This study consisted of genome-wide DNA methylation analyses of extreme phenotypes within two consecutive studies that identified novel epigenetic modifications of DR. In the discovery stage, a large number of novel candidate sites were identified. To validate the results and assess whether these methylation modifications could be used as biomarkers to predict DR, we analyzed DNA methylation in an independent longitudinal study. Our results suggest that aberrant methylation may predict DR, as hypermethylation at cg04026387 and cg12869254 identified in the cross-sectional study was successfully replicated in the longitudinal study. Finally, the methylation degree of these two sites negatively correlated with macular RNFL thickness and DCP VD, which further confirmed their potential for indicating DR [23, 24]. Therefore, methylation at these two sites may contribute to neural and vascular damage in the early stages of DR and serve as novel biomarkers for predicting the disease.
In this study, an extreme phenotypic design was adopted to identify the specific DMSs, which enabled us to identify novel methylation modifications, despite the relatively small yet sufficient sample size, as in studies of similar design [25,26,27,28]. Conversely, the results of previous studies on DR were skewed by the incomplete phenotypes of included participants, that is, healthy individuals and/or inclusion of individuals with diabetes diagnosis of < 20 years as DR control groups, and/or DR patients with random DR-free T2DM duration as the DR case group. Current evidence suggests that an extreme phenotypic design is necessary for screening and identifying DR markers [29, 30]. The extreme phenotypic design is a recent concept that is used to increase homogeneity and better distinguish signals from noise in genome-wide analysis. Under this approach, the decreased variation in homogeneous samples results in a robust ability to identify signals, even in a relatively small number of samples. Concurrently, it enables the discovery of rare methylation alternations, which is another advantage of this design. Thus, we included patients diagnosed with T2DM for at least 20 years without any signs of DR as an extreme control group, and patients diagnosed with DR within 4 years of T2DM diagnosis were included as an extreme case group. As most patients typically develop DR within 17 years of T2DM diagnosis , while epigenetic changes typically initiate retinal impairment within a short period, the enrolled patients may represent extreme phenotypes enriched in “protective” and “risk” epigenetic alternations, respectively. Therefore, this model increased the candidate site identification in this study.
To our knowledge, this is the first study to longitudinally investigate the relationship between DNA methylation and T2DM in patients with DR. Thus far, all previous studies on T2DM patients have been cross-sectional, which limits causal inferences of their results. However, DNA methylation has been suggested to be dynamic rather than rigidly fixed in various diseases and aging process . Therefore, it is important to clarify the time-ordered relationships between such aberrant methylation and DR pathogenesis. Our cross-sectional study revealed significant hypomethylation in sites within S100A13 in the gene-positive DR group, which is in agreement with the results obtained by Li et al.  Demethylation of this gene purportedly increases hyperglycemia-induced damage through calcium signaling and the RAGE pathway. However, these results were not replicated in the longitudinal study. Similarly, many other candidate sites identified in the cross-sectional discovery stage exhibited no significant differences in the longitudinal study. This highlights the limitations of previous cross-sectional studies in which the chronological relationship between exposure and outcome was unclear. The changes in methylation at these sites may be explained as concomitant consequences of DR development rather than as primary contributing factors to DR pathogenesis or progression. Cg12869254 and cg04026387 were successfully replicated in this longitudinal study, suggesting that methylation alterations at these two sites precede DR, supporting their role in the early pathogenesis of DR.
The two novel sites identified in this study are included in the ZDHHC23 and SLC25A21 genes, which have not been identified in previous studies. A recent animal study demonstrated the neurotoxic effects mediated by the downregulation of the ZDHHC23 gene . Decreased ZDHHC23 levels dysregulated interactions between palmitoyl-protein thioesterase-1 and acyl protein thioesterase-1, leading to increased plasma membrane H-Ras and subsequent microglial proliferation in mouse brains. Activation of microglia leads to increased secretion of tumor necrosis factor-alpha, interleukin-6, monocyte chemoattractant protein-1, and complement component C1q, promoting the transformation of astrocytes to the neurotoxic A1 phenotype . The astrocytes of the A1 phenotype not only lose their original neurotrophic role, but also actively produce neurotoxins that degrade neurons. Retinal neurodegeneration is considered an early component of DR, which can precede visible vasculopathy [23, 24]. In this study, the degree of methylation of cg12869254 negatively correlated with macular RNFL thickness in the superior and nasal subregions, while the RNFL thickness represented the axons of retinal ganglion cells. Our teams and other groups have demonstrated that reduced RNFL thickness was significantly associated with a higher risk of development and progression of DR . Therefore, we hypothesize that ZDHHC23 gene-mediated neurotoxicity caused by hypermethylation of cg12869254 may also contribute to the retinal neurodegeneration in DR. Additional studies are warranted to further elucidate the potential role of ZDHHC23 in DR.
SLC25A21 is involved in intracellular organic acid catabolism and metabolism . Within mitochondria, it transports 2-oxoadipate and 2-oxoglutarate to be converted into acetyl-CoA. Boczonadi et al.  found that the dysfunction of SLC25A21 may lead to impaired transport of 2-oxoadipate, which in turn disrupts the degradation of lysine and tryptophan, leading to the accumulation of 2-oxoadipate, piperidinic acid, and quinolinic acid. Overexposure to these cytotoxic substances leads to irreversible mitochondrial and tissue damage mediated by free radicals. Consistently, the methylation degree of cg04026387 locus negatively correlated with macular DCP VD in our study, suggesting that it is a more sensitive indicator than VD in other plexus . Therefore, we hypothesize that similar mechanisms of mitochondrial damage and vascular damage may also occur in the pathogenesis of early DR. However, SCP VD in the nasal subregion appears to be increased with higher methylation of cg04026387. Similar results have been reciprocated in studies on rats and humans [38, 39]. One possible explanation for this is that the deeper capillary layer is more sensitive to mitochondrial damage. In the early stages of DR, the above-mentioned factors may act mainly on the DCP, whereas the SCP exhibits a compensatory flow increase secondary to the reduced DCP VD. However, further studies investigating these genes are required to validate our hypothesis.
DKD is another major microvascular complication of diabetes mellitus . Given the commonalities in pathogenesis and frequent co-development of DR and DKD, we further analyzed the relationship between these two sites and renal function indexes [20, 41,42,43]. Hypermethylation at these two identified sites was significantly associated with elevated ACR and reduced eGFR, further increasing our confidence in the involvement of these two sites in diabetic microvascular complications. We hypothesize that inflammatory responses, mitochondrial dysfunction, and oxidative stress discussed above may also damage the endothelial cells of glomerular capillaries [41, 43,44,45], thus leading to urinary protein leakage and decreased glomerular filtration function.
To our knowledge, few studies have collected systemic and ocular factors, such as renal function and AL, while analyzing the molecular etiology of DR. These factors have been independently associated with DR onset and progression [16,17,18,19]. Thus, the differentiated sites identified in previous studies might have been confounded by the presence of these factors. This study obtained various systemic and ocular factors from the participants, and most characteristics were well matched in each comparison group. This increased the reliability of our results.
This study had certain limitations. First, since most patients typically develop DR within 17 years of T2DM diagnosis , patients diagnosed with T2DM for 20 years without DR were included in the gene-negative control (NDR) group. However, even if the odds are low, these patients may still develop DR in the future. Future longitudinal studies are needed to determine whether these protective factors are still effective when the diabetes duration is over 20 years. Second, the sample size of this study was relatively small, and therefore the results should be treated with care. However, the adopted extreme phenotypic design greatly increases genetic homogeneity, which provides a robust ability to distinguish signals from noise. Despite the relatively small sample size, two sites were successfully replicated in the longitudinal study and were correlated with OCT/A parameters.