P38 inhibition reverses TGFβ1 and TNFα-induced contraction in a model of proliferative vitreoretinopathy

Proliferative vitreoretinopathy (PVR) is a metaplasia in the vitreous of the eye manifested by the transformation of retinal pigment epithelial (RPE) cells and the development of contracting epiretinal membranes (ERM), which lead to retinal detachment and vision loss. While TGFβ1 and TNFα have been associated with PVR, here we show that these cytokines act synergistically to induce an aggressive contraction phenotype on adult human (ah)RPE. Connected RPE detach upon contraction and form motile membranes that recruit more cells. TGFβ1 and TNFα (TNT)-induced contracting membranes uniquely express muscle and extracellular rearrangement genes. Whole transcriptome RNA sequencing of patient-dissected PVR membranes showed activation of the p38-MAPK signaling pathway. Inhibition of p38 during TNT treatment blocks ahRPE transformation and membrane contraction. Furthermore, TNT-induced membrane contractility can be reversed by p38 inhibition after induction. Therefore, targeting the p38-MAPK pathway may have therapeutic benefits for patients with PVR even after the onset of contracting ERMs.

E piretinal membranes (ERMs) are metaplasia near the retina that occur due to activation of quiescent retina cell populations including Muller's glia, macrophages, and RPE cells 1 . ERMs can be benign and asymptomatic or can be aggressive and exhibit myocontractile behavior, leading to retinal detachments and vision loss 2 . The most aggressive form of ERM formation, Proliferative Vitreoretinopathy (PVR), is thought to derive its severity from activation of RPE, since membranes are composed of up to 95% RPE 3 . Similarly, when RPE are not present, a less aggressive ERM phenotype is observed 4 . We previously discovered a putative multipotent stem cell population in RPE, which may contribute to this observed cellular plasticity 5 .
Development of PVR occurs following a retinal tear and complicates rhegmatogenous retinal detachment, resulting in epiretinal and subretinal membrane formation that can re-detach the retina and lead to blindness. The risk factors for PVR, the primary cause of retinal re-detachment after retinal surgery, include the size of the retinal tear, presence of vitreous hemorrhage, intraocular inflammation and trauma [6][7][8][9] . During a retinal tear dislodged RPE come in contact with the vitreous which stimulates RPE migration on the surface of the retina. Inflammatory mediators and blood released from the retinal tear triggers the RPE to become metaplastic and produce collagen, resulting in ERM formation and contraction. Clinically, the retina develops wrinkling and folds which overwhelm the forces keeping the retina attached 10,11 . PVR is also characterized by intraretinal changes caused by retinal gliosis where Müller and other glial cells play an important role 10 .
Although many cytokines, including transforming growth factor beta (TGFβ) 12 and tumor necrosis factor alpha (TNFα) 13 , have been implicated in retinal tear-induced PVR, what behavioral changes are induced in RPE is not well understood. Both TGFβ and TNFα have been implicated in fibrotic diseases [14][15][16][17][18] . Furthermore, nucleotide polymorphisms in TGFβ1 and TNF loci are known to confer an elevated risk of PVR, suggesting a genetic component [19][20][21] . TNFα alone has been found to stimulate the production of EMT-associated fibrotic aggregates in ARPE-19, a RPE immortalized cell line 22 and can affect RPE barrier function via p38 signaling pathway 23 . However, using the same immortalized ARPE-19, another group found both TGFβ1 and TNFα co-stimulation was necessary to produce similar aggregates 24 . The use of ARPE-19 to faithfully predict RPE behavior has come under scrutiny due to reports of not responding like or nor possessing elements of native physiology, leading to reproducibility issues [25][26][27][28][29] .
We sought to characterize the impact of TGFβ and TNFα on a functional ahRPE monolayer with native electrophysiology 30 and evaluate whether this treatment in vitro evoked changes observed in PVR in vivo. Only upon TGFβ and TNFα co-treatment (TNT), ahRPE form aggregates and their formation is contraction dependent with similar protein composition to patient-dissected PVR membranes, a distinct phenotype not evoked by either TGFβ1 or TNFα treatment alone. Additionally, we identified enrichment of the p38-MAPK pathway in both TNT-treated ahRPE and patient-dissected PVR samples. Lastly, we found inhibition p38-MAPK not only prevents, but can also reverse contractility, potentially the most damaging acquired RPE behavior in PVR.
We tested ahRPE lines from 11 different, non-genetically matched donors, and found that 100% of the lines exhibited the formation of aggregates exclusively upon TNT treatment (Fig. 1e). However, in some experiments we only found 42 aggregates, while in others we counted 1200. To determine whether the variability was due to genetic differences between lines or experimental variability, we compared interline variability in membrane number formed upon TNT treatment between all donor ahRPE lines to intraline variability, i.e., variability of the number of membranes formed within individual lines in different experiments. We graphed the interline variability for the line for which we conducted the highest number of experiments (n = 6), Line A. A list of the CV for each line is presented in Table S1. The CV across all experiments was 86.7% (n = 26), whereas intraline CV for line A was 70.67% (n = 6), suggesting non-genetic influences account for at least 81.5% of the observed membrane formation variability between experiments ( Fig. 1e and Supplementary Table 1). By qPCR we also observed reduced expression of the known RPE identity markers, MITF 0.069 ± 0.01, P ≤ 0.001, (n = 3), OTX2 0.527 ± 0.07, P ≤ 0.01, (n = 3), and RPE65 0.013 ± 0.01, P ≤ 0.001, (mean ± S.E.M; n = 3) compared to control (Fig. 1f). Taken together, the data suggest that TNT inhibits RPE identity and drives aggregates formation in RPE independently of genetic background.
TGFβ and TNFα induced aggregate formation are contraction dependent. Due to the perceived contractile nature of the membranes, we evaluated changes in extracellular matrix reorganization and motility-associated changing genes. The muscle associated actin ACTG2 and the extracellular matrix protein TENASCIN C were both significantly up regulated in TNTtreated ahRPE compared to control, TGFβ1, or TNFα treated RPE (ACTG2 19.7 ± 4.4, P ≤ 0.01, (n = 10) and TENASCIN C 6540.6 ± 4063.1, P ≤ 0.01, (mean ± S.E.M; n = 10) (Fig. 2a) and protein levels (Fig. 2b). We further hypothesized contraction underlies TNT induced aggregation and tested inhibitors of mechanisms involved in contraction. Upon TNT treatment in the presence of (S)-nitro-Blebbistatin, an inhibitor of myosin II ATPases, ahRPE were unable to generate contractile masses (Figs. 2c, d). Additionally, cells treated with TNT in combination with W-7 hydrochloride, a calmodulin antagonist which inhibits myosin light chain kinase, were similarly unable to generate contractile masses (Figs. 2c, d). These results suggest ahRPE aggregation depends on contraction.
To better observe changes between untreated ahRPE, and ahRPE treated with TGFβ1 or TNFα or TNT, we conducted a 72 h-long time-lapse imaging, taking one frame per half hour. In FBS-basal media, ahRPE proliferate and begin to re-establish a monolayer ( Fig. 1 TGFβ1 and TNFα work synergistically to induce transformation of RPE. RPE cultured in the presence of 10 ng/ml TGFβ1 and/or TNFα for 5 days then assayed morphologically and transcriptionally. a RPE dissection schematic. Scale bar = 50 μM. b Phase images of RPE 5 days after treatment with TGFβ1 and/or TNFα. Scale bar = 50 μM. c Time-lapse image frames highlighting stages of RPE transformation when stimulated with TGFβ1 and TNFα for 5 days. Scale bar = 300 μM. d Quantification of contractile membranes with varying concentrations of FBS after 5 days treatment with TGFβ1 and/or TNFα. e Coefficient of variation (CV) for total RPE lines (n = 11) and one RPE line. f RT-qPCR of RPE markers. Statistical significance was calculated using Student's t-test, *P < 0.05; **P < 0.01; and ***P < 0.001 Movie 3) appear more mesenchymal. In TNT supplemented media, ahRPE undergo dramatic morphological and behavioral changes (Supplementary Movie 4). RPE transformation can be divided into four distinct stages (Fig. 1c). First, ahRPE lose high phase edges representative of epithelial cells and begin to appear more mesenchymal (stage 1). Then, cells begin to condense into discrete centers (stage 2). A ridge forms at the edge of these groups suggesting multiple layers we refer to as a membrane (stage 3). Finally, the cluster synchronously contracts to form a contracting membrane (stage 4). These membranes develop more quickly at the edge of the well and contract in discrete, almost rhythmic spurts, eventually ripping cells off the plate (Supplementary Movie 5).
Differences between TGFβ and TNF cytokine family members. Both TGFβ2 and TNFβ have been implicated in PVR and may also induce a similar response in RPE 12,20 . To determine if cytokine family members behaved similarly, we analyzed ahRPE upon treatment with TGFβ1 + TNFα (TNT), TGFβ2 + TNFα, TGFβ1 + TNFβ and TGFβ2 + TNFβ. Upon 5 days of treatment, all conditions produce robust contracting membranes (Supplementary Fig. 1A) and express PVR associated EMT proteins ( Supplementary Fig. 1B). While there was no statistical difference in quantity of contractile membranes produced between treatments, TGFβ1 + TNFα tended to produce the most membranes and express PVR associated EMT genes compared to other combinations ( Supplementary Fig. 1C, D). Due to these minimal differences and the implication of TGFβ1 and TNFα in the literature, we continued to explore the effect of TGFβ1 and TNFα.
p38-MAPK pathway participates in transducing the signal mediated by TNT. While expression of selected proteins and genes showed similarity between TNT-treated ahRPE and patient-dissected PVR samples, we sought to identify gene networks enriched in PVR and determine whether they are also present in TNT treated RPE. To do so, we conducted the first whole transcriptome RNA sequencing of patient-dissected PVR membranes, to our knowledge, and compared them to control and TNT-treated RPE. In TNT vs. PVR samples, there were 4388 significantly changing genes, while 2088 significantly changing genes overlapped in both comparisons ( Fig. 4a), with PVR and TNT samples more similar in gene expression than RPE and TNT samples. Considering the number of significantly changing genes between the RPE vs. PVR comparison is the smallest, we suggest PVR samples may still contain a population of RPE with gene expression relatively similar to cultured RPE.
To investigate this further, significantly changing genes were then analyzed by singular value decomposition (SVD) 34,35 using the biosvd package 36 . In fact, when we graph the significantly changing genes individually in eigenspace, we find individual genes from PVR samples in close proximity to RPE (Supplementary Fig. 2). Another interesting point to note is the majority of genes from patient-dissected PVR samples are closer to genes from TNT-treated samples. The first dimension, SVD1 represents 82.5% variance of the data and the second dimension, SVD2 captures 10.7% of the variance, highlighting that SVD1 and SVD2 captures~93% of the total gene variance (Fig. 4b).
When looking for downstream signaling convergence between TGFβ1 and TNFα by evaluating the KEGG pathway database, we identified the p38-MAPK pathway as the predominant signaling node. We therefore examined whether this pathway is enriched in our significantly changing gene comparisons. We analyzed significantly enriched signaling pathways present in both patient-dissected PVR samples and TNT-treated ahRPE. Utilizing the same gene list from the dimensional reduction described above, significantly changing genes were identified. The Bayes factor of log 46.10 highlights the confidence in the probability of the alternative hypothesis that there is a role for p38-MAPK pathway being true. Similar to the significantly changing genes above, there are 67 significantly changing involved in the p38 pathway in RPE vs. TNT and 43 significantly changing p38 implicated genes in RPE vs. PVR. Between those two comparisons, there is overlap of 7 significantly changing p38 genes (Fig. 4c). Based on the significantly changing genes identified by the intersection of control ahRPE vs. patient-dissected PVR membranes comparison and the p38 network, a sub network was constructed (Fig. 4d). The sub network highlights how significantly changing genes involved in the p38 pathway are connected. Additionally, communities can be discerned, suggesting connectedness of related genes. For instance, genes in green tend to be muscle-specific, while genes in pink are directly implicated in p38-MAPK pathway. RNA-seq analysis suggests that the p38 pathway is driving the TNT-mediated transformation of ahRPE and was enriched in patient-dissected PVR membranes. Additionally, gene ontology analysis in which significantly changing genes were upregulated in both TNT-treated ahRPE and patientdissected PVR samples compared to control ahRPE show that both TGFβ1 and TNFα are upstream regulators of p38 and that molecular functions include cell movement and cell-to-cell signaling coupled with an inflammatory immune response (Fig. 4e). We also identified additional genes involved in muscle contraction by having at least a 2-fold increase in expression in PVR and TNT samples compared to control ahRPE by KEGG pathway analysis using the online DAVID software (Supplementary Fig. 3).
To confirm the role of p38, we evaluated candidate pathways downstream of TGFβ1 and/or TNFα by determining which signals upon TGFβ1 or TNFα alone or TNT treatment translocate to the nucleus, as they will more likely play a role in gene expression changes. p38 was the only pathway in which we observed a significant increase in nuclear localization preferentially upon TNT treatment (Supplementary Fig. 4 and Supplementary Table 2).
Inhibition of p38 prevents TNT-induced contractile mass formation. Considering that TNT induced nuclear localization of p38-MAPK in ahRPE and that this pathway was also upregulated in patient-dissected PVR samples, we hypothesized that p38 inhibition would prevent membrane contraction. Upon TNT treatment in the presence of SB 202190, an inhibitor of the p38-MAPK pathway, ahRPE were unable to generate contractile membranes (Fig. 5a). Additionally, cells treated with TNT and SB 202190 showed decreased expression of PVR markers and EMTassociated proteins including αSMA, SNAIL, COL1A1, COL1A2, and LAMININ (Fig. 5b). Next, we hypothesized that p38 is phosphorylated and activates the p38-MAPK signaling pathway upon TNT treatment, so we evaluated p38 phosphorylation over time. Upon TNT treatment for .5, 1, 1.5, 2, 48, 72, 96, and 120 h, total protein was extracted for Western blot analysis. After 30 min, we observed a rapid and transient increase in p38 phosphorylation, which decreased by 60 min and maintained a steady state level of phosphorylation above control for up to five days (Fig. 5c). To determine whether p38 phosphorylation is essential for TNTinduced changes in ahRPE, we tested the efficacy of SB 202190, since it does not interfere with p38 phosphorylation, but instead prevents transfer of the phosphate group to downstream targets 37,38 . Heat shock protein 27 (HSP27) is a known downstream effector of p38 and has been shown to play a role in contractility 39,40 . Therefore, we tested whether inhibition of p38 by SB 202190 prevents the phosphorylation and activation of HSP27. We confirmed that upon treatment with TNT and SB 202190, HSP27 fails to be phosphorylated by p38, while p38 maintains its own phosphorylation (Fig. 5d). We evaluated mRNA expression of PVR associated genes and found treatment with TNT and SB 202190 resulted in decreased gene expression of COL1A1 0.034 ± 0.013, P ≤ 0.001, (n = 3), COL1A2 0.039 ± 0.002, P ≤ 0.001, (n = 3), LAMB2 0.190 ± 0.035, P ≤ 0.001, (n = 3), and SNAI1 0.016 ± 0.006, P ≤ 0.001, (mean ± S.E.M; n = 3) compared to TNT alone (Fig. 5e). Furthermore, BIRB 796, a p38α/β inhibitor that is structurally unrelated to SB compounds, similarly prevents TNT-induced gene expression changes. AhRPE treated with TNT and BIRB 796 showed decreased gene expression of COL1A1 60.7 ± 30.2, P ≤ 0.01, (n = 8), COL1A2 9.3 ± 4.9, P ≤ 0.01, (n = 8), LAMB2 1.4 ± 0.4, P ≤ 0.01, (n = 3), and SNAI1 1.5 ± 0.2, P ≤ 0.01, (mean ± S.E. M; n = 8) compared to TNT-treated ahRPE (Fig. 5f).
To further verify that TNT-induced mass formation is working specifically via p38 and there are no additional off-target The data structure was examined by singular value decomposition. SVD1 and SVD2 explain~93% of the variance in the data. c A p38 gene network was constructed using STRINGdb and each of the pairwise comparisons between samples was intersected with this network and an enrichment was calculated using the BayesFactor package in R (Higher BF means more evidence for enrichment).
Overlaps between comparisons were identified. d Using the genes identified by the intersection of control ahRPE vs. patient-dissected PVR membranes comparison and the p38 network, a sub network was created. The betweenness centrality of each gene and the betweenness of each edge of this sub network was calculated and determines the size and thickness of the nodes and lines, respectively. Communities in the network were identified by the GLay algorithm (identified by color), and the shape of each node indicates the group from the Venn diagram displayed in (c). e The top categories shared between patient-dissected PVR membranes and TNT-treated ahRPE by Ingenuity pathway analysis. Sig signaling, Int interaction, Maint maintanence, Org organization, Metab metabolism h RT-qPCR of extracellular matrix and EMT genes in cells treated in the presence of TGFβ1 and TNFα co-supplemented basal media after transfection without p38 siRNA, and TGFβ1, TNFα co-supplemented basal media after ahRPE were transfected with p38 siRNA. Statistical significance was calculated using Wilcxon signed-rank test, *P < 0.05; **P < 0.01. i Quantification of wound healing scratch assay. j Phase images of wound healing scratch assay on ahRPE treated with basal media alone, TGFβ1 and TNFα supplemented media or TGFβ1 and TNFα and SB 202190 supplemented media for 72 h. Scale bar = 300 μm. White arrowheads indicate areas of preliminary mass formation. Statistical significance was calculated using Student's t-test, *P < 0.05; **P < 0.01; and ***P < 0.001 pathways affected by p38 inhibitors 41,42 , p38 was knocked down in ahRPE using a p38 small interfering (si) RNA. p38 was knocked down approximately in half after p38 siRNA transfection for 48 h P38 0.59 ± 0.1, P = 0.0038, (mean ± S.E.M; n = 3) (Fig. 5g). Upon treatment with TNT and transfection with a p38 siRNA, ahRPE were unable to generate contractile membranes and showed decreased expression of COL1A1 10.4 ± 3.0, P = 0.0325, (n = 3), COL1A2 4.9 ± 1.9, P = 0.0062, (n = 5), LAMB2 1.7 ± 0.3, P = 0.0226, (n = 3), and SNAI1 6.1 ± 1.6, P = 0.0141, (mean ± S.E.M; n = 4) compared to cells treated with TNT and vehicle siRNA, further demonstrating that membrane formation and contractility are acting through a p38-MAPK signaling cascade (Fig. 5h).
In addition to contractility, another behavior hypothesized to be important in the contribution of TNT-mediated ahRPE transformation in PVR-like masses is their ability to migrate. Therefore, we tested whether p38 also plays a role in TNTinduced migration of ahRPE using a scratch-wound healing assay. TNT-treated ahRPE showed an increased rate of scratch closure 38.6 h ± 2.1, P ≤ 0.0001, (n = 43, n = 3 biological replicates) compared to control ahRPE 50.7 ± 2.2, P ≤ 0.0002, (n = 49, n = 3 biological replicates) and SB 202190 and TNT-treated ahRPE 73.1 ± 2.0, P ≤ 0.0001, (mean ± S.E.M; n = 41, n = 3 biological replicates) determined by the amount of time it took for cells to close the scratch wound (Figs. 5i, j).
Modeling the role of RPE in PVR. Based on the collection of results gathered, we propose a model of the initiation of ahRPE transformation by the synergistic action of TGFβ1 and TNFα (Fig. 7). TGFβ1 and TNFα bind to their respective receptors stimulating the phosphorylation of p38, which then indirectly phosphorylates HSP27. These responses lead to the translocation of phosphorylated p38 into the nucleus and transcriptional inhibition of RPE genes MITF, OTX2, and RPE65, and activation of an EMT program, production of membrane associated proteins and activation of contraction machinery.

Discussion
Contractility has long been thought of as the mechanical force generator that brings PVR membranes to cause retinal detachment 43 . Identification of αSMA 44 and Tenascin C 45 in ERMs supported this hypothesis. We provided here additional evidence to suggest a potential contribution from RPE is in fact their ability to transdifferentiate into myocontractile-like cells. We previously discovered a putative multipotent stem cell population in RPE, which may contribute to this RPE observed cellular plasticity (Salero et al., 2012). Here we show the synergistic effect of TGFβ1 and TNFα through p38 is sufficient to stimulate RPE transformation into contractile membranes. Either TGFβ1 or TNFα alone can induce RPE to lose epithelial characteristics including the loss of E-Cadherin and the increased expression of N-Cadherin, yet only when combined do the downstream signaling of these two factors synergize to foster additional changes in ahRPE, including the expression of the mesenchymal gene SNAI1 and of muscleassociated genes, including ACTG2 and Tenascin C, as well as CDON and MEF2C, transcription factors known for muscle differentiation. Moreover, we have conducted the first whole transcriptome analysis, to our knowledge, of patient-dissected PVR samples and identified p38 MAPK central to the PVR signature network.
EMT is a multifaceted process that has been implicated in a variety of biological transformations including embryogenesis 46 , organ development 8 , fibrosis 47,48 , regeneration 49 , and cancer metastasis 50 . In this study, we find that in ahRPE, TGFβ1 and TNFα interact synergistically to drive an EMT program that mimics a step in the pathology of the blinding disease PVR. While RPE EMT is only one step in the process, the participation of RPE in ERM growth is one of the more damaging contributors, indicated by the fact that when RPE are not found in ERMs, PVR is easier to treat 3,4,51 .
Synergism between TGFβ1 and TNFα to initiate EMT has been shown in both the lung 52,53 and intestine 54 epithelia, but this is the first demonstration in primary ahRPE isolated from donor eyes, to our knowledge. Similarly, it has been reported that TGFβ1 and TNFα leads to the phosphorylation of p38 in colon organoids 55 suggesting a similar signaling cascade. We demonstrate that co-stimulation by TGFβ1 and TNFα not only leads to an EMT, but also induces ahRPE cells to contract.
The role of p38 and the fibrotic response of RPE have been studied in a mouse model, in which RPE multilayering was observed in response to the removal of the /lens, vitreous, and retina 56   c Quantification of membrane formations between cultures of ahRPE in contraction reversal conditions described in 6a. d Western blot comparison of ACTG2 and COL1A2 between cultures of ahRPE in contraction reversal conditions. e RT-qPCR of extracellular matrix and EMT genes from ahRPE in contraction reversal conditions. (f) Anti-SNAIL, αSMA, COL1A2, and P38 immunostaining after ahRPE were placed in contraction reversal conditions. Scale bar = 75 μm. Student's t-test, *P < 0.05; **P < 0.01; and ***P < 0.001 reduction in RPE layers was observed, yet the transformation from a monolayer to multiple layers did still occur. While this model may not faithfully recapitulate PVR, the fact that DN p38 could reduce RPE response suggests p38 inhibitors may be effective in vivo at minimizing the activation of RPE observed in PVR.
In response to stress p38 translocates to the nucleus 57 . We observed nuclear localization of p38 upon TNT stimulation and during contraction. The phosphorylation of p38 occurs within 30 min of TNT stimulation, then phospho-p38 levels decline within 30 min, but stay elevated compared to control, establishing a new steady-state level. Only upon treatment with the p38 inhibitor SB 202190 p38 nuclear localization decrease, and that is concomitant with contraction reversal. Therefore, phosphorylation may transmit the switch and the nuclear localization, but maintenance of the signal may be continued by the maintenance of p38 in the nucleus, which may require TGFβ1. Based on nuclear localization experiments, the candidate pathways transmitting this additional signal may be AKT and SMAD signaling. We find both AKT and SMAD3 nuclear localization increases when both TGFβ1 and TNFα are provided compared to when either cytokine is administered alone.
Inhibiting p38 as a therapy for PVR is a promising avenue to explore. A lot of effort was placed in p38 inhibitors as treatment for many inflammatory associated diseases, including COPD 58 , rheumatoid arthritis 59 , and cancer 60 . However, many clinical trials have not progressed past Phase II due to unacceptable safety profiles. Multiple side effects have been reported, including elevated liver enzymes, skin rash, cardiotoxicity, infections, and CNS and GI toxicity 61 . However, in these cases the p38 inhibitors were provided orally and therefore the systemic, repeat exposure of p38 inhibitors likely contributed to the observed side effects. If, on the other hand, a p38 inhibitor is administered locally instead of systemically, then perhaps the severity of the side effects will be sufficiently mitigated. Although current PVR treatment does not involve the use of systemic p38 inhibition, PVR is a disease whose treatment provides the opportunity for a local, acute administration during surgery. Severe PVR is a common problem following complex retinal detachments and trauma and the only treatment is ERM removal during vitrectomy 62 . Unfortunately, PVR often returns, sometimes more aggressively. If during the initial PVR membrane removal a p38 inhibitor was administered locally, perhaps the incidence of recurrence will be reduced. Future work will explore the potential clinical approaches in utilizing local p38 inhibitors for treatment of PVR to prevent vision loss.
Clinical trials using cell replacement therapies to treat blinding diseases such as Age-Related Macular Degeneration are already underway in the United States 63 , Israel 64 , and China 65 . However, it is important to note that upon transplantation, a small incision in the retina is made to deliver the cell replacement therapy. As a result, RPE may be exposed to external cytokines, which may trigger abnormal proliferation and metaplasia as seen in ERM formation and PVR. Recently, a phase I clinical trial using hESCderived RPE on a synthetic basement membrane resulted in one of two patients developing retinal detachment and PVR membrane formation 66 . Another clinical trial using patient specific hiPSC-derived RPE also reported pre-retinal membrane formation 67 . In each case, ERMs have never been determined to derive from the transplanted RPE. Regardless of whether ERMs are produced from the host or donor, managing p38 signaling during RPE replacement therapies may aid in improved outcomes.
To conclude, we examined the synergistic effect of TGFβ1 and TNFα on adult human RPE cultures which exhibit native physiology and compared these effects to PVR membranes dissected from patients by whole transcriptome RNA-seq. We have found RPE treated with the combination of TGFβ1 and TNFα share a gene signature with PVR membranes and central to this signature is the p38 network. Moreover, through time-lapse imaging we have identified RPE contraction upon TNT treatment. By using small molecule inhibitors and siRNA technology we were able to not only inhibit, but also reverse membrane formation and contraction. Future steps will be to test whether p38 inhibitors can prevent retinal detachment in an animal model of PVR.  Table 3) at room temperature for 1 h. Transwell inserts were then mounted on glass slides with Aqua-Poly/Mount (Polysciences, Inc.) and imaged were taken using a Leica microscopy. DAPI fluorescence was used to demarcate the nucleus. In order to measure nuclear fluorescence intensity, immunofluorescent images and their corresponding DAPI images were loaded into Image J. The DAPI image was used to make a mask onto the fluorescence image. The fluorescence signal intensity within each nucleus was measured and averaged for the pixels within each nucleus of each image using a scale of 0-255. Mean fluorescence measurements were then compared between control ahRPE, TGFβ1, TNFα, and TNT-treated ahRPE conditions using a paired t-test.

Methods
RNA isolation and qPCR. AhRPE cells were incubated in RNAprotect Cell Reagent (Qiagen) to attenuate endogenous RNAse activity and scraped off the plate into individual 1.5 ml tubes. Cells were centrifuged at 10,000 rpms for 10 min and the pellet was re-suspended in buffer RLT plus (Qiagen Using the ×10 objective, a phase-contrast image, which fully captured the scratch wound and surrounding cellular environment was taken per well per time point. Data was analyzed using the IncuCyte software using a processing definition with the following parameters: Segmentation Adjustment-1.3; Hole fill (µm 2 )-2500; Area Min (µm 2 )-300. Wound width was calculated as an average of each vertical line of resolution of a scratch wound mask generated at a given time point.
siRNA knockdown of p38. AhRPE cells were seeded into a 24-well plate at a density of 100,000 cells/well in media containing DMEM/F12, 3% Heat Inactivated Fetal Bovine Serum, 1X GlutaMAX, 1X MEM Non-Essential Amino Acids Solution, 1X Penicillin-Streptomycin, 1X Sodium Pyruvate and incubated overnight at 37°C. Two days after initial plating, cells were transfected with 20pmol SignalSi-lence® p38α MAPK siRNA II (Cell Signaling Technology, Cat #:6277 S) using Lipofectamine TM RNAiMAX Transfection Reagent (Thermo Fisher Scientific, Cat #: 13778-075). Following 48 h, media containing Lipofectamine Reagent and p38α MAPK siRNA II was removed. Cells were fed with basal media or media containing TGFβ1 and TNFα (TNT). Feeding was repeated every other day until visible mass formation was observed. Cells were subsequently analyzed by RT-qPCR.
RNA sequencing. Total RNA was isolated with the Qiagen RNAeasy kit according to manufacturer instructions, and libraries were prepared using standard Illumina protocols and sequencing was performed with the Illumina HiSeq 2500 platform. After base-calling with Illumina Realtime Analysis Software, the STAR aligner was used to map read to the UCSC hg19 sparse genome (NCBI GRh37/hg19) and annotated with gencode version 19 and the counts were calculated. Counts were then imported into R and analyzed with the EdgeR and DESeq2 packages for differential gene expression. The bioSVD package was then used to provide a global picture of the significantly changing genes and generate polar plots. The p38 network was identified by taking the first and second-degree neighbors of p38 in the StringDB (Table S5). The odds of selecting a p38 network gene in our gene list was calculated by sampling over 1000 iterations. The Bayes Factor was calculated using the proportion BF function from the BayesFactor package. (Code fully available in supplement). To identify genes with altered expression between the groups, we utilized the edgeR and DEseq2 packages coupled with a 2-fold change threshold to uncover significantly changing genes. We conducted unbiased, pairwise comparisons to identify points of variation. For example, when comparing the expression of genes with a 2-fold difference between RPE and TNT-treated cells there were 7062 genes identified in either direction. SVD is a linear transformation of the expression data from the genes × arrays (samples) space to the reduced "eigengenes" × "eigenarrays" space. SVD is a useful mathematical framework for processing genome-wide expression data in which operations may be assigned a biological meaning. Therefore, SVD1 may capture cell heterogeneity as we predict this would be higher in TNT and in PVR samples while SVD2 may be capturing the EMT (Fig. 4b).
The data are diagonalized such that each eigengene is expressed only in the corresponding eigenarray with the corresponding "eigenexpression" indicating their relative significance. Using this method we can evaluate the individual genes from the different samples and map them in the "eigenarrays" space. We also took the sum of the significantly changing genes from each sample and mapped them in SVD, which is also known as Karhunen-Loeve expansion in pattern recognition 70 and as principal-component analysis in statistics 71 . Doing so captures the variation attributed to the genes and conditions per sample as a whole and allows for an interpretation of the similarities in gene expression between the groups.
Reporting summary. Further information on experimental design is available in the Nature Research Reporting Summary linked to this article.

Data availability
All raw data is available upon reasonable request. Sequencing data has been deposited on https://www.ncbi.nlm.nih.gov/geo/subs/, accession number GSE126633.