All animal procedures were under an experimental protocol approved by Sun Yat-sen University’s Standing Committee on Animals. As described previously, Cynomolgus monkeys with spontaneous type 2 diabetes were screened in southern china, and diagnosed as diabetes with the parameters of fasting glucose and HbA1c levels as described before. Non-diabetic age-matched controls were selected randomly from the colony (Fan et al., 2021). All animals were maintained at 25°C on a 12 h light:12 h dark schedule.

Animals were anesthetized and perfused with physiological saline, the whole retinal layer were dissected out, and digested in the solution with papain (10 units/ml, Sigma cat. P3125) and DNase at 37°C for 5 min, and dissociated into single cell suspensions with 0.04% bovine serum albumin (BSA) in PBS solution. Single cell suspensions were diluted at a concentration of 500–1800 cells/uL in 0.04% BSA/Ames for loading into 10× Chromium Single Cell Chips.

Single cell libraries were constructed using commercial Chromium 3′ v2 and Gel Bead Kit (10× Genomics) following to the manufacturer’s protocol. Single cells were partitioned into Gel beads in EMulsion (GEMs) in the GemCode instrument followed by cell lysis and barcoded reverse transcription of RNA, amplification. Sequencing libraries were sequenced on Illumina HiSeq 2000.

Microglia cells (BV2) were cultured in DMEM (Dulbecco’s modified Eagle’s medium; Gibco, United States) supplemented with 10% fetal bovine serum (FBS), and passaged with 0.05% Trypsin/EDTA. Microglia cells were activated by treating with 100 ng/mL lipopolysaccharide (LPS, Sigma) for 24 h and subsequently with LPS + TGFB2 (50 ng/mL) or LPS for 24 h.

Total RNA was extracted using TRIzol reagent (Invitrogen) following to the manufacturer’s protocol. Exactly 1 μg of total RNA was reverse transcribed into cDNA using Reverse Transcription Master Mix (Tiangen). qPCR reactions were performed in 384-well plates on a CFX-384 thermocycler (Bio-Rad) using iTaq Universal SYBR Green Supermix (Vazyme) as illustrated by the manufacturer’s instructions. Primer sequences for the experiment are listed in Supplementary Table 1.

We integrated two datasets using canonical correlation analysis (CCA) of Seurat and group cells with similar transcriptomics into the same cluster. Briefly, the top 2000 highly variable genes from two datasets were selected, and then normalized and integrated. Next, principal component analysis (PCA) was performed on integrated data. The significant principal components (p-value < 1E-5) were selected to construct k-nearest neighbor graphs (k = 30) based on the Euclidean distance. Cluster these cells into cluster using the Louvain–Jaccard. Cluster-specific genes were identified using the FindAllMarker function in Seurat and applied to annotate cluster through comparison with the well-known markers of retinal cell types.

To identify the differentially expressed genes (DEGs) between the control group and diabetes group, we compared gene expression levels of each cell types in the two groups using Seurat FindMarker function with options: test = “MAST” and logfc.threshold = 0.25. Genes with p_val_adj < 0.05 and the absolute value of avg_logFC ≥ 0.5 were considered as DEGs.

Gene ontology analysis of DEGs was performed in R package clusterProfiler (version 3.14), which using annotation information provided by the R package org.Hs.eg.db (version 3.10.0).

Cell-cell communication analysis was performed using CellChat (Version 0.5), based on the known ligand-receptor pairs in cell types. Briefly, the normalized genes expression matrix and cell labels generated by Seurat acted as input for CellChat. The overexpressed ligands and receptors in each cell types were calculated and then projected into the protein-protein interaction network. Next, to infer the biologically significant cell–cell communication, we assigned each interaction with a probability value and performed a permutation test. Cell-cell communication altered under DM was analyzed by joint manifold learning and quantitative contrasts of multiple cell-cell communication networks. Finally, communication networks were visualized using circle plot and signaling pathways were illustrated by using bubble plot.

All statistical data were analyzed using R software (version 3.6.0). Statistical significance was calculated with unpaired two-tailed Student t test. P < 0.05 was considered statistically significant.

All sequencing data and expression metric for this study are available in Gene Expression Omnibus (GEO) with the accession number GSE168908.

Consistent with our previous study, the fasting blood glucose levels was 32.7 mmol/L in spontaneous type 2 diabetes monkey versus 4.46 mmol/L in control monkey, and the glycated hemoglobin (HbA1c) was 11% in diabetes versus 3.8% in control. To elucidate the cell-type-specific gene expression alterations associated with non-proliferative diabetic retinopathy, we isolated the entire neural retina from diabetic and age-matched healthy cynomolgus monkey (M. fascicularis), and performed droplet-based scRNA-seq. In total, 10263 high-quality cell profiles were retained for downstream analysis, including 8708 retinal cells from the control and 1555 retinal cells from diabetic monkey.

Unbiased clustering identified 13 clusters that existed in both groups (Supplementary Figures 1A,B). We annotated six cell types (Figure 1A), including two glial cell types (microglia and Muller glia cells) and four neuronal cell types (rod and cone photoreceptors, amacrine cells, and bipolar cells) according to the expression pattern of well-known marker genes for each cell type –Rods expressing SAG, NRL, GNAT1, and RHO; Cone expressing GNGT2, GUCA1C, ARR3, and OPN1WM; Amacrine expressing TFAP2A, TFAP2B, and GAD1; Bipolar expressing TMEM215, VSX2, and CABP5; Muller expressing APOE, GLUL, RLBP1, and CRYAB; Microglia expressing C1QA, C1QB, and C1QC (Supplementary Figure 1C). Function annotation of cell-type-specific genes revealed the functional characteristics of non-human primates (NHPs) retinal cells as following: “phototransduction” for Rod and Cone, “Vesicle-mediated transport in synapse” for bipolar and “neutrophil mediated immunity” for microglia (Supplementary Figures 1D,E), which were concurred with the known function of retinal cells.

Through differentially expression analysis of each retinal cell type, microglia showed the largest number of up-regulated genes under diabetic mellitus (Figures 1B,C), indicating that microglia are significantly affected by hyperglycemia. GO analysis of DEGs revealed that “neutrophil activation,” “regulation of neuron death,” and “epithelial cell migration” were enriched in up-regulated genes within microglia (Figure 1D), indicating that hyperglycemia activated retinal microglia. Some crucial genes related to inflammatory regulation were highly expressed, including CEBPB, CEBPD, JUN, and FOS (Figure 1E). CEBPB and CEBPD belong to CCAAT/enhancer binding protein (c/EBP) family, which is implicated in inflammatory and required for activation of microglia (Straccia et al., 2011), JUN and FOS formed AP-1 transcription factor, which induces pro-inflammatory cytokine expression (Qiao et al., 2016). Furthermore, we analyzed the marker genes related to pro-inflammatory (M1) and anti-inflammatory (M2) (Leng and Edison, 2020). Both M1 and M2 marker genes were elevated by hyperglycemia (Figures 1F,G), a similar finding was reported in other neuroinflammatory diseases, such as spinal cord injury (Kigerl et al., 2009). Collectively, these results demonstrate that hyperglycemia induces microglial activation and elevates inflammation levels.

To reveal the communication between activated microglia and other retinal cells, we performed cell-cell communication analysis using R package CellChat (Jin et al., 2021). The overall cell-cell communication strength was enhanced by hyperglycemia, whereas the largest strength of outgoing and incoming signaling was found in microglia (Figure 1H).

Next, we focused on the signaling pathway sent and/or received by microglia (Figure 2A), including CD45, TNF, THBS, and PSAP. TNF signaling, known pro-inflammatory signaling in multiple diseases, was predicated to mediate the self-communication of microglia (Figures 2B,C). The communication was mediated by ligand TNF-α and its two receptors, TNFR1 and TNFR2, which were elevated by hyperglycemia in microglia (Figures 2D,E). Previous studies had revealed that TNF-α and TNF receptors could trigger the activation of NF-κB to activate the downstream inflammatory genes (Hayden and Ghosh, 2014; Bras et al., 2020). Our analysis demonstrated that the NF-κB targeted genes were significantly enriched in the upregulated genes in the microglia (P values < 4.05e-10), including CEBPD, JUNB, and FOS, and TNFRSF1B receptor (Figures 2F,G and Supplementary Figures 3A,B). These findings suggested that TNF potentially mediates the activation of NF-κB signaling in an autocrine manner; this consequently activates microglia under hyperglycemia.

OPN (SPP1) signaling is also enhanced under hyperglycemia. OPN is a proinflammatory cytokine that regulates the activation and function of microglia (Yu et al., 2017). We found OPN was mainly expressed in both Muller and microglia (Figure 3A), whereas its receptor, CD44, was up-regulated in microglia by hyperglycemia (Figures 3B,C). A previous study have demonstrated that OPN could down-regulation the expression of TNF in microglia (Rabenstein et al., 2016). Our data suggested that OPN secreted by Muller and microglia could suppress TNF activation to impede excessive activation of microglia. Similarly, our data showed that GAS6 mediated cell-cell communication between amacrine and microglia was enhanced by hyperglycemia, and negatively regulated the pro-inflammatory via TAM receptors, Axl and Mer (Gilchrist et al., 2020; Figure 3E). CellChat predicted that hyperglycemia-induced GAS6 signaling was mediated by ligand GAS6 and both two receptors (Figure 3D), Ligand GAS6 was up-regulated in microglia and amacrine, whereas the receptors were up-regulated in microglia (Figure 3F).

Moreover, we reported enhanced communication between cone and microglia through TGF-β signaling pathways (Figures 4A,B). TGF-β is a major anti-inflammatory factor, which triggers the downstream expression of target genes through TGFBR1 and TGFBR2 receptors. Previous studies confirmed that overexpression of TGFB1 in the cones can reduce the cone degeneration caused by retinitis pigmentosa through microglia (Wang et al., 2020) and ablation of TGFBR2 in retinal microglia lead to the microglia activation and promote the pathological microglial gene expression profile (Ma et al., 2019). Our study revealed that TGFB2 potentially mediates the communication between cone and microglia under hyperglycemia (Figures 4C,D).

To investigate the function of TGFβ signaling in microglia, microglial BV2 cell lines were cultured for in vitro recombination. Microglia treated with LPS significantly increased expression of Tnf and Mcp1 (Figures 4E,F), and decreased the expression of Arg1 (Figure 4G), indicating the inflammation activation in microglia. TGFβ2 treatment resulted in the significantly decreased the LPS-induced Mcp1 and Tnf expression (Figures 4E,F), and augmented Arg1 expression in microglia (Figure 4G). In summary, these data demonstrated that TGFβ2 suppress microglia activation and potentially reduce the neurotoxicity.

These findings demonstrated that communication between microglia and other retinal cells, including photoreceptors, is a potential feedback mechanism to suppress the excess activation of microglia under diabetes mellitus.