Metabolomic signatures for diagnosis and clinical severity in Parkinson's disease

Summary

Background

Parkinson's disease (PD) remains challenging to diagnose at early stages owing to subtle and heterogeneous clinical manifestations and the lack of reliable biomarkers. Metabolomics offers a powerful approach to capture disease-related biochemical alterations that reflect underlying pathophysiology. This study aimed to identify robust plasma metabolic signatures for PD diagnosis and to elucidate metabolic alterations associated with clinical severity.

Methods

We performed untargeted plasma metabolomic profiling using ultra-high performance liquid chromatography-tandem mass spectrometry in a large Chinese population comprising two independent early-stage PD groups, one of which consisted of drug-naïve, de novo patients. Integrative statistical analyses, pathway enrichment, and machine learning-based diagnostic modelling were applied to identify discriminative metabolites and characterise disease- and treatment-related metabolic changes.

Findings

In the two case–control datasets, 111 metabolites were consistently altered in early-stage PD, among which 12-hydroxyeicosatetraenoic acid, spermine, and niacinamide emerged as key differential metabolites. Pathway enrichment analysis highlighted sphingolipid metabolism as a major dysregulated pathway in early-stage PD. Using machine learning-based models, a classification model based on six metabolites achieved strong performance (area under the receiver operating characteristic curve [AUC] = 0.976), while individual metabolites also demonstrated good discriminative ability, with the highest AUC reaching 0.916. We further observed that antiparkinsonian medication was significantly associated with metabolic alterations in tyrosine, tryptophan, and polyamine pathways. In addition, gut microbiota-derived metabolites, particularly phenylacetylglutamine and p-Cresol glucuronide, were markedly elevated in PD and associated with both motor and non-motor symptom severity, suggesting a potential contribution to clinical heterogeneity.

Interpretation

These findings indicate reproducible plasma metabolic differences associated with early PD and suggest the potential diagnostic value of internally validated classifiers for disease diagnosis. Alterations in gut microbiota-derived metabolites correlate with clinical severity, highlighting the need for further mechanistic and translational research.

Funding

This study was supported by the National Natural Science Foundation of China (82271281 and 82471267), the Science and Technology Major Project of Hunan Provincial Science and Technology Department (2021SK1010), and the National Key Research and Development Program of China (2021YFC2501204).

Research in context.

Evidence before this study

Increasing evidence implicates metabolic dysregulation and gut microbiota-derived metabolites in PD pathophysiology. To contextualise the current work, we searched PubMed and Web of Science for blood-based metabolomics studies of PD published up to April 2026 using combinations of the terms “Parkinson's disease”, “metabolomics”, “plasma”, “serum”, and “whole blood”, and further selected original studies reporting diagnosis- or symptom-related metabolite findings in clinical patients.

Previous metabolomic studies have reported alterations in amino acid, fatty acid, bile acid, and caffeine metabolism in PD. However, most studies are limited by sample sizes, heterogeneous disease stages, and restricted metabolite coverage. Notably, a large proportion of studies were conducted in relatively small samples, with approximately 45% including fewer than 100 participants and ∼70% fewer than 200. Only a limited number of studies included drug-naïve patients or well-defined early-stage populations, which are essential for identifying biomarkers relevant to early disease rather than secondary effects. Although phenylacetylglutamine (PAGln) and p-Cresol derivatives have been implicated in PD in several studies, their systematic evaluation in large, well-phenotyped cohorts remains scarce.

Added value of this study

This study provides comprehensive plasma metabolomic profiling of PD using high-resolution untargeted metabolomics, covering a broad range of metabolites and lipid species. We identified reproducible metabolic differences across two independent datasets of early-stage PD, several of which showed strong diagnostic performance in distinguishing PD from healthy controls, supporting the potential utility of metabolomics-based biomarkers for PD diagnosis.

Beyond biomarker discovery, we identified gut microbiota-derived metabolites, particularly PAGln and p-Cresol glucuronide (Pcg), that were significantly associated with PD status and clinical severity. These metabolites correlated with multiple clinical domains, especially autonomic dysfunction and gastrointestinal symptoms. Although PAGln and Pcg were not primary contributors to the diagnostic model, they were identified as key metabolic features associated with disease status and clinical manifestations. Integration of high-resolution metabolomics with detailed clinical phenotyping underscores the relevance of gut microbiota-derived metabolites in PD clinical heterogeneity.

Implications of all the available evidence

Together with existing evidence, our findings support an important role of metabolic dysregulation in PD and highlight the contribution of gut microbiota-derived metabolites. Reproducible metabolic signatures provide a foundation for developing non-invasive biomarkers for PD, while associations of PAGln and Pcg with clinical severity support their potential value for disease stratification and mechanistic investigation. These results provide a rationale for future studies exploring biological pathways linking gut microbiota, host metabolism, and neurodegeneration in PD.

Introduction

Parkinson's disease (PD) is a common, progressive, and clinically heterogeneous neurodegenerative disorder, with incidence increasing markedly with age. With the ongoing global ageing trend, the number of individuals affected by PD is projected to rise sharply, reaching approximately 25.2 million worldwide by 2050.^1,2 Current diagnosis remains predominantly clinical and largely relies on the recognition of overt motor manifestations, with no widely standardised, easily deployable objective biomarkers for reliable early detection. Although dopamine transporter imaging provides valuable supportive evidence for diagnosis, its use as a routine screening tool is constrained by cost and accessibility. Epidemiological data suggest that the rate of early PD diagnosis has shown limited improvement over recent decades, underscoring the urgent need for simple and broadly applicable diagnostic markers.3, 4, 5

Clinically, PD is characterised by cardinal motor symptoms, including bradykinesia, resting tremor, rigidity, and postural instability. In parallel, non-motor symptoms—such as rapid eye movement sleep behaviour disorder (RBD), constipation, hyposmia, cognitive impairment, and autonomic dysfunction—are highly prevalent and substantially impair quality of life.⁶ Pronounced heterogeneity in clinical manifestations and disease progression poses major challenges to mechanism-based stratification and precision medicine. Accordingly, the identification of quantifiable molecular markers that capture underlying pathophysiological processes is essential for earlier diagnosis, prognostic assessment, and refined disease subtyping.

Metabolomics systematically profiles small-molecule metabolites as downstream readouts of interactions among the genome, transcriptome, proteome, and environment, making it particularly well suited for identifying disease-related biochemical perturbations and candidate biomarkers.^7,8 Metabolomic studies have been conducted using a variety of biological matrices, including cerebrospinal fluid (CSF), serum, plasma, urine, and faecal samples. Among these, peripheral blood represents an especially attractive source for biomarker discovery due to its minimally invasive collection and suitability for large-scale cohort studies. Moreover, circulating metabolite profiles can, at least partially, reflect biochemical processes occurring in the central nervous system.^9,10

Several metabolomic studies in PD have consistently revealed perturbations in metabolic pathways including tryptophan, polyunsaturated fatty acids (PUFAs), short-chain fatty acids, bile acids, caffeine, and sphingolipids. These metabolic alterations may participate in disease pathophysiology and also provide clues for candidate diagnostic and prognostic biomarkers.¹¹ Previous studies have proposed several blood-based metabolite panels for PD diagnosis or clinical outcome prediction, with some models externally validated and reported area under the receiver operating characteristic curve (AUC) values spanning approximately 0.69–0.95.12, 13, 14, 15, 16, 17

However, these investigations frequently suffer from limited sample sizes, cohort heterogeneity, and confounding effects related to antiparkinsonian medications. As a result, the metabolites identified to date are highly heterogeneous and poorly reproducible, limiting the ability to draw consistent conclusions regarding key metabolic alterations and dysregulated pathways underlying PD. Furthermore, the lack of systematic integration of metabolomic signatures and multidimensional clinical profiles has hindered a comprehensive understanding of how metabolic changes relate to clinical heterogeneity. Consequently, larger-scale studies with well-designed cohorts and comprehensive clinical phenotyping are still needed to identify robust and generalisable metabolite biomarkers, as well as to elucidate the key metabolic pathways involved in PD pathophysiology.

In this study, we performed untargeted ultra-high performance liquid chromatography-tandem mass spectrometry (UPLC-MS/MS) plasma metabolomic profiling in a well-characterised Chinese study population to systematically investigate metabolic signatures of early-stage PD and their associations with clinical phenotypes. We aimed to identify and preliminarily validate plasma metabolite biomarkers for early PD diagnosis and to construct a diagnostic model. Furthermore, by integrating multidimensional clinical data, we sought to elucidate relationships between metabolic features and clinical heterogeneity. Ultimately, this work supports metabolite-based approaches for diagnosis and clinical subtyping of PD, with potential implications for clinically translatable diagnostic and patient stratification strategies.

Methods

Study population

Patients with PD and healthy controls (HCs) were recruited between January 2021 and March 2024. Patients were enrolled primarily at Xiangya Hospital, Central South University, with additional recruitment through collaborating centres within the Parkinson's Disease & Movement Disorders Multicenter Database and Collaborative Network in China (PD-MDCNC), including the Second Xiangya Hospital, the Third Xiangya Hospital, and the First Affiliated Hospital of University of South China. HCs were recruited primarily from local communities and residential areas, with additional controls enrolled from spouses of patients attending the neurology outpatient clinic. Demographic information, including age, biological sex, and ethnicity, was self-reported by participants and verified against medical records. To ensure methodological consistency across sites, all clinical assessments, diagnostic confirmation, and biospecimen collection were centralised at Xiangya Hospital and performed according to standardised protocols. Patients with PD aged 30–80 years were diagnosed according to the Movement Disorder Society (MDS) criteria for clinically established or clinically probable PD.¹⁸ Patients with atypical or secondary parkinsonism, other neurological or psychiatric disorders, a history of deep brain stimulation, acute stress conditions at the time of sampling, or clinically significant systemic diseases that could substantially affect metabolic profiles were excluded. HCs aged 30–80 years were matched for age, sex, and body mass index (BMI). Individuals with neurological or psychiatric disorders, first-degree relatives with PD, acute stress conditions at the time of sampling, or clinically significant systemic diseases were excluded. Detailed inclusion and exclusion criteria are provided in the Supplementary Methods.

Ethics

The study was conducted in accordance with the Declaration of Helsinki and was approved by the Ethics Committee of Xiangya Hospital, Central South University (No. 202005124). Written informed consent was obtained from all participants prior to enrolment in the study.

Study design and participant stratification

The overall study design and analytical workflow are summarised in Fig. 1. Participants were enrolled according to study objectives, including the identification of metabolic signatures and potential biomarkers, evaluation of medication effects, and phenotype–metabolite association analyses (Supplementary Fig. S1).

Fig. 1.

Study overview and analytical workflow. Schematic overview of the plasma metabolomics study. The workflow depicts study population recruitment and sample collection, LC-MS/MS-based untargeted metabolomic profiling, data preprocessing and quality control, identification of differential metabolites, diagnostic model construction, and integrative analyses linking metabolic alterations to clinical phenotypes. Created with Biorender.com.

Metabolic signatures and potential biomarkers

This cross-sectional case–control metabolomics study included two independent datasets of patients with early-stage PD and matched HCs to characterise plasma metabolic alterations and identify potential diagnostic biomarkers. Sample size considerations in untargeted metabolomics are inherently challenging due to the high dimensionality of metabolite features and the multiple testing burden. Sample size planning was informed by prior metabolomics studies in PD reporting moderate-to-large between-group differences in key metabolites.12, 13, 14, 15, 16, 17 Under a conventional two-sample comparison framework assuming a moderate effect size (Cohen's d approximately 0.5–0.6), α = 0.05, and 80% power, approximately 40–60 participants per group are generally considered sufficient to detect metabolite-level differences. We therefore adopted sample sizes exceeding this minimum benchmark, consistent with previously published metabolomics studies in PD that have identified reproducible metabolic signatures using comparable or smaller cohorts. Importantly, given the high-dimensional nature of the data, the study design further relies on replication across independent datasets to enhance robustness beyond single-test power considerations.

Demographic and clinical characteristics are summarised in Table 1 and Supplementary Table S1. Cohort 1 consisted of 153 patients with drug-naïve, de novo PD, with a disease duration of ≤2 years and Hoehn and Yahr (H–Y) stage ≤2, together with 153 age-, sex-, and BMI-matched HCs. This population represents a more stringently defined subset of early-stage PD, characterised by medication-naïve status and shorter disease duration. Cohort 2 consisted of a non-overlapping group of 135 patients with early-stage PD, with a disease duration of ≤3 years and H–Y stage ≤2.5, together with 135 matched HCs.

Table 1.

Demographic and clinical characteristics of study participants in Cohort 1 and Cohort 2.

Characteristics	PD (n = 288)	Control (n = 288)	p value
Cohort 1
Number of individuals	153	153	–
Age (years)	60.25 ± 10.40	60.11 ± 10.48	0.9948
Sex			1.0000
Male	85 (55.6%)	85 (55.6%)
Female	68 (44.4%)	68 (44.4%)
BMI (kg/m2)	23.52 ± 3.01	23.75 ± 2.93	0.3657
Hypertension	40 (26.1%)	56 (36.6%)	0.0646
Diabetes mellitus	13 (8.5%)	18 (11.8%)	0.4486
Disease duration (years)	1.22 ± 0.59	–	–
Hoehn and Yahr stage	1.52 ± 0.44	–	–
LEDD	0	0	–
Cohort 2
Number of individuals	135	135	–
Age (years)	63.18 ± 7.24	64.18 ± 7.80	0.2926
Sex			0.6254
Male	65 (48.1%)	60 (44.4%)
Female	70 (51.9%)	75 (55.6%)
BMI (kg/m2)	23.82 ± 3.12	23.69 ± 2.86	0.8884
Hypertension	33 (24.4%)	53 (39.3%)	0.0131
Diabetes mellitus	10 (7.4%)	21 (15.6%)	0.0563
Disease duration (years)	1.92 ± 0.88	–	–
Hoehn and Yahr stage	1.80 ± 0.52	–	–
LEDD	256.27 ± 176.65	0	–

Abbreviations: PD, Parkinson's disease; BMI, body mass index; LEDD, levodopa equivalent daily dose. Note: p values were calculated using the Mann–Whitney U test for continuous variables and the chi-square test for categorical variables, comparing patients with PD and healthy controls.

Antiparkinsonian medication-related metabolic signatures

To investigate the impact of antiparkinsonian medications on metabolic profiles, we selected patients with PD who had a disease duration ≤3 years and H–Y stage ≤2. This H–Y cutoff was applied to ensure comparable disease severity across treatment groups. Patients were stratified into three groups based on their treatment status: unmedicated (n = 175), l-DOPA (n = 85), and l-DOPA combined with dopamine agonists (DA) (n = 50). Comprehensive medication histories were recorded and the levodopa equivalent daily dose (LEDD) was calculated based on established methods reported in the literature.¹⁹ No significant differences were observed among the groups with respect to age, sex, BMI, or H–Y stage (p > 0.05; Supplementary Table S2).

Phenotype-metabolite associations

Building upon the patients included in Cohorts 1 and 2, the patient group was further expanded to encompass both patients with early-stage and those with more advanced-stage PD. All patients underwent comprehensive clinical assessments including motor and non-motor domains conducted by trained neurologists or qualified researchers (details provided in the Supplementary Methods). In brief, motor severity and disease stage were evaluated using the H–Y stage and the Unified Parkinson's Disease Rating Scale (UPDRS). Non-motor symptoms were assessed using the Non-Motor Symptoms Scale (NMSS), autonomic function with the Scale for Outcomes in Parkinson's Disease for Autonomic Symptoms (SCOPA-AUT), and cognitive performance using the Mini-Mental State Examination (MMSE). These clinical assessments were performed prior to metabolomic profiling and were independent of metabolomic data acquisition and downstream analyses. In total, 477 patients with PD at different stages of severity were included, enabling comprehensive analyses of associations between plasma metabolite profiles and multidimensional clinical phenotypes. Clinical characteristics are summarised in Supplementary Table S3, while missing values for each clinical scale and sample sizes for each analysis are summarised in Supplementary Table S4.

Blood sample collection and preparation

All participants underwent venous blood collection after an 8–10 h overnight fast between 07:00 and 11:00 AM. A total of 5–10 mL of blood was collected into EDTA anticoagulant tubes and immediately transported on ice to the laboratory to minimise haemolysis. Within 2 h of collection, samples were centrifuged at 3000 rpm for 10 min at 4 °C. Plasma supernatants were then aliquoted (approximately 400 μL per tube) and stored at −80 °C until analysis, with freeze–thaw cycles limited to a maximum of one. Samples exhibiting visible haemolysis were excluded from subsequent analyses. To comprehensively capture metabolites with diverse physicochemical properties, both polar and non-polar fractions were extracted from each plasma sample for downstream metabolomic analyses, as described in the Supplementary Methods. Quality control (QC) samples for each extract type were generated by pooling equal volumes of extracted solutions from randomly selected study samples.

Metabolomic profiling by UPLC-MS/MS

Metabolomic analyses were conducted following analytical protocols described in previous publications.²⁰ Briefly, untargeted metabolomic profiling was performed using an Ultimate 3000 UPLC system coupled to a Q Exactive quadrupole Orbitrap high-resolution mass spectrometer (Thermo Scientific). Reverse-phase liquid chromatography was employed for pre-separation of compounds in the extracts. Ionisation was achieved using a heated electrospray ionisation (ESI) source operated in both positive (ESI⁺) and negative (ESI⁻) ion modes to maximise metabolite coverage. Tandem mass spectrometry (MS/MS) data were acquired in data-dependent acquisition mode, generating full-scan MS1 spectra and corresponding MS2 spectra for ions with high signal intensities. All samples were injected in a randomised order to minimise potential batch effects. A total of 38 QC samples were evenly distributed throughout the analytical sequence to monitor instrument stability and analytical reproducibility (Additional details are provided in the Supplementary Methods). Sample preparation and metabolomic profiling were performed by personnel blinded to clinical group assignments.

Metabolite annotation

Raw mass spectrometry data were processed as previously described.²⁰ Briefly, full-scan MS1 and data-dependent MS2 spectra were processed using Compound Discoverer software and Xcalibur software (Thermo Scientific). Data processing included peak detection, retention time alignment, peak deconvolution, and removal of isotope peaks and background signals, resulting in a final feature matrix for downstream analysis. Metabolite structural annotation was performed by matching acquired MS2 spectra against two reference resources: (1) a proprietary high-resolution MS/MS spectral library (iPhenome SMOL) constructed from authentic chemical standards analysed under identical experimental conditions, and (2) the mzCloud online spectral library (Thermo Scientific). Lipid species were annotated by matching MS1 and MS2 spectra against the LipidSearch-integrated (Thermo Scientific) and LIPID MAPS databases.²¹ In addition, accurate mass-to-charge ratios (m/z, ±5 ppm) from MS1 spectra were queried against a local Human Metabolome Database (HMDB).²² Candidate identities were rigorously confirmed by evaluating both retention time (RT) and high-resolution MS/MS spectral similarity.

All metabolite and lipid annotations were classified according to the confidence levels defined by the Metabolomics Standards Initiative (MSI).²³ Lipids were predominantly annotated at level 2 due to limited availability of reference standards, allowing determination of lipid class and fatty acyl composition based on MS/MS spectra, while structural isomers could not be resolved. Metabolite nomenclature and pathway mapping were referenced to the HMDB,²² Kyoto Encyclopedia of Genes and Genomes (KEGG),²⁴ and LIPID MAPS²¹ databases. Detailed lipid nomenclature is provided in the Supplementary Methods.

Data preprocessing

Following metabolite annotation, a total of 1146 compounds were initially detected across all confidence levels. To minimise technical variability introduced during LC-MS analysis, data normalisation was performed using MetNormalizer.²⁵ Analytical stability was evaluated using pooled QC samples. The signal intensities of stable isotopically labelled internal standards were monitored across QC samples, yielding relative standard deviations of 1.07%–16.37%, indicating stable instrument performance. Principal component analysis (PCA) showed tight clustering of QC samples, and high Spearman correlation between the first and last QC injections suggested minimal signal drift. Distributions of coefficient of variation (CV) and missing values further supported the robustness of the dataset (Supplementary Fig. S2A–D).

Data cleaning was subsequently conducted based on predefined quality criteria.^26,27 Outlier samples identified using PCA-based Mahalanobis distance (95% confidence interval) were excluded, and metabolites with a CV >20% in QC samples were also removed. Missingness was systematically evaluated at both the sample and metabolite levels. Most samples exhibited very low missingness (<2%), whereas metabolite-level missingness showed a right-skewed distribution, with only 19 of 1146 metabolites (1.66%) exceeding a 20% missingness threshold and being excluded from subsequent analyses. Missingness showed a strong inverse association with metabolite abundance and no evidence of batch- or group-specific clustering, suggesting a missing-not-at-random (MNAR) mechanism driven by values below the detection limit (Supplementary Fig. S3). Based on these observations, remaining missing values were imputed using a half-minimum approach, in which each missing value was replaced with half of the minimum observed value for the corresponding metabolite, with a small random perturbation applied to reduce value ties. Sensitivity analyses were performed using alternative QC thresholds (CV >30% and missingness >30%) to assess the robustness of the preprocessing pipeline. In addition, only metabolites annotated with confidence levels 1 or 2 according to the MSI were retained for downstream analyses. Detailed descriptions of missingness assessment, imputation procedures, and sensitivity analyses are provided in the Supplementary Methods.

After quality control and filtering, 749 metabolites remained for subsequent statistical analyses, representing diverse biochemical classes including fatty acids, amino acids, and glycerophosphocholines (Supplementary Fig. S2E). To account for differences in metabolite abundance scales, data were log10-transformed and Z-score standardised as appropriate for downstream analyses.

KEGG pathway analysis

Over-representation analysis was performed on significantly altered metabolites using the MetaboAnalyst 6.0 platform to identify enriched KEGG metabolic pathways.²⁸ Statistical significance was evaluated using the hypergeometric test, and pathways with a False discovery rate (FDR)-adjusted p-value (p_adj) < 0.05 were considered significantly enriched. In addition, pathway topology analysis was conducted based on the relative-betweenness centrality algorithm. The topological importance of each metabolite within a pathway was assessed, and pathway impact was estimated by integrating the relative positions and centrality of all altered metabolites within the pathway.

Statistics

All statistical analyses and data visualisation were performed using Python and R. Machine learning analyses were implemented using the scikit-learn library.

Comparison of participant characteristics

Continuous variables are presented as mean ± standard deviation (SD), and categorical variables as counts (percentages). Group comparisons were conducted using t-tests or analysis of variance (ANOVA) for normally distributed continuous variables, and Mann–Whitney U or Kruskal–Wallis tests for non-normally distributed variables, with appropriate post hoc analyses when applicable. Categorical variables were compared using the chi-square test.

Differential metabolite analysis

Differential metabolite analyses were performed independently in Cohort 1 and Cohort 2. Metabolite abundances were compared between patients with PD and HCs using logistic regression, adjusting for age, sex, BMI, hypertension status, and diabetes status. Effect sizes were reported as odds ratios (ORs) with 95% confidence intervals (CIs). Fold changes (FC) were calculated to quantify the magnitude of group differences, and their 95% CIs were estimated using bootstrap resampling (1000 iterations). Multiple testing correction was applied using the Benjamini-Hochberg FDR method, with adjusted p-values (p_adj < 0.05) considered statistically significant. Sensitivity analyses were performed using progressively expanded adjustment models that additionally accounted for lifestyle factors (smoking and alcohol consumption) and biomarkers reflecting inflammatory status and liver and renal function, to assess the robustness of the findings (Supplementary Methods).

Multivariate analyses were further conducted using PCA and orthogonal partial least squares-discriminant analysis (OPLS-DA) to characterise global metabolic patterns and group separation. PCA was performed as an unsupervised exploratory approach to assess global metabolic variation and potential separation between groups. OPLS-DA model performance was evaluated using R²X, R²Y, and Q² statistics, and permutation testing (n = 999) was conducted to assess model robustness and the risk of overfitting. The variable importance in projection (VIP) score was used to indicate each metabolite's contribution to group discrimination, and 95% CIs for VIP scores were estimated using bootstrap resampling (1000 iterations). In each iteration, the model was refitted to estimate the distribution of VIP scores, and the proportion of iterations in which VIP exceeded 1 was calculated to assess the stability of metabolite importance.

Feature selection and diagnostic model construction

All metabolites (n = 749) were considered as candidate features for model development. As shown in Supplementary Fig. S4, samples were stratified by diagnosis and randomly split into a training set (80%) and an independent test set (20%), with the test set strictly held out for final evaluation. Metabolite intensities were log10-transformed and Z-score standardised. Feature selection was performed using least absolute shrinkage and selection operator (LASSO) regression with key covariates (age, sex, BMI, hypertension status, and diabetes status) included as unpenalised variables. To improve robustness, a stability selection framework based on repeated subsampling was applied to identify consistently selected metabolites, where features were considered stable if selected in ≥85% of subsampling iterations (50% subsampling, n = 200). Based on the selected features, diagnostic models were constructed within a supervised learning framework using logistic regression, support vector machine (SVM), and extreme gradient boosting (XGBoost), with hyperparameters optimised via nested cross-validation (5-fold inner loop for tuning and 5-fold outer loop for performance estimation). Model performance was assessed in the test set, and comparisons were performed using DeLong's test. The discriminative performance of individual metabolites was also evaluated using the same analytical framework.

In brief, this study used plasma metabolite profiles as input features to develop classification models for distinguishing patients with PD from HCs, with the primary objective of achieving accurate disease status classification based on metabolic signatures. Given the high dimensionality of the data, feature selection was first applied to identify a reduced subset of informative metabolites from all detected metabolites to improve model stability and interpretability. A repeated subsampling strategy was used to perform feature selection multiple times, retaining only metabolites consistently selected across iterations to construct a robust feature set. Using this feature set, multiple classification algorithms were trained to discriminate between PD and control groups. Model training and optimisation were conducted within the training set, with cross-validation applied to assess stability and reduce the risk of overfitting. Finally, the trained models were evaluated on the test set, where predicted class labels were generated based on the same input features and compared against the true diagnostic labels to quantify classification performance and generalisability.

An independent group of patients with essential tremor (n = 64) was used for post hoc external validation (details provided in the Supplementary Methods). To assess robustness to potential medication effects, sensitivity analyses were conducted by evaluating model performance separately in drug-naïve and medicated subgroups within the test set, and by restricting model development to the drug-naïve population. In addition, sex-stratified sensitivity analysis was performed in the independent test set to assess potential sex-related differences in model performance. Detailed procedures for LASSO regression, stability selection, and hyperparameter optimisation are provided in the Supplementary Methods.

Association analysis of clinical phenotypes and metabolites

Associations between metabolites and clinical phenotypes were evaluated using a two-step analytical strategy. For continuous clinical variables (e.g., UPDRS, H–Y stage, NMSS, SCOPA-AUT, and MMSE), Pearson correlation analysis was first performed. Candidate metabolites were subsequently reassessed using multivariable linear regression models adjusting for age, sex, BMI, hypertension status, diabetes status and LEDD, with additional adjustment for education level in MMSE analyses. For binary clinical outcomes (e.g., functional constipation, freezing of gait, and wearing off), multivariable logistic regression models with the same covariate adjustments were applied. For each association analysis, we performed complete-case analysis, including only participants with non-missing values for the specific outcome variable (Supplementary Table S4). Multiple testing correction was performed using the Benjamini-Hochberg procedure, and adjusted p-values (p_adj < 0.05) were considered statistically significant. To assess the robustness of the results, sensitivity analyses were additionally performed with extended adjustment for potential confounders (Supplementary Methods).

Role of funders

Funders had no role in study design, data collection, data analyses, data interpretation, or writing of this report.

Results

Differential metabolic signatures and cross-dataset reproducibility

In Cohort 1, a total of 154 metabolites were significantly altered between patients with PD and healthy controls (p_adj < 0.05, logistic regression). Among these, six metabolites exhibited the most robust and pronounced changes (|log₂FC| >1 and 95% CIs excluding the null value), including 12-hydroxyeicosatetraenoic acid (12-HETE), spermine, niacinamide, glycodeoxycholate 3-glucuronide, 4-Acetylphenol sulfate, and trigonelline (Fig. 2A; Table 2). Specifically, 12-HETE, spermine, niacinamide, and glycodeoxycholate 3-glucuronide were elevated in PD, whereas 4-Acetylphenol sulfate and trigonelline were decreased. These metabolites were annotated to several metabolic pathways, including arachidonic acid metabolism (12-HETE), arginine and proline metabolism (spermine), nicotinate and nicotinamide metabolism (niacinamide and trigonelline), phenolic acid metabolism (4-Acetylphenol sulfate), and bile acid metabolism (glycodeoxycholate 3-glucuronide).

Fig. 2.

Metabolomic signatures distinguishing patients with early-stage PD from healthy controls. (A) Volcano plot of differential metabolites comparing drug-naïve, de novo PD patients (n = 153) with healthy controls in Cohort 1 (n = 153). The horizontal blue dashed line indicates the significance threshold (p_adj < 0.05), and the vertical green dashed lines denote |log₂FC| = 1. Variable importance in projection (VIP) scores were derived from the orthogonal partial least squares-discriminant analysis (OPLS-DA) model. (B) Principal component analysis (PCA) score plot for Cohort 1. The dashed ellipse represents the 95% confidence interval. (C) OPLS-DA score plot for Cohort 1. The dashed ellipse represents the 95% confidence interval. (D) Venn diagram illustrating the overlap of significantly differential metabolites between Cohort 1 (n = 306) and Cohort 2 (n = 270). (E) Bubble plot of KEGG pathway enrichment analysis for the 111 overlapping differential metabolites. Bubble size reflects the pathway impact value, and bubble colour represents the adjusted p-value.

Table 2.

Differential metabolites showing the most pronounced changes in Cohort 1 and Cohort 2.

Metabolite	FC (95% CI)	log2FC	β	OR (95% CI)	p value	padj	VIP
Cohort 1
12-HETE	2.66 (2.04–3.56)	1.41	2.05	7.80 (5.01–12.15)	<0.0001	<0.0001	3.81
Niacinamide	2.10 (1.84–2.38)	1.07	1.97	7.16 (4.69–10.94)	<0.0001	<0.0001	3.54
Spermine	2.57 (2.0–3.38)	1.36	2.07	7.95 (5.06–12.50)	<0.0001	<0.0001	3.80
4-Acetylphenol sulfate	0.39 (0.19–0.76)	−1.36	−0.69	0.50 (0.38–0.66)	<0.0001	0.0001	1.51
Trigonelline	0.46 (0.28–0.76)	−1.12	−0.60	0.55 (0.41–0.73)	<0.0001	0.0016	1.48
Glycodeoxycholate 3-glucuronide	2.87 (1.46–5.45)	1.52	0.44	1.56 (1.22–1.99)	0.0004	0.0144	1.14
Cohort 2
12-HETE	3.14 (2.36–4.28)	1.65	2.20	9.05 (5.53–14.82)	<0.0001	<0.0001	3.58
Adenosine	3.03 (2.60–3.55)	1.60	3.06	21.26 (10.70–42.25)	<0.0001	<0.0001	4.02
Niacinamide	2.27 (1.99–2.60)	1.18	2.01	7.46 (4.70–11.85)	<0.0001	<0.0001	3.49
Spermine	3.64 (2.86–4.67)	1.86	2.50	12.21 (6.84–21.81)	<0.0001	<0.0001	3.63
Ile–Pro	2.08 (1.85–2.33)	1.05	2.27	9.69 (5.61–16.73)	<0.0001	<0.0001	3.40
Spermidine	2.17 (1.91–2.47)	1.11	2.27	9.70 (5.44–17.28)	<0.0001	<0.0001	3.26
3-Allylphenol sulfate	0.38 (0.28–0.54)	−1.38	−0.80	0.45 (0.33–0.61)	<0.0001	<0.0001	1.78
Erucic acid (FFA (22:1n9))	2.13 (1.35–3.35)	1.09	0.89	2.44 (1.72–3.47)	<0.0001	<0.0001	1.90
p-Cresol glucuronide	2.72 (1.86–3.90)	1.45	0.67	1.96 (1.48–2.59)	<0.0001	0.0002	1.43
4-Acetylphenol sulfate	0.45 (0.25–0.83)	−1.16	−0.53	0.59 (0.45–0.78)	0.0003	0.0086	1.10
Chenodeoxycholic acid glycine conjugate	0.47 (0.34–0.66)	−1.08	−0.48	0.62 (0.47–0.80)	0.0004	0.0107	1.10
Glycocholic acid	0.34 (0.16–0.78)	−1.57	−0.47	0.63 (0.48–0.82)	0.0007	0.0184	1.03

Abbreviations: FC, fold change; CI, confidence interval; log₂FC, log2-transformed fold change; OR, odds ratio; p_adj, FDR-adjusted p value for multiple comparisons (Benjamini-Hochberg); VIP, variable importance in projection. Note: The 95% CI for FCs were derived from 1000 bootstrap resamples. β coefficients, OR, and p value were obtained from logistic regression models adjusted for age, sex, body mass index, hypertension status, and diabetes status. Metabolite levels were log-transformed and standardised to Z-scores before being entered into the models.

Using the same analytical framework, differential metabolite analysis was then performed in Cohort 2, independently identifying 211 metabolites (p_adj < 0.05), of which 12 exhibited large effect sizes (|log₂FC| >1 and 95% CIs excluding the null value) (Table 2). Cross-dataset comparison revealed 117 shared differential metabolites between Cohort 1 and Cohort 2, of which 111 exhibited robust directional consistency across datasets. The log₂FC of shared metabolites showed a high degree of concordance between the two datasets, supporting the robustness and reproducibility of these alterations (Spearman r = 0.96, p < 0.0001). Notably, 12-HETE, spermine, niacinamide, and 4-Acetylphenol sulfate were consistently among the most strongly altered metabolites in both datasets, representing key replicated metabolites (Fig. 2D). Sensitivity analyses using progressively expanded covariate adjustment models yielded consistent results, supporting the robustness of the identified differential metabolites (Supplementary Table S5).

Multivariate analyses provided complementary support for these findings. In both datasets, PCA showed partial overlap between PD and controls, reflecting the high-dimensional and heterogeneous nature of metabolomics data. In contrast, supervised OPLS-DA demonstrated clear separation, with robust model performance (Cohort 1: R²Y = 0.841, Q² = 0.725; Cohort 2: R²Y = 0.869, Q² = 0.778), indicating the presence of distinct metabolic profiles between PD and controls (Fig. 2B–C; Supplementary Fig. S5; Supplementary Table S6). Robustness of the models was confirmed by permutation testing, in which all permuted Q² values were lower than the original model Q², confirming a low risk of overfitting. Metabolites contributing most strongly to group discrimination, as ranked by VIP scores, were highly consistent between datasets, including taurine, 12-HETE, spermine, adenosine, niacinamide, lactosylceramide (d40:1), spermidine, two phosphatidylethanolamine species (PE (40:4)-PE (18:0/22:4) and PE (40:4p)-PE (18:0p/22:4)), and dipeptide Ile–Pro (Supplementary Table S7). Bootstrap-based stability analysis further supported their robustness, with stable VIP estimates across resampling iterations. Notably, among the four key replicated metabolites identified above, 4-Acetylphenol sulfate showed relatively lower stability in VIP ranking across models (Supplementary Fig. S5F).

KEGG pathway enrichment analysis of shared differential metabolites

KEGG pathway enrichment analysis of the 111 shared differential metabolites identified across both datasets revealed significant enrichment in six metabolic pathways (p_adj < 0.05; Fig. 2E; Supplementary Table S8). These pathways included sphingolipid metabolism, arginine biosynthesis, citrate cycle (TCA cycle), beta-alanine metabolism, alanine-aspartate-glutamate metabolism, and glyoxylate and dicarboxylate metabolism. Among them, sphingolipid metabolism showed the strongest enrichment signal, with the lowest adjusted p-value (p_adj = 0.0002) and the highest pathway impact score (Impact = 0.54). A total of seven differential metabolites were mapped to this pathway, suggesting a prominent perturbation in sphingolipid-related metabolic processes in early-stage PD.

Diagnostic performance of metabolite-based classification models in PD

Diagnostic models were developed using the combined dataset of the two study populations (Cohort 1 and Cohort 2). Using LASSO with stability selection, six metabolites were identified as robust features with high selection stability across resampling iterations, including N6-Acetyl-L-lysine, taurine, Ile–Val, Ile–Pro, spermine, and LysoPC(18:0)-LysoPC(0:0/18:0) (Supplementary Fig. S6A–C). Models based on individual metabolites showed variable discriminative performance (Supplementary Fig. S7; Supplementary Table S9), with taurine achieving the highest AUC (0.909–0.916 across models), followed by spermine (AUC: 0.848–0.886) and Ile–Pro (AUC: 0.840–0.857). When combined, the six-metabolite model showed substantially improved performance, with AUCs of 0.966, 0.976, and 0.975 for logistic regression, SVM, and XGBoost, respectively (Fig. 3). DeLong's test indicated no significant differences in AUC between models (all p > 0.05). Considering overall performance metrics, XGBoost showed a slight performance advantage; however, given its model complexity, its interpretability is limited compared with logistic regression.

Fig. 3.

Classification performance for distinguishing patients with PD from healthy controls. (A–C) Receiver operating characteristic (ROC) curves of (A) logistic regression, (B) support vector machine (SVM) and (C) extreme gradient boosting (XGBoost) models constructed using the six metabolites selected by the LASSO algorithm. Curves depict the mean ROC obtained from 5-fold cross-validation in the training set (n = 460), together with model performance in the independent test set (n = 116). (D) Radar plot summarising the overall performance of the models.

To contextualise the selected features, we examined their differential profiles in Cohort 1 and Cohort 2. All LASSO-selected metabolites showed statistically significant differences between PD and controls in both datasets, although several exhibited relatively modest effect sizes (Supplementary Table S10). In addition, we evaluated the diagnostic value of key replicated metabolites identified above (spermine, 12-HETE, niacinamide and 4-Acetylphenol sulfate). Among these, 12-HETE and niacinamide demonstrated good discriminative ability, with AUCs ranging from 0.858 to 0.878 and 0.845 to 0.856, respectively, whereas 4-Acetylphenol sulfate showed limited performance (AUC: 0.381–0.610), consistent with its relatively low VIP stability in OPLS-DA models. A combined model incorporating spermine, 12-HETE, and niacinamide further improved classification performance, with AUCs ranging from 0.878 to 0.914 across models (Supplementary Table S9).

Exploratory evaluation of differential diagnostic performance

To examine the model's performance in an exploratory differential diagnostic setting, it was applied to an independent essential tremor group. The model retained moderate discriminative ability for distinguishing PD from essential tremor (AUC: 0.764–0.841), indicating partial generalisability but reduced specificity for distinguishing phenotypically related movement disorders (Supplementary Fig. S8). These findings suggest that the LASSO-selected metabolites capture disease-associated biological variation relevant to PD, but are not sufficiently specific to fully differentiate PD from clinically similar disorders. It is also important to note that the essential tremor group differed in age distribution compared with the PD and control groups, which may introduce residual confounding and partially influence model performance (Supplementary Methods). In addition, as the model was trained exclusively on PD versus healthy controls, it was not optimised for differential diagnosis among movement disorders. Taken together, these results indicate partial generalisability of the metabolic signature beyond the original classification task, while highlighting limitations in disease specificity and dataset comparability.

Sensitivity analyses of model robustness to medication status and sex

To evaluate the robustness of the diagnostic model with respect to medication status, two complementary sensitivity analyses were performed. First, model performance was assessed separately in drug-naïve and medicated subgroups within the test set. All three models maintained good discriminative ability in both subgroups. In the drug-naïve subgroup, AUCs ranged from 0.929 to 0.957, while in the medicated subgroup, performance was consistently higher, with AUCs ranging from 0.988 to 0.993 (Supplementary Fig. S9A). Second, all analyses were restricted to the drug-naïve population. Under the same feature selection framework, LASSO selected four metabolites (taurine, N6-Acetyl-L-lysine, spermine, and Ile–Pro; Supplementary Fig. S6D–F), yielding models with AUCs ranging from 0.873 to 0.922 (Supplementary Fig. S9B). Notably, the top six metabolites ranked by stability remained identical to those identified in the full dataset, and when these six metabolites were directly applied without re-selection, classification performance was further improved, with AUCs ranging from 0.923 to 0.953 across models (Supplementary Fig. S9C–D). Together, these findings demonstrate that the diagnostic model is robust to potential confounding by medication status and that the identified metabolic signatures are stable and reproducible.

Sex-stratified sensitivity analysis was performed by stratifying the test set into male and female subgroups, and diagnostic models were evaluated separately in each sex. The models showed consistently high and comparable performance across both sexes, with AUCs ranging from 0.963 to 0.980 in males and 0.969 to 0.983 in females, indicating that classification performance was not influenced by sex (Supplementary Fig. S10).

Influence of antiparkinsonian medications on metabolomic features

Using the nonparametric Kruskal–Wallis test, nine metabolites were identified as significantly associated with antiparkinsonian medication status (p_adj < 0.05; Supplementary Table S11). Post hoc pairwise comparisons using the Steel–Dwass test revealed significant differences in metabolite levels between the drug-naïve group and either the L-DOPA or L-DOPA + DA treatment groups (p < 0.05), whereas no significant differences were observed between the two medicated groups (Fig. 4A). These results suggest that the observed medication-associated metabolic alterations were primarily attributable to L-DOPA therapy. Functionally, these medication-related metabolites were mainly involved in spermidine, tyrosine, and tryptophan metabolism pathways. Notably, several metabolites identified in this analysis, including tyrosine and phenol sulphate, exhibited alteration patterns consistent with previous reports.¹⁴

Fig. 4.

Metabolic alterations associated with antiparkinsonian medication use. (A) Box plots illustrating the relative abundances of representative metabolites that differed significantly across medication status groups, including patients with drug-naïve PD (n = 175), patients receiving L-DOPA monotherapy (n = 85), and those receiving combined L-DOPA plus dopamine agonist (DA) therapy (n = 50). (B) Regression plots showing the associations between plasma metabolite levels and levodopa equivalent daily dose (LEDD). Shaded areas represent 95% confidence intervals. Spearman's correlation coefficients (r) were calculated within the L-DOPA and L-DOPA + DA subgroups. Significance levels are denoted as ∗∗∗p < 0.001; ∗∗p < 0.01; ∗p < 0.05; ns, p > 0.05 (Post hoc pairwise comparisons using the Steel–Dwass test).

Further analyses restricted to medicated patients (L-DOPA and L-DOPA + DA groups) revealed that four metabolites were significantly correlated with LEDD, including N1-Acetylspermidine/N8-Acetylspermidine (the analytical method did not distinguish between isomeric forms), 2-O-Methylascorbic acid, gamma-Glu-Tyr, and desaminotyrosine (p < 0.05, linear regression; Fig. 4B; Supplementary Table S11), indicating potential dose-dependent relationships. Among these, N1-Acetylspermidine/N8-Acetylspermidine demonstrated the strongest association with LEDD (r = 0.28, p < 0.05), particularly within the L-DOPA group (r = 0.33, p < 0.05). This increase was accompanied by a concomitant decreasing trend in its downstream degradation product, N-(3-acetamidopropyl) pyrrolidin-2-one. Collectively, by disentangling medication-related effects from disease-associated metabolic alterations, these analyses help to delineate the endogenous metabolic signature of PD and underscore the importance of explicitly accounting for antiparkinsonian treatments in metabolomics-based PD research.

Metabolites associated with disease stage and motor symptom burden in PD

We identified multiple plasma metabolites significantly associated with disease stage and motor symptom burden in PD, as assessed by H–Y stage and UPDRS scores. Correlation analysis identified 12 metabolites significantly associated with H–Y stage (|r| > 0.20, p < 0.05), among which N1-Acetylspermidine/N8-Acetylspermidine exhibited the strongest correlation (r = 0.37, p < 0.0001; Table 3). Given that several of these metabolites were also associated with antiparkinsonian medication use in prior analyses, these findings should be interpreted with caution. After adjustment for age, sex, BMI, hypertension status, diabetes status and LEDD, four metabolites remained significantly associated with H–Y stage after FDR correction (p_adj < 0.05, linear regression). Specifically, 4-Acetamidobutanoic acid, p-Cresol glucuronide (Pcg) and ceramide trihexosides (d34:1) (Hex3Cer(d34:1)-Hex3Cer(d18:1/16:0)) were positively associated with advancing disease stage, whereas methionine sulfoxide showed an inverse association (Supplementary Fig. S11).

Table 3.

Metabolites associated with Hoehn and Yahr stage in patients with Parkinson's disease.

Metabolite	Pearson r	β (95% CI)	p value	padj	padja	padjb
N1-Acetylspermidine/N8-Acetylspermidine	0.37, <0.0001	0.070 (0.004–0.137)	0.0390	0.0898	0.3913	<0.0001
4-Acetamidobutanoic acid	0.34, <0.0001	0.080 (0.019–0.141)	0.0104	0.0383	0.0015	0.7757
2-O-Methylascorbic acid	0.31, <0.0001	0.049 (−0.012 to 0.111)	0.1151	0.1535	0.0054	<0.0001
Desaminotyrosine	0.28, <0.0001	0.060 (0.001–0.119)	0.0474	0.0898	0.6138	0.1883
p-Cresol glucuronide	0.25, <0.0001	0.072 (0.015–0.129)	0.0128	0.0383	0.0550	0.0002
Phenylacetylglutamine	0.25, <0.0001	0.057 (−0.002 to 0.116)	0.0599	0.0898	0.0540	0.0002
Hex3Cer(d34:1)-Hex3Cer(d18:1/16:0)	0.24, <0.0001	0.083 (0.024–0.142)	0.0060	0.0361	0.9592	0.7655
Methionine sulfoxide	−0.23, <0.0001	−0.094 (−0.149 to −0.039)	0.0008	0.0097	0.0326	0.4872
Pregnandiol sulfate isomer 1	−0.22, <0.0001	−0.030 (−0.089 to 0.028)	0.3084	0.3700	0.9988	0.6027
N-Butyrylglycine	0.22, <0.0001	0.021 (−0.036 to 0.078)	0.4692	0.5118	0.9304	0.1051
Hex3Cer(d42:2)-Hex3Cer(d18:1/24:1)	0.21, <0.0001	0.059 (−0.001 to 0.119)	0.0556	0.0898	0.0040	0.0026
N-(3-acetamidopropyl) pyrrolidin-2-one	−0.20, <0.0001	0.011 (−0.050 to 0.072)	0.7165	0.7165	0.3913	<0.0001

Abbreviations: CI, confidence interval. Note: Pearson's correlation coefficient (r) and corresponding p values represent univariate associations between each metabolite and Hoehn and Yahr stage. β coefficients with 95% CI are derived from multivariable linear regression models adjusted for age, sex, body mass index, levodopa equivalent daily dose, hypertension status, and diabetes status. p values and FDR-adjusted p values (p_adj) are reported for the regression analyses. In addition, FDR-adjusted p values from differential abundance analyses in Cohort 1 and Cohort 2 are provided separately (p_adj^a and p_adj^b). Associations with FDR-adjusted p values < 0.05 in the regression analysis are highlighted in bold.

Although these metabolites showed limited statistical significance in the drug-naïve PD subgroup, several demonstrated convergent but sub-threshold evidence across analyses. In particular, Pcg was significantly elevated in patients with early-stage PD (p < 0.0001, p_adj = 0.0002) and remained nominally significant in drug-naïve patients (p = 0.0019, p_adj = 0.0550), while also showing a positive association with H–Y stage (p = 0.0128, p_adj = 0.0383). Phenylacetylglutamine (PAGln) exhibited a similar pattern across analyses. These findings suggest that Pcg and PAGln may emerge early in the disease course and exhibit dynamic changes associated with disease progression, warranting further validation in longitudinal studies.

Expanding the analysis to motor symptom severity, six metabolites were significantly associated with UPDRS total scores after covariate adjustment (p_adj < 0.05, linear regression; Supplementary Table S12). Among these, Pcg and PAGln demonstrated the most robust associations, correlating positively with both UPDRS total and Part Ⅲ (motor) scores (r > 0.20, p < 0.0001), reinforcing their link to motor symptom burden. No metabolites were significantly associated with specific motor complications, likely reflecting the low prevalence of these features in this predominantly early-stage population and the stringent multiple-testing correction applied.

Metabolites associated with non-motor symptoms and gastrointestinal dysfunction in PD

For non-motor symptoms, several metabolites showed significant associations with symptom burden across different clinical domains (p_adj < 0.05, linear regression; Supplementary Table S12). With respect to overall non-motor symptom severity, 4-Acetamidobutanoic acid exhibited the strongest association with NMSS total score (β = 4.175, p_adj = 0.0002), followed by 2-Pyrrolidinone, N-Butyrylglycine, and desaminotyrosine. For autonomic dysfunction, as assessed by SCOPA-AUT, a consistent pattern was observed, characterised by the involvement of gut microbiota-derived metabolites. In particular, Pcg, PAGln, and p-Cresol sulfate (Pcs) showed strong positive associations with symptom severity (all p_adj < 0.0005). Notably, these metabolites also showed associations beyond autonomic dysfunction, as both Pcg and PAGln were significantly associated with probable RBD (pRBD, assessed by RBD-HK), indicating shared metabolic involvement across symptom domains.

Extending these findings, we further focused on gastrointestinal manifestations, a key component of autonomic dysfunction in PD. Several metabolites including PAGln, Pcg, Pcs, and trimethylamine N-oxide (TMAO) were significantly associated with worse gastrointestinal symptom scores and were elevated in patients with functional constipation (Table 4; Supplementary Fig. S12A). Collectively, these findings suggest that microbiota-related metabolic alterations are closely associated with non-motor domains, particularly autonomic and gastrointestinal dysfunction.

Table 4.

Metabolites associated with gastrointestinal symptom severity in patients with Parkinson's disease.

Metabolite	Pearson r	β (95% CI)	p value	padj
SCOPA-AUT gastrointestinal domain
p-Cresol glucuronide	0.37, <0.0001	0.592 (0.385–0.798)	<0.0001	<0.0001
Phenylacetylglutamine	0.37, <0.0001	0.569 (0.354–0.784)	<0.0001	<0.0001
p-Cresol sulfate	0.31, <0.0001	0.519 (0.319–0.720)	<0.0001	<0.0001
4-Methylcatechol 2-sulfate	0.22, <0.0001	0.390 (0.192–0.588)	0.0001	0.0004
Trimethylamine N-oxide	0.21, <0.0001	0.365 (0.166–0.564)	0.0003	0.0009
Glu–Trp	0.25, <0.0001	0.322 (0.113–0.531)	0.0026	0.0057
Hex3Cer (d34:1)-Hex3Cer (d18:1/16:0)	0.21, <0.0001	0.310 (0.089–0.531)	0.0061	0.0113
5,6-Dihydrouridine	0.20, <0.0001	0.283 (0.076–0.490)	0.0075	0.0121
NMSS Domain 6 (gastrointestinal symptoms)
p-Cresol glucuronide	0.32, <0.0001	0.801 (0.457–1.146)	<0.0001	0.0001
Phenylacetylglutamine	0.32, <0.0001	0.760 (0.401–1.120)	<0.0001	0.0002
p-Cresol sulfate	0.27, <0.0001	0.662 (0.326–0.999)	0.0001	0.0004
Hex3Cer (d34:1)-Hex3Cer (d18:1/16:0)	0.21, <0.0001	0.577 (0.214–0.941)	0.0019	0.0043
4-Acetamidobutanoic acid	0.27, <0.0001	0.508 (0.131–0.886)	0.0084	0.0152
7-Methylguanine	0.21, <0.0001	0.404 (0.061–0.746)	0.0210	0.0270
N-Butyrylglycine	0.23, <0.0001	0.419 (0.069–0.769)	0.0191	0.0270
Glu–Trp	0.22, <0.0001	0.389 (0.041–0.737)	0.0286	0.0321

Abbreviations: CI, confidence interval; SCOPA-AUT, Scale for Outcomes in Parkinson's Disease for Autonomic Symptoms; NMSS, Non-Motor Symptoms Scale. Note: Pearson's correlation coefficient (r) and corresponding p values represent univariate associations between each metabolite and gastrointestinal symptom scores. β coefficients with 95% CI are derived from multivariable linear regression models adjusted for age, sex, body mass index, levodopa equivalent daily dose, hypertension status, and diabetes status. p values and FDR-adjusted p values (p_adj) are reported for the regression analyses.

Phenylacetylglutamine and p-Cresol glucuronide: exploratory links to multidimensional clinical severity in PD

Building on the above findings, PAGln and Pcg demonstrated the most consistent and broad-spectrum associations across motor and non-motor symptom domains (Fig. 5). Given their close relationship with gut microbiota metabolism, we further evaluated whether the associations were confounded by gastrointestinal status. After stratification by functional constipation, both metabolites remained significantly correlated with UPDRS scores in patients without functional constipation (PAGln: r = 0.23; Pcg: r = 0.24; p < 0.0001), whereas no significant associations were observed in patients with constipation (Supplementary Fig. S12B). This finding suggests that the associations between PAGln, Pcg, and motor severity are unlikely to be solely attributable to gut dysmotility.

Fig. 5.

Exploratory associations of phenylacetylglutamine and p-Cresol glucuronide with clinical phenotypes in patients with PD (n = 477). Associations with binary clinical phenotypes were evaluated using covariate-adjusted logistic regression, with odds ratios (ORs) and corresponding p-values reported. Associations with continuous clinical phenotypes were assessed using covariate-adjusted linear regression, with standardised regression coefficients (β) and p-values provided. Metabolite levels were log-transformed and standardised to Z-scores prior to analysis to ensure comparability across metabolites. False discovery rate (FDR) correction was applied across all phenotype associations for the two metabolites. Horizontal lines indicate 95% confidence intervals for ORs or β coefficients. Abbreviations: UPDRS, Unified Parkinson's Disease Rating Scale; TD, tremor-dominant; PIGD, postural instability-gait difficulty; NMSS, Non-Motor Symptoms Scale; SCOPA-AUT, Scale for Outcomes in Parkinson's Disease for Autonomic Symptoms; RBD, rapid eye movement sleep behaviour disorder; RBD-HK, RBD Questionnaire-Hong Kong; ESS, Epworth Sleepiness Scale; HRS, Hyposmia Rating Scale; HAMD-17, 17-item Hamilton Depression Rating Scale; MMSE, Mini-Mental State Examination; PDQ-39, Parkinson's Disease Questionnaire-39.

The observed associations remained largely consistent in sensitivity analyses with additional adjustment for extended covariates (including lifestyle factors and laboratory proxies of liver and renal function), supporting the robustness of the findings (Supplementary Table S13). Specifically, the association between Pcg and H–Y stage was attenuated to borderline significance (p_adj = 0.0577), whereas associations of Pcg and PAGln with UPDRS motor score, SCOPA-AUT, gastrointestinal symptom scores from the NMSS and SCOPA-AUT, and RBD-HK remained statistically significant. Taken together, these findings delineate distinct yet partially overlapping metabolite clusters associated with motor severity, non-motor symptoms, and gastrointestinal dysfunction, providing convergent evidence for a prominent role of gut microbiota-related metabolism in the clinical heterogeneity of PD.

Discussion

Over the past two decades, metabolomics has emerged as a powerful systems biology approach for capturing disease-associated biochemical alterations and identifying potential biomarkers in PD. However, heterogeneity across cohorts, medication effects, and limited reproducibility have constrained clinical translation. In this study, we systematically profiled plasma metabolic alterations in well-characterised patients with PD, providing a comprehensive view of metabolic remodelling and a basis for biomarker discovery and clinical stratification.

By integrating two independent datasets of early-stage PD spanning drug-naïve and treated populations, we identified consistently altered metabolites, represented by 12-HETE, spermine, and niacinamide, reflecting key metabolic disturbances at disease onset. Prioritising untreated subjects and evaluating medication-associated metabolites allowed partial separation of disease-related metabolic signatures from antiparkinsonian medication-related effects. Beyond univariate signatures, data-driven feature selection enabled the construction of metabolite-based diagnostic models, which showed stable discriminative performance across algorithmic frameworks in internal validation. Importantly, a compact metabolite panel maintained consistent performance in sensitivity analyses stratified by medication status, suggesting its potential as a clinically interpretable biomarker set for PD diagnosis. Compared with previous metabolite-based diagnostic models in PD, our model showed performance within the upper range of reported results (Supplementary Table S14). However, such comparisons should be interpreted with caution given inter-study heterogeneity.

12-HETE, a downstream product of arachidonic acid metabolism, was significantly elevated in patients with PD, suggesting enhanced lipid oxidation and inflammatory activity.²⁹ This finding is broadly consistent with previous reports implicating dysregulated polyunsaturated fatty acid metabolism and eicosanoid signalling in PD, although studies of hydroxyeicosatetraenoic acids (HETEs) have yielded heterogeneous results across cohorts and biological matrices.^14,30, 31, 32, 33 Similarly, altered spermine and spermidine levels indicate dysregulation of polyamine metabolism, a pathway linked to neuronal stress adaptation and autophagy regulation, both of which are central to neurodegenerative processes in PD.^34,35 Despite heterogeneous findings across studies and biological matrices,^13,36, 37, 38 converging evidence collectively supports polyamine metabolism as a diagnostic biomarker source, mechanistic pathway, and potential therapeutic target. In addition, changes in niacinamide, trigonelline and related metabolites point to perturbations in nicotinamide adenine dinucleotide (NAD⁺) metabolism and energy homoeostasis, which aligns with prior observations of mitochondrial dysfunction and impaired bioenergetic regulation in PD.³⁹ Collectively, these findings converge with existing literature to support coordinated dysregulation of lipid signalling, polyamine metabolism, and cellular energy pathways in PD pathophysiology. Further studies are warranted to elucidate the mechanisms underlying these metabolic pathways in PD and to determine whether they represent upstream drivers or downstream consequences of neurodegeneration.

KEGG pathway analysis further revealed significant perturbations in sphingolipid metabolism, TCA cycle, arginine biosynthesis, and several amino acid metabolism pathways. These pathways are closely linked to pathogenic processes in PD, such as mitochondrial energy dysfunction and neuroinflammatory regulation.40, 41, 42 Among these, sphingolipid metabolism showed the strongest enrichment, consistent with previous metabolomic and mechanistic studies.^40,43 Notably, several PD-associated genes such as GBA1, PLA2G6, GALC, and SMPD1 are directly involved in sphingolipid metabolism, supporting its central role in PD pathogenesis.44, 45, 46 These observations support sphingolipid dysregulation as a central metabolic feature of idiopathic PD, not restricted to genetic forms.

L-DOPA treatment significantly influenced metabolic pathways including tyrosine, tryptophan, and polyamines. Elevated levels of L-tyrosine and desaminotyrosine with increasing LEDD likely reflect feedback regulation of dopamine biosynthesis under exogenous L-DOPA supplementation. In the polyamine pathway, we observed a potential dose-dependent accumulation of N1-acetylspermidine/N8-acetylspermidine accompanied by reduced degradation products. This pattern suggests an inhibition of spermidine/spermine-N1-acetyltransferase activity, which may reduce oxidative byproducts from polyamine oxidation.³⁵ In parallel, reductions in indole metabolites indicate a shift toward kynurenine metabolism in tryptophan pathway, which is consistent with previous evidence of L-DOPA-induced kynurenine pathway activation.⁴⁷ Collectively, these findings suggest that dopaminergic therapy is associated with broad alterations in amino acid and polyamine metabolism, and may partially influence metabolites previously proposed as PD biomarkers. These observations highlight the importance of accounting for medication status in metabolomic studies and support the inclusion of drug-naïve cohorts to better delineate disease-related metabolic alterations.

One of the most salient findings is the elevation of gut microbiota-derived metabolites PAGln and Pcg in PD plasma, which correlated with clinical severity. Our findings are consistent with prior metabolomic studies,^14,48, 49, 50 and extend them by demonstrating that these alterations are detectable at early disease stages, and are robustly associated with a broad range of motor and non-motor clinical features. Notably, several studies have linked these metabolites to gut microbiota composition, supporting their microbial origin.^48,50 Together, these observations place PAGln and Pcg within a growing body of evidence linking gut dysbiosis to PD pathogenesis.⁵¹

The observed metabolite changes suggest a PD-associated dysbiosis in which proteolysis replaces carbohydrate fermentation as the dominant microbial metabolic pathway, a shift corroborated by metagenomic analyses.^48,52,53 This metabolic reprogramming is driven by an enrichment of mucin-degrading and proteolytic taxa, which generate phenylacetic acid and p-Cresol that are conjugated in the liver into PAGln, Pcg, and Pcs.^48,54 Excessive mucin degradation weakens the protective gut barrier, while the toxic derivatives directly disrupt epithelial integrity.54, 55, 56 This cascade of microbial metabolic change and barrier breakdown likely underlies the prevalent gastrointestinal dysfunction in PD. From another perspective, the strong association of these metabolites with constipation hints at a self-reinforcing cycle: slowed colonic transit increases protein retention, which further fuels proteolytic metabolism and metabolite production, exacerbating mucosal injury and dysfunction.^48,57 Beyond the gut, accumulating evidence indicates that these metabolites may exert systemic and neurobiological effects. Experimental studies suggest that PAGln and p-Cresol derivatives may disrupt blood–brain barrier integrity and modulate neuroinflammation, oxidative stress, and mitochondrial function.58, 59, 60, 61, 62 While causal inference cannot be established in this study, the broad associations observed here support the hypothesis that gut microbiota-derived metabolites may be involved in disease-relevant biological processes rather than merely reflecting downstream effects of PD. In particular, PAGln and Pcg emerge as key components of the microbiota-metabolite-brain axis linked to clinical heterogeneity in PD, warranting further mechanistic investigation.

A key strength of this study is its cohort design focussing on early-stage PD, with a substantial proportion of drug-naïve patients, which reduces confounding from disease progression and antiparkinsonian medication effects. Medication-associated metabolites were also evaluated to help disentangle disease-related metabolic alterations from treatment effects. The use of two independent datasets, together with a cross-dataset consensus strategy, enhances the robustness and reproducibility of the identified metabolic signatures. In addition, integration of complementary LC-MS/MS platforms enabled broad metabolome and lipidome coverage, providing a more comprehensive characterisation of PD-associated metabolic disturbances. Finally, systematic integration of metabolomic profiles with multidimensional clinical phenotyping strengthens the biological interpretability and translational relevance of the findings.

Several limitations of this study highlight priorities for future work. First, model development relied on pooled random splitting, which enables internal validation but does not capture cohort-specific heterogeneity. As such, performance estimates may be optimistic and require external validation in independent, multicentre cohorts with diverse populations. Second, the cross-sectional design precludes inference on temporality and causality. The identified metabolites should therefore be interpreted as correlates of established disease rather than predictive biomarkers. Longitudinal studies, particularly in prodromal and at-risk populations, are needed to evaluate early diagnostic potential. In addition, studies with extended follow-up duration and repeated assessments are required to establish the utility of baseline metabolites as prognostic biomarkers of disease progression. Third, the clinical specificity of the metabolomic signatures remains limited. Although an independent essential tremor group was included for exploratory evaluation, it does not adequately represent disease controls for differential diagnosis. Future studies should incorporate atypical parkinsonism and other neurodegenerative conditions and adopt multi-class modelling strategies to better reflect real-world diagnostic scenarios. Fourth, residual confounding from medication use, diet, comorbidities, and environmental exposures cannot be fully excluded, as these factors may influence circulating metabolites and exhibit overlapping metabolic signatures with PD. In addition, plasma metabolomics reflects systemic rather than brain-specific processes. Integration with complementary modalities, such as cerebrospinal fluid profiling and other multi-omics approaches, may improve biological specificity and mechanistic resolution. Finally, mechanistic interpretation remains limited. Associations involving gut microbiota-derived metabolites should be considered hypothesis-generating and require experimental validation. Despite these limitations, the consistent findings across datasets provide a foundation for future biomarker validation and mechanistic investigation.

In conclusion, this study delineates a distinctive plasma metabolic signature of early PD, highlighting sphingolipid dysregulation and other pathway-level perturbations. Metabolite-based diagnostic models showed promising performance, supporting their potential utility for diagnosis pending further external validation. Moreover, gut microbiota-derived metabolites such as PAGln and Pcg were consistently associated with clinical heterogeneity, suggesting a potential link between microbial metabolism and both motor and non-motor symptoms. Together, these findings advance our understanding of PD-related metabolic alterations and provide a foundation for future biomarker and mechanistic studies.

Contributors

Jifeng Guo and Yige Wang conceived and designed the study. Jifeng Guo, Beisha Tang, Lu Shen, Jinchen Li, Xinxiang Yan, Junling Wang, Hong Jiang, Qian Xu, Qiying Sun, Zhenhua Liu, Jieqiong Tan, Chunyu Wang, Lifang Lei, and Heng Wu organised the recruitment, managed the cohorts, and supervised the project. Yige Wang, Yaqin Xiang, Xiurong Huang, Lizhi Li, Taole Li, Xuxiang Zhang, Yuxuan Hu, and Jiabin Liu collected blood samples, performed clinical assessments, and managed the datasets. Yige Wang carried out data processing, statistical analyses, and wrote the first draft of the manuscript. Yige Wang, Jifeng Guo, Jinchen Li, and Beisha Tang interpreted the results. Jifeng Guo, Yige Wang, and Beisha Tang had full access to the original data and verified the underlying data. All authors critically revised the manuscript, provided feedback on the analyses, read and approved the final version of the manuscript.

Data sharing statement

The metabolomics datasets supporting the findings of this study have been deposited in the China National GeneBank Sequence Archive (CNSA) under accession number CNP0009726 (https://db.cngb.org/data_resources/project/CNP0009726). De-identified participant data underlying the results reported in this article will be available from the corresponding author upon reasonable request and after approval of a data access proposal. Data requests should be directed to Jifeng Guo (guojifeng@csu.edu.cn).

Code availability

The code that supports the findings of this study is publicly available at https://github.com/ygwang1/PD-metabolomic-signatures.

Declaration of interests

The authors declare no competing financial or non-financial interests.