Abstract
Purpose :
To determine differences among the metabolite profile of aqueous humor from control and glaucoma eyes. To determine whether machine-learning algorithms allow accurate classification of the aqueous humor metabolite profiles into control, pseudoexfoliation (PEX) and primary open angle glaucoma (POAG) categories.
Methods :
All human samples were collected adhering to tenets of declaration of Helsinki and ARVO guidelines under IRB approved protocols. Aqueous humor (AH) samples were obtained from human donors [control, POAG and PEX glaucoma (each n=20)]. AH samples were subjected to two-dimensional nuclear magnetic resonance (NMR) analysis on a Bruker Avance 600 MHz NMR device. The same samples were then analyzed on a Q-Exactive orbitrap mass spectrometer after chromatography. We implemented isotopic ratio outlier analysis for mass spectrometry and used ClusterFinderTM software for identification and quantification. Control and glaucoma donors were all Caucasian, age 71±5.0 and 70.3± 7.5 (equal distribution of both genders), respectively. Ophthalmic clinical history of donors were recorded and POAG, PEX patients were diagnosed. Control subjects were those who underwent cataract surgery. Bioinformatics analysis of metabolites was performed using MetaboAnalyst 4.0/STATA 14.2. Three different machine-learning algorithms were used to build prediction models.
Results :
Significant differences in metabolites in PEX and POAG compared to controls as well as between PEX and POAG were found. Principal component analysis indicated clear grouping based on the metabolomes of the three conditions: PEX, POAG and control. We used machine-learning algorithms and a percentage set of data to train and utilized a larger dataset to test whether a trained model is able to correctly classify the test dataset as PEX, POAG or control. All algorithms were able to accurately classify the test datasets based on the AH metabolome of the sample. We next compared the AH metabolome with known AH and trabecular meshwork (TM) proteome/genome datasets to gain insight into metabolic pathways. We found putative protein/gene pathways associated with observed significant metabolite changes in PEX.
Conclusions :
We found significant differences in metabolites in glaucoma compared to controls. The machine-learning algorithms were able to predict PEX as a distinctive glaucoma group based on AH metabolite profiles when trained with a 30% subset of the total data.
This is a 2020 ARVO Annual Meeting abstract.