**Background:** Macular edema (ME) is a common complication of retinal diseases such as age-related macular degeneration (AMD), diabetic retinopathy (DR), and branch retinal vein occlusion (BRVO). Current imaging techniques reveal structural changes but not underlying molecular pathophysiology. Metabolomics of aqueous humor (AH) may provide insights into disease mechanisms and identify biomarkers for early diagnosis and treatment stratification. This study aimed to characterize the AH metabolome in ME of different etiologies and develop a diagnostic model.
**Methods:** AH samples were collected from 60 ME patients (20 AMD-ME, 20 DME, 20 BRVO-ME) and 20 age- and sex-matched controls during routine intravitreal injection or anterior chamber puncture. Samples were analyzed by liquid chromatography-tandem mass spectrometry (LC-MS/MS). Data preprocessing used XCMS and an in-house MS2 database (BiotreeDB). Differentially expressed metabolites (DEMs) were identified using Mann-Whitney U test and partial least squares-discriminant analysis (PLS-DA) with criteria VIP ≥ 1, log2(fold change) > 1, and FDR < 0.05. Pathway enrichment was performed on MetaboAnalyst. A gradient boosting machine diagnostic model was developed using recursive feature elimination (RFE) on a training cohort (60% of samples) and validated on a test cohort (40%).
**Key Results:** Compared to controls, 93 DEMs (58 upregulated, 35 downregulated) were identified in DME, 19 (7 up, 12 down) in BRVO-ME, 21 (7 up, 14 down) in AMD-ME, and 30 (17 up, 13 down) in all ME combined. Nine DEMs were common across all three ME types, including one upregulated (clavulanic acid) and eight downregulated. Etiology-specific DEMs included 66 for DME, 6 for AMD-ME, and 1 for BRVO-ME. Pathway analysis revealed that DEMs were mainly involved in lipid metabolism (nicotinate/nicotinamide, linoleic/linolenic acid, sphingolipid, arachidonic acid, glycerophospholipid, and steroid metabolism) and amino acid metabolism (tryptophan, tyrosine, alanine, aspartate, glutamate). Lipid metabolism pathways were significantly upregulated in all ME patients. Among ME etiologies, 40 DEMs were found in BRVO-ME vs AMD-ME, 128 in DME vs AMD-ME, and 84 in BRVO-ME vs DME. Carbohydrate metabolism (citric acid cycle, glyoxylate/dicarboxylate, pyruvate, glycolysis/gluconeogenesis) was upregulated in BRVO-ME compared to DME and AMD-ME. The diagnostic model using 60 selected metabolites achieved an AUC of 1.0 in the training cohort for distinguishing ME from controls and for each etiology. In the test cohort, AUCs were 0.83 for ME vs controls, 0.79 for AMD-ME, 0.94 for DME, and 0.77 for BRVO-ME. The model showed robust accuracy, sensitivity, specificity, positive predictive value, and negative predictive value.
**Clinical Implications:** This study demonstrates that AH metabolomics can differentiate ME of different etiologies, revealing distinct metabolic profiles and pathways. The identified DEMs, particularly those involved in lipid and amino acid metabolism, may serve as biomarkers for early diagnosis and help guide treatment selection, especially for cases unresponsive to anti-VEGF therapy. The machine-learning model shows promise for clinical application, with highest accuracy for DME detection. These findings support the potential of AH metabolomics in precision medicine for retinal diseases.