Development and multi-center validation of machine learning models based on targeted metabolomics for rheumatoid arthritis

Abstract

Background

Rheumatoid arthritis (RA) remains in urgent need of more effective biomarkers to improve diagnostic accuracy.

Methods

In this study, we conducted a comprehensive analysis of 2,863 blood samples obtained from seven cohorts comprising RA, osteoarthritis (OA), and healthy control (HC) subjects, recruited across five medical centers spanning three geographically diverse regions. Candidate biomarkers were first identified through untargeted metabolomic profiling, and subsequently validated using targeted approaches. Metabolite-based classification models were then developed employing a range of machine learning algorithms.

Results

Six metabolites were ultimately identified as promising diagnostic biomarkers, including imidazoleacetic acid, ergothioneine, N-acetyl-L-methionine, 2-keto-3-deoxy-D-gluconic acid, 1-methylnicotinamide and dehydroepiandrosterone sulfate. Based on these metabolites, we constructed classification models to differentiate RA from both HC and OA groups, and evaluated their performance across multiple independent validation cohorts. In three geographically distinct cohorts, RA vs. HC classifiers demonstrated robust discriminatory power, with an area under the receiver operating characteristic curve (AUC) ranging from 0.8375 to 0.9280, while RA vs. OA classifiers achieved moderate to good accuracy (AUC range: 0.7340–0.8181). Importantly, analysis of the seronegative RA subgroup indicated that the classifier’s performance was independent of serological status. Furthermore, validations conducted across different sample types and analytical platforms confirmed the reproducibility and stability of the models.

Conclusions

Taken together, these findings highlight the utility of metabolomics as a complementary approach for improving RA diagnosis and establish a broadly applicable framework for the development of metabolite-based classifiers across diverse and clinically heterogeneous disease contexts.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12967-025-07265-w.

Introduction

Rheumatoid arthritis (RA) is a chronic autoimmune disease that primarily affects multiple joints, with an estimated prevalence of approximately 0.28% in China [1]. It is characterized by symmetrical joint pain, swelling, and stiffness, which can progressively lead to joint destruction, functional impairment, and permanent deformity [2]. Although immunosuppressive therapies help alleviate symptoms and attenuate disease progression, a curative treatment remains unavailable, highlighting the critical need for early and accurate diagnosis to prevent irreversible joint damage and improve long-term outcomes [3–5]. Clinically, RA diagnosis is based on clinical symptoms, imaging findings, and serological markers. According to the 2010 classification criteria established by the American College of Rheumatology and the European League Against Rheumatism (ACR/EULAR) [6], rheumatoid factor (RF) and anti-cyclic citrullinated peptide (anti-CCP) antibodies are the primary serological indicators. When both markers are negative, a condition referred to as seronegative RA, the criteria require the involvement of more than ten joints to establish a definitive diagnosis. This reliance on serological indicators may hinder the early identification of patients with seronegative RA, resulting in delayed diagnosis, postponed treatment initiation, and consequently, accelerated disease progression and irreversible joint damage [7]. Therefore, there is an urgent need to develop novel, reliable, and accessible biomarkers that can improve diagnostic precision and facilitate timely intervention, particularly in patients lacking conventional serological evidence.

In the field of biomarker discovery, multi-omics approaches using blood-derived samples such as serum, plasma, and peripheral blood mononuclear cells have gained increasing prominence in biomarker discovery. Genomic, transcriptomic, and proteomic analyses are widely employed to identify disease-associated molecular signatures [8–10]. In parallel, metabolomics, a multidisciplinary field focused on the comprehensive profiling of small-molecule metabolites in biological systems, has gained prominence for its capacity to directly reflect biochemical changes under both physiological and pathological conditions [11–13]. The metabolome comprises the full spectrum of low-molecular-weight metabolites generated through cellular metabolic processes, including endogenous compounds, exogenous xenobiotics, and microbiota-derived products. As downstream products of genomic processes, metabolites integrate genetic and environmental factors, offering a dynamic snapshot of cellular function. Systematic metabolomic profiling enables the dissection of complex gene–environment–metabolism interactions and supports the identification of key disease-associated pathways and molecular signatures [14–16].

In recent years, liquid chromatography–tandem mass spectrometry (LC-MS/MS) has become the cornerstone of metabolomics research due to its high sensitivity, broad dynamic range, and extensive metabolite coverage. Metabolomic analyses based on this platform are typically classified as either untargeted or targeted. Untargeted metabolomics aims to detect a broad range of metabolites without prior selection, enabling the discovery of novel compounds, biomarkers, and metabolic pathways. In contrast, targeted metabolomics focuses on the quantitative analysis of predefined metabolites, offering high specificity and suitability for absolute quantification and biomarker validation [11, 17]. Among these approaches, untargeted metabolomics–driven global metabolomic profiling has been instrumental in identifying disease-specific metabolic signatures and facilitating biomarker discovery for diagnostic and prognostic applications [18–21]. However, translating untargeted metabolomics into clinical diagnostics remains challenging due to insufficient quantification accuracy, lack of standardized controls, limited cross-platform reproducibility, high costs, time-consuming procedures, and technical complexity. In contrast, targeted metabolomics employs chemical standards and stable isotope-labeled internal standards to achieve precise and reproducible absolute quantification, thereby enhancing its suitability for clinical implementation [17, 22]. Notably, several targeted metabolomics-based biomarker assays have been implemented in clinical laboratories for the diagnosis and monitoring of specific diseases. Therefore, integrating untargeted and targeted metabolomics offers a practical framework for diagnostic model development, involving initial screening and candidate identification through untargeted metabolomics, followed by quantitative validation of selected metabolites using targeted assays, and culminating in external validation across large-scale, multi-center cohorts to ensure the robustness and generalizability of the models, ultimately facilitating successful clinical translation.

In this study, following the proposed research framework, we conducted metabolomic profiling on 2,863 samples, primarily plasma with a small proportion of serum, from patients with RA, osteoarthritis (OA), and healthy controls (HC). Comparative analyses identified a panel of plasma metabolites capable of distinguishing RA from both OA and HC. Based on these metabolites, we developed and validated classification models with the potential to substantially improve the clinical diagnostic accuracy of RA.

Methods

Study cohorts

A total of seven cohorts were enrolled in this study, including one exploratory cohort, one discovery cohort, and five independent validation cohorts. The exploratory cohort consisted of 30 patients with RA, 30 with OA, and 30 healthy individuals. The discovery cohort included 450 RA, 450 OA, and 450 HC participants. Validation cohort 1 comprised 106 RA, 102 OA, and 106 HC participants. These three cohorts were all recruited from The First Affiliated Hospital of Fujian Medical University, Fuzhou, China. Validation cohort 2 included 62 RA, 67 OA, and 62 HC participants and was recruited from The First Hospital of Lanzhou University, Lanzhou, China. Validation cohort 3 consisted of 108 RA, 98 OA, and 108 HC participants from Guanghua Hospital Affiliated to Shanghai University of Traditional Chinese Medicine, Shanghai, China. Validation cohort 4 comprised 82 RA, 77 OA, and 82 HC participants from Xinhua Hospital Affiliated to Shanghai Jiao Tong University School of Medicine, Shanghai, China. Validation cohort 5 included 121 RA, 91 OA, and 151 HC participants and was recruited from Tongren Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China. RA patients were diagnosed based on the 2010 ACR/EULAR classification criteria [6]. OA patients fulfilled the 1987 ACR clinical guidelines for OA diagnosis [23]. HC participants were recruited during routine physical examinations, with medical records confirming no clinical evidence of disease at the time of enrollment. Participants were excluded if they were pregnant at the time of recruitment or had a history of malignancies, major organ dysfunction, psychiatric disorders, or autoimmune diseases other than RA. For RA patients, only those with RA and no coexisting autoimmune conditions were included. In each cohort, OA and HC participants were matched to RA patients by age and sex as closely as possible. Because p values are highly sensitive to sample size, we did not use them to judge matching adequacy. Instead, at enrollment in each independent cohort we applied pre-specified pragmatic thresholds: a median age difference versus RA within ±3 years and a male proportion difference within ±5% points. The actual differences for each cohort are reported in Supplementary Table 1.

Sample collection

For plasma collection, venous blood was drawn into EDTA-coated tubes, while serum samples were collected using clot-activator serum separator tubes. All samples were processed promptly and stored at −80 °C or in liquid nitrogen until analysis. Sample collection and storage were standardized and identical across all cohorts and platforms. Plasma was used in all cohorts except for validation cohort 4, where serum samples were analyzed. Clinical information, including anti-CCP, RF, C-reactive protein (CRP), and erythrocyte sedimentation rate (ESR), was collected as completely as possible. All assays were performed using clinically approved methods and commercially available diagnostic kits. For RA and OA participants with partially missing clinical data, additional testing was performed using leftover serum samples from routine clinical examinations, which were collected concurrently with the corresponding plasma specimens. The clinical characteristics of the study population are summarized in Supplementary Table 1.

Untargeted metabolomics

Sample extraction and preparation

Each biological sample (50 μL) was mixed with 200 μL of prechilled extraction solvent consisting of methanol and acetonitrile (1:1, v/v) containing deuterated internal standards. The mixture was vortexed for 30 seconds, followed by sonication in a 4 °C water bath for 10 minutes. To precipitate proteins, samples were incubated at −40 °C for 1 hour and subsequently centrifuged at 12,000 rpm (13,800 × g, rotor radius = 8.6 cm) for 15 minutes at 4 °C. The resulting supernatants were carefully transferred into glass autosampler vials for subsequent analysis. Quality control (QC) samples (n = 10) were prepared by pooling equal aliquots from all individual specimens.

LC-MS/MS analysis

Polar metabolites were separated using a Vanquish UHPLC system (Thermo Fisher Scientific) equipped with a Waters ACQUITY BEH Amide column (2.1 mm × 50 mm, 1.7 μm). The mobile phases consisted of 25 mmol/L ammonium acetate and 25 mmol/L ammonium hydroxide in water (pH 9.75) as phase A, and acetonitrile as phase B. The autosampler was kept at 4 °C, and each run involved a 2 μL injection. The UHPLC system was interfaced with an Orbitrap Exploris 120 mass spectrometer (Thermo Fisher Scientific), operated in both positive and negative electrospray ionization (ESI) modes. Data were acquired in an information-dependent MS/MS mode using Xcalibur software (version 4.4). Instrument parameters were set as follows: sheath gas flow rate at 50 arbitrary units, auxiliary gas at 15, capillary temperature at 320 °C, full MS resolution of 60,000, MS/MS resolution of 15,000, stepped normalized collision energies (NCE) of 20, 30, and 40, and spray voltages of +3.8 kV (positive mode) and −3.4kV (negative mode).

Data processing and quality control

Raw LC-MS/MS data were initially converted to mzXML format using ProteoWizard, and subsequently processed through a custom R-based pipeline (version 4.4.3) built on the XCMS package (version 4.6.0). This workflow facilitated comprehensive peak picking, retention time alignment, and signal integration. A total of 14,746 metabolite features were detected across all experimental and quality control (QC) samples. To minimize technical variation and enhance biological interpretability, a multistep preprocessing strategy was applied. Features with high relative standard deviation (RSD) across QC replicates were excluded to eliminate instrumental noise. Peaks with more than 50% missing values within any group or across all samples were also removed. For the remaining data, missing values were imputed using one-half of the minimum non-zero value. Subsequently, all peak intensities were normalized to the internal standard to correct for inter-sample variability. After quality filtering and normalization, 13,110 metabolite features were retained for downstream analysis. Metabolite identification was performed by matching accurate mass and MS/MS fragmentation patterns to reference spectra in BiotreeDB (version 3.0) [24].

Targeted metabolomics

Targeted LC-MS/MS analysis on the primary platform

Plasma or serum samples (100 µL) were extracted using 300 µL of ice-cold acetonitrile containing stable isotope-labeled internal standards. Following vortexing and centrifugation at 13,800 × g for 5 minutes at 4 °C, the supernatants were collected for LC-MS/MS analysis. Metabolites were quantified using a Shimadzu LCMS-8050CL triple quadrupole mass spectrometer equipped with an electrospray ionization (ESI) source operated in multiple reaction monitoring (MRM) mode. Chromatographic separation was performed on a Phenomenex Kinetex C18 column (2.1 × 100 mm, 2.6 µm) using 0.1 mM ammonium fluoride in water and methanol as mobile phases. The column temperature was maintained at 40 °C, and other chromatographic parameters followed standard conditions for the Phenomenex Kinetex C18 column. MRM transitions were individually optimized for each target metabolite, using the most sensitive transition for quantification and additional transitions for structural confirmation. Calibration curves were generated by serial dilution of standard mixtures and fitted using 1/x weighted linear regression based on analyte-to-internal standard peak area ratios. Analytical performance metrics, including linearity (r ≥ 0.99), carryover (≤20% of the lower limit of quantification), and precision (coefficient of variation ≤ 15%), were validated for all metabolites.

Targeted LC-MS/MS analysis on the cross-platform

Plasma samples from validation cohort 5 (95 µL) were extracted using 380 µL of a pre-chilled acetonitrile:methanol mixture containing stable isotope-labeled internal standards. Samples were vortexed, sonicated, incubated, and centrifuged at 13,800 × g for 15 minutes at 4 °C. The resulting supernatants were analyzed using an Agilent 1290 Infinity II UHPLC system coupled to an Agilent 6460 triple quadrupole mass spectrometer operated in MRM mode. Chromatographic separation was conducted on a Waters ACQUITY BEH Amide column (2.1 × 100 mm, 1.7 µm), using a mobile phase composed of 10 mM ammonium formate and 0.1% formic acid in water and acetonitrile. The column temperature was maintained at 35 °C, and other chromatographic parameters followed standard conditions for ACQUITY BEH Amide columns. MRM transitions were optimized for each metabolite, with the most sensitive transition used for quantification and others for confirmation. Calibration curves were generated by serial dilution of standard mixtures and fitted using 1/x weighted linear regression based on analyte-to-internal standard peak area ratios. Analytical validation metrics, including linearity, carryover, and precision, were consistent with those of the analysis on the primary platform to ensure cross-platform comparability.

Bioinformatics and statistical analysis

All statistical analyses and data visualizations were performed using R software (version 4.4.3), unless otherwise specified. Following metabolite annotation, untargeted metabolomics data were log-transformed and scaled to reduce technical noise and variable variance prior to downstream analyses. Principal component analysis (PCA) and orthogonal partial least squares discriminant analysis (OPLS-DA) were conducted using the PCAtools (version 2.16.0) and ropls (version 1.36.0) packages, respectively. Differential metabolites were identified using Student’s t-test. Metabolites clearly identified as exposure-specific were manually excluded prior to downstream analysis, based on reference annotations from the Human Metabolome Database (HMDB; http://www.hmdb.ca). Volcano plots were generated using the ggplot2 package (version 3.5.1), and metabolite set enrichment analysis (MSEA) was performed via MetaboAnalyst 6.0. Upset plots were created using the UpSetR package (version 1.4.0).

For targeted metabolomics data, group comparisons were performed using the Wilcoxon rank-sum test. Prior to model training, all quantified metabolite values were log-transformed to reduce variance across variables. Five machine learning algorithms were applied, including logistic regression, Least Absolute Shrinkage and Selection Operator (LASSO) regression, random forest, support vector machine (SVM), and eXtreme Gradient Boosting (XGBoost). Logistic regression was implemented via the stats package (version 4.4.3), LASSO via glmnet (version 4.1–8), random forest via randomForest (version 4.7–1.2), SVM via kernlab (version 0.9–33), and XGBoost via xgboost (version 1.7.9.1). All models were trained and evaluated using a unified framework based on the caret package (version 7.0–1), which enabled 10-fold cross-validation and hyperparameter tuning. After cross-validation, the final model for each algorithm was retrained using the entire discovery cohort and subsequently applied to external validation cohorts. Receiver operating characteristic (ROC) curve analysis and visualization were performed using the pROC package (version 1.18.5) or ggplot2 (version 3.5.1). To assess the robustness of the models, potential batch effects were visualized through hierarchical clustering and heatmaps using the ComplexHeatmap package (version 2.20.0).

To reduce overfitting, we constrained the predictor set to a pre-specified panel of six MSI Level-1 metabolites and trained all algorithms within a unified 10-fold cross-validation framework with hyper-parameter tuning (caret). Penalized regression (LASSO) was used with the penalty parameter (λ) selected by cross-validated minimum error (λ.min). To specifically probe potential optimism in the seronegative subgroup, we conducted two analyses: (i) 100 iterations of class-matched down-sampling of SNRA and comparators while applying the original models; and (ii) 100 iterations of class-matched down-sampling with model re-training in each iteration using the same six-metabolite feature set. External evaluation was additionally performed on exclusively seronegative cases pooled from Validation Cohorts 1–3, applying the discovery-trained models without recalibration.

All the statistical tests were two-tailed, and a p value < 0.05 was considered significant (*p < 0.05, ** p < 0.01, *** p < 0.001, **** p < 0.0001, ns not significant).

Results

Study cohorts and participant characteristics

The clinical characteristics of participants from all cohorts are summarized in Supplementary Table 1. Participants were recruited from five independent centers. Specifically, the exploratory cohort was enrolled in Fuzhou, comprising 30 patients with RA, 30 with OA, and 30 HC. The discovery cohort and validation cohort 1 were also recruited from Fuzhou, including 450 RA, 450 OA, and 450 HC participants in the discovery cohort, and 106 RA, 102 OA, and 106 HC participants in validation cohort 1. The validation cohort 2 was recruited from Lanzhou, comprising 62 RA, 67 OA, and 62 HC participants. Additionally, validation cohorts 3–5 were recruited from three independent hospitals in Shanghai, including validation cohort 3 (108 RA, 98 OA, and 108 HC), validation cohort 4 (82 RA, 77 OA, and 82 HC), and validation cohort 5 (121 RA, 91 OA, and 151 HC). Age and sex distributions were matched within each cohort to the greatest extent possible. An overview of the experimental and analytical workflow is provided in Fig. 1.

Fig. 1.

Overview of workflow and study cohorts. A total of 2,863 participants were recruited from seven cohorts across five centers. The exploratory cohort underwent untargeted metabolomics analysis to identify potential biomarkers, which were subsequently validated in the discovery cohort using targeted metabolomics. These biomarkers were then utilized to construct metabolite-based classifiers, which were further validated in cohorts 1–5. The graphics in this figure were created using Biorender.com. Abbreviations: RA, rheumatoid arthritis; OA, osteoarthritis; HC, healthy controls; VIP, variable importance in projection; MSI, metabolomics standards initiative

Plasma metabolomic profiles in RA, OA, and healthy control subjects

To identify potential candidate metabolic biomarkers, we first performed untargeted metabolomic profiling using LC-MS/MS in the exploratory cohort. A total of 13,110 metabolic peaks were detected, including 6,270 in positive ion mode and 6,840 in negative ion mode. PCA was initially applied to the combined dataset from both positive and negative electrospray ionization (±ESI) modes, followed by separate analyses of +ESI and −ESI data to validate consistency across ionization conditions. Notably, QC samples clustered tightly at the center of the PCA plot, indicating high instrumental stability and reliable performance of the metabolomic workflow (Fig. 2A–C). In addition to QC samples, HC exhibited relatively tight clustering, while RA and OA samples were more dispersed and partially overlapped, reflecting both patient heterogeneity and shared clinical features. (Fig. 2A–C). Moreover, a subtle grouping tendency was observed among CRP-positive and ESR-positive samples within the PCA space, suggesting a potential link between inflammatory burden and metabolic variation, despite the absence of distinct clustering (Fig. 2D–E). After assessing global metabolic variation using unsupervised PCA, we applied supervised OPLS-DA to further discriminate RA from control groups, which demonstrated clear separation and generated variable importance in projection (VIP) scores (Fig. 2F–G and Supplementary Fig. 1A–B).

Fig. 2.

Plasma metabolomic profiles in RA, OA, and HC subjects. (A–C) PCA of untargeted metabolomics data. (A) combined PCA results for positive and negative ion modes; (B) PCA for positive ion mode; (C) PCA for negative ion mode. (D–E) comparison of PCA profiles between indicated groups based on the metabolic classifier of 27 metabolites. (D) CRP-positive vs. CRP-negative; (E) ESR-positive vs. ESR-negative. (F–G) OPLS-DA for (F) RA vs. HC and (G) RA vs. OA. (H–I) volcano plots for differential metabolites identified in the (H) RA vs. HC and (I) RA vs. OA comparisons. (J–K) MSEA for (J) RA vs. HC and (K) RA vs. OA, highlighting enriched biological pathways. Abbreviations: RA, rheumatoid arthritis; OA, osteoarthritis; HC, healthy controls; QC, quality control; PCA, principal component analysis; OPLS-DA, orthogonal partial least squares discriminant analysis. ESI, electrospray ionization; OPLS-DA, orthogonal partial least squares discriminant analysis; CRP, C-reactive protein; ESR, erythrocyte sedimentation rate; MSEA, metabolite set enrichment analysis

Metabolite annotation was standardized according to the Metabolomics Standards Initiative (MSI) [25], which defines four confidence levels based on identification criteria: Level 1 (confirmed identification with reference standards), Level 2 (putative annotation based on spectral similarity), Level 3 (putative compound class), and Level 4 (unknown compounds without structural information). Unannotated peaks (MSI Level 4) were excluded from all downstream analyses. Differential metabolites were identified from the remaining annotated dataset using VIP scores > 1 and statistical significance (p < 0.05), and were visualized using volcano plots (Fig. 2H–I). In total, 414 and 96 differential metabolites were identified in the RA vs. HC and RA vs. OA comparisons, respectively. To gain functional insight into these metabolic alterations, MSEA was performed for each group comparison. In RA patients compared to HC, notable enrichment was observed in pathways such as phosphatidylcholine biosynthesis, oxidation of branched-chain fatty acids, estrone metabolism, ubiquinone biosynthesis, and arginine and proline metabolism (Fig. 2J). In contrast, the comparison between RA and OA patients identified distinct pathway shifts, including spermidine and spermine biosynthesis, thiamine metabolism, riboflavin metabolism, sulfate/sulfite metabolism, and steroid biosynthesis, among others (Fig. 2K). Collectively, these results underscore distinct metabolic signatures associated with RA, relative to both HC and OA groups.

Identification and validation of potential metabolic biomarkers

In order to establish a clinically deployable metabolic diagnostic model, we screened potential metabolic biomarkers using untargeted metabolomics data. The selection criteria were as follows: 1. Metabolites with p < 0.05 and VIP > 1 were selected to maximize the inclusion of relevant metabolites, ensuring all potentially significant features were captured; 2. Only metabolites annotated as MSI Level 1 were included, ensuring clear identification and absolute quantification using reliable authentic standards for subsequent targeted metabolomics validation; 3. Metabolites exhibiting consistent upregulation or downregulation in both RA vs. OA and RA vs. HC comparisons were selected to minimize confounding factors. Specifically, when a metabolite shows inconsistent regulation across comparisons, and the differential expression between OA and HC exceeds that between RA and HC, it may introduce classification bias, increasing the risk of OA being misclassified as RA in the RA vs. HC model. Finally, 11 metabolic biomarkers were selected (Fig. 3A-B), including: imidazoleacetic acid, 1 H-imidazole-1-acetic acid, ergothioneine, lysine, N-acetyl-L-methionine, PI (18:1(9Z)/18:1(9Z)), 2-keto-3-deoxy-D-gluconic acid, alpha-ketoisovaleric acid, 1-methylnicotinamide, GlcCer(d18:1/16:0), and dehydroepiandrosterone sulfate (Supplementary Table 2). Figure 3B illustrates the log₂ fold change with 95% confidence intervals (CI) for each metabolite.

Fig. 3.

Identification and validation of potential metabolic biomarkers. (A) UpSet plot showing the overlap of potential biomarkers identified from the exploratory cohort through untargeted metabolomics. (B–C) Log₂ Fold change with 95% CI of 11 metabolites in RA vs. HC and RA vs. OA from the exploratory cohort using untargeted metabolomics (B) and the discovery cohort using targeted metabolomics (C). (D–N) violin plots showing the absolute quantification levels of the 11 metabolites in the discovery cohort for RA, OA, and HC. Abbreviations: RA, rheumatoid arthritis; OA, osteoarthritis; HC, healthy controls; CI, confidence interval

As a next step to confirm the diagnostic value of the selected metabolites, absolute quantification was performed using targeted metabolomics in the discovery cohort, which included 450 RA, 450 OA, and 450 HC participants. The log₂ fold changes with 95% CI for each metabolite are presented in Fig. 3C. Group comparisons among RA, OA, and HC are shown in Fig. 3D–N. We further selected metabolites that were differentially expressed and exhibited consistent regulation patterns in both RA vs. HC and RA vs. OA comparisons. As a result, six metabolites were ultimately identified as potential biomarkers for the development of diagnostic models, including imidazoleacetic acid, ergothioneine, N-acetyl-L-methionine, 2-keto-3-deoxy-D-gluconic acid, 1-methylnicotinamide, and dehydroepiandrosterone sulfate. Subsequently, LASSO-based feature selection was applied to the six candidate metabolites to refine the biomarker panel for diagnostic model development. The optimal penalty parameter (λ) was determined via 10-fold cross-validation, with the value minimizing the mean cross-validated error (λ.min) selected. The resulting coefficient profiles and cross-validation curves supported the inclusion of all six metabolites in both RA vs. HC and RA vs. OA models, as shown in Supplementary Fig. 2A–D.

Development and validation of metabolite-based classifiers for RA diagnosis

To evaluate the diagnostic potential of the metabolite-based panel for RA, we chose five machine learning algorithms to balance simplicity and interpretability with the ability to capture linear and non-linear interactions [26]: LASSO regression, logistic regression, random forest, SVM, and XGBoost. In the initial comparison between RA patients and healthy individuals, who served as a baseline for assessing disease-associated metabolic alterations, these models were trained on the discovery cohort. The resulting area under the receiver operating characteristic curve (AUC) values are presented in Fig. 4A. LASSO regression and logistic regression each achieved an AUC of 0.9336, while random forest, SVM, and XGBoost yielded AUCs of 0.9682, 0.9609, and 0.9655, respectively. To further evaluate model robustness, 10-fold cross-validation was performed for each algorithm, with nine folds used for training and one for validation. The cross-validated mean AUCs and corresponding 95% CI are summarized in Fig. 4B, demonstrating that all five models maintained stable and reliable performance. To assess generalizability, logistic regression, selected for its high interpretability, and random forest and XGBoost, chosen for their superior predictive performance, were further evaluated in three independent external cohorts from different regions. As shown in Fig. 4C–E, logistic regression achieved AUCs of 0.8806, 0.8676, and 0.9156 across cohorts 1 to 3, respectively. Random forest attained AUCs of 0.8474, 0.8375, and 0.9280, while XGBoost yielded 0.8413, 0.8723, and 0.9205.

Fig. 4.

Development and validation of metabolite-based classifiers for RA diagnosis. (A, F) ROC curves showing the classification results of training five machine learning models for RA vs. HC (A) and RA vs. OA (F) in the discovery cohort. (B, G) Ten-fold cross-validation AUC with 95% CI showing the performance of five machine learning models for distinguishing RA from HC (B) and RA from OA (G) in the discovery cohort. (C–E) ROC curves for the classification results of three machine learning models in validation cohorts 1–3, comparing RA vs. HC in cohorts 1 (C), 2 (D), and 3 (E). (H–J) ROC curves for the classification results of three machine learning models in validation cohorts 1–3, comparing RA vs. OA in cohorts 1 (H), 2 (I), and 3 (J). Abbreviations: RA, rheumatoid arthritis; OA, osteoarthritis; HC, healthy controls; ROC, receiver operating characteristic; AUC, area under the curve; LASSO, least absolute shrinkage and selection operator regression; logistic, logistic regression; SVM, support vector machine; RandomForest, random forest; XGBoost, extreme gradient boosting; CI, confidence interval

To further assess the panel’s ability to distinguish RA from OA, the same five machine learning algorithms were applied to the discovery cohort. The corresponding AUCs are shown in Fig. 4F. LASSO regression and logistic regression produced AUCs of 0.8278 and 0.8277, respectively, whereas random forest, SVM, and XGBoost achieved 0.8726, 0.8666, and 0.8708. Model stability was again evaluated using 10-fold cross-validation, with results summarized in Fig. 4G. All five models exhibited consistently high classification performance. To assess classification performance for OA, logistic regression, random forest, and XGBoost were applied to the same three validation cohorts. As shown in Fig. 4H to Fig. 4J, logistic regression produced AUCs of 0.7991, 0.7340, and 0.8181. Random forest achieved 0.8128, 0.7838, and 0.7753, while XGBoost yielded 0.8156, 0.7771, and 0.7648. Overall, despite a decline in classification performance when differentiating RA from OA relative to HC, the six-metabolite panel continued to demonstrate reliable discriminative ability.

Diagnostic performance of metabolite-based classifiers in seronegative RA

Seronegative RA has long posed a diagnostic challenge due to the absence of conventional serological markers. To explore whether our metabolite-based classification models could reliably identify seronegative RA, we assessed their diagnostic performance within this clinically distinct subgroup. To assess whether our metabolite-based models could effectively identify this subgroup, we evaluated their performance in seronegative RA patients (n = 46) from the discovery cohort. The full model, trained on the entire RA and HC cohort, was subsequently tested for its ability to distinguish seronegative RA from HC. The results showed that both XGBoost and random forest classifiers achieved higher AUCs than those observed in the complete RA classification model, with RF notably reaching an AUC of 1.0 (Fig. 5A). To rule out the potential influence of sample size imbalance, we performed 100 iterations of random resampling, matching seronegative RA cases with equal numbers of HC in each iteration. However, classification performance remained consistently high after resampling, comparable to that obtained using the full HC cohort (Fig. 5B). Given that the original model was trained on data including seronegative RA samples, the observed high performance may reflect overfitting rather than true generalizability to this subgroup. To more accurately evaluate the model’s ability to generalize to seronegative RA and to eliminate potential effects of sample size imbalance, we conducted 100 iterations of matched downsampling between seronegative RA and HC samples, retraining the model in each iteration using the six selected metabolites. This approach enabled a direct assessment of the panel’s true discriminative capacity in distinguishing seronegative RA from healthy controls. As shown in Fig. 5C, the mean AUCs of logistic regression, random forest, and XGBoost in the downsampled seronegative RA vs. HC classification were 0.9229, 0.9392, and 0.9110, respectively. Subsequently, validation cohorts 1 to 3 were combined to increase the number of seronegative RA cases (n = 23), and the original models were applied for evaluation. The analysis demonstrated that the resulting AUCs were 0.8929 for logistic regression, 0.8894 for random forest, and 0.8856 for XGBoost (Fig. 5D).

Fig. 5.

Diagnostic performance of metabolite-based classifiers in seronegative RA. (A, E) ROC curves showing the classification performance of the original model for distinguishing seronegative RA from HC (A) and seronegative RA from OA (E) in the discovery cohort. (B, F) original model trained on downsampled data (100 iterations) for seronegative RA vs. HC (B) and seronegative RA vs. OA (F) in the discovery cohort. (C, G) Retrained model based on downsampled data (100 iterations) for seronegative RA vs. HC (C) and seronegative RA vs. OA (G) in the discovery cohort. (D, H) ROC curves showing the classification performance of the original model for distinguishing seronegative RA from HC (D) and seronegative RA from OA (H) in the combined validation cohorts 1–3. Abbreviations: RA, rheumatoid arthritis; OA, osteoarthritis; HC, healthy controls; ROC, receiver operating characteristic; AUC, area under the curve; logistic, logistic regression; RandomForest, random forest; XGBoost, extreme gradient boosting

A similar pattern was observed in the classification of seronegative RA vs. OA. As shown in Fig. 5E–F, both random forest and XGBoost models exhibited signs of overfitting regardless of whether resampling was applied. To evaluate the true discriminative capacity of the six-metabolite panel in distinguishing seronegative RA from OA, we adopted the same downsampling and retraining strategy as described above. The analysis revealed that the mean AUCs of logistic regression, random forest, and XGBoost were 0.8213, 0.8559, and 0.8281, respectively (Fig. 5G). Finally, validation cohorts 1 to 3 were combined, and external evaluation using the original models yielded AUCs of 0.8012 for logistic regression, 0.8278 for random forest, and 0.8142 for XGBoost (Fig. 5H). Collectively, these findings indicate that the metabolite-based classifiers maintain robust diagnostic performance regardless of serological status of RA patients, thereby reinforcing their potential value in clinical applications.

Robustness evaluation of metabolite-based classifiers

The stability of classification performance is critical for clinical deployment. To assess the robustness of our metabolite-based classifiers, we evaluated their diagnostic performance for distinguishing RA from HC in two additional external validation settings: validation cohort 4, which comprised serum samples, and validation cohort 5, derived from a different analytical platform. In validation cohort 4, the AUCs for logistic regression, random forest, and XGBoost were 0.8715, 0.8509, and 0.8450, respectively (Fig. 6A). In validation cohort 5, the corresponding AUCs were 0.8572, 0.8609, and 0.8590 (Fig. 6B). The clustering results revealed only mild aggregation based on sample type, suggesting a limited impact of sample type on classifier performance (Fig. 6C). In comparison, platform-specific clustering was more apparent, indicating stronger batch effects (Fig. 6D). Nevertheless, the metabolite-based classifiers consistently maintained strong classification performance across these technical variations. This stability may be attributed to the consistent relative expression patterns of the selected metabolites across cohorts, which preserved the discriminative signals necessary for classification despite differences in sample source or analytical platform.

Fig. 6.

Robustness evaluation of metabolite-based classifiers. (A-B) ROC curves showing the classification results of three machine learning models in validation cohorts 4 and 5, comparing RA vs. HC in cohort 4 (A) and cohort 5 (B). (C, D) heatmap and clustering results of six metabolites in RA vs. HC in the discovery cohort and validation cohort 4 (C), or validation cohort 5 (D). (E-F) ROC curves showing the classification results of three machine learning models in validation cohorts 4 and 5, comparing RA vs. OA in cohort 4 (E) and cohort 5 (F). (G, H) heatmap and clustering results of six metabolites in RA vs. OA in the discovery cohort and validation cohort 4 (G), or validation cohort 5 (H). Abbreviations: RA, rheumatoid arthritis; OA, osteoarthritis; HC, healthy controls; ROC, receiver operating characteristic; AUC, area under the curve; logistic, logistic regression; RandomForest, random forest; XGBoost, extreme gradient boosting. Disc. Cohort, discovery cohort; Val. Cohort, validation cohort

A similar evaluation was conducted in validation cohorts 4 and 5 to assess the classifiers’ ability to distinguish RA from OA. In validation cohort 4, the AUCs for logistic regression, random forest, and XGBoost were 0.7474, 0.8051, and 0.8084, respectively (Fig. 6E). In validation cohort 5, the corresponding AUCs were 0.7939, 0.7455, and 0.7618 (Fig. 6F). To further investigate potential sources of variability, clustering analyses were performed and revealed stronger batch effects associated with detection platform than with sample type, consistent with the patterns observed in the RA vs. HC comparisons. Importantly, these variations did not substantially affect classifier performance (Fig. 6G–H). Collectively, these findings support the robustness and potential generalizability of the metabolite-based classifiers across different sample types and analytical platforms.

Discussion

According to current classification criteria [6], the diagnosis of RA is based on the type and number of affected joints, levels of serological markers such as RF and anti-CCP, acute-phase reactants including CRP and ESR, and symptom duration. However, given the substantial overlap in clinical manifestations between RA and other forms of arthritis, laboratory-based assessments, particularly serological markers, are essential for improving diagnostic precision. RF and anti-CCP remain the most widely utilized serological indicators, although both may also be detected in other autoimmune conditions such as systemic lupus erythematosus and systemic sclerosis [27, 28]. When both RF and anti-CCP fail to confirm diagnosis, patients with seronegative RA face a markedly increased risk of misdiagnosis or delayed recognition. The lack of highly sensitive and specific biomarkers remains a major barrier to accurate diagnosis, prognostic assessment, and risk stratification in RA [7, 29]. Therefore, identifying novel biomarkers independent of existing classification criteria, particularly those capable of differentiating RA, including seronegative cases, from other inflammatory arthritides, is critical.

Metabolites, as end products of cellular processes, offer an integrative readout of physiological and pathological states [14]. In RA, affected joints exhibit characteristic synovial changes, including hyperplasia and immune cell infiltration [2, 4]. The increased demand for energy and biosynthetic precursors in inflamed synovium underscores metabolic dysregulation as a key contributor to RA pathogenesis. Advances in high-resolution techniques, especially metabolomics, have enabled deeper exploration of these metabolic alterations and facilitated biomarker discovery [30–32]. Accumulating evidence shows that RA patients display distinct metabolic profiles compared to healthy individuals and patients with other inflammatory or rheumatic diseases, even when clinical features overlap [33–35].

Multivariate diagnostic models based on metabolomic data have shown promise in RA. For example, one study integrating metabolomic and lipidomic profiles distinguished seronegative RA from psoriatic arthritis with an AUC of 0.845 [36]. Another untargeted metabolomics study identified a biomarker panel that differentiated RA from primary Sjögren’s syndrome and healthy controls with high accuracy (sensitivity 93.3%, specificity 95.2%) [37]. Similarly, analysis of serum samples from 49 RA patients and 43 healthy individuals yielded six RA-associated metabolites, with an AUC of 0.890 [38]. While these collectively underscore the considerable diagnostic potential of metabolite-based models in RA, their clinical translation remains limited by small sample sizes, lack of validation in large, independent cohorts, and challenges in metabolite identification in untargeted analyses. Multicenter studies validating clinically applicable RA metabolomic classifiers remain relatively limited. Therefore, further efforts are warranted to develop robust, reproducible, and translatable diagnostic models across diverse populations and clinical settings.

In the present study, we conducted metabolomic profiling of 2,863 biological samples collected from five centers across three regions, including RA, OA, and HC groups. To ensure clinical applicability, several practical factors were considered in study design. Drawing on clinical experience, we recognized that effective implementation requires not only predictive performance but also laboratory standardization. Key components include standardized pre-analytical procedures, validated reference standards, comprehensive quality control systems, and well-defined standard operating procedures. Plasma was chosen over serum to improve reproducibility and minimize pre-analytical variability. Unlike serum, plasma preparation avoids coagulation and reduces variability due to enzymatic activity during clotting. Differences in anticoagulants and collection materials may also influence ionization and metabolite detection [39, 40]. To ensure that downstream analyses were based on reliable and biologically meaningful features, candidate metabolites were selected through a multi-step filtering strategy. First, untargeted metabolomics in the exploratory cohort applied permissive criteria (p < 0.05, VIP > 1) to capture potential biomarkers. Second, only MSI Level 1 metabolites identified using reference standards were retained. Third, metabolites showing consistent directionality in both RA vs. OA and RA vs. HC were selected. This approach balanced sensitivity and specificity, reduced confounding, and prioritized reliability for downstream validation.

We performed targeted metabolomics validation in the discovery cohort and applied LASSO regression for feature selection, yielding six informative metabolites, including imidazoleacetic acid, ergothioneine, N-acetyl-L-methionine, 2-keto-3-deoxy-D-gluconic acid, 1-methylnicotinamide, and dehydroepiandrosterone sulfate. These metabolites were then used to construct diagnostic models. Notably, several of these metabolites have previously been implicated in RA pathophysiology, providing biological support for their inclusion. For example, ergothioneine levels were found to be lower in RA patients and associated with poor methotrexate response [41]. In addition, it has been shown to protect chondrocytes from oxidative stress by reducing reactive oxygen species production and maintaining mitochondrial integrity, which may contribute to cartilage preservation in RA [42]. 1-methylnicotinamide has attracted attention as well, as it is upregulated in RA synovial fibroblasts under estrogen stimulation and may exert anti-inflammatory effects by regulating RA-related genes [43]. Likewise, dehydroepiandrosterone sulfate has been consistently reported to be downregulated in RA patients across multiple studies [44–46], suggesting that hormonal regulation may be associated with RA pathophysiology. By contrast, although several of the remaining metabolites have been reported in other biological contexts, their relevance to RA pathogenesis remains unclear and warrants further investigation. Imidazoleacetic acid, a principal histamine catabolite detectable in human biofluids, rises in allergic or asthma conditions, indicating increased histamine turnover [47]. By analogy, shifts in imidazoleacetic acid may reflect altered histamine metabolism in RA [48]. N-acetyl-L-methionine, an acetylated methionine that feeds the methionine–S-adenosylmethionine one-carbon pathway, shows antioxidant and protein-protective effects [49]. Changes in N-acetyl-L-methionine may mark redox and methylation imbalance in RA. Finally, 2-keto-3-deoxy-D-gluconic acid, a central intermediate of the Entner–Doudoroff pathway that is abundant in bacteria and detectable in human biofluids [50, 51], may indicate microbial carbohydrate metabolism and altered host–microbe carbon flux relevant to RA inflammation. Together, these hypotheses provide testable leads for future mechanistic and longitudinal studies.

Based on six candidate metabolites, we developed integrated diagnostic classifiers to distinguish RA from HC and OA. To balance interpretability and predictive performance, five machine learning algorithms were employed: LASSO, logistic regression, random forest, SVM, and XGBoost [26, 52]. All models showed excellent performance for RA vs. HC (AUC: 0.9336–0.9682) and good performance for RA vs. OA (AUC: 0.8277–0.8726). Based on 10-fold cross-validation, logistic regression, random forest, and XGBoost were selected for external validation. In validation cohorts 1–3, these models maintained strong discriminatory power for RA vs. HC (AUC: 0.8375–0.9280) and moderate-to-good accuracy for RA vs. OA (AUC: 0.7340–0.8181), supporting their generalizability and the biological relevance of the selected metabolites. Of note, the models’ relative performance varied across cohorts. Logistic regression outperformed in some cohorts, while random forest or XGBoost performed better in others. These differences likely reflect cohort-specific characteristics, feature distributions, and noise levels. Logistic regression excelled with stronger linear associations and minimal interaction, while non-linear models performed better in complex settings [26]. Together, these findings underscore the importance of a multi-model strategy to enhance the robustness and adaptability of diagnostic models across heterogeneous clinical settings. The model’s utility was further demonstrated in seronegative RA patients, where it achieved performance comparable to that observed in the overall RA population. This independence from serological status highlights its potential as a valuable complementary biomarker. Although plasma samples were prioritized as the primary testing matrix, serum samples were also analyzed to assess the impact of matrix background on analytical signals. To further evaluate cross-platform robustness, we also assessed the model’s performance across different mass spectrometry platforms. Notably, despite the presence of batch effects, predictive accuracy remained stable across sample types and analytical platforms, as long as relative expression patterns were preserved. Taken together, the evidence indicates that the model performs reliably and transfers well beyond the development setting. Clinical mass spectrometry workflows are standardized and operationally mature, and laboratories follow established standard operating procedures (SOPs) and quality controls, often with automation, producing consistent results at reasonable cost and with short turnaround times. These conditions enable practical, scalable clinical deployment of our model.

This multicenter, cross-sectional, case–control study has several limitations. We did not assess how disease activity/severity or current treatment status influences model performance, Therefore, prognostic utility and treatment-monitoring capability cannot be inferred and should be examined in prospective cohorts. We also did not evaluate whether the model can identify high-risk individuals who are RF or anti-CCP positive, a subgroup that warrants dedicated assessment. Likewise, we did not compare performance against other autoimmune diseases, which would be informative for specificity and clinical utility. Finally, because participants were recruited mainly in China, generalizability to other ethnic groups and regions remains to be established. Accordingly, prospective longitudinal studies are needed to validate real-world performance and to determine whether serial metabolite measurements provide prognostic insight and practical utility for monitoring treatment response and informing clinical decision-making, which is clinically meaningful given the extensive involvement of metabolic pathways in RA pathogenesis.

In summary, this study demonstrates that a metabolite-based clinical classifier can effectively differentiate RA from both healthy controls and OA, including seronegative cases where conventional biomarkers are insufficient. Its consistent performance across multiple cohorts, specimen types, and analytical platforms underscores its strong clinical utility. These findings highlight the value of metabolomics as a complementary diagnostic approach for RA and offer a generalizable framework for the development of metabolite-based classifiers in other complex diseases.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Author contributions

All authors were involved in drafting the manuscript or revising it critically for important intellectual content and approved the final version for publication. Jinpiao Lin, Huiming Sheng and Fang Xie had full access to all study data and assume responsibility for the integrity of the data and the accuracy of the analysis. They also serve as corresponding authors. Jifeng Tang, Renquan Jiang, and Huali Gao contributed equally to this work and are recognized as co-first authors. Together, they conceived the study, designed the research protocol, and oversaw its implementation. Specifically, Jifeng Tang was primarily responsible for data analysis and led the drafting of the final manuscript. Renquan Jiang led the collection of clinical samples and the corresponding clinical data. Huali Gao established the inclusion and exclusion criteria for sample collection, contributed to clinical data acquisition, and supervised the patient enrollment review. All three authors were actively involved in biospecimen processing and experimental procedures. Jinfang Xia, Yanhui Ma, Zhenge Han, Haitao Yu and Yizhong Zhang contributed to the collection of clinical samples and relevant clinical data. Jinpiao Lin, Huiming Sheng and Fang Xie reviewed the study protocol, validated the data, and critically revised the manuscript.

Funding

This work was supported by grants from the Shanghai Municipal Health Commission (2024ZZ2008), the National Natural Science Foundation of China (82302019), and the Natural Science Foundation of Fujian Province (2022J02035).

Data availability

The untargeted metabolomics data generated and analyzed in this study, along with de-identified clinical information and targeted metabolomics results are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

This study was approved by the Medical Ethics Committee of the First Affiliated Hospital of Fujian Medical University, Fuzhou, China (Approval No. [2022]082), and by the Medical Ethics Committee of Tongren Hospital, Shanghai Jiao Tong University School of Medicine, Shanghai, China (Approval No. AF/SC-11/04.0). All other participating institutions were included under the ethics approval of Tongren Hospital through formal institutional collaboration agreements. All procedures involving human participants complied with the Declaration of Helsinki, and informed consent was obtained from all subjects.

Consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing financial or non-financial interests related to this work.