Predicting Dry Eye Disease and Discovering Serum Metabolic Signatures with Machine Learning

Background: Dry eye disease (DED) is a highly prevalent inflammatory ocular surface disease. It is believed to have at least some root in impaired metabolism, but knowledge of changes associated with DED is limited. Metabolomics produces complex high-dimensional information well suited to machine learning. We describe a pipeline for preprocessor and model benchmarking for DED prediction using metabolomics data from twins, and model interpretation to identify new serum metabolite associations with DED.<br><br>Methods: Untargeted serum Metabolon profiles from TwinsUK participants and two definitions of DED from the Women’s Health Study (WHS) DED questionnaire were used: WHS DED and Highly Symptomatic (HS) DED. After 80/20 train-test split, benchmarking was performed by training 32 algorithms on 21 differently preprocessed training sets, using Area Under the Receiver Operating Characteristic Curve (ROC AUC) as the main metric. Statistically significant metabolite associations with DED were found by using L2 penalized logistic regression for cluster-aware bootstrap-based coefficient inference.<br><br>Findings: Adding metabolite levels to covariates significantly improved test-set discrimination versus covariates alone (permutation p&lt;0·05 for both outcomes). Peak test-set performance reached ROC AUC 0·70 for WHS DED and 0·75 for HS DED. Preprocessing materially influenced performance, particularly for linear and stochastic gradient descent models; the sample-wise L2 normalized dataset performed best. Thirteen serum metabolites were significantly associated with HS DED, including lower 5-methyluridine, guanosine, monoacylglycerols, lysophosphatidylethanolamines, and isobutyrylcarnitine, and higher N-formylphenylalanine and nicotinamide. No significant associations were found for WHS DED.<br><br>Interpretation: Metabolomics contribute to DED prediction. Optimal choice of data preprocessing, and model selection, is critical to maximize predictive performance. Significant metabolite associations were observed for HS DED, suggesting a metabolically distinct phenotype and pointing to systemic alterations in nucleotide-related metabolites, selected lipid classes, and nicotinamide/acylcarnitine-related metabolism.