Integrative multi-omics analysis reveals host–microbiome metabolic alterations and candidate biomarkers in Parkinson’s disease Authors: Qiang Zhao, Shujia Zuo, Sufang Liu, Juan Wang, Jie Tang, Xuanxuan Zou, Yahui Leng, Xianhong Li, Min Zhou, Jing Tian, Puqing Wang Published: 2026 ABSTRACT AbstractAlterations in the gut microbiome have been increasingly implicated in Parkinson’s disease (PD), but the associated metabolic changes remain incompletely understood. Here, we applied an integrative multi-omics approach combining shotgun metagenomic sequencing and untargeted LC–MS-based plasma metabolomics to investigate host–microbiome alterations in PD. Fecal and plasma samples were collected from 30 PD patients and 30 healthy spouse controls. Significant differences in microbial diversity and taxonomic composition were observed between the two groups. Taxonomic profiling revealed marked gut microbial dysbiosis in PD, including altered abundances of Phocea massiliensis, Bacteroides sp900766005, and Alistipes_A indistinctus. Metabolomic analysis identified 86 significantly altered plasma metabolites, including glycerophospholipids, indoleacetic acid, and kynurenic acid. Integrative pathway analysis suggested links between microbial functional alterations and host metabolic changes. Machine-learning analyses identified three biomarker panels that distinguished PD patients from controls in validation datasets, with the highest area under the curve (AUC) reaching 0.92. In silico molecular docking further suggested potential interactions between several metabolite biomarkers and alpha-2-macroglobulin (A2M) or the human Bact spliceosome. Overall, these findings provide an integrative view of host–microbiome metabolic alterations associated with PD and highlight candidate biomarkers and exploratory host–metabolite associations for further investigation.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12866-026-05168-4.Keywords: Parkinson’s Disease, Gut Microbiome, Gut-Brain Axis, Metabolomics, Biomarker Introduction Parkinson’s disease (PD) is the second most prevalent neurodegenerative disorder, characterized by the progressive degeneration of dopaminergic neurons within the substantia nigra of the brain [1, 2]. The clinical manifestations of PD encompass a broad spectrum of motor and non-motor symptoms, including tremors, rigidity, bradykinesia, cognitive impairment, and constipation [3, 4]. Despite significant research efforts, the precise etiology and pathogenesis of PD remain unclear, posing substantial challenges in the development of effective therapeutic strategies. Recent studies have highlighted the involvement of gut microbiome alterations in various diseases, including PD [5–11]. Our previous research, utilizing 16 S rRNA amplicon sequencing and shotgun metagenomic sequencing, demonstrated significant changes in the composition and function of the gut microbiome in PD patients compared to healthy controls [6, 9, 11]. These alterations in the gut microbiota are associated with increased intestinal permeability, chronic inflammation, and the production of neurotoxic metabolites, suggesting a potential link to PD pathogenesis. The gut microbiome influences host metabolism, particularly through its impact on metabolite abundance. Several studies have investigated the metabolic changes driven by gut microbiota alterations. Hertel, Harms et al. were the first to use metabolic modeling of shotgun metagenomic sequencing data to explore the metabolomic changes associated with gut microbiota alterations in PD [12]. Their study identified that Bilophila wadsworthia and Akkermansia muciniphila contribute to disruptions in sulfur metabolism, influencing the levels of PD-associated taurine-conjugated bile acids. Another study employing a 16 S rRNA-based metabolic model found that Akkermansia muciniphila impacted the levels of nine metabolites, including the neurotransmitter gamma-aminobutyric acid (GABA), which was significantly altered between PD patients and healthy controls [13]. Furthermore, Rosario et al. used shotgun metagenomics-based metabolic modeling, revealing multiple host metabolites associated with gut microbiota [14]. These studies have begun to reveal potential host–microbiome metabolic associations relevant to the gut-brain axis in PD. While predictive modeling has provided preliminary insights into gut–brain axis-related metabolites, its limited resolution underscores the need for untargeted metabolomic approaches to more comprehensively characterize PD-associated metabolic perturbations. To investigate host–microbiome interactions in PD and their effects on host metabolic alterations, we employed shotgun metagenomic sequencing and untargeted LC–MS-based metabolomics in both hypothesis-driven and exploratory analyses. Our primary hypothesis-driven objectives were to identify specific gut microbiota and plasma metabolite alterations associated with PD compared to healthy controls, and to construct predictive models capable of accurate disease classification based on these microbial and metabolic signatures. As an exploratory aim, we performed comprehensive in silico molecular docking analysis to investigate potential metabolite-protein interactions that may be associated with PD-related metabolic alterations, thereby providing exploratory structural context for the observed multi-omics findings. In this study, we investigated the gut metagenomic and plasma metabolomic profiles of 30 PD patients and 30 matched healthy spouse controls recruited from the Department of Neurology at Xiangyang No. 1 People’s Hospital. Fecal and plasma samples were collected for analysis. Through these high-throughput omics approaches, we aimed to characterize host–microbiome metabolic alterations associated with PD and to generate hypotheses for future mechanistic studies. Our findings could provide insights that may contribute to novel therapeutic strategies, though further research is needed to validate these results and explore their clinical implications in larger cohorts. Materials and methods Study participants This study involved 2 independent cohorts recruited from the Department of Neurology, Xiangyang No. 1 People’s Hospital, Xiangyang city, Hubei Province, China. Cohort A consisted of 30 patients with Parkinson’s disease (PD) and their 30 healthy spouse controls. PD patients were included if they met the 2015 International Movement Disorders Society diagnostic criteria for PD, with diagnoses confirmed by two neurologists. Exclusion criteria for PD patients were: (1) secondary Parkinsonism, Parkinson-plus syndromes, or other neurodegenerative diseases; (2) serious comorbidities such as cancers, gastrointestinal diseases, liver or kidney dysfunction, gout, etc.; (3) use of antibiotics or probiotics within the previous three months. Healthy controls were recruited from the spouses of PD patients to minimize variability in environmental factors such as diet and lifestyle. The exclusion criteria for healthy controls were: (1) Serious disease such as neurodegenerative diseases, cancers, gastrointestinal diseases, liver or kidney dysfunction, gout; and (2) use of antibiotics or probiotics within the previous three months. Cohort B consisted of an additional 25 PD patients and their 25 healthy spouse controls, recruited under the same criterion as Cohort A. No significant differences in PD-related clinical characteristics were observed between cohort A and cohort B (Supplementary Table S1). Demographic information, including age and gender, was collected for all participants. Parkinson’s disease clinical symptoms were assessed using the Unified Parkinson’s Disease Rating Scale III (UPDRS-III) [15], Hoehn and Yahr Scale [16], Roland Morris Disability Questionnaire (RMDQ) [17], Montreal Cognitive Assessment (MoCA) [18], Hamilton Depression Rating Scale (HAM-D) [19], and Hamilton Anxiety Rating Scale (HAM-A) [20]. All assessments were conducted by an experienced neurologist. Fecal sample collection and shotgun metagenomic sequencing Participants collected fecal samples, which were promptly transported to the hospital. Samples were initially flash-frozen in liquid nitrogen and then stored at -80 °C. DNA extraction and metagenomic sequencing were performed by Novogene Co., Ltd. (Beijing, China). A total of 0.2 µg DNA was used for sequencing library preparation. Shotgun metagenomic sequencing was performed on the DNBSEQ-T7 platform with a 150 bp paired-end read length. Plasma sample collection and untargeted LC–MS metabolomic profiling Plasma samples were prepared and stored according to the manufacturer’s procedures. Untargeted LC–MS-based metabolomic profiling was performed at the Beijing Genomics Institute (BGI-Shenzhen, Shenzhen, China). Metabolite extraction followed established protocols [21], using 100 µL of plasma. The extraction involved adding 300 µL of precooled methanol and acetonitrile (2:1, v/v), followed by the addition of internal standards for quality control. Samples were vortexed, incubated at -20 °C for 2 h, centrifuged, and the supernatants were freeze-dried. The metabolites were resuspended in 150 µL of 50% methanol and centrifuged before LC–MS analysis. Quality control samples were prepared by pooling equal volumes from all samples. Chromatographic separation was achieved using a Waters 2D UPLC system (Waters, USA) coupled to a Q-Exactive mass spectrometer (Thermo Fisher Scientific, USA). Separation was performed on a Waters ACQUITY UPLC BEH C18 column (1.7 μm, 2.1 × 100 mm; Waters, USA) maintained at 45 °C. In positive ion mode, the mobile phases consisted of 0.1% formic acid in water (A) and acetonitrile (B), whereas in negative ion mode, the mobile phases consisted of 10 mM ammonium formate in water (A) and acetonitrile (B). The flow rate was set at 0.35 mL/min with a 5 µL injection volume. Mass spectrometric settings included a spray voltage of 3.8/-3.2 kV and a full scan range of 70–1050 m/z with a resolution of 70,000. Raw LC–MS/MS data were processed using Compound Discoverer software (version 3.3, Thermo Fisher Scientific, USA), including peak extraction, retention time alignment, adduct grouping, missing value imputation, background peak annotation, and metabolite identification. Metabolite annotation was performed based on multiple databases, including HMDB, KEGG, LipidMaps, mzCloud, and ChemSpider, with the following parameters: precursor mass tolerance < 5 ppm, fragment mass tolerance < 10 ppm, and retention time tolerance < 0.2 min. The processed data were further analyzed using metaX for data preprocessing. Signal intensities were normalized using probabilistic quotient normalization (PQN), followed by batch effect correction using quality control-based robust LOESS signal correction (QC-RLSC). To ensure data quality, features with a coefficient of variation (CV) greater than 30% in quality control samples were excluded from downstream analyses. Shotgun metagenomic sequencing analysis Quality control of sequencing data involved removing reads containing adapter sequences or with more than 10% unidentified bases. Reads were processed using Trimmomatic (version 0.39) with specified parameters [22]. Host and viral DNA contamination were removed by aligning reads to the human hg38 genome and phiX174 reference genome using bowtie2 (version 2.4.4) [23]. Unaligned reads were used for downstream analysis. Taxonomic profiling of the gut microbiome was performed using Kraken2 [24] (version 2.1.2) with the Unified Human Gastrointestinal Genome [25] (UHGG, version 2.0) database. Bracken (version 2.5.3) was used to estimate abundance at species, genus, and family levels [26]. Rare taxa were filtered using the permutation filtering method with the PERFect [27] R package. Alpha diversity indices (Shannon, Simpson, Invsimpson) and beta diversity based on Bray-Curtis distance were analyzed using the vegan R package (version 2.6-4). Differential abundance analysis was conducted using the ALDEx2 [28] R package with parameters “test = ‘t’, mc.samples = 1000, paired.test = TRUE”. The Benjamini-Hochberg correction was applied to Wilcoxon results for significant taxa. For gene and pathway level analysis, Wilcoxon signed-rank test and Spearman’s rank correlation were used, with analysis performed in the Scipy Python package (version 1.10.0) [29]. Functional profiling was performed by using HUMAnN3 [30] (version 3.0.1) with UniRef90 gene-family database (uniref90_v201901b). The gene-family level RPKs (reads per kilobase) value was normalized to CPMs (copies per million) value before the downstream analysis and procedures. By following the HUMAnN3 pathway group methods, we first regroup the gene level abundance to MetaCyc reaction level, and then regroup reaction level abundance to MetaCyc pathway level. Finally, we obtained the latest version of MetaCyc (version 26) pathway abundance data. Gene-reaction-pathway information was extracted from the MetaCyc website on February 1, 2022. Metabolomic data processing and analysis LC–MS data processing was performed by the Beijing Genomics Institute (BGI-Shenzhen, Shenzhen, China). Raw metabolomic data were first processed using Compound Discoverer software (version 3.3, Thermo Fisher Scientific, USA) with the following parameters: parent ion mass deviation < 5 ppm, fragment ion mass deviation < 10 ppm, and retention time deviation < 0.2 min. The results were then imported into metaX for further analysis [31]. This included normalization using PQNmethod, batch effect correction via QC-RLSC method, and calculation of the CV. Metabolites with a CV exceeding 30% were excluded. Finally, metabolites were annotated using mzCloud, HMDB (version 4.0), KEGG, and ChemSpider databases. The log2 peak intensity data were used for subsequent analysis. Because all PD participants were receiving anti-parkinsonian medications, and drug ADME (absorption, distribution, metabolism, and excretion) may have a greater impact on metabolite profiles than on the gut microbiome, we converted medication dosages to levodopa equivalent daily dose (LEDD), then applied generalized linear mixed-effects model (GLMM) with LEDD dosage (mg/day), sex and age as random effects to eliminate the influence of irrelevant variables. The Benjamini-Hochberg method was used to control for false discovery rate in multivariable corrections. Microbial association network construction Differential abundance microbiota at species and genus level were used for network construction. We applied NetCoMi [32] package (version 1.1.0) with SparCC [33] association measurement for network construction and comparison between groups. Clusters were identified using ‘cluster_optimal’ algorithm and eigenvector measurement. Construction and validation of metabolite-based PD prediction models Two independent cohorts were recruited. Cohort A included 30 PD patients and 30 healthy spouse controls, while cohort B included 25 PD patients and 25 healthy spouse controls. Untargeted LC–MS-based metabolomic profiling was performed on both cohorts. Feature selection for model construction involved GLMM filtering mentioned above and OPLS-DA (Orthogonal Projections to Latent Structures Discriminant Analysis) with VIP (the variable importance in projection) values over 1. Random forest and logistic regression models were used, with five-fold cross-validation to evaluate model performance. Metagenome gene abundance inference and microbial gene-based disease prediction model construction Clean reads were aligned to Global Microbial Gene Catalog v1.0 Human Gut all gene catalog using salmon aligner [34]. The catalog contains 52,544,908 genes, which represent up-to-data the most comprehensive gut microbial gene catalog. The TPM (Transcripts Per Kilobase Million) value output from salmon was used for the next analysis. We applied two methods to select microbial gene biomarkers. First, we conducted differential gene abundance analysis by using Wilcoxon signed-rank test and corrected multiple comparison by using false discovery rate. Then, we applied LassoNet [35] method to differential abundance genes and selected top 25 genes according to the feature importance value evaluated by LassoNet. The 25 microbial genes were used as features for construction of PD prediction models. Next, we applied random forest classification to predict the PD status. To evaluate the model performance, a fivefold cross-validation approach was applied to internal validation. External shotgun metagenomic sequencing data with 40 PD and pairing healthy spouses based on previous research [7] were downloaded with NCBI BioProject accession PRJNA433459. The same procedure was adopted for preprocessing the sequence of the external data. Then, we use our data as train set and test the model performance with external dataset. Blind molecular docking Protein structures were retrieved from the RCSB website as of March 27, 2024, yielding a total of 67,440 structure files [36]. We used pymol (version 2.5.0) to remove the hetero atoms including nucleic acid, solvent, metals from the structure and determine the coordinate center of the protein. The metabolites 3D structures were retrieved from PubChem database [37]. All structures were converted to pdbqt files by using ADFRsuite (version 1.0) with default parameters [38]. We applied Autodock vina (version 1.2.5) to perform blind molecular docking with parameters “--spacing 1, --exhaustiveness 16 --seed 99 --size_x 126 --size_y 126 --size_z 126” [39]. Protein structures were prioritized using two criteria: (i) a predicted binding affinity lower than − 10 kcal/mol, which is commonly used as a threshold for strong ligand–protein interactions in virtual screening studies, or (ii) inclusion among the top ten highest-affinity docking hits for a given metabolite. These criteria were used to prioritize protein structures with relatively higher predicted docking affinity from the large set of screened proteins. All analyses were conducted in R (version 4.1.3) and Python 3.10.9 (Anaconda 202303 distribution). The Benjamini-Hochberg false discovery rate was applied for multivariable correction, with a significance threshold of P < 0.05. A schematic overview of the study design and analytical workflow is shown in Fig. 1. Schematic overview of the study design and analytical workflow Results Demographic characteristic of participants A total of 30 Parkinson’s disease (PD) patients and their healthy spouse controls were enrolled in the study. Age was well-matched between the two groups, while a significant difference in sex distribution was observed between the PD and healthy spouse controls (HP) groups. Due to the frequent use of levodopa among PD patients, a notable difference in drug intake was also detected between the two groups. The detailed demographic characteristics of the participants are presented in Table 1. Table 1Participant demographics at the time of plasma and stool samplingCharacteristicOverall(N = 60)aPD(N = 30)aHP(N = 30)aP-valueGender0.004 Female31 (52%)10 (33%)21 (70%) Male29 (48%)20 (67%)9 (30%)Age69 (64, 73)69 (64, 74)69 (64, 73)0.8UPDRS-III35 (28, 42)35 (28, 42)NANAHoehn-Yahr stage2.50 (2.00, 3.00)2.50 (2.00, 3.00)NANARBDQ score13 (7, 27)13 (7, 27)NANAMoCA21 (18, 24)21 (18, 24)NANAHAM-A7 (3, 12)7 (3, 12)NANAHAM-D7 (4, 12)7 (4, 12)NANADisease Duration8 (5, 11)8 (5, 11)NANALEDD63 (0, 437)458 (375, 719)0 (0, 0)< 0.001Demographic data are presented along with P-values derived from Student’s t-test for continuous variables and chi-square tests for categorical variablesAbbreviations: UPDRS- III The Unified Parkinson’s Disease Rating Scale III, RBDQ Rapid eye movement sleep behavior disorder questionnaire, MoCA Montreal Cognitive Assessment, HAM-A Hamilton Anxiety Rating Scale, HAM-D Hamilton Depression Rating Scale, LEDD Levodopa Equivalent Daily DoseaMedian (IQR) Participant demographics at the time of plasma and stool sampling Demographic data are presented along with P-values derived from Student’s t-test for continuous variables and chi-square tests for categorical variables Abbreviations: UPDRS- III The Unified Parkinson’s Disease Rating Scale III, RBDQ Rapid eye movement sleep behavior disorder questionnaire, MoCA Montreal Cognitive Assessment, HAM-A Hamilton Anxiety Rating Scale, HAM-D Hamilton Depression Rating Scale, LEDD Levodopa Equivalent Daily Dose aMedian (IQR) Alterations in microbial diversity and taxonomic changes in PD After filtering out low-abundance taxonomies, a total of 991 species, 380 genera, and 44 families were identified. Alpha-diversity indices, including Shannon’s index, Simpson index, and Invsimpson index, were significantly higher in PD compared with healthy spouse controls (P = 1.9e-4, 9.9e − 5, 5.1e-4; effect size r = 0.46, 0.42, 0.42 respectively) (Fig. 2a). Additionally, Bray-Curtis distance analysis indicated substantial differences in beta-diversity between the PD and control groups (P = 0.00042) (Fig. 2b). Distinct gut microbiome and plasma metabolome profiles in parkinson’s disease and healthy individuals. a The boxplot of α-Diversity indices including Shannon, Simpson, and Invsimpson between the Parkinson’s Disease (PD) and healthy control (HP). (***P < 0.001, Wilcoxon rank-sum test; box represents median [25th-75th percentile]). b Principal Coordinates Analysis (PCoA) based on Bray-Curtis distance at the species level demonstrates significant differences in microbiome composition between HP and PD individuals (P = 0.005, F-value = 2.0777, df = 59; Adonis test). c The class counts of differential plasma metabolites. d Clustering analysis of the altered 109 species across all participants is presented, with the red lines indicating negative associations between species and green lines indicating positive associations. e Metabolite pathway network derived from differential profiling and MetaCyc database records. Circles denote metabolites, while rectangles represent metabolic pathways. Red lines indicate consumption, whereas grey lines indicate both consumption and production relationships between pathways and metabolites Using ALDEx2 to compare microbiome abundance between PD and HP groups, 109 species exhibited significant alterations in abundance. Specifically, 30 species showed increased abundance, while 79 species exhibited decreased abundance in the PD group compared to controls (Supplementary Table S1). Approximately half of the altered species belong to four genera: Bacteroides, Phocaeicola, Prevotella, and CAG-83. Some of these taxonomic changes align with previous findings. For instance, at the species level, Faecalibacterium prausnitzii was decreased in PD, while Anaerotruncus rubiinfantis and Alistipes indistinctus were increased, consistent with a meta-analysis of PD metagenomic studies. Additionally, Christensenella hongkongensis was found to be increased in PD, in agreement with a Russian study [40]. At the genus level, Prevotella showed decreased abundance in PD, contrasting with findings from other studies [40–43]. Furthermore, ALDEx2 analysis indicated that gender did not contribute to significant taxonomic alterations in PD (Supplementary Table S1). Microbial association network Network analysis revealed three clusters of differentially abundant species (Fig. 2d, Supplementary Fig. 1). Taxa within each cluster exhibited positive correlations with one another, while opposing trends were observed between taxa in different clusters. Species in Cluster-1 were more abundant in PD, whereas species in Cluster-2 and Cluster-3 exhibited decreased abundance. Cluster-1 species were predominantly associated with CAG-83 at the genus level and Oscillospiraceae at the family level. Cluster-2 species were more abundant in Phocaeicola and Bacteroides at the genus level and Bacteroidaceae at the family level. Cluster-3 species were associated with Lachnospiraceae at the family level. These findings align with the differential abundance results at the genus and family levels, suggesting that specific species may contribute to taxonomic shifts at genus or family level in PD. Notably, Phocaeicola and Bacteroides, which are dominant genera in Cluster-2, possess complete pathways for short-chain fatty acid (SCFA) production and glycan decomposition, processes critical for gut health that are dysregulated in PD [44–46]. Plasma metabolome characterization in PD The effects of the gut microbiome on the host are thought to be mediated, at least in part, through microbial and host metabolic interactions. To explore these effects, we applied LC–MS-based untargeted metabolite profiling on plasma samples. After adjusting for the potential influence of anti-parkinsonian drugs, 86 metabolites were significantly altered in PD patients compared to controls (26 with lower abundance and 66 with higher abundance) (Supplementary Table S1). Some identified altered metabolites, such as Furaltadone and Tulathromycin A, were excluded from further analysis due to potential drug-induced biases caused by medication use. Among the 86 altered metabolites, 16 were glycerophospholipids (Fig. 2c), which have been implicated in PD-related metabolic disturbances [47, 48]. Cluster analysis revealed that phosphatidylcholine (PC) and lysophospholipids (LyP) metabolites tended to cluster together. Notably, plasma levels of diphenyl sulfone (P < 0.001, fold change = 1.089), a compound commonly used as an ovicide and acaricide, were significantly elevated in PD patients. While various pesticides have been implicated as potential risk factors for PD, to our knowledge, this compound has not been widely reported in PD and warrants further investigation [49, 50]. Additionally, indoleacetic acid (IAA, P < 0.001, fold change = 0.989) and kynurenic acid (KYNA, P < 0.001, fold change = 0.978), located in brain tissue, were found to be decreased in PD, aligning with previous studies [51, 52]. Interestingly, trehalose (P < 0.001, fold change = 1.019), which has been discussed as a potential therapeutic candidate in PD-related studies, was increased in PD [53]. Potential regulatory mechanisms of differential metabolites To further explore the metabolic associations linked to gut microbiome alterations in PD, we performed extended pathway analysis using HUMAnN3 on shotgun metagenomic data from fecal samples of 30 PD patients and 30 healthy spouse controls. This analysis identified 369 differentially abundant pathways (P < 0.05, FDR-corrected), spanning processes such as energy metabolism, biosynthesis, and degradation. Related metabolic pathways for significantly altered plasma metabolites were retrieved from the MetaCyc database, after excluding exogenous compounds from downstream analyses to reduce their impact on biological interpretation. We then intersected these metabolite-associated pathways with the 369 differential microbial pathways to identify potential links between microbial functional alterations and host metabolic changes in PD. Notably, the reduction in Citrate, a key tricarboxylic acid (TCA) cycle intermediate, was associated with the upregulation of five pathways, including PWY-6574, a citrate-consuming pathway. Some metabolites showed complex patterns. For example, trehalose, which was increased in PD, intersected with two pathways showing opposite directions of change: PWY-7900 was upregulated, whereas PWY-7269 was downregulated (Fig. 2e). In addition, these metabolite-related pathways showed significant correlations with differentially abundant species, suggesting potential microbiome–metabolite associations. A comprehensive summary of these findings, including metabolite-pathway interaction types and direction of change, is presented in Fig. 2e and Supplementary Table S1. Overall, these findings highlight the complex and potentially heterogeneous associations between microbial functional alterations and host metabolic changes. Association of the altered microbes and metabolites with clinical parameters To further investigate the relationships among gut microbiota, altered metabolites, and clinical features in PD, we assessed correlations between clinical parameters and differentially abundant species and metabolites. Our analysis revealed that the PD severity indices, specifically the Hoehn-Yahr stage and the UPDRS-III scale, were positively correlated with most species in Cluster-1, while some species in Cluster-3 were negatively correlated with these clinical parameters. Notably, species in Cluster-1 were more abundant in PD, while those in Cluster-3 were less abundant, reflecting their association with disease severity. These findings suggest that species in Cluster-1 and Cluster-3 may be associated with PD severity. Furthermore, four age-related species were identified, all of which were grouped in Cluster-2 (Fig. 3a). Correlation of disease-related species and metabolites with clinical parameters. a Heatmap displaying the correlation between altered species and clinical parameters, including age, UPDRS-III, Hoehn-Yahr stage, RBDQ score, MoCA, HAM-A, and HAM-D. b Heatmap illustrating the correlation between altered metabolites and clinical indices. Correlations are assessed using Spearman rank-order analysis, with *P < 0.05 considered significant. Positive correlations are indicated in red, while negative correlations are shown in blue. Metabolites marked with a hash (#) represent abbreviated names, with the corresponding full names provided in Supplementary Table S1 Regarding plasma metabolites, three metabolites — Cer(d15:1/16:0), PE(0:0/22:1(13Z)), and LysoPC(P-16:0/0:0) — were positively correlated with the MoCA scale, indicating their potential relevance to cognitive function in PD. Moreover, five metabolites showed significant correlations with the UPDRS-III score. Notably, Citrate, which was reduced in PD, showed a positive correlation with UPDRS-III, whereas, Cer(d15:1/16:0), which was elevated in PD, showed a negative correlation with UPDRS-III. Additionally, the metabolites diphenyl sulfone, calycin, and thione# were all elevated in PD and positively correlated with UPDRS-III. A similar pattern was observed concerning the Hoehn-Yahr stage, where four metabolites showed significant correlations. Two metabolites, calycin and thione#, were positively correlated with Hoehn-Yahr stage, while 2-mercaptoethanol and 1-(2-methoxy-nonadecanyl)-sn-glycero-3-phosphoserine exhibited negative correlations (Fig. 3b). These findings highlight potential links between specific gut microbiota species, metabolites, and disease severity in PD, although further validation in larger cohorts is required. Gut microbiome gene and plasma metabolites biomarker of PD Feature selection methods were applied to identify potential PD biomarkers from gut microbiome genes and plasma metabolites, and predictive performance was assessed using different machine-learning algorithms. For gut microbiome genes, a two-step feature selection method was adopted. We selected the top 25 genes as features for constructing PD prediction models (Supplementary Table S1). Four machine-learning algorithms were used to construct PD prediction models, and an external validation dataset was downloaded from Gene Expression Omnibus (GEO) database with accession ID PRJNA433459 [7]. The random forest algorithm demonstrated the highest AUC score in both discovery (0.85 ± 0.11) and validation (0.75) datasets (Fig. 4d), with a sensitivity of 0.61 and specificity of 0.65 in the external validation dataset. Disease prediction based on gut microbiome gene and metabolite signatures. a Feature importance values for the top 25 gut microbiome genes identified using LassoNet. b–c Metabolite biomarkers with a VIP score greater than 1, identified from the cohort A and cohort B discovery groups separately. d–f Receiver Operating Characteristic (ROC) curves generated separately for the 25 gut microbiome genes, 8 metabolites, and 14 metabolite biomarkers. The best-performing machine learning results are presented For plasma metabolites, two independent cohorts were analyzed using the same LC–MS-based analytical methods. Firstly, cohort A (30 PD patients and 30 healthy spouse controls) was used as the discovery dataset, whereas cohort B (25 PD patients and 25 healthy spouse controls) served as the validation dataset. In cohort A, 86 significantly altered metabolites were identified, with 28 also found in cohort B. Eight of these metabolites, with VIP scores over 1, were used to construct PD predictors. The neural network method yielded the highest AUC score (0.91) in cohort B, whereas random forest method provided more consistent performance across both datasets (AUC = 0.83 ± 0.11 and 0.86) (Fig. 4b and Supplementary Fig. 2b). We then reversed the discovery and validation sets, using cohort B as the discovery dataset and cohort A as the validation dataset. 76 significantly altered metabolites were identified in cohort B, with 41 found in cohort A. Among these, 15 metabolites with VIP scores greater than 1 were selected. One metabolite, Nitrofurazone, was excluded from further analysis because it is an anti-infective drug and is not a naturally occurring human metabolite. The remaining 14 metabolites were identified as potential biomarkers. Logistic regression and neural network methods provided the best predictive performance, with both achieving an AUC of 0.92 in the validation dataset (Fig. 4f and Supplementary Fig. 3c). The results represent improved predictive performance compared to previous prediction models and demonstrate more consistent results across both discovery and validation datasets. These results indicate that the identified biomarkers have promising discriminative performance in the current datasets. However, further validation in larger and independent cohorts will be necessary to confirm their robustness and generalizability. Integrated microbiome and metabolome analysis Spearman correlation analysis was conducted to explore microbiome-metabolite associations. A total of 621 significant correlations (P < 0.05) were identified (Supplementary Table S1 and Supplementary Fig. 2). Most metabolites exhibited consistent directional correlations across multiple species. For instance, IAA showed negative correlations with 13 of the 14 species with which it was associated, whereas KYNA showed positive correlations with all six associated species. It is worth noting that Phocea massiliensis showed significant positive correlations with almost all glycerophospholipid metabolites, suggesting that it may be associated with host glycerophospholipid metabolism. Furthermore, Phocea massiliensis showed a positive correlation with calycin (R = 0.285, P = 0.027), a metabolite that most effectively discriminated between the two groups (Fig. 4b). To provide exploratory structural context for the metabolite biomarkers, we performed blind molecular docking using Autodock Vina against experimentally determined human protein structures available in the RCSB Protein Data Bank. Based on the predefined prioritization criteria, multiple metabolite biomarkers were frequently prioritized for two structures, 1BV8 (alpha-2-macroglobulin, A2M) and 6FF7 (the human Bact spliceosome complex) [54, 55]. Specifically, 5 of the 8 metabolite biomarkers in the first panel and 12 of the 14 metabolites in the second panel were computationally prioritized for these structures under the defined docking criteria (Supplementary Fig. 4 and Supplementary Fig. 5). Because these findings were derived from blind docking without active-site validation, RMSD assessment, or enrichment analysis, they should be interpreted with caution and regarded as exploratory and hypothesis-generating only. Further experimental validation of metabolite–protein interactions will be required to assess their potential biological relevance in PD-related contexts. To better contextualize these findings, we constructed an exploratory gut microbiome–metabolite–protein association network (Supplementary Fig. 6). Additional correlation analysis identified 29 species that were significantly correlated with metabolites computationally prioritized in the docking analysis for association with A2M and/or the human Bact spliceosome. This integrative network summarizes exploratory associations among the gut microbiome, host metabolites, and computationally prioritized protein structures in PD. Discussion In this study, we performed an integrative multi-omics analysis combining shotgun metagenomic sequencing and untargeted LC–MS-based metabolomics to investigate host–microbiome alterations associated with Parkinson’s disease (PD). We identified significant differences in gut microbial composition and circulating metabolite profiles between PD patients and healthy spouse controls. Integrative analyses further revealed associations among microbial species, plasma metabolites, and clinical features, and supported the development of predictive models based on microbial genes and metabolite biomarkers. Together, these findings provide a systems-level view of host–microbiome metabolic alterations associated with PD and generate hypotheses regarding interactions along the gut–brain axis. Previous PD microbiome studies have mainly focused on genus-level taxonomic shifts or targeted metabolite analyses. By integrating shotgun metagenomics with untargeted metabolomics, our study provides a broader characterization of microbial and metabolic alterations associated with PD. We observed significant differences in both α-diversity and taxonomic composition between PD patients and healthy spouse controls. Because the control group consisted of spouses of PD patients, shared environmental factors were partially controlled, which may reduce, but not eliminate, the influence of shared environmental variation when interpreting disease-associated differences. Several microbial taxa previously implicated in PD also showed altered abundance in our dataset, further supporting the growing evidence linking intestinal microbial dysbiosis to neurodegenerative disorders. However, given the cross-sectional design, these findings should be interpreted as associative rather than causal. Our metabolomic analysis revealed substantial alterations in circulating metabolites, particularly lipid-related compounds such as glycerophospholipids. Although the fold changes of individual metabolites were relatively modest, the consistent directional changes across multiple metabolites may reflect coordinated metabolic perturbations at the system level. Previous studies have shown that disruption of the homeostatic crosstalk between sphingolipids and glycerophospholipids can induce metabolic stress, which has been implicated in PD-associated metabolic disturbances [56]. Consistent with this, the glycerophospholipid-related alterations observed in our study may reflect disturbed lipid homeostasis in PD. We also identified changes in tryptophan-related metabolites, including indoleacetic acid and kynurenic acid, which are involved in immune regulation and neuroactive signaling and have previously been implicated in neuroinflammatory and neuroprotective processes in PD [52, 57–59]. These findings are consistent with a link between microbial metabolism and host metabolic alterations, although their functional significance remains to be established. To further explore microbiome–metabolome relationships, we intersected differentially abundant microbial pathways with metabolite-associated pathways derived from the MetaCyc database. Several overlapping pathways were identified, suggesting that microbial functional alterations may be associated with plasma metabolic changes. For example, decreased plasma citrate levels corresponded with enrichment of microbial pathways involved in citrate utilization. However, citrate also showed a positive correlation with UPDRS-III scores within the PD group, indicating that between-group differences and within-group clinical associations may not follow the same pattern. This apparent discrepancy may reflect differences between group-level alterations and within-group variability, potentially influenced by heterogeneity in disease progression or systemic metabolic regulation. Likewise, although trehalose abundance increased and one trehalose-related biosynthesis pathway was upregulated, another trehalose-related pathway showed the opposite trend. Because trehalose has been reported to exhibit chaperone-like activity and to prevent protein misfolding and aggregation [53], its elevation in PD could reflect a compensatory or stress-related metabolic response. Taken together, these observations indicate that the citrate- and trehalose-related findings are complex and not yet explained by a single coherent mechanism. Therefore, they should be regarded as preliminary and hypothesis-generating rather than mechanistically resolved. Machine-learning analyses showed that metabolite-based models performed relatively well in distinguishing PD patients from healthy spouse controls, with an AUC of 0.92, suggesting that circulating metabolites may represent candidate biomarkers. However, the sample size of this study was modest relative to the dimensionality of the multi-omics data, which may affect model stability and increase the risk of overfitting. Although cross-validation, independent validation cohorts, reverse discovery–validation analysis, and external validation for the gene-based model were used to improve reliability, formal feature stability analyses based on repeated resampling were not performed. Therefore, further validation in larger and more diverse populations will be necessary to confirm the robustness, reproducibility, and generalizability of these models. To provide exploratory structural context for the metabolite biomarkers, we conducted large-scale in silico molecular docking against experimentally determined human protein structures. Among the computationally prioritized protein structures, alpha-2-macroglobulin (A2M) and the human Bact spliceosome complex were notable because multiple metabolite biomarkers independently showed relatively higher predicted docking affinity to these structures under the predefined criteria. This convergence across multiple metabolites suggests that these protein structures may warrant further exploratory investigation in the context of metabolite–protein association analysis. In addition, genetic variants in the A2M gene have been associated with PD, and dysregulation of RNA splicing has also been reported in PD-related studies [60–63]. However, these observations were derived from computational docking analysis and will require further experimental validation to assess their potential biological relevance in PD-related contexts. Several limitations should be acknowledged. First, the cross-sectional design limits causal inference regarding the relationships among gut microbiota, circulating metabolites, and PD. Second, the relatively limited sample size may affect the robustness of high-dimensional multi-omics analyses and machine-learning models. Although cross-validation, independent validation cohorts, reverse discovery–validation analysis, and external validation for the gene-based model were applied to improve reliability, further validation in larger and more diverse populations will be necessary. Third, although metabolomic analyses were adjusted using LEDD, the influence of other medications commonly used in PD could not be fully accounted for and may have affected plasma metabolite profiles. In addition, while untargeted LC–MS provides broad metabolite coverage, it has limited structural resolution for certain metabolite classes, particularly lipids and carbohydrates; targeted lipidomics or glycomics approaches will therefore be important for further validation. Moreover, some metabolites in this dataset were annotated at lower confidence levels (e.g., Level 3–4), and the observed fold changes were generally modest. These factors may affect the reliability of interpretations at the level of individual metabolites. Therefore, downstream analyses, including pathway analysis, machine-learning modeling, and docking, should be interpreted cautiously and regarded as exploratory. Importantly, our main conclusions are based on consistent patterns observed across multiple metabolites rather than on any single feature. Finally, the docking analysis was based on a large-scale blind screening strategy rather than active-site-directed docking, and neither redocking of native ligands, RMSD-based validation, nor enrichment analysis was performed. Accordingly, the docking results should be interpreted with caution and regarded as exploratory and hypothesis-generating only. Future studies focusing on PD-related proteins, validated binding sites, and experimental validation of metabolite–protein interactions will be necessary to obtain more specific and reliable structural insights. In summary, this multi-omics study identified coordinated alterations in gut microbial composition, microbial metabolic pathways, and circulating plasma metabolites in PD. These findings provide an integrative overview of host–microbiome metabolic alterations associated with PD, highlight candidate biomarkers for disease classification, and generate hypotheses for future investigation. Further experimental and longitudinal studies will be necessary to assess the biological relevance of these associations. Supplementary Information Supplementary Material 1. Supplementary Material 2. Supplementary Material 3. Supplementary Material 4. Supplementary Material 5. Supplementary Material 6. Supplementary Material 7. Authors’ contributions PQW, JT, MZ and QZ designed the study. PQW, JT and MZ supervised the study. QZ did the data analysis. SFL, SJZ, JW, JT and XHL were responsible for participant recruitment, clinical diagnosis and assessment of disease scales. SFL, SJZ and JW collected the data of demographic and clinical samples. QZ, PQW, JT and MZ drafted the manuscript. XXZ and YHL polished and edited the language of manuscript. All authors had full access to all the data in the study and approved the final version of the manuscript. Funding This study was founded by the Hubei Provincial plan of science and technology key research project (grand number: 2023BCB140). Hubei Provincial Natural Science Foundation / Joint Fund Project Cultivation Project (grand number: 2023AFD045). The Advantages Discipline Group (Medicine) Project in Higher Education of Hubei Province (2021–2025) (grand number: 2023XKQT1). Data availability Metagenomic sequencing data of our dataset are available at the CNGB Sequence Archive (CNSA) of China National GeneBank DataBase (CNGBdb) with accession number CNP0006416. The metabolomics data (LC–MS raw files) can be obtained upon reasonable request by contacting the corresponding author Puqing Wang. Declarations Ethics approval and consent to participate The study was approved by the Ethical Research Committee of Xiangyang No. 1 People’s Hospital (Approval No. 2021KYLX09). This study was conducted in full compliance with the ethical principles outlined in the Declaration of Helsinki, ensuring adherence to standards for research involving human subjects. Informed consent was obtained from all participants prior to their involvement in the study. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Qiang Zhao and Shujia Zuo contributed equally to this work. Contributor Information Min Zhou, Email: 591974569@qq.com. Jing Tian, Email: 15030662@qq.com. Puqing Wang, Email: wpq20110328@qq.com. References Associated Data Supplementary Materials Supplementary Material 1. Supplementary Material 2. Supplementary Material 3. Supplementary Material 4. Supplementary Material 5. Supplementary Material 6. Supplementary Material 7. Data Availability Statement Metagenomic sequencing data of our dataset are available at the CNGB Sequence Archive (CNSA) of China National GeneBank DataBase (CNGBdb) with accession number CNP0006416. The metabolomics data (LC–MS raw files) can be obtained upon reasonable request by contacting the corresponding author Puqing Wang.