Decoding health from NMR spectra: machine learning models for metabolic health

Decoding health from NMR spectra: machine learning models for metabolic health Alain Ibáñez de Opakuaa,b, Rubén Gil-Redondob, Maider Bizkarguenagab, Ángela de Diegob, Ricardo Condeb, Tammo Diercksb, Beatriz González-Valleb, Nieves Embadeb, José María Matob, Oscar Milleta,b Introduction Methods Results Metabolic age Disease classification Conclusions Ref. Metabolism offers a rich and dynamic window into human health, reflecting both physiological balance and pathological disruptions. Nuclear Magnetic Resonance (NMR) spectroscopy, with its reproducibility and non-destructive nature, provides a powerful platform for metabolic profiling. In recent years, the integration of machine learning techniques with NMR data has opened new avenues for deciphering complex biochemical signatures associated with aging and disease. In this study, we introduce a comprehensive computational framework designed to extract clinically relevant insights from NMR metabolomics data. By quantifying metabolite concentrations from J-resolved spectra and deriving clinical parameters from 1D ¹H NOESY spectra, we generate a multi-layered feature set that captures diverse aspects of metabolic health. These features are subsequently used to train predictive models aimed at estimating biological age and classifying disease states. Our approach achieves high accuracy in age prediction, enabling the detection of individuals whose metabolic profiles deviate from normative trajectories—potentially signaling accelerated or decelerated aging. Furthermore, we demonstrate the utility of our models in distinguishing between multiple health and disease conditions, highlighting the potential of NMR-based machine learning as a scalable and interpretable tool for precision health monitoring. 1Measuring Biological Age via Metabonomics: The Metabolic Age Score. Hertel et al. J Proteome Res, 2016. 2Metabolic age based on the BBMRI-NL 1H-NMR metabolomics repository as biomarker of age-related disease. Van den Akker et al. Circ Genom Precis Med, 2020. 3NMR metabolomic modelling of age and lifespan: a multi-cohort analysis. Lau et al. medRxiv, 2023. ¹H-NMR spectra were acquired from over 30,000 serum samples using both 1D NOESY and fastacquisition 2D J-resolved (JRES) experiments. The 2D JRES spectra were used to quantify 49 metabolites through a dedicated pipeline designed to minimize signal overlap. In parallel, 1D NOESY spectra were used to estimate 25 clinical parameters via supervised regression models, covering both directly observable markers (e.g., CRP, albumin) and inferred physiological indices (e.g., calcium, eGFR). While models based on raw 1D spectra generally offer higher predictive performance, the use of derived features enhances explainability and supports downstream clinical interpretation. To estimate chronological age—referred to as metabolic age—, we developed a stacking ensemble machine learning model trained on a subset of ~8,000 individuals selected from an initial population of ~28,000. This reduction was performed to ensure a uniform distribution across the age range and to mitigate regression to the mean effects during model training. The model based on 1D NOESY spectra shows a strong correlation between chronological and metabolic age (R=0.92), outperforming previous NMR-based metabolic age models1,2,3, which didn't reach 0.8. Consistent results were obtained with other NMR datasets: CPMG (R=0.91), NOESY FID (R=0.87), and quantified metabolites plus clinical parameters (R=0.88). Additionally, distributions of metabolic distortion (the difference between metabolic and chronological age) show significant alterations across disease groups. For example, individuals with prostate cancer exhibit a positive distortion, suggesting accelerated metabolic aging. In contrast, patients with liver diseases display a broader distribution, likely due to the heterogeneous nature of liver conditions, each associated with distinct metabolic consequences. These findings underscore the utility of metabolic age as a biomarker capable of capturing diseaserelated physiological alterations. Using quantified metabolites and predicted clinical parameters, we trained a multiclass XGBoost classifier to differentiate between nine health categories. The confusion matrix (colors normalized per row) shows excellent overall performance with an accuracy of 0.80. Misclassifications occur predominantly between expected neighboring classes, such as younger vs old adults (with the threshold set at 50 years), older adults vs long CoVid patients (due to mild residual effects observed in long CoVid), and older adults vs metabolic syndrome (a condition closely associated with aging). ROC curves for each class demonstrate robust discrimination, with AUC values ranging from 0.91 to 1.00, confirming the model’s high sensitivity and specificity across conditions. SHAP values revealed that variables associated with systemic inflammation and metabolic stability are among the most relevant predictors of metabolic age. GlycA and erythrocyte sedimentation rate showed the strongest positive contributions, consistent with their roles as markers of inflammation. In contrast, GlycB and albumin contributed negatively, likely reflecting lower chronic inflammation and better nutritional status, both associated with a younger metabolic profile. 49 quantified metabolites: 25 predicted clinical paremeters: For disease classification, we implemented a multiclass model based on eXtreme Gradient Boosting (XGBoost) to distinguish nine health categories: seven disease conditions and two age-defined healthy groups (young and older). The healthy reference population was selected to match the age and sex distribution of the combined disease cohorts, enabling partial class balance and improving generalizability. 1,5-Anhydrosorbitol, 2-Aminobutyric acid, 2-Hydroxybutyric acid, 2-Oxoglutaric acid, 3-Hydroxybutyric acid, 3-Hydroxyisobutyric acid, Acetic acid, Acetoacetic acid, Acetone, Alanine, Arginine, Asparagine, Aspartate, Betaine, Choline, Citric acid, Creatine, Creatinine, Cystine, D-Galactose, Dimethylamine, Dimethylsulfone, Ethanol, Formic acid, Glucose, Glutamic acid, Glutamine, Glycerol, Glycine, Histidine, Isoleucine, Lactic acid, Leucine, Lysine, Methanol, Methionine, Myo-inositol, N,N-Dimethylglycine, Ornithine, Phenylalanine, Proline, Pyruvic acid, Sarcosine, Serine, Succinic acid, Threonine, Trimethylamine-N-oxide, Tyrosine, Valine Albumin, Apolipoprotein B, Bilirrubin, Calcium, C reactive protein, Erythrocyte sedimentation rate, Erythrocytes, Estimated Glomerular Filtration Rate, Fructosamine, Glyc A, Glyc B, HDL cholesterol, Hemoglobin, Iron, LDL cholesterol, Leukocytes, Lipoprotein(a), Platelets, SPC, Total cholesterol, Total protein, Transferrine, Triglycerides, Urate, Urea Leveraging NMR-based metabolomics and integrated machine learning pipelines enables robust estimation of metabolic age and accurate classification of multiple disease states. The combination of raw spectral data with derived metabolite and clinical features enhances both model performance and interpretability. Distinct shifts in metabolic age and feature patterns across disease groups demonstrate the framework’s ability to capture physiological heterogeneity and early signs of metabolic dysregulation. These advances support the development of scalable, non-invasive tools for precision health monitoring and underscore the relevance of metabolic profiling in personalized medicine. aATLAS Molecular Pharma, Bizkaia Science and Technology Park, 48160 Derio, Spain bPrecision Medicine and Metabolism Laboratory, CIC bioGUNE, Bizkaia Science and Technology Park, 48160 Derio, Spain [email protected] Machine Learning ML 1H (ppm) 1D-NOESY 1H Fast 2D J-Resolved 1H Serum samples Metabolic age classification Disease targets input output Metabolite quantification Estimated clinical parameters Age Disease Clinical parameters Metadata Metabolite quantification from J-RES Intensity (a. u.) J-RES 1H (serine) Gaussian fit Clinical parameter prediction with ML Measured values Measured values Measured values Predicted values HDL cholesterol (mg/dl) R=0.97 Albumin (g/dl) R=0.94 ESR (mm/h) R=0.78 R=0.92 RMSE=7.2 years Chronological age Metabolic age 0.0±6.8 p-value= 3.2E-30 14.5±11.1 0.0±8.2 p-value= 1.0E-19 4.9±9.2 Liver diseases Prostate cancer Individual report (precision medicine) To showcase the potential for personalized diagnostics, we present an individuallevel report. This includes a spider plot displaying the probabilistic class predictions, modulated using a temperature scaling factor of 10 to provide a smoother representation of classification confidence (e.g., probability of 1.00 for prostate cancer in the example). Below, a corresponding SHAP value plot identifies the most influential features driving the model’s decision. Together, these outputs illustrate the capacity of the framework to support precision medicine applications based on individualized metabolic phenotyping. COVID-19 long COVID Observed labels Predicted labels COVID-19 long COVID COVID-19 long COVID COVID-19 long COVID