Plasma metabolomic signatures enable the diagnosis and prognosis of chronic obstructive pulmonary disease Authors: Chen Li, Jie Yao, Fan Wu, Leqing Zhu, Xiaowen Zhang, Lanzhu Peng, Jieda Cui, Su Fang, Li Lin, Wanli Peng, Miao Lei, Mingmeng Fang, Cuihua Wang, Bin Chen, Zongzheng Gui, Xianju Wang, Yubei Jin, Biying Zhu, Chunyan Lu, Qi Wan, Xiaohui Wu, Gaoying Tang, Xinru Ran, Wei Rao, Lei Wang, Yumin Zhou, Pixin Ran, Wenxiang Hu Published: 2026 ABSTRACT Chronic obstructive pulmonary disease (COPD) is a progressive heterogeneous lung disease driving global illness and death, yet reliable biomarkers for its early diagnosis, molecular subtyping and prognosis are lacking. Here, we conduct targeted plasma metabolomic profiling in two independent, deeply phenotyped cohorts comprising 1344 participants across the full spectrum of COPD severity, with 3-year longitudinal follow-up for 651 subjects. Machine learning integration screens metabolite signatures associated with disease presence, clinical subtypes, and longitudinal outcomes. We identify unique plasma metabolic profiles distinguish COPD patients from healthy people and separate emphysema and small-airway-dominant subtypes with high accuracy in test and validation cohorts. A 27-metabolite panel detects early COPD by capturing metabolic shifts prior to lung function loss. Moreover, specific metabolite subsets independently predicted lung function decline, occurrence of acute exacerbations, and dyspnoea severity. Together, plasma metabolomics yields robust multi-purpose COPD biomarkers, offering an accessible strategy for early screening and personalized clinical risk stratification. Introduction Chronic obstructive pulmonary disease (COPD) is a major global health challenge and is projected to become the third leading cause of death worldwide by 20301. It is a complex respiratory disorder characterized by persistent airflow limitation and symptoms such as dyspnea, chronic cough, and sputum production2. COPD encompasses a spectrum of clinical phenotypes and biological endotypes, ranging from emphysema-predominant to bronchitis-dominant subtypes, each with distinct pathophysiology, treatment responses, and risk of exacerbation2–4. This clinical and molecular heterogeneity poses significant challenges for diagnosis and management, highlighting the urgent need for precision medicine approaches that enable early detection, molecular stratification, and tailored therapeutic interventions. Timely identification of COPD at an early or even preclinical stage offers a critical opportunity to delay or prevent irreversible lung function decline. Indeed, our previous studies have shown that early therapeutic intervention in patients with mild COPD yields favorable outcomes5,6. However, identifying at-risk individuals before the onset of overt airflow obstruction remains a major unmet clinical need. Current golden diagnostic criteria that primarily based on spirometry-defined airflow limitation (FEV1/FVC <0.70), often fail to detect the disease until substantial structural and functional damage has occurred7,8. Despite being the gold standard, spirometry remains underutilized in routine clinical practice due to several barriers, including procedural complexity, the need for stringent quality control, lengthy examination times, and a lack of trained personnel and equipment in many healthcare settings9–11. As a result, screening often relies on suboptimal tools such as symptom questionnaires, peak expiratory flow measurements, and imaging modalities, which present challenges for the detection of early-stage disease12–14. Several protein-based biomarkers, including C-reactive protein (CRP), fibrinogen, and surfactant protein D, have been proposed for risk stratification and prognosis in COPD15–17. However, their limited sensitivity, specificity, and predictive value have hindered widespread clinical adoption. Recent studies have revealed metabolic reprogramming as a central feature of COPD pathogenesis, with alterations in amino acid, lipid, and redox metabolism observed in both the lungs and systemic circulation18–21. These metabolic shifts not only reflect underlying pathophysiological processes but may also serve as early indicators of disease activity and potential therapeutic targets. Metabolomics, a high-throughput approach for profiling small-molecule metabolites, has emerged as a promising tool for uncovering such signatures22–24. Initial studies applying metabolomics to COPD have suggested its utility for identifying candidate biomarkers predictive of disease development25–28. However, these investigations have been hampered by small cohorts, lack of replication in independent populations, and limited longitudinal follow-up. Consequently, large-scale metabolic profiling in well-characterized COPD populations remains scarce, and the potential of metabolite-based signatures for early detection, molecular subtyping, and risk stratification is far from fully realized. Meanwhile, rapid advances in computational science, particularly in machine learning and integrative modeling, are revolutionizing our ability to interpret complex biological data29,30. These approaches excel at capturing nonlinear relationships and high-dimensional interactions, making them uniquely suited for analyzing the intricate biochemical networks underlying chronic diseases such as COPD. When combined with high-throughput metabolomics, machine learning enables the systematic interrogation of thousands of metabolites simultaneously, facilitating the discovery of robust, generalizable biomarkers that outperform traditional univariate or hypothesis-driven methods23,31,32. Beyond classification, these integrative frameworks can uncover latent molecular patterns, define clinically relevant subtypes, and predict disease trajectories, offering unprecedented opportunities to move beyond one-size-fits-all management toward precision medicine33,34. In COPD, where pathophysiologic heterogeneity and unpredictable progression have long challenged clinical practice, the synergy between metabolomics and computational modeling holds particular promise for enabling early detection, stratifying risk, and identifying novel therapeutic targets. Here, we integrated targeted plasma metabolomics with machine learning-based modeling in two independent, well-characterized cohorts spanning the full spectrum of COPD severity, with longitudinal follow-up. This approach enabled the identification of reproducible metabolite signatures for early disease detection, molecular subtype classification, and outcome prediction. Collectively, our findings establish a clinically actionable metabolite-based framework for noninvasive diagnosis, precision stratification, and therapeutic target discovery in COPD. Results Study design, patient cohorts, and data collection To comprehensively characterize the circulating metabolic landscape of COPD, we leveraged two independent and deeply phenotyped cohorts with longitudinal follow-up (Fig. 1). The discovery cohort was drawn from the Early COPD (ECOPD) study (Chinese Clinical Trial Registry: ChiCTR1900024643) and included 787 participants (321 with COPD and 466 without COPD) recruited across multiple clinical centers between 2018 and 2024 (Supplementary Fig. 1a)35. Among these, 651 individuals (82.7%) completed a 3-year follow-up with repeated assessments of lung function, symptom scores, quality of life, and acute exacerbations. The external validation cohort comprised 557 participants (372 with COPD and 185 without COPD) from the Respiratory Health study enrolled in 2023, with baseline plasma samples and 1-year follow-up data for independent model validation (Supplementary Fig. 1i).Fig. 1Study design and analytical workflow for COPD diagnosis and prognosis.Targeted plasma metabolomics was performed to profile metabolic alterations associated with COPD. Machine learning models were trained using metabolite and clinical features to classify COPD stage, COPD subtype in the discovery cohort and subsequently validated in an independent cohort. Prognostic models were developed using longitudinal follow-up data to predict rapid lung function decline, acute exacerbations of COPD (AECOPD), and deterioration in health-related quality of life. Created in BioRender. Hu, W. (2026) https://BioRender.com/bkxso50. Study design and analytical workflow for COPD diagnosis and prognosis. Targeted plasma metabolomics was performed to profile metabolic alterations associated with COPD. Machine learning models were trained using metabolite and clinical features to classify COPD stage, COPD subtype in the discovery cohort and subsequently validated in an independent cohort. Prognostic models were developed using longitudinal follow-up data to predict rapid lung function decline, acute exacerbations of COPD (AECOPD), and deterioration in health-related quality of life. Created in BioRender. Hu, W. (2026) https://BioRender.com/bkxso50. COPD diagnosis and classification were established using the post-bronchodilator ratio of forced expiratory volume in one second to forced vital capacity (FEV1/FVC <0.70), in accordance with GOLD recommendations at the time of blood sample collection36. Disease severity was staged based on spirometric thresholds and symptom burden. Individuals at heightened risk without airflow limitation (early COPD and COPD-E) were identified by established criteria, including respiratory symptoms, imaging evidence of emphysema or airway wall thickening, and preserved spirometry37. Normal individuals were free of chronic respiratory disease and had normal lung function on spirometry. Demographic and clinical characteristics of these participants are summarized in Tables S1–S3. Consistent with previous reports, COPD stage correlated with older age, higher smoking index, occupational dust and biomass exposure, family history, and male predominance across both cohorts (Supplementary Fig. 1b–h, j–p). Plasma metabolomic profiles were obtained using targeted liquid chromatography-mass spectrometry (LC-MS), enabling precise quantification of 206 metabolites spanning multiple classes, including fatty acids, amino acids, organic acids, and carnitines (Supplementary Fig. 1q). The discovery cohort was used to identify differentially expressed metabolites and to investigate associations between metabolic signatures and clinical phenotypes using machine learning-based approaches. Diagnostic models for COPD stage and subtype were developed by integrating metabolite profiles with clinical parameters and were subsequently validated in the external cohort (Fig. 1). In addition, prognostic models were constructed to predict lung function decline, worsening quality of life, and risk of acute exacerbations over a 3-year follow-up (Fig. 1). Plasma metabolic reprogramming during COPD development and progression To investigate systemic metabolic reprogramming associated with COPD, we performed targeted plasma metabolomic profiling using the Q300 platform, which covers key metabolic pathways as previously described (Table S4)38. Comparative analyses were conducted between patients with COPD and non-COPD (NCOPD) individuals in the discovery cohort. Among quantified metabolites, 55 exhibited significant differential abundance between groups (false discovery rate, FDR <0.05) (Fig. 2a, b and Table S5). These metabolites spanned multiple biochemical classes, including amino acids and derivatives, fatty acids, organic acids, and intermediates of energy metabolism, underscoring widespread metabolic dysregulation in COPD (Fig. 2c). To explore the biochemical interrelationships among COPD-associated metabolites, we constructed a debiased sparse partial correlation (DSPC) network, which revealed three functional subnetworks (Fig. 2d). Two subnetworks were centered on glycolysis and the tricarboxylic acid (TCA) cycle, incorporating metabolites such as lactic acid, pyruvic acid, oxoglutaric acid (α-KG), and aspartic acid, underscoring disrupted central carbon metabolism as a hallmark of COPD pathogenesis. The largest subnetwork indicated extensive reprogramming of amino acid and lipid metabolism, including branched-chain amino acids (BCAAs) and ω−3/ω−6 polyunsaturated fatty acids (PUFAs) (Fig. 2d), which were previously implicated in inflammation and muscle wasting in COPD39–42. KEGG pathway enrichment analysis confirmed broad perturbations in metabolic pathways, with arginine biosynthesis emerging as the most enriched (Fig. 2e).Fig. 2Characteristic plasma metabolites associated with COPD.a Composition of the study population used for differential metabolite analysis, comprising 321 patients with COPD and 466 non-COPD (NCOPD) participants from the discovery cohort, including both healthy individuals and those with early-stage COPD without spirometric impairment (normal FEV1/FVC). Created in BioRender. Hu, W. (2026) https://BioRender.com/opgm3xh. b Volcano plot illustrating 55 significantly altered plasma metabolites in COPD patients compared to non-COPD individuals. Differential metabolites were identified using two-sided t-tests or Mann–Whitney U-tests, depending on distribution normality, followed by multiple testing correction using the Benjamini–Hochberg method. Metabolites with a false discovery rate (FDR) <0.05 were considered significant. c Heatmap showing relative abundance patterns of the 55 differential metabolites across disease states. Values represent z-scores of mean metabolite abundance within each disease state.d Debiased sparse partial correlation (DSPC) network constructed from the 55 differential metabolites. Each node represents a metabolite, and edges reflect partial correlation coefficients between pairs of metabolites. Edge thickness indicates the strength of correlation, and node colors denote distinct metabolite classes (e.g., amino acids, fatty acids, organic acids, carnitines). e Pathway enrichment analysis of the 55 differential metabolites using the Kyoto Encyclopedia of Genes and Genomes (KEGG) database. The reference background was defined as all metabolites quantified in the targeted metabolomics panel. The p values were calculated by a two-sided hypergeometric test. f Differential metabolites in the plasma of COPD patients were clustered into three groups through unsupervised clustering. Shaded areas represent the 95% confidence intervals. g Line plot illustrates the progressive abundance increase of metabolites within cluster 1 throughout disease progression. h Line plot illustrates the progressive abundance decrease of metabolites within cluster 2 throughout disease progression. i, j KEGG enrichment analysis of metabolites in cluster 1 (i) and cluster 2 (j). The reference background was defined as all metabolites quantified in the targeted metabolomics panel. The p values were calculated by a two-sided hypergeometric test.k Representative metabolites in cluster 1 (Glutamine) and cluster 2 (Erythronic acid). Statistical significance was assessed using one-way ANOVA. Normal: n = 267, COPD-E: n = 199, COPD-I: n = 122, COPD-II: n = 102, COPD-III: n = 83, COPD-IV: n = 14. Boxes represent the interquartile range (IQR), horizontal lines indicate the median, and whiskers denote 1.5× IQR. Source data are provided as a Source Data file. Characteristic plasma metabolites associated with COPD. a Composition of the study population used for differential metabolite analysis, comprising 321 patients with COPD and 466 non-COPD (NCOPD) participants from the discovery cohort, including both healthy individuals and those with early-stage COPD without spirometric impairment (normal FEV1/FVC). Created in BioRender. Hu, W. (2026) https://BioRender.com/opgm3xh. b Volcano plot illustrating 55 significantly altered plasma metabolites in COPD patients compared to non-COPD individuals. Differential metabolites were identified using two-sided t-tests or Mann–Whitney U-tests, depending on distribution normality, followed by multiple testing correction using the Benjamini–Hochberg method. Metabolites with a false discovery rate (FDR) <0.05 were considered significant. c Heatmap showing relative abundance patterns of the 55 differential metabolites across disease states. Values represent z-scores of mean metabolite abundance within each disease state.d Debiased sparse partial correlation (DSPC) network constructed from the 55 differential metabolites. Each node represents a metabolite, and edges reflect partial correlation coefficients between pairs of metabolites. Edge thickness indicates the strength of correlation, and node colors denote distinct metabolite classes (e.g., amino acids, fatty acids, organic acids, carnitines). e Pathway enrichment analysis of the 55 differential metabolites using the Kyoto Encyclopedia of Genes and Genomes (KEGG) database. The reference background was defined as all metabolites quantified in the targeted metabolomics panel. The p values were calculated by a two-sided hypergeometric test. f Differential metabolites in the plasma of COPD patients were clustered into three groups through unsupervised clustering. Shaded areas represent the 95% confidence intervals. g Line plot illustrates the progressive abundance increase of metabolites within cluster 1 throughout disease progression. h Line plot illustrates the progressive abundance decrease of metabolites within cluster 2 throughout disease progression. i, j KEGG enrichment analysis of metabolites in cluster 1 (i) and cluster 2 (j). The reference background was defined as all metabolites quantified in the targeted metabolomics panel. The p values were calculated by a two-sided hypergeometric test.k Representative metabolites in cluster 1 (Glutamine) and cluster 2 (Erythronic acid). Statistical significance was assessed using one-way ANOVA. Normal: n = 267, COPD-E: n = 199, COPD-I: n = 122, COPD-II: n = 102, COPD-III: n = 83, COPD-IV: n = 14. Boxes represent the interquartile range (IQR), horizontal lines indicate the median, and whiskers denote 1.5× IQR. Source data are provided as a Source Data file. To characterize dynamic metabolic changes during COPD progression, we performed unsupervised clustering, which revealed three distinct patterns of metabolite perturbation across disease stages (Clusters 1–3) (Fig. 2f). Cluster 1 metabolites, exemplified by glutamine and dimethylglycine, displayed a progressive increase from healthy individuals to advanced COPD, suggesting an adaptive or compensatory role in nitrogen and arginine metabolism (Fig. 2g, i, k and Supplementary Fig. 2c). Cluster 2 metabolites, including erythronic acid, creatine, and glyceric acid, declined markedly with disease onset and progression (Figs. 2h, j, k and Supplementary Fig. 2c), consistent with impaired fatty acid biosynthesis, energy homeostasis and disruption of the creatine-phosphate buffering system43,44. Cluster 3 metabolites were transiently elevated in early-stage COPD (COPD-E) but stabilized in moderate-to-severe disease (Supplementary Fig. 2a, b), suggesting early metabolic stress responses that normalize with chronic adaptation45. Metabolic features associated with COPD severity To delineate metabolomic signatures linked to COPD severity, we performed weighted gene co-expression network analysis (WGCNA) on all quantified metabolites in the discovery cohort. This approach identified modules of highly correlated metabolites, each representing coordinated biochemical changes associated with disease status and clinical measures. Hierarchical clustering yielded four distinct modules (Fig. 3a). Module 3, enriched in branched-chain amino acid (BCAA) metabolism, showed strong positive associations with smoking status, forced expiratory volume in one second (FEV1), and forced vital capacity (FVC) (Fig. 3b, c). These findings suggest that preservation of BCAA-related metabolic activity may be linked to better lung function in smokers and COPD patients39. In contrast, Module 4 was negatively correlated with the FEV1/FVC ratio and positively associated with COPD Assessment Test (CAT) score, modified Medical Research Council (mMRC) dyspnea score, and GOLD stage (Fig. 3b and Supplementary Fig. 3a), indicating that this metabolic cluster is strongly linked to symptom burden and disease severity. Pathway enrichment analysis revealed that Module 4 was enriched for the pentose phosphate pathway, tyrosine metabolism, and tryptophan metabolism.Fig. 3Correlation analysis between plasma metabolites and clinical parameters in COPD.a Weighted gene co-expression network analysis (WGCNA) for metabolite module identification in the discovery cohort. b Module-trait relationship heatmap for COPD-associated clinical parameters. c KEGG pathway enrichment analysis of module 3. The p values were calculated by a two-sided hypergeometric test. d Individual metabolite-clinical trait correlation heatmap. The p values were calculated using the two-sided Wald test. e Identification of FEV1/FVC-associated characteristic metabolites using multiple linear modeling approaches. f, g Random forest-based linear model performance across discovery (f) and validation (g) datasets based on characteristic metabolites in (e). Shaded areas represent the 95% confidence intervals. The p values were calculated by two-sided Pearson correlation analysis. h Linear regression fitting curves showing metabolite-FEV1/FVC associations in discovery and validation cohorts. Shaded areas represent the 95% confidence intervals. The p values were calculated by two-sided Pearson correlation analysis. Source data are provided as a Source Data file. Correlation analysis between plasma metabolites and clinical parameters in COPD. a Weighted gene co-expression network analysis (WGCNA) for metabolite module identification in the discovery cohort. b Module-trait relationship heatmap for COPD-associated clinical parameters. c KEGG pathway enrichment analysis of module 3. The p values were calculated by a two-sided hypergeometric test. d Individual metabolite-clinical trait correlation heatmap. The p values were calculated using the two-sided Wald test. e Identification of FEV1/FVC-associated characteristic metabolites using multiple linear modeling approaches. f, g Random forest-based linear model performance across discovery (f) and validation (g) datasets based on characteristic metabolites in (e). Shaded areas represent the 95% confidence intervals. The p values were calculated by two-sided Pearson correlation analysis. h Linear regression fitting curves showing metabolite-FEV1/FVC associations in discovery and validation cohorts. Shaded areas represent the 95% confidence intervals. The p values were calculated by two-sided Pearson correlation analysis. Source data are provided as a Source Data file. Because prior analyses revealed significant associations between demographic and clinical parameters such as sex, age, and smoking status with COPD (Supplementary Fig. 1b, j), we employed partial correlation analysis to account for these covariates. This analysis uncovered that lower levels of α-ketoglutaric acid (α-KG), erythronic acid, and glyceric acid were strongly associated with reduced FEV1/FVC (Fig. 3d), consistent with impaired glycolytic flux and disruption of the TCA cycle46. Notably, erythronic acid and glyceric acid were also significantly associated with additional COPD traits, including FEV1, mMRC scores, and GOLD stage (Fig. 3d and Supplementary Fig. 3b). To evaluate the predictive value of plasma metabolomic profiles for lung function impairment, we developed regression models for the FEV1/FVC ratio using metabolite features alone. Feature selection was performed by integrating results from multiple algorithms, including least absolute shrinkage and selection operator (Lasso), random forest, and extreme gradient boosting (XGBoost), yielding 15 consensus metabolite predictors (Fig. 3e). The model built on these 15 metabolites demonstrated strong correlations with measured FEV1/FVC values in both the discovery and validation cohorts (Fig. 3f, g). Notably, the most influential predictors included palmitoleic acid, citrulline, erythronic acid, and creatine (Fig. 3h and Supplementary Fig. 3c), which were previously implicated in lipid metabolism, nitrogen balance, redox homeostasis, and energy buffering associated with COPD pathogenesis41,44,47,48. Similarly, metabolite-based models were also developed to predict absolute FEV1 values in both the discovery and validation cohorts (Supplementary Fig. 3d–f). Together, these analyses highlight coordinated metabolic remodeling in association with COPD severity. Development of a plasma metabolite-based diagnostic model for COPD using machine learning To develop a robust diagnostic model for COPD using plasma metabolites, we implemented a machine learning pipeline optimized through a systematic feature selection and model evaluation workflow on the discovery cohort (Fig. 4a). Candidate metabolites were first identified using univariate analysis (FDR <0.05) and ranked by variable importance in projection (VIP) scores from orthogonal partial least squares-discriminant analysis (OPLS-DA) (Supplementary Fig. 4a). Model-based feature selection was then performed through repeated random sampling to ensure robustness and reproducibility across multiple algorithms, including XGBoost, Lasso, and Random Forest (Supplementary Fig. 4b). This approach yielded 12 key metabolites strongly associated with COPD status, encompassing intermediates of amino acid metabolism (e.g., erythronic acid, citrulline), lipid metabolism (e.g., palmitoleic acid), and energy metabolism (e.g., creatine) (Fig. 4b–d and Supplementary Fig. 4c).Fig. 4Machine learning-derived diagnostic models based on COPD-associated metabolites.a Schematic workflow for constructing diagnostic models using metabolomic features. b Venn diagram showing the overlap between metabolites identified by statistical testing and those selected by model-based feature ranking. c, d Box plots illustrating the distribution of selected metabolite levels in COPD versus non-COPD (NCOPD) groups in two independent cohorts. “Disc” denotes the discovery cohort; “Valid” denotes the validation cohort. Disc-NCOPD: n = 466, Disc-COPD: n = 321; Valid-NCOPD: n = 185, Valid-COPD: n = 372. Boxes represent the interquartile range (IQR), horizontal lines indicate the median, and whiskers denote 1.5× IQR. Statistical significance was assessed using two-sided t-tests or Mann–Whitney U-tests, depending on distribution normality. e Receiver operating characteristic (ROC) curves for a support vector machine (SVM) model with a radial basis function (RBF) kernel, trained using 12 metabolites alone or in combination with clinical parameters, in the discovery cohort test set. f ROC curves showing corresponding model performance in the independent validation cohort. g Bar plot ranking the relative importance of the 12 metabolites in the COPD diagnostic model. h–j Scatter plots illustrating the ability of the diagnostic models to distinguish COPD from non-COPD (NCOPD) individuals in the test and validation cohorts. A classification cutoff of 0.5 was applied. Source data are provided as a Source Data file. Machine learning-derived diagnostic models based on COPD-associated metabolites. a Schematic workflow for constructing diagnostic models using metabolomic features. b Venn diagram showing the overlap between metabolites identified by statistical testing and those selected by model-based feature ranking. c, d Box plots illustrating the distribution of selected metabolite levels in COPD versus non-COPD (NCOPD) groups in two independent cohorts. “Disc” denotes the discovery cohort; “Valid” denotes the validation cohort. Disc-NCOPD: n = 466, Disc-COPD: n = 321; Valid-NCOPD: n = 185, Valid-COPD: n = 372. Boxes represent the interquartile range (IQR), horizontal lines indicate the median, and whiskers denote 1.5× IQR. Statistical significance was assessed using two-sided t-tests or Mann–Whitney U-tests, depending on distribution normality. e Receiver operating characteristic (ROC) curves for a support vector machine (SVM) model with a radial basis function (RBF) kernel, trained using 12 metabolites alone or in combination with clinical parameters, in the discovery cohort test set. f ROC curves showing corresponding model performance in the independent validation cohort. g Bar plot ranking the relative importance of the 12 metabolites in the COPD diagnostic model. h–j Scatter plots illustrating the ability of the diagnostic models to distinguish COPD from non-COPD (NCOPD) individuals in the test and validation cohorts. A classification cutoff of 0.5 was applied. Source data are provided as a Source Data file. We next integrated these metabolites with core clinical covariates (age, sex, body mass index (BMI), and smoking status) to construct COPD diagnosis models. Ten machine learning algorithms, including logistic regression, random forest, gradient boosting, and support vector machines (SVMs), multilayer perceptron (MLP) neural networks, were trained in the discovery cohort using nested cross-validation (10 iterations of fivefold cross-validation). Model performance was assessed using area under the receiver operating characteristic curve (AUROC) and F1 score (Supplementary Fig. 4d). Among these, the Logistic Regression model demonstrated superior accuracy, achieving an AUROC of 0.774 (95% confidence interval (CI): 0.712–0.836) in the test set of discovery cohort 1 (Fig. 4e). External validation in cohort 2 yielded an AUROC of 0.751 (95% CI: 0.709–0.791) (Fig. 4f), reflecting good generalizability despite potential batch-related variability. Feature importance analysis identified erythronic acid and creatine, which decreased progressively with disease severity, as the top-ranked contributors (Fig. 4g). Notably, incorporating clinical parameters alongside metabolite features significantly improved classification accuracy, yielding an AUROC of 0.883 (95% CI: 0.839–0.928) in the test set of the discovery cohort and 0.813 (95% CI: 0.776–0.849) in the validation cohort (Fig. 4e, f). Calibration plots comparing predicted versus observed COPD status confirmed strong discriminatory performance and model calibration using a cutoff probability of 0.5, which correctly identified 75.0% of COPD patients in the test set and 70.1% in the validation cohort (Fig. 4h–j). Notably, in the above analysis, early-stage COPD (COPD-E) was grouped with non-COPD (NCOPD) individuals, as these subjects do not meet GOLD diagnostic criteria for established COPD. This classification strategy may attenuate effect sizes and reduce discrimination between cases and controls. To address this, we performed a sensitivity analysis using only normal individuals as controls (excluding COPD-E), which yielded consistent results (Supplementary Fig. 5a–f). Collectively, these results underscore the diagnostic utility of integrating metabolomics with clinical features and highlight specific metabolites as candidate biomarkers for COPD diagnosis. Metabolite-based diagnosis of early COPD Accurate identification of COPD-E remains a critical unmet clinical need, as our previous work and others suggest that timely intervention at this stage can significantly slow or prevent disease progression5,6. To investigate whether circulating metabolites could facilitate early detection, we applied OPLS-DA combined with lenient model-based feature selection, identifying 27 metabolites with reproducible discriminatory potential across multiple resampling iterations (Fig. 5a). A diagnostic model based on this 27-metabolite panel with clinical parameters demonstrated moderate discriminatory performance, achieving an AUROC of 0.750 (95% CI: 0.656–0.841) in the test set and 0.638 (95% CI: 0.559–0.721) in the validation cohort (Fig. 5b, c). Among these, creatine and glycocholic acid (GCA) ranked among the top predictors (Fig. 5d). Using a classification cutoff of 0.5, the model achieved 63.8% accuracy in identifying COPD-E cases in the validation cohort, a subgroup that is notoriously challenging to detect using spirometry alone37 (Supplementary Fig. 6a–c). These findings highlight the potential role of plasma metabolomics in complementing traditional lung function testing to enable earlier identification of high-risk individuals who may benefit from timely intervention.Fig. 5Precision diagnostic models for early COPD and COPD subtype classification.a Venn diagram showing the overlap between metabolites identified by OPLS-DA and those selected by model-based feature ranking for the COPD-E model. b, c ROC curve for the early COPD diagnostic model, trained using eight metabolites alone in the discovery cohort test set (B) and validation set (C). d Bar plot ranking the relative importance of the 27 metabolites in the COPD-E diagnostic model. e Volcano plot illustrating significantly altered plasma metabolites in patients with the small-airway disease (SAD) subtype compared with all other individuals. f Venn diagram showing the overlap between metabolites identified by statistical testing and those selected by model-based feature ranking for the SAD diagnostic model. g Bar plot ranking the relative importance of the 16 metabolites in the SAD diagnostic model. h, i ROC curves for the SAD diagnostic model built with 16 metabolites alone or in combination with clinical parameters, in the discovery cohort test set (h) and the independent validation cohort (i). j Volcano plot illustrating significantly altered plasma metabolites in patients with emphysema subtype compared with all other individuals. k KEGG pathway enrichment analysis of differential metabolites in the emphysema subtype compared with all other individuals. The reference background was defined as all metabolites quantified in the targeted metabolomics panel. The p values were calculated by a two-sided hypergeometric test. l Venn diagram showing the overlap between metabolites identified by statistical testing and those selected by model-based feature ranking for emphysema subtype diagnostic model. m Bar plot ranking the relative importance of the 14 metabolites in the emphysema diagnostic model. n, o ROC curves for the emphysema subtype diagnostic model built with 14 metabolites alone or in combination with clinical parameters, in the discovery cohort test set (n) and the independent validation cohort (o). Differential metabolites were identified using two-sided t-tests or Mann–Whitney U-tests, depending on distribution normality, followed by multiple testing correction using the Benjamini–Hochberg method. Metabolites with a false discovery rate (FDR) <0.05 were considered significant (e, j). Source data are provided as a Source Data file. Precision diagnostic models for early COPD and COPD subtype classification. a Venn diagram showing the overlap between metabolites identified by OPLS-DA and those selected by model-based feature ranking for the COPD-E model. b, c ROC curve for the early COPD diagnostic model, trained using eight metabolites alone in the discovery cohort test set (B) and validation set (C). d Bar plot ranking the relative importance of the 27 metabolites in the COPD-E diagnostic model. e Volcano plot illustrating significantly altered plasma metabolites in patients with the small-airway disease (SAD) subtype compared with all other individuals. f Venn diagram showing the overlap between metabolites identified by statistical testing and those selected by model-based feature ranking for the SAD diagnostic model. g Bar plot ranking the relative importance of the 16 metabolites in the SAD diagnostic model. h, i ROC curves for the SAD diagnostic model built with 16 metabolites alone or in combination with clinical parameters, in the discovery cohort test set (h) and the independent validation cohort (i). j Volcano plot illustrating significantly altered plasma metabolites in patients with emphysema subtype compared with all other individuals. k KEGG pathway enrichment analysis of differential metabolites in the emphysema subtype compared with all other individuals. The reference background was defined as all metabolites quantified in the targeted metabolomics panel. The p values were calculated by a two-sided hypergeometric test. l Venn diagram showing the overlap between metabolites identified by statistical testing and those selected by model-based feature ranking for emphysema subtype diagnostic model. m Bar plot ranking the relative importance of the 14 metabolites in the emphysema diagnostic model. n, o ROC curves for the emphysema subtype diagnostic model built with 14 metabolites alone or in combination with clinical parameters, in the discovery cohort test set (n) and the independent validation cohort (o). Differential metabolites were identified using two-sided t-tests or Mann–Whitney U-tests, depending on distribution normality, followed by multiple testing correction using the Benjamini–Hochberg method. Metabolites with a false discovery rate (FDR) <0.05 were considered significant (e, j). Source data are provided as a Source Data file. Metabolite-driven stratification of COPD subtypes COPD represents a clinically heterogeneous condition, comprising subtypes such as emphysema-predominant and small-airway disease (SAD)-dominant phenotypes, each associated with distinct pathophysiological mechanisms and clinical trajectories49. Accurate molecular stratification of these subtypes could facilitate precision medicine approaches. However, current classification relies largely on imaging and spirometry, which may fail to capture early or systemic changes50. In our cohort, subtype annotations were available and showed a moderate correlation with COPD severity grading, although notable discrepancies were observed (Supplementary Fig. 6d, e). Accordingly, we developed diagnostic models for emphysema and SAD subtypes. To assess whether plasma metabolomic signatures could discriminate between COPD subtypes, we performed classification modeling using features derived from differential metabolite analysis and model-based selection. In total, 28 metabolites were significantly altered in small-airway disease (SAD) subtype and 52 metabolites in emphysema-predominant COPD (Fig. 5e, j, k). Subtype-specific diagnostic models incorporating 16 or 14 metabolite features demonstrated strong performance in the discovery cohort test set, with AUROCs of 0.825 (95% CI, 0.757–0.891) for SAD subtype (Fig. 5f, h) and 0.841 (95% CI: 0.786–0.894) for emphysema subtype (Fig. 5l, n). External validation showed moderate performance, with AUROCs of 0.725 (95% CI: 0.675–0.776) for SAD subtype (Fig. 5i) and 0.649 (95% CI: 0.574–0.721) for emphysema subtype (Fig. 5o). Feature contribution analysis identified eicosapentaenoic acid (EPA), creatine and erythronic acid as major contributors for the SAD model (Fig. 5g), whereas asparagine, creatine and erythronic acid were the strongest discriminators for the emphysema model (Fig. 5m). Incorporating clinical variables (age, sex, BMI, smoking status) significantly improved performance. The optimized SAD model achieved AUROCs of 0.908 (95% CI: 0.864–0.946) in test set of discovery cohort and 0.849 (95% CI: 0.808–0.888) in validation cohort (Fig. 5h, i), whereas the emphysema model reached 0.935 (95% CI: 0.898–0.965) in test set of discovery cohort and 0.760 (95% CI: 0.691–0.823) in validation cohort (Fig. 5n, o). Using a classification cutoff of 0.5, the COPD-subtype models accurately assigned patients to their respective subtypes, demonstrating the potential for metabolic phenotyping of COPD subtype (Supplementary Fig. 6f–i). Metabolite-informed models deliver accurate COPD outcome predictions Accurate prognostic assessment is essential for personalized management of COPD, as it enables clinicians to identify high-risk individuals, implement early interventions, and optimize treatment strategies33. Current prognostic approaches rely heavily on spirometry and symptom scores, which often fail to capture the systemic changes that precede functional decline or exacerbation events51. In our longitudinal cohorts, most healthy individuals and those with COPD-E remained stable over time, with only a small proportion progressing to overt COPD (Supplementary Fig. 7a). We therefore focused subsequent analyses on clinically relevant outcomes among patients with established COPD, including lung function decline and deterioration in quality of life (Supplementary Fig. 7b, c). To predict lung function decline in COPD patients, we applied multiple feature selection algorithms, including Random Survival Forest, Cox Proportional Hazards, and Gradient Boosting Survival Analysis, to identify the most prognostic 11 metabolites (Fig. 6a). Prognostic models were trained using three-year longitudinal data from Cohort 1, with the primary endpoint defined as rapid lung function decline (ΔFEV1 ≥60 mL/year)5. The survival-based metabolite model demonstrated strong prognostic accuracy, achieving an AUROC of 0.768 (95% CI: 0.696–0.831) (Fig. 6b). High-risk patients, as identified by the model, exhibited substantially greater FEV1 decline over the 3-year follow-up (Fig. 6c, d). Kaplan–Meier analysis further confirmed effective stratification of patients into high- and low-risk groups, with significant differences in event-free survival for lung function preservation (Fig. 6e). In addition to the survival-based approach, we also developed classification models to predict lung function decline within a 3-year follow-up period using binary outcomes (decline vs. no decline) derived from the same longitudinal dataset. These models achieved AUROCs of 0.811 (95% CI: 0.729–0.888) in the test set using only four key metabolites, including hydroxyphenyllactic acid, creatine, glutarylcarnitine, and proline (Supplementary Fig. 7d–i). These metabolites implicate pathways related to amino acid turnover, energy homeostasis, and oxidative stress, consistent with mechanisms driving COPD progression43,44,47. Notably, integration with clinical parameters further enhanced predictive performance compared with metabolite-only or clinical-parameter-only models (Fig. 6b and Supplementary Fig. 7g).Fig. 6Disease progression prediction models based on metabolites.a Venn diagram illustrating overlapping metabolites identified by feature ranking for the lung function rapid decline prognostic model. b ROC curve for the lung function rapid decline prognostic model during the 3-year follow-up in the test set. c Scatter plot showing the classification of individuals into high- and low-risk groups for rapid lung function decline based on the predefined classification threshold. d Bar plot showing the predictive performance of the model for identifying patients at risk of rapid lung function decline. e Survival curves comparing high-risk and low-risk groups for rapid lung function decline. f Venn diagram showing overlapping metabolites for the prognostic model predicting CAT score deteriorates. g ROC curve for the model predicting CAT score >10 during the 3-year follow-up in the test set. h Scatter plot showing the classification of individuals into predicted CAT score >10 and ≤10 groups based on the predefined classification threshold. i Bar plot showing the model’s accuracy for identifying patients at risk of CAT score deterioration. j Survival curves comparing high-risk vs. low-risk populations for CAT score progression. k Venn diagram showing overlapping metabolites for the prognostic model predicting mMRC score deteriorates. l ROC curve for the model predicting mMRC score >2 during the 3-year follow-up in the test set. m Scatter plot showing the classification of individuals into predicted mMRC score >2 and ≤2 groups based on the predefined classification threshold. n Bar plot showing the model’s accuracy for identifying patients at risk of mMRC score deterioration. o Survival curves comparing high-risk vs. low-risk populations for mMRC score progression. p ROC curve for 1-year mMRC score outcomes in the independent validation cohort. The p values were calculated by two-sided log-rank test (e, j, o). Source data are provided as a Source Data file. Disease progression prediction models based on metabolites. a Venn diagram illustrating overlapping metabolites identified by feature ranking for the lung function rapid decline prognostic model. b ROC curve for the lung function rapid decline prognostic model during the 3-year follow-up in the test set. c Scatter plot showing the classification of individuals into high- and low-risk groups for rapid lung function decline based on the predefined classification threshold. d Bar plot showing the predictive performance of the model for identifying patients at risk of rapid lung function decline. e Survival curves comparing high-risk and low-risk groups for rapid lung function decline. f Venn diagram showing overlapping metabolites for the prognostic model predicting CAT score deteriorates. g ROC curve for the model predicting CAT score >10 during the 3-year follow-up in the test set. h Scatter plot showing the classification of individuals into predicted CAT score >10 and ≤10 groups based on the predefined classification threshold. i Bar plot showing the model’s accuracy for identifying patients at risk of CAT score deterioration. j Survival curves comparing high-risk vs. low-risk populations for CAT score progression. k Venn diagram showing overlapping metabolites for the prognostic model predicting mMRC score deteriorates. l ROC curve for the model predicting mMRC score >2 during the 3-year follow-up in the test set. m Scatter plot showing the classification of individuals into predicted mMRC score >2 and ≤2 groups based on the predefined classification threshold. n Bar plot showing the model’s accuracy for identifying patients at risk of mMRC score deterioration. o Survival curves comparing high-risk vs. low-risk populations for mMRC score progression. p ROC curve for 1-year mMRC score outcomes in the independent validation cohort. The p values were calculated by two-sided log-rank test (e, j, o). Source data are provided as a Source Data file. We next evaluated whether metabolite-based models could predict deterioration in health-related quality of life among COPD patients, using the COPD assessment test (CAT) score >10 and Modified Medical Research Council Dyspnea Scale (mMRC) score >2 as endpoints over a 3-year follow-up period. These models relied on only five features for CAT (Fig. 6f) and four features for mMRC (Fig. 6k), yet maintained robust predictive performance compared to the more complex lung function model. AUROCs were 0.880(95% CI, 0.804–0.950) for CAT prediction (Fig. 6g) and 0.779 (95% CI, 0.679–0.868) for mMRC prediction (Fig. 6l). Risk stratification effectively separated patients with and without quality-of-life deterioration (Figs. 6h–j, m–o). External validation using 1-year follow-up data of the validation cohort confirmed these findings that the mMRC model achieved an AUROC of 0.732 (95% CI, 0.605–0.845) (Fig. 6p). Finally, we assessed the ability to predict the occurrence of acute exacerbations of COPD (AECOPD) within this follow-up period by developing a classification model to forecast exacerbation risk within 3 years. Using 17 metabolite features (Supplementary Fig. 7j-k), the model achieved an AUROC of 0.775 (95% CI: 0.711–0.836) (Supplementary Fig. 7l) and correctly identified 69.4% of AECOPD cases at a 0.5 cutoff (Supplementary Fig. 7m). Incorporating clinical parameters did not improve survival models, and only sex conferred a marginal benefit. Discussion In this study, we performed large-scale, targeted plasma metabolomic profiling across two independent, deeply phenotyped COPD cohorts to identify robust metabolite signatures associated with disease presence, clinical subtype, and progression. Using integrative machine learning frameworks, we developed predictive models that accurately discriminated individuals with COPD from healthy controls, resolved clinically relevant subtypes, and stratified patients by risk of longitudinal decline. Importantly, a parsimonious panel of 27 metabolites demonstrated strong performance for early-stage detection. Our findings suggest that COPD-associated metabolites are not merely biomarkers but active modulators of key pathogenic processes, including epithelial remodeling and inflammatory responses, highlighting their potential as therapeutic targets. Comparative analysis of feature metabolites across all models revealed that creatine, glutarylcarnitine, and erythronic acid were most frequently selected (Supplementary Fig. 8a), suggesting their potential pathophysiological relevance to COPD. Reduced creatine levels are consistent with disruption of the creatine-phosphate energy buffering system, a process critical for maintaining cellular bioenergetics and skeletal muscle function, both of which are impaired in COPD52. Carnitine metabolism has been linked to improved exercise tolerance and inspiratory muscle strength, as well as reduced lactate production in COPD patients53. In addition, erythronic acid is closely associated with oxidative stress and redox imbalance, a hallmark of COPD pathogenesis47. Notably, creatine and erythronic acid were consistently retained in diagnostic models for both COPD and its clinically relevant subtypes (Supplementary Fig. 8b). Based on these findings, we constructed a parsimonious two-metabolite model comprising creatine and erythronic acid. This simplified model demonstrated predictive performance comparable to that of the full-feature model (Supplementary Fig. 8c–e). To further enhance model simplicity, we also evaluated a reduced set of clinical parameters, which similarly achieved robust diagnostic performance (Supplementary Fig. 8f, g). Collectively, these optimization strategies facilitate the potential clinical translation of metabolite-based signatures. Metabolic reprogramming reflects a convergence of local pulmonary injury and systemic alterations during COPD development. Circulating metabolites offer unique advantages as diagnostic biomarkers, as they can be rapidly assessed in a minimally invasive manner, with potential for integration into clinical workflows54–57. Previous studies have demonstrated the potential of metabolomics to predict COPD; however, these efforts have typically been limited by relatively small cohort sizes58,59. In the future, paired analyses of metabolites in lung tissue, bronchoalveolar lavage fluid, or sputum alongside plasma may provide deeper mechanistic insight and clarify the tissue origins of COPD-associated metabolic alterations. Our machine learning-based framework enabled integration of multidimensional clinical and omics data, allowing for the identification of complex and nonlinear relationships. While this study focused on metabolomics, future work incorporating multi-omics layers, including proteomics, transcriptomics, epigenomics, and genomics, will likely enhance disease modeling, subtype discovery, and therapeutic prediction, ultimately advancing the goal of personalized COPD management. Beyond diagnosis and prognosis, metabolite-based classifiers could eventually guide therapeutic decision-making by predicting drug responsiveness or identifying patients most likely to benefit from anti-inflammatory or antioxidant strategies, which needs further investigation. Furthermore, several of the metabolites identified here are amenable to pharmacological or nutritional modulation, opening the door to future interventional trials aimed at modifying metabolic pathways to halt or reverse COPD progression. Our findings demonstrated that lipid metabolism is dynamically remodeled during COPD development and progression. Several lipids, including palmitoleic acid and EPA, consistently emerged as robust diagnostic and prognostic markers, underscoring their biological and clinical relevance. Palmitoleic acid, a monounsaturated fatty acid with anti-inflammatory properties60, was elevated in COPD patients, suggesting a potential compensatory response to chronic inflammation and metabolic stress61. Conversely, EPA, an ω−3 polyunsaturated fatty acid, has well-established immunomodulatory and anti-inflammatory effects and has been shown to attenuate lung function decline in both experimental models and human studies62,63. The consistent contribution of these lipids to disease prediction highlights lipid metabolism as a central axis of COPD pathogenesis. In addition, we also observed dysregulation of the TCA cycle during COPD progression. The TCA cycle plays a pivotal role in cellular energy production and redox homeostasis, and its impairment has been implicated in mitochondrial dysfunction and oxidative stress, both key features of COPD pathology47,64. Altered levels of intermediates such as α-KG suggest disruption of central carbon metabolism, which may compromise bioenergetic capacity and exacerbate oxidative damage in the lung microenvironment. We also identified significant alterations in amino acid metabolism, including creatine, glutamine, and tryptophan. Elevated glutamine levels may reflect compensatory regulation of nitrogen balance and arginine metabolism, pathways closely linked to nitric oxide production and airway inflammation65. Dysregulated tryptophan metabolism, particularly through activation of the kynurenine pathway, has been implicated in immune tolerance and chronic inflammation66, suggesting a mechanistic link between amino acid remodeling and the persistent inflammatory environment in COPD. Several aspects of our methodology deserve emphasis. Firstly, the use of targeted metabolomics enabled precise quantification and facilitated biological interpretability; however, this approach inherently limits metabolome coverage. Untargeted metabolomics or integration with other omics modalities, such as transcriptomics, epigenomics, and proteomics, may provide complementary insights into upstream regulatory pathways and downstream functional consequences67. Previously reported protein biomarkers, including CRP, fibrinogen, lipocalin-1, IL-8 and MMP9, have been associated with COPD diagnosis68. In addition, PXDN and EFNA2 have been linked to emphysema-predominant phenotypes, while proteins such as sRAGE, IL-6 and CC16 are associated with both cross-sectional and longitudinal COPD outcomes68–70. However, protein biomarkers have shown limited consistency across cohorts. These observations underscore the potential complementary value of integrating metabolomic and protein biomarkers to improve predictive accuracy and enhance the robustness of COPD risk stratification. Secondly, the inclusion of two large, independent cohorts with harmonized phenotyping strengthens the robustness and generalizability of our findings. However, several feature metabolites exhibited discordant directions of effect between cohorts, which may reflect cohort-specific biological differences or potential batch effects across sample sets. Nonetheless, the relatively short follow-up period limits our ability to capture long-term trajectories or rare clinical events such as mortality and acute exacerbations. Notably, during the 3-year follow-up period, most control individuals remained clinically stable, with only a small proportion progressing to COPD. This likely reflects both the relatively short duration of follow-up and the typically slow progression of the disease. Future studies with longer follow-up and larger cohorts will be necessary to determine the utility of metabolite-based signatures in predicting incident COPD among initially unaffected individuals. Moreover, the current analysis was limited to cohorts of Chinese ancestry, and validation in populations of diverse ethnic backgrounds will be critical to ensure broader applicability. Finally, as this was a cross-sectional study, longitudinal profiling of the dynamic metabolic landscape during disease progression would offer deeper insight into the mechanisms driving COPD development and heterogeneity. In conclusion, our study establishes a metabolite-based framework for precision phenotyping in COPD. By integrating targeted metabolomics with machine learning in large, clinically annotated cohorts, we identified metabolite signatures that not only reflect disease biology but also hold translational promise for early detection, stratification, and therapeutic targeting. Future efforts should focus on expanding metabolic coverage, integrating multi-omics data, extending longitudinal follow-up, and evaluating the predictive value of these biomarkers in the context of treatment response and personalized care. Methods Ethics All participants provided written informed consent. The ECOPD study was approved by the Ethics Committee of the First Affiliated Hospital of Guangzhou Medical University (No. 2018-53), and the Respiratory Health study received approval from the Ethics Committee of Guangzhou National Laboratory (No. 2023-0002 and 2023-0009). Participant recruitment and clinical data collection This study utilized data from two prospective cohorts: the Early COPD (ECOPD) cohort and the Respiratory Health cohort. The ECOPD study was a community-based, prospective, observational cohort conducted across three Chinese cities (Guangzhou, Shaoguan, and Heyuan) in Guangdong Province35. All participants completed standardized questionnaires, spirometry, impulse oscillometry, and chest CT scans at baseline, undergoing annual spirometry and acute exacerbation assessments. The Respiratory Health study was a prospective cohort conducted in multi-centers in China, and the samples of this study are specifically from Qingyuan, Guangdong. All participants completed standardized questionnaires, spirometry, and chest CT scans at baseline. All participants provided written informed consent. The ECOPD study was approved by the Ethics Committee of the First Affiliated Hospital of Guangzhou Medical University (No. 2018-53), and the Respiratory Health study received approval from the Ethics Committee of Guangzhou National Laboratory (No. 2023-0002 and 2023-0009). Exclusion criteria included: (1) Respiratory infections or acute exacerbations within 4 weeks prior to screening; (2) History of lung resection surgery; (3) Significant pulmonary diseases other than asthma (e.g., lung cancer and active tuberculosis). The reporting of this study follows the STARD guidelines. Standardized respiratory epidemiology questionnaires were finished by well-trained staff to collect the following data: demographic characteristics, COPD risk factors, smoking history (including pack-years), exposure to biomass fuels or occupational dusts, family history of respiratory diseases, chronic respiratory symptoms, previous history of acute exacerbations, and respiratory medications. Well-trained staff performed spirometry using portable spirometers in strict compliance with the American Thoracic Society or European Respiratory Society ATS/ERS standards71. Plasma sample collection All metabolomic data were obtained from blood samples collected at baseline, at which time participants were classified as cases or controls. Participants were subsequently followed for 3 years to monitor COPD symptoms and lung function. Peripheral venous blood was collected from participants in both cohorts after an overnight fast. Samples were maintained at room temperature for at least 30 min, and all processing was completed within 8 h of collection. For each participant, one tube was collected in an inert separation gel procoagulant tube and one in an EDTA anticoagulation tube. Tubes designated for routine clinical testing were sent to the hospital laboratory for biochemical analyses. The remaining EDTA tubes were centrifuged at 850×g for 10 min, and plasma was separated and stored at −80 °C until metabolite extraction. Clinical diagnostic criteria and assessment protocols Radiological assessment Chest computed tomography (CT) imaging was performed using standardized acquisition protocols on 128-slice multidetector scanners. All participants underwent standardized breathing training prior to image acquisition to ensure optimal lung inflation and reproducibility72. Image interpretation was conducted independently by two board-certified radiologists blinded to all clinical and laboratory data, with discrepancies resolved through consensus review. Quantitative image analysis was performed using a validated, dedicated imaging platform. Emphysema was quantified as the percentage of inspiratory lung volume with attenuation values below −950 Hounsfield units (LAA-950), while small-airway disease was assessed using the percentage of expiratory lung volume with attenuation values below −856 Hounsfield units (LAA-856). Consistent with established radiological criteria, emphysema and small-airway disease were defined as LAA-950 > 0.5% on inspiratory scans and LAA-856 > 15% on expiratory scans, respectively. Definition of acute exacerbations of COPD (AECOPD) Acute exacerbations of chronic obstructive pulmonary disease (AECOPD) were defined in accordance with the global initiative for chronic obstructive lung disease (GOLD) criteria as an acute worsening of respiratory symptoms, including cough, sputum production, wheezing, dyspnea, chest tightness, necessitating hospitalization, or escalation of pharmacological therapy73. The frequency of AECOPD events was prospectively documented over a 3-year follow-up period. The metabolite-based diagnostic models were subsequently employed to stratify participants according to their predicted risk of experiencing AECOPD within this observation window. Definition of early-stage COPD Early-stage COPD was operationalized as the presence of individuals at elevated risk for disease progression who do not yet exhibit post-bronchodilator airflow limitation (FEV1/FVC ≥0.70). This phenotype was characterized by: (i) persistent respiratory symptoms attributable to chronic airway inflammation; (ii) radiological evidence of emphysematous changes or airway wall thickening on quantitative CT analysis; and (iii) preserved ratio impaired spirometry (PRISm), defined as FEV1 < 80% predicted with FEV1/FVC ≥0.7037. Assessment of lung function decline Longitudinal decline in pulmonary function was evaluated based on annualized changes in forced expiratory volume in 1 s (FEV1). Rapid lung function decline was defined as a mean annual decrease in FEV1 exceeding 0.06 L, a threshold derived from population-based cohort studies and recognized as clinically significant for disease progression monitoring. Modified Medical Research Council (mMRC) dyspnea scale The mMRC scale was utilized to grade the severity of breathlessness during daily activities: Grade 0: Dyspnea only during strenuous exercise Grade 1: Dyspnea when hurrying on level ground or walking up a slight incline Grade 2: Walks slower than contemporaries on level ground due to breathlessness, or requires cessation for breathlessness when walking at own pace Grade 3: Stops for breath after walking ~100 m or after a few minutes on level ground Grade 4: Too breathless to leave the house, or breathless during minimal activities such as dressing. COPD assessment test (CAT) The COPD assessment test (CAT) is a validated, patient-reported outcome measure comprising eight items assessing the multidimensional impact of COPD on health status, including cough, phlegm, chest tightness, breathlessness during activities, activity limitation, confidence, sleep, and energy levels. Scores are interpreted as follows: 0–10 (low impact), 11–20 (medium impact), 21–30 (high impact), and 31–40 (very high impact). Metabolite extraction Plasma samples were thawed on ice to minimize degradation. About 20 μL of plasma was added to a 96-well plate and processed using the Eppendorf epMotion Workstation (Eppendorf Inc., Hamburg, Germany). About 120 μL of ice-cold methanol with internal standards was added, followed by vigorous vortexing for 5 min. The plate was centrifuged at 4000×g for 30 min (Allegra X-15R, Beckman Coulter, Inc., Indianapolis, IN, USA). About 30 μL of supernatant was transferred to a new plate, and 20 μL of derivative reagents were added to each well. Derivatization occurred at 30 °C for 60 min. Subsequently, 330 μL of ice-cold 50% methanol was added to dilute the sample, followed by storage at −20 °C for 20 min and centrifugation at 4 °C for 30 min at 4000×g. About 135 μL of supernatant was transferred to a new plate with 10 μL internal standards. Serial dilutions of derivatized stock standards were added to designated wells. The plate was then sealed for LC-MS analysis. Targeted metabolomics analysis We conducted metabonomics analysis using the Q300 Kit (MetaboProfile, Shanghai, China)38. The targeted panel was developed by Metabo-Profile Biotechnology Co., Ltd. and is based on a 3-nitrophenylhydrazine (3-NPH) derivatization platform 2, enabling simultaneous quantification of chemically diverse metabolites across a wide polarity range in a single LC-MS/MS run. The panel was designed to cover key metabolic pathways, including glycolysis, the tricarboxylic acid cycle, and fatty acid metabolism, with enhanced coverage of microbiota-related metabolites such as bile acids, fatty acids, and short-chain fatty acids. Standards were sourced from Sigma-Aldrich (St. Louis, MO, USA), Steraloids Inc. (Newport, RI, USA), and TRC Chemicals (Toronto, ON, Canada). Stock solutions (50 mg/mL) were prepared in water, methanol, sodium hydroxide, or hydrochloric acid. Calibration solutions were created by mixing appropriate amounts of each stock solution. An ACQUITY UPLC-Xevo TQS system (Waters Corp., Milford, MA, USA) was used for quantification by Metabo-Profile Biotechnology (Shanghai). HPLC settings: ACQUITY HPLC BEH C18 columns (2.1 × 5 mm precolumn, 2.1 × 100 mm analytical, 1.7 μm), column temp: 40 °C, sample temp: 10 °C, mobile phases: A = water with 0.1% formic acid, B = acetonitrile/IPA (70:30), gradient: 0–1 min (5% B), 1–11 min (5–78% B), 11–13.5 min (78–95% B), 13.5–14 min (95–100% B), 14–16 min (100% B), 16–16.1 min (100-5% B), 16.1–18 min (5% B), flow rate: 0.40 mL/min, injection: 5.0 μL. Mass spectrometer: capillary: 1.5 kV (ESI+), 2.0 kV (ESI−), source temp: 150 °C, desolvation temp: 550 °C, desolvation gas: 1000 L/h. Data preprocessing The raw data files generated by UPLC-MS/MS were processed using the TMBQ software (v1.0, Metabo-Profile, Shanghai, China) to perform peak integration, calibration, and quantitation for each metabolite as previously described38. Missing values were imputed by substituting one-tenth of the minimum detected value for each respective metabolite. This strategy was employed to mitigate potential biases introduced by zero values in downstream statistical analyses, particularly during log-transformation and multivariate modeling. The self-developed platform iMAP (v1.0, Metabo-Profile, Shanghai, China) was used for statistical analyses, including OPLS-DA, univariate analysis and pathway analysis, et al. Mass spectrometry-based quantitative metabolomics refers to the determination of the concentration of a substance in an unknown sample by comparing the unknown to a set of standard samples of known concentration (i.e., calibration curve). The calibration curve is a plot of how the analytical signal changes with the concentration of the analyte (the substance to be measured). For most analyses, a plot of instrument response vs. concentration will show a linear relationship. This yields a model described by the equation \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$y={ax}+b$$\end{document}y=ax+b, where y is the instrument response, e.g., peak height or area, a represents the slope/sensitivity, and b is a constant that describes the background. The analyte concentration (x) of unknown samples may be calculated from this equation. Prior to model construction and performance evaluation, batch effects across different datasets were corrected using the PyComBat algorithm, with principal component analysis (PCA) serving as the diagnostic metric. For data preprocessing, a RobustScaler was applied based on the discovery cohort to ensure consistent scaling during modeling. In the analysis of unified batches, global mean and variance were calculated to perform uniform standardization across both the discovery and validation cohorts. Samples with missing clinical indicators were excluded from the joint modeling of metabolites and clinical parameters. WGCNA analysis Metabolite abundances were normalized and analyzed using weighted gene co-expression network analysis (WGCNA v1.73) implemented in the Bioconductor R package (v4.4.1)74. Pairwise similarity between metabolites was calculated using Pearson’s correlation coefficients, and the resulting similarity matrix was transformed into an adjacency matrix. A topological overlap matrix (TOM) and its corresponding dissimilarity measure (1-TOM) were then computed. Metabolite modules were identified using the dynamic tree cut (DTC) algorithm, with a cut height of 0.25. Correlation analysis Spearman correlation coefficients were calculated to assess associations between individual metabolites and COPD-related clinical parameters, including lung function indices (FEV1, FVC, and FEV1/FVC), COPD Assessment Test (CAT) score, modified Medical Research Council (mMRC) dyspnea score, and GOLD stage. To account for potential confounding effects, partial correlation analysis was performed using the Python package pingouin (v0.5.5), adjusting for age, sex, body mass index (BMI), and smoking status. Metabolic differential analysis and pathway enrichment analysis To identify differentially abundant metabolites for downstream feature selection, we performed both univariate and multivariate analyses using data from participants in the discovery cohort. Univariate analysis Metabolite distributions were first assessed for normality using the Kolmogorov–Smirnov test. For normally distributed metabolites, group comparisons were conducted using two-sided t-tests; otherwise, the Mann–Whitney U-test was applied. To correct for multiple hypothesis testing, p values were adjusted using the Benjamini–Hochberg method to control the false discovery rate (FDR). Metabolites with FDR <0.05 were considered significantly differentially abundant. Multivariate analysis Orthogonal partial least squares-discriminant analysis (OPLS-DA) was conducted using the pypls package (v1.0.1), incorporating tenfold cross-validation. Metabolites with a variable importance in projection (VIP) score >1 were defined as discriminative features. Pathway enrichment and network analysis To contextualize the differential metabolites, Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment was defined as all metabolites in targeted metabolomics as background and debiased sparse partial correlation (DSPC) network analyses were performed using MetaboAnalystR v6.0 (https://www.metaboanalyst.ca)63. Diagnostic prediction model In addition to statistical analyses, model-based feature ranking was employed for feature selection in the discovery cohort. Three machine learning algorithms, including least absolute shrinkage and selection operator (Lasso), Random Forest, and extreme gradient boosting (XGBoost), were implemented using scikit-learn (v1.7), XGBoost (v3.0.3), and LightGBM (v4.6.0). Feature selection involved 10,000 random sampling iterations, with each iteration utilizing two-thirds of the samples for training. Feature importance was recorded per iteration, and features with a stability frequency >50% were deemed robustly significant. The intersection of features identified across these methods constituted the final feature set for model training. To implement a more inclusive feature selection strategy for the COPD-E diagnostic prediction model, a relaxed stability threshold was applied, retaining all variables with a stability frequency exceeding 40% to preserve a broader set of candidate predictors. We evaluated ten classification algorithms: Random Forest, support vector machine (SVM) with linear, radial basis function (RBF), and polynomial kernels, XGBoost, LightGBM, logistic regression, K-nearest neighbors (KNN), naïve Bayes, and a multilayer perceptron (MLP) neural network. Models were trained using ten iterations of stratified fivefold cross-validation by GOLD stage, with hyperparameter tuning performed during each round. Performance metrics included accuracy, precision, recall, F1 score, and area under the receiver operating characteristic curve (AUROC). The best-performing model, based on average cross-validation performance, was selected for downstream analysis. Final model validation was conducted in an independent cohort. To incorporate clinical parameters, we employed the same model-based feature ranking approach. Characteristic metabolites were first identified for model construction, and these were then integrated with all available clinical variables. Stable clinical features were selected using the ranking methodology and combined with metabolite features to train the final optimized predictive model. Feature importance was calculated using a hierarchical approach customized to the model type: for tree-based models (Random Forest, XGBoost, and LightGBM), built-in feature_importances_ attributes were employed, based on impurity reduction; for linear models (logistic regression and linear SVM), the absolute values of coefficients were used; and for all remaining models, permutation importance was applied. The latter was computed on the optimal validation fold by quantifying the mean decrease in AUROC following random shuffling of each feature (ten repetitions). Prognostic model For prognostic classification model development, we adapted the diagnostic model training pipeline. The models for CAT and mMRC were designed to predict longitudinal deterioration. Specifically, the models estimate the likelihood that patients develop a CAT score >10 or an mMRC score >2 during the 3-year follow-up period. In the acute exacerbation of COPD (AECOPD) analysis, patients were stratified into two groups: those who experienced at least one exacerbation during the 3-year follow-up and those who did not. For models of lung function decline, patients were classified according to whether their average annual decline in FEV1 over 3 years exceeded the predefined threshold for rapid deterioration. For survival modeling, feature selection was performed exclusively using model-based ranking approaches, including random survival forest, Cox proportional hazards (CoxPH) regression, and gradient boosting survival analysis. Survival models were constructed using random survival forest, CoxPH regression, gradient boosting survival analysis, and fast survival support vector machine (SVM), implemented in scikit-survival (v0.24.1). Model performance was evaluated using the concordance index (C-index). Risk scores generated by each model were dichotomized at the median to stratify patients into high- and low-risk groups. Clinical parameters were also evaluated using the same model-based feature ranking approach. However, subsequent testing demonstrated that inclusion of clinical variables provided minimal improvement in prognostic model performance. Statistical analysis The analysis and modeling evaluation methods are described in the Results, figure legends, and corresponding “Methods” subsections. Specifically, the Shapiro–Wilk normality test was applied prior to all comparisons to determine whether parametric or non-parametric tests should be used. For comparisons between two groups with unpaired samples, two-sided t-tests or Mann–Whitney U-tests were employed. For comparisons among three or more groups, two-sided analysis of variance (ANOVA) or Kruskal–Wallis H-tests were utilized. p values were adjusted using the Benjamini–Hochberg (BH) method to control for False Discovery Rate (FDR). A p value or FDR less than 0.05 was considered statistically significant. All statistical tests were conducted using Python (v3.12). Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Reporting Summary Transparent Peer Review file Supplementary Information Reporting Summary Transparent Peer Review file Source data Source Data Source Data