Gut-retina axis has been proposed according to the mice experiments  and the hypothesize that diabetic-associated microbiome could lead to increased inflammation and vascular permeability, which influence the development and progression of DR . However, alterations of gut microbiota have not been directly linked to DR in human studies. A previous study using fecal colony culture and PCR strategy did not find significant difference in the abundance of Bacteroides between diabetic patients with and without DR . This is a pioneering study to explore the DR-associated alterations of human gut microbiome and metabolome. We enrolled PDR patients as the DR group, which were compared with NDR group, to maximize the diversity between DR and NDR. We demonstrated sharply decreased bacteria abundance and gut microbiota diversity in diabetic patients with PDR as compared to those without DR. The loss of microbial taxonomic diversity is frequently observed in many human diseases such as diabetes and cancer and is recognized to be associated with systemic inflammation [10, 11], which play a clear role in the pathogenesis of DR . Therefore, we suppose the microbial diversity might reflect the severity of DR.
Moreover, we showed the change of microbiota composition and specific population of bacterial species in diabetic patients with PDR as compared to those without DR. There was no significant difference in bacterial abundance at phylum level between the two groups. However, at family level, PDR was found to be associated with significantly decreased abundant of a series of bacteria including Coriobacteriaceae, Veillonellaceae and Streptococcaceae. Several bacteria have known function particularly in metabolic diseases. For instance, Coriobacteriaceae may regulate host glucose homeostasis via liver energy metabolism and protect against hyperglycemia . Veillonellaceae are the key organisms in human gut that metabolize lactate , thereby reducing the risk of developing diabetic complication including PDR . In addition, Clostridiales_unclassified, Ruminococcaceae, Firmicutes_unclassified, Clostridiaceae and Rikenellaceae were the TOP bacterial taxa at the family level contributing to the ClpB-like gene function that leads to reduced fat mass . In contract to the bacteria with decreased abundance, the family Burkholderiaceae was the only bacterial taxa that enriched in the gut of PDR patients and the key discriminative microbial marker as identified by LEfSe analysis. Burkholderiaceae is a known heterotrophic bacteria that was reported to colonize in the gut of patients with immunosuppression  and positively correlate with chemokine IP-10, inducing systemic inflammation . Moreover, Burkholderiaceae, along with Coriobacteriaceae and Streptococcaceae, were closely correlated with altered glutamate metabolism , which has been proven to be an early pathogenic event in the development of DR .
We further demonstrated that not only the abundance and composition of gut microbiome but also the gut-derived metabolites displayed PDR-specific biosignature. The significantly differential expressed metabolites were enriched in metabolic pathways such as linoleic acid metabolism, purine metabolism, tyrosine metabolism and carbohydrate metabolism. Some metabolic pathways including arachidonic acid metabolism and purine metabolism were predicted to be altered by microbiome and further proved by metabolome. We found that the arachidonic acid metabolites such as hydroxyeicosatetraenoic acids (HETEs) and leukotriene, which are known mediators for DR development [20, 21], were increased in the fecal of PDR patients, making them as the potential diagnostic markers and therapeutic targets. Our results were consistent with previous studies that HETEs were also increased in the serum of DR patients as compared to NDR patients . On the other hand, there were 34 fecal metabolites including Vanillate, D-galactonate, D-gluconic acid and Aerobactin that were significantly enriched in microbial metabolism, showing an extensive interplay between the gut microbiota and the host through metabolic exchange and substrate co-metabolism .
We next revealed an integrated cross-omics framework to better understand the link between gut microbiome and metabolome particularly under the circumstance of PDR. We enriched coexpression-based clusters to classify metabolites with similar physicochemical properties. The method greatly shortened the numbers of metabolomics parameters from thousands of metabolite features to dozens of metabolite clusters, which made the risk analysis using logistic regression available. We then identified several metabolite clusters that were significantly associated with PDR risk. For instance, the metabolic pattern of cluster 8, including 33 up-regulated (e.g., Desogestrel and Acylcarnitine 21:2) and 13 down-regulated metabolite features (e.g., LysoPA 21:0 and Linoleic acid) in PDR patients as compared to NDR patients, sharply increased the risk of PDR by 18.5-fold in diabetic patients. In contrast, the metabolic pattern of cluster 20, including 12 up-regulated (e.g., Succinic anhydride) and 25 down-regulated metabolite features (e.g., Acylcarnitine 22:2 and (-)-Riboflavin) in PDR patients, dramatically reduced the PDR risk by 17.6-fold. The results not only confirmed the previously identified PDR-associated metabolites such as the arginine and carnitine metabolites [24, 25] but also provided new candidates from unlabeled metabolites. Furthermore, we correlated the gut microbiota with the metabolic phenotype and found significant relationships between certain bacteria families and PDR-associated metabolite clusters. For example, cluster 3, the most microbial affected metabolite cluster containing 19 organic acids and derivatives, was significantly correlated with 9 microbial families including Clostridiaceae, Lachnospiracea, Lactobacillaceae and Bifidobacteriaceae. The PDR-negative metabotype cluster 32 was only positively correlated with Coriobacteriaceae. The results shed light on the PDR-linked microbe-metabolite interaction.
There were limitations in this study. First, sample size was relatively small. Therefore, we used propensity-score matched cohort to minimize the selection bias. Second, there were some confounding factors. Although the diabetic complications and medications, which have impact on the gut microbiome and metabolomics, were not differ significantly between the two groups, the diet habit and lifestyle were not controlled and may also affect the results. Third, the casual relationship between gut microbe-metabolite and PDR could not been determined by this study. Last, most co-abundance metabolite clusters remained largely uncharacterized and need further exploration. The therapeutic significance of restoration of gut microbiota in PDR also needs to be proved.