Matched Samples Hold the Key to Making Metabolomics Models Travel Across Labs

Machine learning models built on metabolomics data often crumble the moment they are asked to make predictions on samples they have never seen, not because the biology is wrong but because the numbers themselves have shifted. Differences in analytical platforms, sample handling, storage conditions and metabolite quantification procedures introduce what researchers call domain shift, a technical drift that can make a model trained in one laboratory nearly useless in another. A new study published in the journal Metabolomics tackles this problem head-on, asking a deceptively simple question: how many matched samples, the same biological specimens measured on two different instruments, does it take to make a metabolomics model travel reliably between cohorts?

The research team, led by A. Ottas with co-last authors A. Armirotti and S. Decherchi, drew on plasma samples from the Estonian Biobank, a volunteer-based cohort designed to reflect the demographic composition of the adult Estonian population. Their source dataset comprised 12,000 non-fasting plasma samples measured on a 500 MHz proton NMR spectrometer at Nightingale Health in Finland, quantified with proprietary software to yield 168 absolute metabolite concentrations plus 81 metabolite ratios, a total of 249 biomarkers spanning lipoproteins, lipids, fatty acids, amino acids, ketone bodies, apolipoproteins and compounds tied to glycolysis, fluid balance and inflammation. The destination dataset was far smaller but deliberately harder: 295 plasma samples from the same biobank, measured on an independent 600 MHz spectrometer at the Istituto Italiano di Tecnologia in Genoa, Italy, with metabolite concentrations quantified using the PULCON approach, a pulse length-based concentration determination method that relies on an entirely different quantification workflow.

The clinical target was heart failure, specifically the ICD-10 code I50.9 for unspecified heart failure. The researchers applied strict filtering criteria to isolate cases relevant to imminent disease onset: diagnoses had to occur within 60 days after blood sample donation, individuals with a cardiovascular diagnosis in the 10 years before recruitment were excluded, and a minimum of 30 documented cases was required for statistical robustness. The training dataset contained 45 heart failure cases among the 12,000 samples, while the destination dataset held 98 cases and 197 controls. Crucially, the team added 108 duplicated biospecimens, 74 controls and 34 cases, that had been measured on both platforms, providing a rare and valuable set of true matched pairs for calibrating alignment between the two measurement domains.

Because only a subset of metabolites could be confidently identified and quantified on both platforms, the cross-platform analysis was restricted to 11 compounds: glucose, lactate, citrate, alanine, glutamine, histidine, isoleucine, leucine, valine, phenylalanine and tyrosine. The selection was based solely on successful spectral assignment using Chenomx NMR Suite, not on downstream agreement metrics or harmonization performance, a design choice meant to avoid circularity. Several of these metabolites carry established biological weight in cardiovascular research. Elevated glucose reflects hyperglycemia and cardiac risk related to diabetes, lactate has prognostic value in acute heart failure, and amino acids such as alanine, glutamine and the branched-chain amino acids have been implicated in heart failure progression and metabolic dysfunction.

To bridge the two domains, the researchers built a software framework, freely available on GitHub, that applies and evaluates multiple harmonization strategies. The simplest baseline was within-domain normalization, which median-imputes missing values and rescales each feature to a zero-to-one range separately in each dataset, using no cross-domain information at all. Above that sat three alignment methods that can exploit matched samples. Optimal Transport learns a linear mapping that aligns the destination distribution with the source distribution, a principled framework for matching distributions across domains. Bridge mapping directly learns a transformation from destination features to source features, either through ordinary least squares regression on true matched pairs or, when pairs are unavailable, through nearest-neighbor pseudo-pairs or Correlation Alignment, which matches second-order statistics at the distribution level. Empirical Bayes harmonization, in the ComBat style, models and corrects feature-wise location and scale differences across batches, shrinking estimates across features to produce stabilized correction parameters.

The experimental design was meticulous. For each value of k, the number of matched pairs used for calibration, the team randomly sampled k pairs without replacement across 30 independent trials, estimated a mapping operator from those pairs, applied it to all destination samples, and then trained machine learning models exclusively on the source dataset before evaluating them on the transformed destination data. Seven algorithms were tested, including Random Forest, Decision Trees, Support Vector Machines, Logistic Regression, Gaussian Naive Bayes, Stochastic Gradient Descent and k-nearest neighbors, with performance measured by balanced accuracy, the mean of sensitivity and specificity, and the Matthews Correlation Coefficient, which accounts for all four confusion matrix values and suits imbalanced datasets.

The results tell a nuanced story. With no matched samples at all, the best test performance came from a Support Vector Machine on normalized data, reaching a balanced accuracy of 0.64 and an MCC of 0.31. When all 108 matched samples were used, the same model on normalized data climbed to a balanced accuracy of 0.72 with an MCC of 0.46, and Logistic Regression on Empirical Bayes-transformed data achieved a balanced accuracy of 0.705 with an MCC of 0.39. Principal component analysis revealed a clear separation between the two datasets for raw, normalized and Empirical Bayes data with zero matched samples, but showed clustering for Optimal Transport and Bridge mapping, suggesting these methods genuinely pulled the domains together in feature space.

Yet the most striking finding is the strength of the humble baseline. Simple within-domain normalization, which never sees a matched pair, remained competitive across the full range of k and often matched or exceeded the more sophisticated alignment approaches. The authors attribute this to a central trade-off in harmonization: reducing distributional differences between domains does not necessarily improve predictive performance. Alignment methods can improve source-destination overlap, but they may also distort or attenuate features that matter for classification, particularly when alignment is based on marginal distributions rather than class-conditional structure. PCA identifies directions of maximal variance without using class labels, so improved visual overlap does not guarantee preserved class separability. The best harmonization method, the authors conclude, is not the one that maximizes distributional similarity but the one that best preserves predictive structure.

On agreement metrics applied to the matched samples, Optimal Transport consistently outperformed all other methods, achieving Pearson correlation of 0.89, Spearman rank correlation of 0.86, an intraclass correlation coefficient of 0.89, Lin’s concordance correlation coefficient of 0.89, a root mean square error of 0.03 and a mean absolute error of 0.01. Performance generally increased as the number of matched samples grew, with gains stabilizing at larger values of k, though variability was pronounced at small k where alignment was unstable. At k equal to one, the cross-domain transformation cannot be estimated reliably at all, illustrating that paired-sample harmonization requires a minimum level of correspondence to produce a stable mapping.

To test whether the framework generalizes beyond the Estonian Biobank setting, the team turned to Metabolomics Workbench, a public repository, and selected independently measured datasets with matching species, tissue and metabolite names. For human gestational diabetes, two LC-MS studies sharing nine common metabolites were harmonized, with models trained on one dataset and tested on the other; raw data yielded a best balanced accuracy of 0.63 while harmonized data reached 0.71. For mouse diabetes models, two studies sharing 12 common metabolites produced an even more dramatic improvement, from a balanced accuracy of 0.67 on raw data to 0.92 on normalized data with an MCC of 0.87. Since none of these external datasets contained matched samples, only unpaired harmonization methods could be applied, yet harmonization still improved classification relative to raw, unaligned evaluation.

The practical recommendations that emerge are refreshingly concrete. When several matched samples are available and dominate over unmatched ones, Optimal Transport is a strong candidate, having significantly outperformed all other strategies on balanced accuracy, MCC, sensitivity, positive predictive value, negative predictive value, ROC-AUC and PR-AUC in paired iterations. When unmatched samples dominate or are the only option, simple within-domain normalization should be preferred. The authors caution that their implementation does not use fully nested cross-validation, so reported cross-validation estimates may be slightly optimistic, though the principal conclusions rest on external destination-domain performance, which provides a substantially less biased estimate of generalizability. They also note the risk of overcorrection: if technical and biological factors are confounded, harmonization may remove biologically meaningful differences, especially in metabolomics where subtle intensity shifts may carry disease-relevant information.

For a field increasingly dependent on biobank-scale data and multi-center collaborations, the message is clear. Matched samples are highly valuable for cross-domain prediction, and even a limited number can provide useful calibration information, but their benefit is not automatic and depends on the availability and representativeness of the reference set. Robust normalization remains a formidable baseline that no laboratory should skip, and the sophisticated machinery of optimal transport and empirical Bayes correction earns its complexity only when enough matched reference data constrains it. As metabolomics models move from discovery to clinical deployment, the ability to preserve biologically relevant signals while erasing technical noise may determine whether the promise of cross-cohort precision medicine survives contact with the real world.

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

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

Article References: Ottas, A., Goldoni, L., Armirotti, A., & Decherchi, S. (2026). Impact of matched samples on domain harmonization in cross-cohort NMR metabolomics. Metabolomics, 22(6), Article 172. https://doi.org/10.1007/s11306-026-02543-0

Image Credits: AI Generated

DOI: 10.1007/s11306-026-02543-0

Keywords: Impact, matched, samples, domain, harmonization, cross-cohort, metabolomics, scientific research