Impact of matched samples on domain harmonization in cross-cohort NMR metabolomics

Abstract

INTRODUCTION: Machine learning models trained on metabolomics data frequently show reduced performance when applied to independent cohorts because of domain shifts introduced by differences in analytical platforms, sample handling, storage conditions, and metabolite quantification procedures. Robust approaches for harmonizing metabolomics datasets across cohorts are therefore needed to improve model transferability and reproducibility. OBJECTIVES: To evaluate the impact of matched samples on domain harmonization and cross-cohort prediction performance in NMR metabolomics data, and to compare multiple harmonization strategies for improving model generalization across independently acquired datasets. METHODS: A machine-learning framework was developed to classify ICD-10 code I50.9 (unspecified heart failure) using plasma NMR metabolomics data from the Estonian Biobank. Models were trained on a source dataset comprising 12,000 samples measured at Nightingale Health and evaluated on an independently measured destination dataset of 295 samples acquired on a different NMR platform. Harmonization approaches included within-domain normalization, Optimal Transport, Bridge mapping, and Empirical Bayes correction. The effect of matched samples was assessed using up to 108 duplicated biospecimens measured in both datasets. Additional validation was performed using independently generated human gestational diabetes and mouse diabetes metabolomics datasets from Metabolomics Workbench. RESULTS: Domain harmonization generally improved cross-cohort prediction compared with unaligned data, particularly when matched samples were available. Predictive performance increased as the number of matched samples used for calibration increased, although gains varied across methods. Nevertheless, simple within-domain normalization consistently provided strong performance and often matched or exceeded more complex alignment approaches. External validation using fully unmatched human and mouse metabolomics datasets demonstrated that harmonization strategies improved classification performance relative to raw data, supporting the broader applicability of the framework. CONCLUSION: Matched samples substantially enhance domain harmonization and improve the transferability of metabolomics prediction models across analytical platforms and cohorts. However, robust normalization remains a strong baseline, and the benefits of more sophisticated alignment methods depend on the availability and representativeness of matched reference samples. These findings provide practical guidance for developing reproducible cross-cohort metabolomics machine-learning models while preserving biologically relevant signals.