, and TGF-b signaling (Table S4). A set of 385 common genes were found to be differentially expressed in both comparisons (Figure 2G). Remarkably, most transcriptional differences were identified between cell types but not between HC-and DMDderived cells, indicating a conserved sequence of transcriptional events driving the development of muscle fibers even in disease, as clearly shown by clustering ( Figure 2H). DEGs are listed in Tables S5 and S6.

RBPs identified by mass spectrometry
Analysis of proteins bound to RNA was performed on comparable amounts of proteins purified across samples and an average of 430 proteins per sample were identified ( Figure 3A; Table S7). For two MT samples (one derived from a healthy subject and one from a DMD subject), a large number of non-RBPs were co-purified, which were largely composed of muscle proteins not known to be RBPs. Out of the 1135 identified proteins, a set of 769 proteins showing five or more peptide counts across samples was further analyzed. Comparison of HC MB, HC MT, DMD MB, and DMD MT replicates showed highly significant correlations in the quantities (peptide counts) of identified proteins across 769 proteins ( Figure 3B, Spearman P ranging between 10 À142 and 10 À9 ). We then proceeded to analyze which proteins were identified in each condition by pooling together the proteins identified in the three replicates. The majority (49.0%) of RBPs were present in all cell types, with a considerable proportion (29.6%) Figure 1 Interactome capture shows enrichment in RBPs in muscle cells A. Schematic representation of the procedure including UV crosslinking of cells, poly-A mRNA capture, and either RNase or proteinase K digestion to analyze proteins and RNA, respectively. B. Representative image of silver staining of SDS-PAGE gel. C. Scatter plot showing the log 2 ratios between peptide counts +1 in a CL vs. noCL sample of two independent replicates both obtained from HC KM155 immortalized MB cells. Red dots represent proteins that show a positive FC in both CL vs. noCL samples and a significant change following differential expression analysis with DeSeq2 (adjusted P values by FDR < 0.1). D. Volcano plot showing the log 2 FC in CL vs. noCL MB samples (x-axis) and the Àlog 10 P value (y-axis). A clear shift towards the right is visible, representing the enrichment of proteins in CL samples compared to noCL samples. Red dots represent proteins that show a positive FC in both CL vs. noCL samples and a significant change following differential expression analysis with DeSeq2. E. Bar graph showing the results of the GO molecular function analysis. Bar height represents the number of proteins mapped to top 10 GO molecular function terms. The red line represents the Àlog of muscle proteins ''contaminating" the MT samples ( Figure 3C). Gene-set enrichment analysis (GSEA) showed that among the top 10 molecular functions per condition, nine were shared across the four groups ( Figure 3D), while biological processes were more conserved within cell types ( Figure 3E). Notably, all pathways point to RNA-related molecular functions. We further characterized the repertoire of identified proteins using Panther ( Figure 4). Molecular function analysis mapped the majority of the proteins to three GO categories: binding, catalytic activity, and structure molecule activity ( Figure S2). Sorting of proteins according to the protein classes showed that the majority of identified proteins are nucleic acid binding proteins, specifically RBPs. In this protein class we identified splicing factors (e.g., SUGP2), ribonucleoproteins (e.g., SNRPD1), and transcription and translation factors (e.g., FUS and PUM2). Proteins with catalytic activity were classified mostly as hydrolases (e.g., UPF1), transferases (e.g., TERT), and helicases (e.g., DHX9).   Proteins with structure molecule activity were mostly structural components of the ribosome (e.g., RPS14). An overview of the protein classification per sample type is shown in Figure S2.

Muscle differentiation affects the composition of RBPs
To understand whether RBPs could actively participate in the muscle formation process by regulating or affecting the observed shift in gene expression needed to convert MB into MT, we studied whether any RBP was differentially represented in the interactome dataset. We found that 32 RBPs were differentially represented in MT compared to MB, with 28 proteins enriched in MT and 4 proteins enriched in MB ( Figure 5A and B; Table 1). Analysis of intensity-based absolute-protein-quantification (iBAQ) data confirmed changes in the same direction as spectral data for all proteins and showed significant differences for 15 RBPs ( Figure S3). Most of the proteins showed RNA and nucleotide binding activities ( Figure 5C). To better understand the specificity of the identified signature, we searched the gene expression atlas for tissues in which these proteins are preferentially expressed. Roughly half showed prominent muscle expression, while the rest were expressed in multiple tissues at variable levels (Figure 5D). Examples of differentially represented RBPs are FXR1 and RBM28; while FXR1 binding to target transcript was increased upon muscle differentiation, RBM28 binding capacity was shown to be reduced after differentiation. To understand whether the observed changes were due to differences in gene expression or preferential target engaging, we compared interactome data and gene expression data for these two genes. The increased presence of FXR1 in differentiated MT was paralleled by increased FXR1 expression, while RBM28 expression was stably low in both MB and MT ( Figure 5E and F). To understand whether the increase of RBM28 binding to mRNA targets in MB was caused by an increased availability of target transcripts, we mapped all RBM28 and FXR1 binding sites in the genome (based on the known binding motifs obtained from the CISBP-RNA database http://cisbp-rna.ccbr.utoronto.ca/index.php) and quantified the expression levels of the binding motifs in the available RNA-seq samples. For FXR1, there were indeed more binding motifs present in genes expressed in MT; however, there was no difference in the total number of RBM28 binding motifs across the four conditions that would explain the reduction of RBM28 in the interactome of MT compared to MB ( Figure S4), suggesting preferential mRNA target engaging by RBM28 in MB compared to MT. No statistically significant differences were found in the comparison of RBPs between HCs and DMD patients. However, certain RBPs showed interesting trends such as AHNAK ( Figure S5), previously reported to play a role in myogenic differentiation [19].

Overlap with interactomes of other cell lines shows common and cell type-specific RBPs
To assess the specificity of the identified interactome, we investigated the overlap of proteins identified in MB and MT with five previously published interactomes in macrophages, HeLa cells, HEK293 cells, mouse HL-1 cardiomyocytes, and mouse embryonic stem cells (mESCs) [12,[14][15][16][17]. A core of 195 proteins was shared across all six cell types, representing a repertoire of conserved RBPs needed to exert basic functions across cells and tissues. A variable number of proteins was shared by groups of two to five cell types. We identified a set of 294 RBPs specific for muscle tissue of which only 19 were reported in a previous census of all RBPs [20] ( Figure 6A; Table S8). To understand whether differences exist between cell type-specific RBPs compared to common RBPs, we tested what molecular functions and biological processes are enriched for RBPs uniquely identified in specific cell types compared to RBPs found in five or more cell types (Table S9). Among the significant GO molecular function terms we found 33 shared between unique and common RBPs (e.g., RNA binding, nucleic acid binding, and heterocyclic compound binding), 74 specific for the common RBPs (e.g., single-stranded RNA binding, mRNA 3 0 -UTR binding, and RNA helicase activity), and 70 specific for RBPs found in unique cell types (e.g., nucleoside phosphate binding, catalytic activity, actin filament binding, and ribonuclease activity) ( Figure 6B). A large group of 556 biological processes was enriched for RBPs found in unique cell types and included terms such as RNA phosphodiester bond hydrolysis, regulation of ATPase activity, termination of DNA-templated transcription, and regulation of actin filament depolymerization ( Figure 6C). Interestingly, the interactome of skeletal muscle cells was more closely related to mouse HL-1 cardiomyocytes with 49 RBPs shared between the two muscle-derived cell lines (Table S10). Both RBP repertoires showed the highest number of unique RBPs (294 for MB and MT cells and 216 for HL-1 cardiomyocytes). Direct comparison of the unique set of proteins identified in our experiments and in cardiomyocytes revealed that skeletal muscle cells and cardiomyocytes showed a remarkably similar distribution of protein function even though different proteins were identified (Figure 7). To understand what protein domains may be enriched in the collected interactome data and in the set of 294 muscle RBPs, we tested the enrichment of Pfam protein domains and compared it to the comprehensive list published by Gerstberger and colleagues [20]. The analysis showed a significant enrichment of 68 Pfam domains in our dataset (Table S11) compared to the 189 Pfam domains mapped to the proteins listed by Gerstberger et al. (Table S12). Forty Pfam domains were Among the proteins exclusively shared by MB and MT cells and HL-1 cardiomyocytes, we identified STAT3 as an interesting hit, as we recently reported an increased concentration of this protein in circulation in DMD patients [21] (Figure S6). STAT3 was also recently shown to affect the pathophysiology in the mouse model for DMD [22].

Discussion
Muscle formation is a coordinated process requiring extensive changes in gene expression driving the fusion of multiple cells into multinucleated muscle fibers. While the gene expression signature resulting in the synthesis of myofibrillar proteins involved in the contractile sarcomeres and driving this physiologic process is known, there is no data on the repertoire of RBPs involved in this transition. It has been widely described how RBPs are needed for mRNA localization, turnover, and translation [4][5][6][7][8]. In this work we studied the repertoire of RBPs in proliferating and differentiated muscle cells in healthy and dystrophic conditions, where the gene expression signature is heavily affected and muscle regeneration and new muscle fiber formation are impaired [23][24][25]. Our work shows the enormous gene expression changes upon muscle differentiation in both healthy and dystrophic conditions. A total of 769 RBPs were identified across proliferating and differentiated muscle fibers obtained from both HC-and DMD-derived cells, showing large enrichment for RBPs.
A core of 195 proteins was shared with another five reported interactome datasets spanning from macrophages to ESCs. These RBPs are required to maintain primary metabolic processes such as nonsense codon surveillance (e.g., UPF1), splicing (e.g., NONO), and RNA metabolism (e.g., XRN1). A total of 741 proteins was found in unique cell types. Comparison of RBPs found in unique cell types and RBPs in five or more cell types showed that cell type-specific RBPs are involved in key functions such as RNA metabolic process as well as other less obvious functions such as actin filament depolymerization. These associations may clarify the role of previously unsuspected RBPs in processes such as mRNA trafficking. It has been shown that the cytoplasmic accumulation of human antigen R (HuR), an RBP known to play a role in stabilization and translation of AU-rich elements, is a process guided by actin and myosin [26].
We observed a large fraction of muscle-specific proteins composed of 294 new hits and a pool of 49 proteins shared with the previously reported HL-1 cardiomyocytes [17]. Interestingly, the repertoire of newly identified RBPs shows an almost identical protein class composition to the RBPs uniquely identified in cardiomyocytes, suggesting that different RBPs with similar function are responsible for muscle-specific physiological tasks in skeletal muscle fibers and cardiomyocytes. Of note, this group of unique muscle-specific proteins was enriched with proteins known to possess catalytic activity. While this is somewhat unexpected, a similar result was obtained in cardiomyocytes by Liao and colleagues [17], who hypothesized that these proteins could have a second role as RNA regulators, as previously shown for mitochondrial proteins [27]. While this may turn out to be the true interpretation, we think that capturing with oligo-dT beads could lead to purification of proteins involved in ATP synthesis, which is one of the major pathways in muscle tissue. Among the set of 49 shared proteins between skeletal and cardiac muscles, we identified STAT3, which we recently reported to be a candidate diagnostic biomarker for DMD as its serum concentration is elevated in DMD patients compared to HCs [21]. This finding shows the potential of interactome studies to not only uncover the background biology supporting cell-specific RNA mechanisms, but also to identify clinically relevant tissuespecific biomarkers.
We observed that the concentration of 32 RBPs was affected during myogenesis. Half of these proteins were muscle-specific according to the gene expression atlas ( Figure 5). We focused on two proteins, FXR1 and RBM28, which showed opposite directional changes during differentiation. FXR1 was found to be more prevalently bound to mRNAs in differentiated MTs compared to proliferating MBs. Fxr1 knock-out (KO) mice died at the neonatal stage; whereas mice with reduced FXR1 expression survived longer than KO mice, it showed reduced limb musculature compared to WT mice, suggesting an important role of FXR1 during muscle formation [28]. The levels of FXR1 in the interactome fraction were mirrored by the gene expression levels, supporting a correlation between the protein levels in the cell and the loading of the protein on its target RNA motif. We observed a completely opposite situation for RBM28, which was reduced during differentiation even though transcriptional levels were lower but constant during myogenesis. A motif search for both FXR1 and RBM28 showed how the transcriptome of muscle cells is enriched for FXR1 binding sites compared to RBM28, further supporting the role of FXR1 in myogenesis compared to RBM28. However, the motif expression levels did not clarify the increased motif engaging by RBM28 in MBs.
Our analysis did not allow us to identify RBPs differentially present between healthy and dystrophic cells. It is possible that further studies including more cases and several differentiation time points could identify RBPs that differ between the two situations. Nevertheless, it was encouraging to see a few proteins with interesting trends, such as an increased representation in dystrophic conditions of AHNAK, known to be highly expressed in muscle tissue and involved in muscle regeneration [29,30].
In conclusion, our study has increased the catalog of known RBPs and allowed us to identify muscle-specific RBPs that play a role in muscle formation. Our data show how we can advance our knowledge of muscle formation regulation as well as lead towards identifying candidate molecules that can be used to develop cell-specific assays with diagnostic potential.

Cell culture
Four distinct immortalized MB cell lines were used for the experiments. Control cell lines were obtained from two healthy individuals (KM155 and AB117). The cell lines 8036 and 6311 were obtained from DMD patients. Cell line 8036 was established from a DMD patient carrying an out-of-frame deletion of exons 48-50, while cell line 6311 carries a deletion of exons 45-50. Both DMD patient lines were the kind gift of Dr. Vincent Mouly (Institute of Myology, Paris, France). MB cells were kept in proliferating medium, while multinucleated MT cells were obtained by serum deprivation of MB cultured to a confluence of 80%, as previously described [31]. Cells were incubated in differentiation medium for 4-5 days. To obtain a sufficient number of cells, we coated 24 Petri dishes (Ø = 15 cm) with 0.5% gelatin and used six dishes per condition (four conditions: MB noCL, MT noCL, MB CL, and MT CL).

Crosslinking and interactome capture
To crosslink RNA and proteins, we followed the procedure as previously described [13]. Briefly, cells were washed twice with PBS and crosslinked with 0.15 J/cm 2 using the Stratalinker UV Crosslinker 2400 (Stratagene, La Jolla, CA). Cells from one dish were lysed using 2.8 ml of lysis buffer [20 mM Tris-HCl (pH 7.5), 1 mM EDTA, 500 mM LiCl, 0.5% LiDS, 5 mM DTT, and 0.025% NP-40] and they were harvested with a rubber policeman. The lysate was then transferred to the next petri dish to harvest cells from all dishes belonging to the same group. Lysates were collected in a 50 ml tube and homogenized by passing the whole lysate at least 10 times through a 29 G or 27 G needle. Samples were incubated on ice for 10 min and then stored at À80°C for up to one week.
For each condition 300 ml of oligo-dT magnetic beads (Catalog No. S1419S, New England Biolabs, Ipswich, MA) were used, which were equilibrated at room temperature in lysis buffer according to the manufacturer's instructions. The equilibrated beads were added to each sample and the samples were incubated for 1 h at 4°C with gentle rotation. Beads were then placed on a magnet for approximately 15 min to capture RNA-protein complexes until the supernatant was clear. The supernatant was collected in a new 15 ml tube for a second round of capturing and kept at 4°C. The beads were resuspended in 3 ml of lysis buffer, split into two 1.5 ml tubes, and incubated for 5 min on ice with consistent agitation (inverted manually~10 times per min). The tubes were placed on a magnet and the supernatant was removed. 1 ml of buffer 1 [20 mM Tris-HCl (pH 7.5), 1 mM EDTA, 500 mM LiCl, 0.1% LiDS, 5 mM DTT, and 0.025% NP-40] was added to each tube, the pellet was resuspended, and the samples were incubated for 5 min on ice with consistent agitation. The supernatant was removed after putting the samples on the magnet. This washing step was repeated once more with buffer 1, then twice with buffer 2 [20 mM Tris-HCl (pH 7.5), 1 mM EDTA, 500 mM LiCl, 5 mM DTT, and 0.025% NP-40] and twice with buffer 3 [20 mM Tris-HCl (pH 7.5), 1 mM EDTA, 200 mM LiCl, and 5 mM DTT]. After washing, captured samples were eluted in a total volume of 200 ll elution buffer [20 mM Tris-HCl (pH 7.5) and 1 mM EDTA] by pooling both samples together. The samples were incubated for 5 min at 55°C. Immediately, the tubes were put on the magnet and the supernatant containing the RNA-protein complexes was transferred into a new tube. For the second round of capture, the previously collected supernatant was treated with equilibrated oligo-dT beads and the whole procedure was repeated. Samples were then frozen at À80°C pending use.

RNA purification and analysis
For RNA analysis, 40 ml of sample were used. The captured RNA-protein complexes were treated with 2 mg proteinase K and 10 ml of 5Â proteinase K buffer [50 mM Tris-HCl (pH 7.5), 750 mM NaCl, 1% SDS, 50 mM EDTA, 2.5 mM DTT, and 25 mM CaCl 2 ]. Samples were incubated for 1 h at 50°C. RNA was then purified using the RNeasy Mini Kit (Catalog No. 74104, Qiagen, Hilden, Germany) according to manufacturer's instructions. Samples were eluted in 40 ml of RNase-free water, and RNA concentrations were measured by NanoDrop spectrophotometer (Catalog No. ND 1000, ThermoFisher Scientific, Waltham, MA). Sample preparation for RNA-seq was performed using the standard poly-A mRNA library preparation of the TruSeq RNA Library Prep Kit v2 (Illumina, San Diego, CA). RNA-seq analysis was performed for MB and MT samples of both KM155 and 8036 cells (total n = 4).

Protein purification
For protein analysis, samples were purified using Amicon ultra centrifugal filters (3 kDa; catalog No. UFC500396, Merck Millipore, Burlington, MA) according to the manufacturer's instructions. The samples were defrosted and loaded onto the filters, and the tubes were centrifuged at 18,000 g for 20 min at 4°C. 400 ll of buffer 4 [10 mM Tris-HCl (pH 7.5) and 50 mM NaCl] was added to sample in the filter and the centrifugation step was repeated. After removing the supernatant, the filter was placed bottom up in a new collection tube and the remaining solution was spun down for 1 min at 18,000 g at 4°C. The recovered sample volumes of different conditions were determined and adjusted to the highest volume by adding buffer 4. RNA digestion was performed by adding an RNase mix consisting of 0.11 ll RNase buffer [100 mM Tris-HCl (pH 7.5), 1.5 mM NaCl, 0.5% NP-40, and 5 mM DTT], 0.04 U RNase A (Catalog No. 19101, Qiagen) and 0.04 U RNase T1 (Catalog No. EN0541, ThermoFisher Scientific) per ll of sample and incubating for 90 min at 37°C followed by 15 min at 50°C with gentle agitation.

SDS-PAGE and silver staining
The obtained protein samples were loaded onto a Criterion XT 4%-12% BisTris precast gel (Catalog No. 3450123, BioRad, Hercules, CA) in a Biorad chamber filled with 1Â XT MOPS Running Buffer (Catalog No. 161-0788, BioRad). A total of 12 ml of sample with 4 ml 4Â NuPage sample buffer (Catalog No. NP0007, Life Technologies, Carlsbad, CA) were first denatured for 5 min at 95°C and then loaded onto the gel together with 5 ml of protein marker (PageRuler; catalog No. 26619, ThermoFisher Scientific). The gel was run for 45 min at 200 V followed by an incubation in fixing solution (40% ethanol, 10% acetic acid, and 50% H 2 O) for 1 h. The gel was washed with water for at least 30 min before sensitization with 0.02% sodium thiosulfate solution for 1 min. The gel was washed three times for 20 s with water and was incubated for 20 min at 4°C in 0.1% silver nitrate solution. Residual silver nitrate was removed by washing the gel three times for 20 s with water. Before imaging the gel was developed with a 3% sodium carbonate / 0.05% formaldehyde solution until the staining was sufficient. The gel was washed again for 20 s with water and left in 5% acetic acid for 5 min to terminate the staining.

Protein analysis
Samples were incubated at 90°C with LDS sample buffer (Life Technologies). Proteins were separated on a 4%-12% PAGE (NuPAGE Bis-Tris Precast Gel, Life Technologies) after a run for 1 cm. Staining was performed with silver (SilverQuest Silver Stain, Life Technologies). Proteomic analysis was performed as previously described [32]. The lane was cut in three parts and subjected to reduction by DTT followed by alkylation by iodoacetamide and in-gel digestion with trypsin with the Proteineer DP digestion robot (Bruker, Billerica, MA). Tryptic peptides were lyophilized, dissolved in 95/3/0.1 (v/v/v) water/acetonitril/formic acid, and analyzed by on-line C18 nanoHPLC MS/MS with an Easy nLC 1000 gradient HPLC system (Thermo, Bremen, Germany) and LUMOS mass spectrometer (Thermo). Peak lists were obtained from raw data and submitted to the Homo sapiens database at Uniprot, using Mascot v. 2.2 for protein identification. Protein with at least two peptides with a threshold of !95% were sorted and compared using Scaffold software version 4.7.5 (www.proteomesoftware.com).

Data analysis
The mass spectrometry proteomics data have been deposited to the ProteomeXchange Consortium via the PRIDE [33] partner repository. Differential expression analysis of RNAseq data was performed on 23,126 genes that showed more than ten counts across the four samples. We tested for differences in cell type (MB vs. MT) and diagnosis (HC vs. DMD). A Benjamini-Hochberg false discovery rate (FDR) correction for multiple testing was applied. We considered P values below 0.05 significant. The GeneTrail2 tool [34] was used to perform overrepresentation analysis for RNA-seq data using GO, Reactome, and WikiPathways as background information. A two-sided test was performed and the Benjamini-Yekutieli multiple testing correction was applied. Bioinformatic analysis to map the Pfam domains was performed using STRING [35]. Protein IDs (Table S7) were compared to the protein IDs provided by Gerstberger and colleagues [20]. Comparison of CL and noCL samples was performed using four samples (two noCL and two CL samples obtained from HC KM155 MB cells). Proteins with common IDs across the replicate experiments were tested. Differential representation was tested using the DeSeq2 package on 262 proteins. Additional data analysis including data from eight samples (HC MB noCL, HC MT noCL, DMD MB noCL, DMD MT noCL, HC MB CL, HC MT CL, DMD MB CL, and DMD MT CL) was performed on a total of 655 proteins. The comparison between CL and noCL samples was in this case performed for cell type (MB and MT) and group (HC and DMD). To compare the interactome across cell types and groups we analyzed samples in triplicate with a total of 12 samples. Samples were either MB or MT from HCs (n = 3 each, of which two replicates of KM155 cells and one replicate of AB117 cells), or, MB or MT of DMD patients (n = 3 each, two replicates of 8036 cells and one replicate of 6311 cells). To analyze the data we started with the more complex and comprehensive model including both cell type (MB or MT) and diagnosis (HC or DMD) as covariates and proceeded to simplify the model by testing only the difference between MB and MT as contribution of diagnosis was not significant. Benjamini-Hochberg FDR correction for multiple testing was applied and P values below 0.05 were considered significant. Analysis of protein data was performed on 12 CL samples. We filtered out proteins with less than five spectral counts across all samples, resulting in the exclusion of 366 proteins. We tested whether differences existed in the remaining 769 proteins. Confirmation of quantitative differences was performed on the identified RBPs using MaxQuant iBAQ quantitation data analyzed using the kruskal.test function in R. Analysis of data was performed in R with DeSeq2 [36], ggplot2 [37], and UpSetR [38] packages.

Data availability
The mass spectrometry proteomics data have been deposited to the ProteomeXchange Consortium via the PRIDE partner repository (ProteomeXchange: PXD015471), and are publicly