Single-Cell Transcriptomics Reveals Novel Role of Microglia in Fibrovascular Membrane of Proliferative Diabetic Retinopathy

Proliferative diabetic retinopathy (PDR) (1), one of the main diabetic complications, remains as one of the primary causes of severe vision loss diseases worldwide. The epiretinal fibrovascular membrane (FVM), as the hallmark of PDR, grows between the retina and posterior hyaloids of vitreous and can cause vitreous hemorrhage and tractional retinal detachment. However, the pathogenesis of the FVM in PDR is now poorly understood (2). Typically, neovascularization is initiated by hyperglycemic toxicity, retinal ischemia, and damage of the blood-retinal barrier, and is accompanied by an accumulation of extracellular matrix, resulting in the formation of FVMs that extend into the vitreous (3,4). Though the cellular composition of FVM has been studied histologically (5,6), the definition of cells was mostly based on traditional immunohistology method. On the other hand, most studies regarding gene expression of FVM have relied on bulk measurements of tissue or cultured cells (7–9). However, the cellular heterogeneity of FVMs cannot be fully explored by the above conventional approaches that have difficulties in revealing cell type–specific changes of gene expression.

Single-cell RNA sequencing (scRNA-seq) is a recently developed approach for molecular separation of different cell populations based on expression profiles (10). This emerging technology has thus opened the door to answering longstanding biological questions on heterogeneity of complex samples, identification of distinct cell subsets, and cell differentiation processes and regulatory signaling networks. In the current study, we used the scRNA-seq on surgically harvested FVMs of PDR patients and successfully identified eight biological cell clusters. Among them, microglia (MG), as the main cell population, was shown to present both profibrotic and fibrogenic properties, contributing to FVM formation. Moreover, cytokines that upregulated in PDR vitreous showed ligand-receptor interactions with FVM-resident MG. Taken together, we systematically provided a cellular classification of FVMs and potential targets for developing novel therapeutic interventions for PDR.

The study was approved by the Institutional Review Boards at Ethic Committee of First Affiliated Hospital of Nanjing Medical University and registered at https://clinicaltrials.gov/ (ID NCT04682054 for single-cell RNA sequencing; ID NCT03506750 for measurement of vitreous cytokines). Informed consent with regard to FVM and vitreous samples was obtained from each patient in accordance with requirements of the First Affiliated Hospital of Nanjing Medical University Ethic Committee. Patients with PDR diagnosis who underwent necessary vitrectomy and with sufficient cell count in FVMs for single-cell sequencing were included. Patients who received preoperative anti-vascular endothelial growth factor (VEGF) or steroid injections were excluded from the study.

Cell count is an important issue for single-cell sequencing. In order to harvest as many cells as possible in the FVM, we performed a four-port, bimanual, pars plana vitrectomy. After excising the posterior hyaloid, FVMs were carefully dissected and removed from the retina with a cutter, forceps, or scissors. The entire FVMs or FVM pieces were then cut into pellets, which were then aspirated through the cutter tub into a 5-mL syringe manually. The pellets were then centrifuged at 800g, 4°C for 3 min, and stored in GEXSCOPE Tissue Preservation Solution (Singleron Biotechnologies, Jiangsu, China) at 4°C for transportation to the processing laboratory. The FVMs pellets were washed three times with PBS (10 mmol/L sodium phosphate, 0.15 mol/L NaCl, pH 7.4) at room temperature, followed by a standard two-step digestion procedure. First, collagenase type IV (Sigma-Aldrich, catalog no. C5138-100MG) was used to digest the FVM pellets for 15 min, and the tubules were sedimented by centrifugation at 200g for 5 min and washed with Hank's balanced salt solution. Second, FVMs were digested with TrypLE Express Enzyme (Thermo Fisher, catalog no. 12604013), sedimented, and washed again as in the first step. Single cells within the FVMs were then obtained by filtering the material through strainers with a mesh size of 70 µm (Falcon, catalog no. 352350). Afterward, the cells were pelleted by centrifugation at 200g for 5 min and washed twice with PBS.

For vitreous cytokines measurements, 31 patients with PDR and 12 patients with idiopathic macular hole (iMH) but no other retinal diseases were included. Approximately 0.5 to 0.8 mL undiluted vitreous fluid samples were collected by a vitreous cutter at the start of vitrectomy before intraocular infusion. Samples were collected into sterile tubes, immediately placed on ice, centrifuged at 1500 rpm for 5 min to remove the cells and debris, and then stored at −80°C until analyzed.

The single-cell suspensions (1 × 10⁵ cells/mL in PBS) were loaded onto microfluidic devices and scRNA-seq libraries were constructed according to the Singleron GEXSCOPE protocol with GEXSCOPE Single-Cell RNA Library Kit (Singleron Biotechnologies, Jiangsu, China) (11). The individual library of each sample was diluted to 4 nmol/L and pooled for sequencing on the platform of the Illumina HiSeq X Ten platform.

The in-house pipeline was used to generate gene expression profiles from raw reads. Briefly, after removing R1 reads without poly-T tails, the cell barcodes and unique molecular identifiers (UMIs) were extracted. Meanwhile, the R2 reads were mapped to reference genome hg19 (STAR v2.5.4a and featureCounts 1.6.2) using STAR (v2.5.2b) reference. Reads with the same cell barcode, UMI, and gene were grouped together to calculate the number of UMIs per gene per cell. The total number of distinct UMI sequences for each gene was reported as the number of transcripts corresponding to that gene in the digital gene expression matrix.

Seurat analysis was performed in R software using Seurat 3.1, tSNE, ggplot2, and dplyr. The data were normalized by scaling gene content by cell and log-normalizing, and variable features were identified. Data sets integrated in Seurat2 were aligned by performing canonical correlation analysis on the combined data sets and reduced into a lower dimensional space using dimensional reduction. Differential expressed genes (DEGs) were calculated using FindMarkers function in the Seurat package. Gene ontology (GO) enrichment analysis was performed on the basis of these DEGs by Metascape (https://metascape.org). The activity of gene signatures per cell was scored by gene set variation analysis (GSVA) (12). Gene coexpression network analysis was performed by GENIE3 (13) and visualized via Cytoscape 3.7.1.

Pseudotime analysis was performed with Monocle 2 (14). During the pseudotime processing, the dimensionality of the data set was reduced by the Discriminative Dimensionality Reduction with Trees (DDRTree) algorithm on the log-normalized data set.

Cellchat was used to infer the cell-cell communication networks from single-cell transcriptome data (15).

Expression matrix of published scRNA-seq data sets of human retina were downloaded from Gene Expression Omnibus (GEO) to extract the physiological retinal MG (Supplementary Fig. 3) (16). Two publicly available single-cell data sets of peripheral blood mononuclear cells (PBMCs) from healthy donors (matched with age and ethnicity) were also obtained from GEO (GSE175499). After creating the Seurat object, “FindIntegrationAnchors” and “IntegrateData” functions were used to perform the integration with our data set.

FVMs were harvested from 10 other PDR patients during surgery, were then fixed with 4% paraformaldehyde overnight, and were placed in 30% sucrose solution at 4°C overnight for cryoprotection. Afterward, FVMs were embedded in optimal cutting temperature compound (Tissue-Tek, Naperville, IL) and cut, kept at thicknesses of 8 μm, before being stored at −80°C. Primary antibodies used in this study were rabbit anti–Iba-1 (1:300; Wako Chemicals, Tokyo, Japan) and rat anti-fibronectin 1 (FN1; 1:100; Santa Cruz Biotechnology, Dallas, TX). Secondary anti-sera were Alexa Fluor 488 of donkey anti-rat (1:1000 dilution) and Alexa Fluor 594 of donkey anti-rabbit (1:1000 dilution). For immunofluorescence staining, the frozen sections were first washed with PBS three times (5 min/time) and blocked with 1% BSA (Sigma-Aldrich) solution at room temperature for 1 h. The sections were then incubated for 2 h with primary antibodies at 25°C. After being washed with PBS with 0.1% Tween 20 four times (10 min/time), sections were then incubated with fluorochrome-conjugated secondary antibodies for 1 h. Cell nuclei were counterstained with DAPI. Images were captured using a confocal microscope (Olympus 1 × 81 microscope).

The expression level of vitreous cytokines in PDR or MH vitreous samples were quantified with commercial kits (Supplementary Table 1). Vitreous Cytometric Bead Array 1 (Bio-Rad) for angiopoietin-2, chemokine (C-X-C motif) ligand (CXCL) 12, CXCL9, endostatin, fibroblast growth factor (FGF)-acidic, interleukin (IL)-18, macrophage colony-stimulating factor (M-CSF), migration inhibitory factor (MIF), matrix metalloproteinase (MMP)-1, MMP-3, MMP-7, MMP-8, platelet-derived growth factor (PDGF)-AB, PDGF-DD, placental growth factor (PLGF), periostin, VEGF-C, and VEGF-D; Cytometric Bead Array 2 (R&D) for IL-1b, IL-1ra, IL-2, IL-4, IL-5, IL-6, IL-7, IL-8, IL-9, IL-10, IL-12, IL-13, IL-15, IL-17A, eotaxin, FGF basic, G-CSF, GM-CSF, interferon-γ (IFN-γ), interferon-inducible protein (IP)-10, C-C motif chemokine ligand 2 (CCL2), macrophage inflammatory protein (MIP)-1a, PDGF-bb, MIP-1b, RANTES, tumor necrosis factor-α, and VEGF-A; and Cytometric Bead Array 3 (Bio-Rad) for transforming growth factor (TGF)-β1, TGF-β2, and TGF-β3. The Cytometric Bead Array experiment was a contract service offered by Wayen Biotechnologies, Inc. (Shanghai, China) with Bio-Plex MAGPIX System (Bio-Rad) in accordance with the manufacturer’s instructions. Cytokines like VEGF-B, NLRP3, and connective tissue growth factor (CTGF) were analyzed using human ELISA kits.

All data supporting the findings of this study are available within the article and its supplementary information files or upon request. The FVMs scRNA-seq data have been deposited in the GEO data sets under series number GSE165784 (https://www.ncbi.nlm.nih.gov).

Single cells were isolated from FVMs during vitrectomy in five patients with PDR. The clinical characteristics of these patients are shown in Table 1. A total of 6,894 cells passed filters with an average of 1,894 genes and 6,561 unique molecular identifier count per cell. Quality control metrics were highly reproducible between individual samples (Supplementary Fig. 1). Eight biological clusters were identified by unsupervised clustering following batch correction (Fig. 1A). The expression of collagen type I (COL1A1, COL1A2), the main components of fibrotic matrix, was restricted to cells of the mesenchymal lineages (fibroblast and pericyte), which account for only 8.2% (Fig. 1B and C and Supplementary Fig. 2A). Unexpectedly, the main cell population in FVM was identified as MG based on the combined expression of TREM2, C1QA, and GPR34 (Fig. 1B and Supplementary Fig. 2B) (17,18). It was further confirmed by the extensive expression of myeloid marker ionized calcium binding adapter molecule 1 (IBA-1) in FVM via immunofluorescence labeling (73.40 ± 6.03%) (Fig. 2). Other immune cells (FCN1⁺ monocyte/macrophage, CD2⁺ T cells, and CD1C⁺ dendrite cells) were also present in FVM (Fig. 1C and Supplementary Fig. 2A). The result implied the important function of immune cells, especially MG, in the pathogenesis of FVM.

Clustering of MG on FVM further identified four subclusters (Fig. 1A). Cluster MG(1) was enriched in the expression of GPNMB, LIPA, and FABP5. This subtype displayed the phenotype analogous to the activated MG accumulated around lipid plaque of diverse neurodegenerative disease (19) (Supplementary Fig. 2D). MG(3) expressed a high level of SELENOP, MRC1, and IGF1, which are often associated with anti-inflammatory activation (Supplementary Fig. 2E). In addition, we observed a CX3CR1⁺ cluster MG(2). CX3CR1 signaling of resident MG was previously demonstrated to maintain the homeostasis of the retina and serves as a primed responder of the surrounding insult (20). We therefore annotated the CX3CR1⁺ cluster as preactive state (Fig. 1D). Gene correlation network further highlighted the three independent gene networks that corresponded to each MG subtype (Fig. 1E). Of note, a cluster that highly expressed multiple cell cycle-related genes (e.g., MKI67, CENPF) was identified, indicating the presence of MG proliferation on FVM (Fig. 1C and D).

Under homeostatic condition, tissue-resident MG predominate in a steady state since the embryonic stage. However, in the diseased state, especially when the interruption of blood-retinal barrier occurs, MG may activate and multiply, relying on the local expansion of MG or the recruitment of monocytes from circulation (21,22). To investigate the ontogeny of FVM-resident MG, we integrated the single-cell transcriptomic 2 data sets of PBMCs and two data sets of retinal MG with ours (Fig. 3A). The transcriptional profiles of the combined data set were mapped along a pseudotemporal trajectory. The analysis suggested the transcriptional similarity between retinal MG and FVM-resident MG, which was segregated from monocyte-to-macrophage differentiation (Fig. 3B).

To characterize the cellular dynamics during the transition of MG, we investigated the distribution of each MG subtype along pseudotime. The cycling MG and CX3CR1⁺ MG(2) were positioned at the starting point of the trajectory, which gradually transitioned to either MG(1) or MG(3) (Supplementary Fig. 4C). Interestingly, the monocyte-derived macrophage was also connected with the cluster of MG(2) on the Uniform Manifold Approximation and Projection (UMAP) plot, indicating the identical fate determined by two differentiation paths (Fig. 3C). We next compared the expression patterns of MG subtype-specific genes between retinal MG and FVM-resident MG. Significant downregulation of the preactive marker CX3CR1 and upregulation of the MG(1) marker GPNMB were observed in FVM-resident MG compared with retinal MG. However, the expression of MG(3) marker MRC1 were altered to a lesser extent (Fig. 3D). Together, these results indicated MG(1) as the major activation state of MG resident on FVM, which highlights the potential role of MG(1) in the pathogenesis of FVM.

Additionally, upregulated genes along the transitioning trajectory were enriched for the ontology terms that are associated with the myeloid cell migration, activation, and proliferation (Supplementary Fig. 4A and B). The result further supported the notion that MG on FVM was recruited from retina and subsequently activated to MG(1) state within the PDR-associated niche.

To further delineate the functional profile of MG(1) on FVM, we performed gene set variation analysis (GSVA) on single-cell transcriptomic data to investigate the effector pathways for the formation of FVM. GSVA revealed that fibrosis-related gene expression was significantly enriched in MG(1) compared with other cell types (Fig. 4A). The fibrosis-related signature included genes such as CTSB, SPP1, and LGALS3, which are known to regulate fibroblast activation in fibrosis (Fig. 4C). This gene signature was also found in the scar-associated macrophage recently identified in a single-cell transcriptome study of liver fibrosis (Supplementary Fig. 5) (23). Furthermore, the gene set that associated with extracellular matrix was upregulated in MG(1) as well as in mesenchymal cells (Fig. 4B). We found that fibronectin (FN1), osteonectin (SPARC), and TGF–β–induced protein (TGFBI), the extracellular matrices that previously were found abundantly in epiretinal membrane (24), were highly expressed in MG(1) (Fig. 4C and E). These associated gene signatures were weakly displayed in the retinal MG under physiological conditions (Fig. 4D). Fluorescence staining of FVM also demonstrated the presence of IBA1⁺ myeloid cells topographically localized in fibronectin-positive regions (Fig. 5). Taken together, our data highlighted that the activated MG on FVM represent both profibrotic and fibrogenic properties.

To explore the effect of PDR niche on the activation of the profibrotic MG, we first evaluated cytokines in vitreous samples from PDR and iMH patients using multiplex liquid-chip assay and ELISA (Table 2). The concentrations of VEGF-B and NLRP3 were below the limit of ELISA detection, thus the two cytokines were not included. The upregulated cytokines in PDR vitreous were also compared with publications (25–29) (Supplementary Table 1). Here, we identified 23 cytokines that significantly upregulated in PDR vitreous compared with control (Fig. 6A). Among them, cytokines ANGPT2, TGFB3, and CTGF were found exclusively expressed in the mesenchymal cells of FVM. In consistent with previous studies (30), MG remained to be the major source of CCL2, CCL3, and CCL4 among all cell types (Fig. 6B). This indicated that the paracrine effect of FVM-resident cells may partially contribute to the pathological microenvironment.

We next extracted the receptors of these 23 cytokines from the CellTalker database and evaluated their expression level among different cell types on FVM. Receptors that specifically expressed in each cell type were identified (Fig. 6C). NRP1, NRP2, TGFBR1, and LRP1 were specifically expressed in both MG(1) and mesenchymal cells, whose cognate ligands VEGFA, TGFB, and CTGF have been demonstrated to induce fibrogenesis in various organs (31,32). In addition, MG(1) expressed a subset of receptors that were also enriched in monocyte and dendrite cells, such as IFNGR, CD44, CD74, CCR1, and CCR5. These receptors generally interact with the ligands IFNG, MIF, and CCL, which are critical regulators of macrophage mobilization and activation (33) (Fig. 6C). Notably, receptors CCR1, CCR5, IFNGR1, CD44, CXCR4, and NRP1 were significantly upregulated in FVM-resident MG compared with the healthy retinal MG, indicating their potential function in the activation of MG(1) state on FVM (Fig. 6D). We therefore inferred their cognate ligands VEGFA, TGFB, MIF, IFNG, and CCL as the extrinsic regulators for the activation of profibrotic MG within the PDR niche.

To further investigate the signaling of profibrotic MG that contribute to fibroblast activation, we applied CellChat to perform an unbiased ligand-receptor interaction analysis between different cell clusters (Supplementary Fig. 6). Among numerous statistically significant interactions between MG(1) and fibroblasts, SPP1 was identified as the ligand mainly expressed in MG(1) and contributed to fibroblast-dominated signaling that participates in tissue remodeling and cell-matrix cross talk (Fig. 6E). Taken together, our dissection of key ligand-receptor interactions of MG highlights the potential regulators of FVM pathogenesis within the PDR niche.

As the important hallmark of PDR, the FVM and its pathophysiological processes might be the key to understanding the mechanism of DR. In this current study, we surveyed the differential gene expression profiles of the cells, especially the MG within the FVM, at single-cell resolution for the first time. Our study also demonstrated the feasibility of scRNA-seq for comprehensive mapping of gene expression signatures of surgically incised, small-sized, human pathological tissue compared with other relatively larger tissues (e.g., retina, outflow tissues) (16,34–39). Here, we identified eight different cell types in FVM tissues, among which MG was the major cell population. Importantly, we further revealed that the MG in FVMs originated from retinal MG, with its MG(1) subtype presenting both profibrotic and fibrogenic properties. These results provide the basis for a better appreciation of cell composition and functions of human FVM as well as future potential development of interventional target.

Because of the small size and limited material availability, few studies concerning the cell compositions of FVMs have been reported. Coltrini et al. (40) recently reported using quantitative RT-PCR analysis to identify two distinct idiopathic epiretinal membrane clusters based on 20 selected genes. However, the cellular heterogeneity of cell components in disease-specific FVMs and in-depth intercellular interactions cannot be fully explored. In our study, we identified eight cell clusters in FVMs: MG, monocytes, dendrite cells, fibroblasts, endothelia, pericytes, and T cells. This finding indicated the high-quality and robust data to uncover in-depth FVM-related gene expressions at single-cell level.

Most of the previous reports individually focused on specific cells on FVMs, such as retinal pigment epithelium cells, glial cells, and fibroblasts (41). It is also generally believed that the glial components of the FVM come from Müller cells and astrocytes (42). Previous histological studies usually identified cells using marker proteins, such as glial fibrillary acidic protein-positive (GFAP), glutamine synthase (GS), and vimentin for Müller cell (43). However, almost all tissues include a heterogeneous mix of cell types. Single-cell gene expression studies have enabled the characteristics of previously known cell types to be more fully defined and facilitated the identification of novel categories of cells (44). For instance, in our study, the MG cell type also expressed Müller cell canonical markers, GS and vimentin, yet the global gene expression profiles of this cell cluster defined them as MG based on their combined expression of AIF1 and C1QA. Thus, this current study may upend people’s understanding of the cellular components in FVMs. Here, MG, identified as the major cell population, was also verified by immunofluorescence staining. Previously, Zeng et al. (45) also found cells in FVMs were heavily labeled with MG markers. Besides, in experimental model of DR (46), elimination of MG and macrophages reduced retinal vascular leakage of DR (46). Of note, we not only verified the MG as the major cell population but further found that the activation of MG(1) subtype was strongly associated with FVM formation. Furthermore, the activated MG also expressed SPP1, which acts on other cell types in FVMs, participating in tissue remodeling and cell-matrix cross talk. In addition, upregulated vitreous cytokines (ligands), such as CCL, MIF, VEGF-A, and TGF-β, were found have high expression of corresponding receptor genes in MG, indicating the specific response of the resident MG to the disease microenvironment. In the future, the downstream consequence of those interactions in the progression of FVMs remains a topic of further investigation.

DR has been diagnosed and classified clinically based on the vasculature changes by fundoscopy. However, recent technological advances have confirmed neuroretina pathology prior to any detectable vascular alteration (47). These, coupled with molecular evidence, have highlighted the role of dysfunctional immune system in the pathophysiology of diabetic complications (48). In this study, we demonstrated that MG cells play a central role in end-stage diabetic retinal disorder and that interfering with the transcriptome profile of MG might modulate their functionality and pathogenicity in DR. Further studies are needed to verify the origin of the FVM-resident MG using the combined analysis of both local tissue and systematic monocytes.

Of note, this study has several limitations. Firstly, despite the difficulty of harvesting FVMs in the clinic, the sample size of FVMs sent for scRNA-seq was relatively small. Secondly, we used iMH as a control in the measurement of vitreous cytokines; vitreous floaters would be a more ideal control in future studies since macular holes have been shown to cause changes in the vitreous proteomics (49). Thirdly, there might be heterogeneity of the included FVMs concerning the varied medical history, hemoglobin A_1c value, and visual prognosis (Supplementary Table 2). Finally, we did not included patients who had received preoperative anti-VEGF or steroid injections, which calls for further investigation.

In conclusion, we have generated a cell atlas of the human FVMs in PDR. Using scRNA-seq, we have emphasized that activated MG participate in the FVM formation. We believe our study highlights the feasibility of in-depth single-cell analysis on other tissues with diabetic complications, such as diabetic neuropathy and diabetic foot. Our detailed scRNA-seq also shed light on better understanding the critical role of MG and the molecular mechanism underlying FVMs and improvement of current preventative and therapeutic strategies for DR.

Acknowledgments. The authors thank Prof. Qianghu Wang (Nanjing Medical University) for the bio-informative assistance.

Funding. This study was supported by the National Key Research and Development Program of China (2017YFA0104101 to Q.L., and 2017YFA0104103 to S.Y.), the National Natural Science Foundation of China (81970821 to Q.L., 81900875 to Z.H., and 81870694 to S.Y.), the Special-funded Program on National Key Scientific Instruments and Equipment Development (12027808 to Z.H.), the Key Research and Development Program of Jiangsu Province (BE2018131 to S.Y.), and the Natural Science Foundation of Jiangsu Province (BK20191059 to Z.H.).

The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.

Duality of Interest. No potential conflicts of interest relevant to this article were reported.

Author Contributions. Z.H. and X.M. wrote the manuscript and researched data. Z.H., X.M., X.W., and T.Z. researched data. Z.H., S.Y., and Q.L. obtained funding and contributed to the design. M.C., Y.L., Z.Z., and W.F. researched data and contributed to discussion. P.X. contributed to resources and visualization. S.Y. and Q.L. reviewed and edited the manuscript. S.Y. and Q.L. contributed to discussion and reviewed and edited manuscript. S.Y. is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.