Widespread Translational Control of Fibrosis in the Human Heart by RNA-Binding Proteins

Supplemental Digital Content is available in the text.

C ardiac remodeling, heart failure, and arrhythmia syndromes are frequently associated with fibrosis, which is a common late-stage pathology in many human diseases. 1 Cardiac fibrosis is seen in numerous cardiac conditions, including atrial fibrillation, 2 hypertrophic cardiomyopathy, 3 dilated cardiomyopathy (DCM), 4 and heart failure with preserved ejection fraction. 5 Fibrosis of the heart is driven primarily by the activation of resident fibroblasts. 6,7 A better understanding of the molecular mechanisms underlying fibroblast activation is of great importance for the development of novel antifibrotic therapies. 8 While various cues initiate the cellular conversion of fibroblasts to myofibroblasts, TGFβ1 (transforming growth factor β1) is considered a master regulator. 9 However, antifibrotic therapeutic approaches based on TGFβ1 inhibition have side effects due to the pleiotropic roles of this cytokine, especially in cancer and inflammation. 10,11 Thus, unraveling the fibroblast-specific footprint of TGFβ1 signaling is an important step toward the identification of novel downstream drivers of cardiac disease.
While RNA expression changes via TGFβ1-induced SMAD (small mothers against decapentaplegic) signaling have been studied previously, 12,13 the independent impact of TGFβ1 on RNA translation remains unknown.
To address this gap in knowledge, we profiled genomewide RNA transcription and translation levels 14 in human primary cardiac fibroblasts at several time points after TGFβ1 stimulation. A tailored computational analysis 15 identified posttranscriptional regulatory networks underlying fibroblasts activation. To corroborate our findings in an independent and disease-relevant context, we performed ribosome profiling of cardiac samples from patients with DCM and found much of the network was active in disease. Finally, silencing of key regulatory hubs from these networks was enough to limit the progression of TGFβ1treated fibroblasts toward the myofibroblast state. This integrative approach provides a detailed perspective on posttranscriptional regulatory hubs in human heart disease.

Data and Software Availability
The raw data are available on the gene expression omnibus (GEO submission: GSE131112, GSE123018, GSE131111). Transcription and translation levels of individual genes are provided as a web resource: http://ribo.ddnetbio.com.

Human Primary Fibroblast Culture
Human primary atrial fibroblasts were prepared from atrial biopsies of patients undergoing coronary artery bypass grafting in keeping with local guidelines (Singhealth Centralized Institutional Review Board 2013/103/C and 2018/2543) and cultured as described previously (Table I in the online-only Data Supplement). 13 All the experiments were carried out at low cell passage (<4). In all experiments, cells were starved in serum-free DMEM for 16 hours prior to TGFβ1 stimulation. Stimulated fibroblasts were compared to unstimulated fibroblasts that had been grown for the same duration under the same conditions (serum-free DMEM), but without the stimuli.

Immunostaining, Operetta High-Content Imaging, and Confocal Microscopy
Operetta phenotyping assay was performed as described previously. 13 Briefly, atrial fibroblasts were seeded in 96-well black CellCarrier plates (PerkinElmer) at a density of 10 4 cells/well. Following TGFβ1 stimulation, cells were fixed in 4% paraformaldehyde and permeabilized with 0.1% Triton X-100, and nonspecific sites were blocked with 0.5% BSA and 0.1% Tween-20 in PBS. Cells were incubated overnight (4°C) with primary antibodies (Table II in the online-only Data Supplement), followed by incubation with the appropriate secondary antibodies and counterstained with DAPI (4',6-diamidino-2-phenylindole, dihydrochloride). Each condition was imaged from duplicate wells (7 fields/well) using Operetta high-content imaging system 1483 (PerkinElmer). The quantification of α-smooth muscle actin (ACTA2; ACTA2 +ve ) cells and fluorescence intensity per area of collagen I and periostin were performed using Harmony

Clinical Perspective
What Is New?
• We show that a third of all genes involved in the transforming growth factor β1-driven fibrotic response of cardiac fibroblasts are regulated at the translational level, independent of RNA transcript abundance. • This is the first study to investigate the overall importance of translational regulatory networks in myocardial fibrosis. • We identify RNA binding proteins such as Quaking or Pumilio RNA binding family member 2 as regulatory hubs for gene translation in the fibrotic hearts of dilated cardiomyopathy patients. • Knockdown of RNA-binding proteins, Pumilio RNA binding family member 2 and Quaking, inhibits the activation of fibroblasts with transforming growth factor β1 in vitro.

What Are the Clinical Implications?
• There are currently no specific therapies for myocardial fibrosis which predict both electrical and pump failure of the heart and this remains poorly understood. • We identify previously unappreciated genes under translational control as novel candidates for disease biology and as potential therapeutic targets. • Critical fibrosis factors impact cellular phenotypes at the protein level which cannot be appreciated using bulk or single cell RNA-sequencing approaches. • RNA binding proteins are central to the fibrosis response and represent unexplored gene expression regulators and potential diagnostic or therapeutic targets. . Immunostaining for confocal microscopy was conducted as previously described. 16 Atrial fibroblasts were cultured in 8-well chamber slides and fixed with 4% formaldehyde. After permeabilization with 0.1% Triton X-100, cells were washed and blocked in 1% BSA in PBS for 30 minutes. Cells were incubated overnight at 4°C with the primary antibodies (Table II in the online-only Data  Supplement). After washing and incubation with secondary antibodies for 30 minutes at room temperature, slides were mounted with ProLong gold antifade. Images were acquired using a confocal laser scanning microscope (LSM710, Zeiss).

ELISA
The amount of MMP2 and TIMP-1 in equal volume of cell culture media was quantified using Total MMP2 (R&D Systems) and TIMP-1 Quantikine ELISA kit (R&D Systems) as per the manufacturer's instructions.

Colorimetric Assay
Quantification of total secreted collagen in the cell culture supernatant was performed using Sirius red collagen detection kit (9062, Chondrex) according to the manufacturer's protocol.

Western Blotting
Atrial fibroblasts were washed with ice-cold PBS and solubilized by gentle rocking in radioimmunoprecipitation assay buffer containing protease and phosphatase inhibitors (Roche). Protein concentrations were determined by Bradford assay (Bio-Rad). After centrifugation, equal amounts of protein lysates were separated by SDS-PAGE, transferred to polyvinylidene difluoride membrane, and subjected to immunoblot analysis (Table II in the online-only Data Supplement).

Statistical Analysis
Statistical analyses of high content imaging were performed using GraphPad Prism software (v6.07). Outliers (ROUT 2%, Prism Software) were removed before analysis. Dunnett's test was used to calculate multiple testing corrected P values for comparison of several time points to baseline.

Ribosome Profiling and RNA Sequencing
For primary cells, primary human atrial fibroblasts outgrown from cardiac tissue biopsies of 4 patients undergoing coronary artery bypass grafting were expanded to reach 80% confluency in several 10-cm dishes. Cells were stimulated with 5 ng/mL TGFβ1 for 45 minutes, 2 hours, 6 hours, and 24 hours. Per condition, 3 10-cm dishes were employed in order to obtain enough material for ribosome profiling (2 dishes) and RNA sequencing (1 dish). For heart tissue, as part of a larger consortium effort to characterize the cardiac translatome, 17 we generated ribosome profiling data of left ventricular tissue samples collected during left ventricular device implantation or cardiac transplantation from patients with end-stage DCM (n=30). We specifically selected these 30 DCM patients as they were obtained from the same site of tissue collection (Cardiovascular Research Centre Biobank at Royal Brompton and Harefield National Health Service Trust) and had previously been subjected to RNA sequencing analysis (30 out of the published 97 DCM patient samples) in order to reduce technical variability and facilitate accurate patient stratification based on the degree of cardiac fibrosis. 18 Ribosome profiling was performed as previously described. 19 Briefly, snap-frozen cell pellets or 50 to 100 mg of tissue, previously powdered under liquid nitrogen, was lysed in 1 mL cold lysis buffer (formulation as in TruSeq Ribosome Profile, Illumina) supplemented with 0.1 mg/mL cycloheximide to stabilize ribosomal subunits and prevent postlysis translocation. Homogenized and cleared lysates were then footprinted with Truseq Nuclease (Illumina) according to the manufacturer's instructions. Ribosomes were purified using Illustra Sephacryl S400 columns (GE Healthcare), and the protected RNA fragments were extracted with a standard phen ol:chloroform:isoamylalcohol technique. Following ribosomal RNA removal (Mammalian RiboZero Magnetic Gold, Illumina), sequencing libraries were prepared out of the footprinted RNA. Ribosome profiling libraries were pooled to perform multiplex sequencing on Illumina Hiseq machines.
We used RNA sequencing data for left-ventricle tissue from DCM patients that we had previously published (108 control and 97 DCM patients). 18 To prepare polyA+ RNA sequencing libraries from primary cardiac fibroblasts, total RNA was extracted with Trizol from 1 10-cm dish per condition. Following cleanup with RNeasy kit (Qiagen), ~500 ng of each sample was further processed with the Truseq Stranded mRNA kit (Illumina). Barcoded RNA sequencing libraries were pooled and sequenced on the Illumina HiSeq platform.
DCM tissue studies complied with UK Human Tissue Act guidelines and were carried out with approval from the Royal Brompton and Harefield local ethical review committee and the National Research Ethics Service Committee South Central, Hampshire B (reference 09/H0504/104).

Data Processing for RNA Sequencing and Ribosome Profiling
Raw sequencing data were demultiplexed with bcl2fastq V2.19.0.316, and the adaptors were trimmed using Trimmomatic V0.36, 20 retaining reads longer than 20 nucleotides postclipping. RNA sequencing reads were further clipped with FASTX Toolkit V0.0.14 to 29 nucleotides, to allow comparison directly with ribosome profiling reads. Reads were aligned using bowtie 21 to known mitochondrial RNA, ribosomal RNA, and transfer RNA sequences (RNACentral, 22 release 5.0); aligned reads were filtered out to retain only ribosome protected fragments (RPFs). Alignment to the human genome (hg38) was carried out using STAR. 23 Gene expression was quantified on the coding sequence region using uniquely mapped reads (Ensembl database release GRCh38 v86 combined with additional transcripts from RefSeq GRCh38, downloaded January 2018) with feature counts. 24 Genes with mean transcripts per million mapped reads (TPM) <1 in either RNA sequencing or ribosome profiling across all conditions were removed prior to downstream analysis. Ribotaper was used to obtain the in-frame reads around the start and stop codon. 25 These peptidyl-sites (P-sites) were then visualized across samples and genes. Heatmap for the ribosome drop-off was generated using the pheatmap, 1.0.8, R package. Principal component analysis was carried out using the prcomp function in R (Table III in

Detection of Differential Translational-Efficiency Genes Using DESeq2
The calculation of change in translational efficiency (ΔTE) for human fibroblast data was done using an interaction term while accommodating for patient effect in the statistical model within DESeq2 (~ Patient + Time + Sequencing + Time:Sequencing). 26 This allows for the identification of significant differences between time points that are discordant between sequencing methodologies; ie, for changes in ribosome occupancy that are not explained by changes in RNA abundance. The ΔTE fold change derived with this approach is comparable to traditional TE, but also accounts for variance and level of expression. In addition, the statistical model reveals if ΔTE is statistically significant.
Since the RNA and RPF fold changes can be obtained by the same process using DESeq2, the fold changes are also directly comparable with ΔTE. In combination, these 3-fold changes can help predict the regulation status of the gene at the transcriptional and translational levels. Results were combined from the Wald test for each time point and likelihood ratio test across all time points (Table III in the online-only Data Supplement).

Classification of Differentially Transcribed Genes and Differential Translational-Efficiency Genes Into Regulatory Classes
A gene's regulation class was first classified using an adjusted P value threshold <0.05 for ΔTE, ΔRPF (change in RPFs), and ΔRNA (change in mRNA) combined with the associated direction of change (as shown in Table IV in the online-only Data Supplement). These first level classifications are then combined in order to fully characterize the regulation of each gene. For instance, Forwarded genes are transcriptionally driven genes that have significant ΔRPF and ΔRNA, but do not have a significant ΔTE. Conversely, exclusive genes are regulated only translationally and hence have a significant ΔRPF and ΔTE, but no change on the mRNA levels (ie, ΔRNA not significant). In order to further classify the genes with respect to time, we also performed hierarchical clustering on the log fold changes within each category. For instance, forwarded genes were clustered using the ΔRPF and ΔRNA values, while exclusive genes were clustered using ΔRPF and ΔTE values. Hierarchical clustering was carried out using euclidean distances and the ward.D method in the hclust function and cutreeDynamic with default settings.

Overrepresentation Analysis
R packages topGO 27 and KEGGrest 28 were used to carry out (GO: BP, MF, and KEGG pathways) overrepresentation tests for each gene cluster. Genes that are classified as either differentially transcribed genes (DTG) or differential translation-efficiency genes (DTEG) were used as background. Peak files from eCLIP experiments on ENCODE were downloaded and filtered for 8-fold enrichment and P value <10 -5 . Peak files were also downloaded from POSTAR 29 and used with default filters. For each RNAbinding protein (RBP), the actual number of targets in each gene class or cluster was determined using these peak files. Expected distribution for RBP targets in each group was calculated by randomly selecting gene sets of the same size and quantifying the targets found within the group. This was repeated 100 000 times to obtain an empirical P value. These P values were further corrected for multiple testing using the Benjamini-Hochberg method. The groups with corrected P value <0.05 for an RBP were considered overrepresented for its targets. Z-scores were calculated to evaluate the effect size of this overrepresentation.

DCM Disease Patient Network Analysis
Spearman ranked correlation was calculated between the RBP's log 10 (TPM RPF ) and target's log 10 (TE) across the patient population (n=30, using all ribosome profiling and RNA sequencing matched samples) with the cor.test R function. A permutation (n=10 000) test was carried out to determine the expected number of correlated pairs (RBP:target) that would be found in a random set. The empirical P value was calculated, and Benjamini-Hochberg correction was applied for multiple testing. RBPs correlating with more pairs than random at an adjusted P value <5% were selected as network hubs. The network was visualized using Cytoscape. 30 Spearman ranked correlation was calculated between the RBP's log 10 (TPM mRNA ) and fibrosis marker's log 10 (TPM mRNA ) with the cor.test R function. To maximize the number of comparisons for the correlation between RBPs and disease severity, we used all 97 patient RNA sequencing data. Patient clustering was carried out using hclust R package in default settings based on marker gene expression levels across all 97 patients. Treecut R package was used to obtain 4 levels of severity in fibrosis. Student t test was used to determine significance of the difference between RBP expression in patients with low and high fibrosis severity.

RNA-Binding Protein Short Interfering RNA Knockdown
For knockdown experiments, cells were transfected using Lipofectamine RNAiMax (Life Technologies), following the manufacturer's instructions for standard forward transfection (6-well plate format) or reverse transfection (96-well plate format). In both cases, short interfering RNAs (Dharmacon) were used at a final concentration of 25 nM. Reverse transfection was the method of choice for high-throughput transfections to be followed by immunostainings on 96-well plates. Per well of a 96-well plate (96-well black CellCarrier, PerkinElmer), 1 × 10 4 human cardiac fibroblasts were transfected with 25 nM On-Targetplus short interfering RNAs (Dharmacon) in a medium consisting of serum-free Opti-MEM and DMEM supplemented with FBS (10%), combined in a 1:9 ratio. Twenty-four hours later, media were changed, and cells were cultured in serum free DMEM overnight, before being subjected to TGFβ1 stimulation (5 ng/mL) for 6 hours.

Translational Profiling During the Activation of Human Cardiac Fibroblasts
During the fibrotic response, resident fibroblasts become profibrotic myofibroblasts that express ACTA2 and secrete extracellular matrix proteins such as collagen I and periostin. 31 To better understand this transition in the human heart, we isolated primary cardiac fibroblasts from atrial biopsies of 4 individuals undergoing coronary artery bypass grafting ( ORIGINAL RESEARCH ARTICLE Supplement). TGFβ1 stimulation (5 ng/mL) resulted in a significant increase of ACTA2 +ve cells and an upregulation of extracellular matrix-related proteins, indicating activation of fibroblasts within 24 hours (Figure 1a through 1h). This cellular transformation was accompanied by rapid phosphorylation of SMAD as well as extracellular signal-regulated kinase (ERK), which is a key factor in noncanonical signaling pathways and known to regulate posttranscriptional processes (Figure 1i). 13 To capture a time-resolved snapshot of the molecular changes that underlie the transition of fibroblasts into myofibroblasts, we performed RNA sequencing and Riboseq at baseline and at 45 minutes, 2 hours, 6 hours, and 24 hours after TGFβ1 stimulation. Ribosome profiling entails deep sequencing of ribosomal footprints, which are short RNA fragments protected from nuclease treatment by RPFs. RPFs therefore quantify both mRNA abundance and ribosome occupancy of protein-coding genes and as such are a superior proxy for protein levels compared to RNA-seq. 19 On average we generated ~51 million (RNA sequencing) and ~12 million (ribosome profiling) uniquely mapped reads per sample. Ribosome profiling reads mapped predominantly to the coding sequence and had an average read length of 29 bp (Figure Ia through Id in the online-only Data Supplement), both of which are characteristic of high-quality RPF data. 14 We then inferred the exact position of the P-site, the site in the ribosome where transfer RNAs recognize their complementary codon, based on RPFs (see Methods). On average, 89% of all P-sites mapped to the coding frame in known genes, indicating that captured ribosomes translate the known reading frame of transcripts (Figure 2a and 2b). Plotting of P-site density around the 3′ location of expressed coding sequence revealed that ribosomes recognized the stop codon and disengaged from RNA transcripts (Figure 2c). High triplet periodicity and a strong dissociation signal reveal the stepwise movement of the ribosome along the coding regions of the transcripts and indicate the positions of actively translating ribosomes being captured at single nucleotide resolution in our ribosome profiling data.

Dynamic Translational Regulation During the Fibrotic Response
DTGs can be detected on a genome-wide scale with RNA sequencing alone. Conversely, genes that are translationally regulated between conditions will display a significant change in TE, ie, the genewise ratio between ribosome occupancy and transcript abundance, requiring both RNA sequencing and ribosome profiling for detection. We recently developed an analytical approach (which we refer to as ΔTE) that integrates RNA sequencing and ribosome profiling data to reveal DTEGs. 15 The output of this approach is a ΔTE value (and associated adjusted P value) for each gene describing the log fold-change of TE at each time point. This analysis allowed reliable detection of DTEGs in our data by accounting for the patientrelated batch effects ( Figure II in the online-only Data Supplement). Using the ΔTE approach, we identified 1691 DTEGs during the fibrotic response. For instance, ribosome occupancy of both FTL (Ferritin Light Chain, ΔTE=3.24; P adj. =2×10 -2 ) and FTH1 (Ferritin Heavy Chain 1, ΔTE=3.12; P adj. =1x10 -2 ) increased significantly upon TGFβ1 stimulation, despite underlying transcript levels remaining the same. Translating ribosomes located on ITGA3 (Integrin Subunit Alpha 3, ΔTE=-1.9; P adj. =2×10 - We detected gradually more DTGs at later time points, up to a total of 4216. Several genes were detected as both DTG and DTEG. This occurs when TGFβ1 affects both RNA levels and TE of the same gene. In order to describe the relationship of this overlap in transcriptional and translational regulation, we categorize each of the DTGs and DTEGs into 1 of 8 regulatory groups (Figure 3c, File II in the online-only Data Supplement). For more than 29% of DTGs, differences in transcription were not forwarded to the translational level but were translationally buffered or intensified. Of these, translational buffering was most prominent, ie, changes in transcript expression detected with RNA sequencing were less pronounced in the ribosome profiling data. This effect can be due to either a less efficient translation of genes whose RNA levels are increasing (568 genes, buffered down) or vice versa, more efficient translation of genes whose RNA levels are decreasing (479 genes, buffered up). Out of these 1047 buffered genes, 419 (231 down, 188 up) had transcriptional regulation that was completely counteracted by translational regulation, resulting in a similar density of translating ribosomes despite underlying TGFβ1-driven transcriptional changes. For example, RNA sequencing suggests a downregulation of the protein kinase PRKG1 upon TGFβ1 stimulation. However, this effect is not forwarded to the translational level, and thus protein levels do not decrease (Figure 3d  ORIGINAL RESEARCH ARTICLE Data Supplement). Interestingly, intensified genes were overrepresented for functions such as "SMAD-protein signal transduction" (P adj. =6.5×10 -3 ) and "Regulation of ERK1 and ERK2 cascade" (P adj. =1.4×10 -2 ) (File III in the online-only Data Supplement). Overall, these results demonstrate that more than one-third of all gene expression changes during fibroblast activation involve translational regulation and that these changes can affect protein levels. Transcriptional and translational regulation downstream of TGFβ1 appears to be closely interlinked and tightly regulated over time. To further stratify these effects over time, we performed unsupervised clustering of the temporal profiles of Forwarded, Buffered, Exclusive, and Intensified genes. This revealed 64 distinct regulatory patterns during the fibrotic response ( Figure IV in the online-only Data Supplement). The clustering highlights the fact that there are a substantial number of transiently regulated genes during the fibrotic response (Figure 3f). Transient differences in expression would not be apparent when quiescent fibroblasts are compared to myofibroblasts at 24 hours but may be crucial for the cellular transition and therefore important in disease. Transient clusters were predominantly enriched for processes involved in the regulation of gene expression (regulation of RNA metabolic process, 2.5×10 -7 ; regulation of transcription, DNA-templated, 1.7×10 -5 ), further substantiating their role in fibroblast transformation (File IV in the online-only Data Supplement). We plot RNA sequencing, ribosome profiling, and ΔTE results for any gene of interest at http://ribo.ddnetbio.com and share the values in File I in the online-only Data Supplement.

RNA-Binding Proteins as Regulators of TGFβ1 Signaling
Having identified different regulatory groups and temporal profiles, we sought to identify potential regulators. It is known that RBPs bind to target transcripts to affect protein production, 19,32 a number of which been previously linked to heart disease. 33,34 To identify key RNA-binding proteins during the fibrotic response, we integrated our expression data with more than 200 available transcriptome-wide RBP-RNA binding datasets. We first identified RBPs that were differentially expressed at 45 minutes, 2 hours, 6 hours, and 24 hours after TGFβ1 stimulation in human primary cardiac fibroblasts, revealing 53 differentially expressed RBPs. Next, to construct reliable RBP-target networks, we utilized experimental evidence for RBP-RNA binding derived from eCLIP (n=117), PAR-CLIP (n=58), 35 HITS-CLIP (n=23), and iCLIP (n=22) data provided by EN-CODE 36 and POSTAR. 29 Following this, for each of the differentially expressed RBPs, a permutation test was used to detect whether there was a significant overrepresentation of their targets in each of the 8 regulatory groups defined above, which was the case for 47 RBPs. We found that the targets of these RBPs were predominantly enriched in DTEGs but not in DTGs, which suggests that RBPs shape the fibrotic response

ORIGINAL RESEARCH ARTICLE
mainly through translational regulation and rarely influence transcript levels (Figure 4a). Most targets of individual RBPs were overrepresented in clusters with similar translational regulation profiles (Figure 4b). For instance, QKI targets genes that have an increase in translation, while DDX24 targets genes that have a decrease in translation. These distinct enrichments in unidirectional ΔTE patterns suggest RBPs tend to act either as translational repressors or activators during the fibrotic response.

RBPs and Their Target Regulatory Networks in the Fibrotic DCM Heart
To ascertain whether the same translational program was active in disease, we utilized RNA sequencing data ORIGINAL RESEARCH ARTICLE from dilated cardiomyopathy and control hearts, which we generated in a previous study (controls=108, DCM patients=97). 18 Analysis of this data showed upregulation of cardiac fibrosis markers in the hearts of end-stage DCM patients, and expression of these markers was shown to be restricted to cardiac fibroblasts in singlecell RNA sequencing data from the mouse heart (Figure 5a, Figure V in the online-only Data Supplement). 6,37 Furthermore, 45 of the 47 RBPs identified previously were expressed in DCM hearts (TPM>5), and of these, To assess whether this dysregulation of RBPs at the RNA level also influenced translation in the fibrotic human heart, we performed ribosome profiling of left ventricular tissue collected from a subset of these DCM patients. If the RBPs are truly regulators of ribosome occupancy in fibrosis, then RBP expression levels should correlate with the translational efficiency of their target transcripts. In total, we were able to generate highdepth and high-quality ribosome profiling data of 30 individuals ( Figure VI in the online-only Data Supplement; see Methods for details of data generation). We observed a total of 3771 RBP:target pairs correlating significantly in our human heart datasets. Fourteen translational regulators (targeting 926 transcripts) remained significantly correlated with their targets based on permutation analyses (Figures 5c, 6a, and 6b; File V in the online-only Data Supplement), identifying them as regulatory hubs controlling ribosome occupancy both during cardiac fibroblast activation and in the hearts of DCM patients. This substantiates further the influence of RBPs on ribosome occupancy in an independent dataset and provides evidence for posttranscriptional regulatory networks in the human heart.
Modules of this posttranscriptional regulatory network are enriched for distinct biological functions related to actin remodeling and other processes important in fibrosis (File VI in the online-only Data Supplement). One module in the heart is controlled by Pumilio RNA binding family member 2 (PUM2), which is known to drive glial scar formation, a fibrotic and TGFβ1dependent process in the brain. 38 Others are regulated by QKI, important in heart development, 39 or PCBP2, which is known to inhibit cardiac hypertrophy driven by the fibrogenic stimulus angiotensin 2. 40 Interestingly, we detected a large overlap between targets of regulators, suggesting that the concerted action of several RBPs determines the translational efficiency of bound transcripts. Both QKI and PUM2 appear to cooperate, with more than 21% of their targets (103 genes) overlapping between the RBPs, both of which appear to be acting as translational repressors. . Taken together, the dysregulation of these RBPs in the activation of atrial fibroblasts and end-stage DCM patients suggests that these genes regulate translation in the context of cardiac fibrosis. Unsupervised clustering revealed different degrees of fibrogenic marker gene expression (low, moderate, high) in the DCM hearts (Figure 6c, Figure  VIIb in the online-only Data Supplement). Nine regulators of the network were significantly elevated in patients with high severity (vs low severity) cardiac fibrosis molecular signature (2-tailed t test significance P value <0.05, Figure 6d).
In order to ascertain whether some of these RBPs play a driving role in the transition toward profibrotic myofibroblasts in response to TGFβ1, we selected QKI and PUM2 as representative RBPs for further in vitro validation. We targeted these RBPs with short interfering RNAs alone and in combination with TGFβ1 treatment followed by high-content imaging after immunostaining for ACTA2. In the presence of TGFβ1, the knockdown of either QKI or PUM2 alone, as well as in combination, was found to significantly reduce ACTA2 levels in 3 different patients (Figure 7a and 7b, Figure VIII in the online-only Data Supplement). The overall scale of the effect was like that of knocking down TGFβR1, and the combined knock-down appears less variable across patients when compared to knock-downs of QKI or PUM2 alone. To investigate this premise using an orthogonal method, we measured the concentration of MMP2, which is secreted by activated fibroblasts, in cell culture supernatants using ELISA. This revealed a

ORIGINAL RESEARCH ARTICLE
similar reduction in MMP2 secretion (Figure 7c). Taken together, these results demonstrate that a reduction in either QKI or PUM2 or both inhibits fibroblast activation downstream of TGFβ1 stimulation.

DISCUSSION
These results shed new light on the extent of posttranscriptional regulation in response to extracellular signaling in human disease. Our data suggest that cytokines such as TGFβ1, which are heavily investigated as potential therapeutic targets, directly impact cellular phenotypes via the regulation of translation. While the transcriptional effects of TGFβ1 are well studied, 13 we show for the first time a genome-wide snapshot of the posttranscriptional effect of the pathway: over onethird of all changing genes are regulated at the translational level. Buffering or intensifying RNA differences and exclusive translational regulation robustly modify protein abundance during the cellular transformation to myofibroblasts. We reveal novel molecular pathways and substantially broaden the scope in the search for urgently needed novel therapeutic targets, which have been missed in previous RNA-based target discovery screens. Discrepancies between transcript and translation levels also have implications for the use of RNAbased biomarkers, which are likely to be less informative than protein-based methods for certain genes. We found that targets of specific RBPs are significantly enriched in translationally but not transcriptionally regulated genes, revealing their posttranscriptional regulatory footprint during cardiac fibroblast activation. Integration of more than 150 protein-RNA binding data sets with our ΔTE analysis revealed an unprecedented view of the RBP-driven landscape of translational regulation. Some of these identified candidates (CELF2, 41 PUM2, 38 KHSRP, 42 MBNL2, 43 and LARP4 44 ) are known to repress/activate the translation of single target genes. Our results go beyond this, exposing the extended translational regulatory network of each RBP in the fibroblast to myofibroblast transition. Furthermore, both Quaking (QKI) and MBNL2 have been previously associated with regulating alternative splicing, 45,46 but our findings suggest that both also regulate the translational efficiency of hundreds of target transcripts during the fibrotic response. This emphasizes the multifunctional ORIGINAL RESEARCH ARTICLE role RBPs can play in regulating many posttranscriptional processes such as mRNA splicing, localization, translation, and turnover in parallel. 47 To further investigate the disease relevance for these networks, we integrated cardiac translatomes of end-stage dilated cardiomyopathy patients. Despite differences between atrial and ventricular fibroblasts, 48 similarities of the TGFβ1 response across cell types revealed common profibrotic signatures in human cardiac fibrosis ( Figure IX in the online-only Data Supplement). We documented highly significant correlation of 14 of these RBPs with the translational efficiency of more than 900 transcripts in the DCM heart, and as such we consider these key regulatory hubs of the diseased cardiac translatome. Regulatory footprints of the networks of these RBP often overlap, suggesting that RBPs act in concert to control the final protein levels of shared targets. This is especially evident with the repressors PUM2 and QKI with more than 21% of their targets overlapping.
Knockdown of PUM2 and QKI blunts the TGFβ1 response in vitro. Translational repression by PUM2 in astrocytes is known to cause astrogliosis and the formation of glial scarring, which is a TGFβ1-dependent process, 38 and we reveal here the target gene network for PUM2-regulated translation in cardiac scarring. Taken together, this indicates an important role for both QKI and PUM2 in fibrotic cardiac disease.
There are more than 1500 RBPs encoded in the human genome, 49 but their role in the translation of target mRNAs remains largely unexplored. Just as transcription factors have been targeted therapeutically to modulate transcription, RBPs represent novel diagnostic, prognostic, and therapeutic targets. Translational control may be particularly pertinent in fibroblasts that produce huge amounts of extracellular matrix protein with limited changes in the expression of the correspondingly extracellular matrix transcript. 50