Large-scale profiling of blood microbial signatures in patients with Parkinson's disease and its association with disease progression: a cross-sectional study

Summary

Background

Emerging evidence supports the presence of microbial signatures in the blood, yet their clinical relevance remains poorly understood. In this study, we profiled blood microbial signatures in patients with Parkinson's disease (PD) and investigated their associations with disease progression.

Methods

We analysed 4018 whole-genome sequencing (WGS) data of blood samples from two independent cohorts. The high-quality non-human reads were extracted for microbial annotation using Kraken 2 and Bracken software with the PlusPF database. To identify PD-associated signatures, we implemented a population-based, cross-cohort filtration process with resequencing validation to minimise noise and putative contaminants.

Findings

Microbial DNA signals, predominantly bacterial, were extensively detected in the sequencing data and were more abundant in individuals with PD than in controls. Across the two cohorts, 126 bacterial species were identified as key signatures, nearly two-thirds of which are known to colonise human body sites. Among these, 19 species exhibited increased abundance and higher prevalence in PD, and could serve as features to discriminate effectively patients from controls. Furthermore, several microbial signatures were correlated with more severe clinical manifestations, such as motor dysfunction and cognitive impairment.

Interpretation

Our findings supported blood microbial signatures as promising biomarkers in PD, although their origin and functional relevance remain to be validated. The analytical framework may facilitate future investigations into the potential clinical implications of blood microbial signatures in disease contexts.

Funding

This work was supported by Hunan Innovative Province Construction Project, National Natural Science Foundation of China, and Natural Science Foundation of Hunan Province.

Research in context.

Evidence before this study

We searched PubMed for studies on microbial signatures in the blood of patients with Parkinson's disease (PD) published from database inception to December 31, 2025 without language restrictions, using the search terms (“Parkinson's disease” OR “paralysis agitans”) AND (“patients” OR “participants” OR “cohort”) AND (“blood microbiome” OR “blood microbiomes” OR “blood microbiota” OR “blood microorganisms” OR “blood bacteria”) in All Fields, and 3 manuscripts were searched. These three articles reported the detection of blood microbial signals and their correlations with PD progression in a small sample of patients, as detected by 16S rRNA sequencing or RNA-seq analysis. However, these findings remain to be verified in large-scale and multi-centre datasets. In addition, whole-genome sequencing (WGS) data of peripheral lymphocytes enables detection of microbial DNA alongside the human genome sequencing, offering a valuable approach to profile microbial landscape in the blood.

Added value of this study

Here, we acquired 4018 WGS data of blood samples from two independent cohorts (PD-MDCNC and PPMI), as well as two re-sequenced data (WGS and RNA-seq) from these two cohorts. The two cohorts were used to explore the shared microbial signatures in the blood, which were subsequently verified using the two re-sequenced datasets. Consequently, a key subset of blood microbial signatures consisting of 126 bacterial species was identified. Of these, nearly two-thirds are also detectable in human body sites, especially in the gut and genitourinary tract. Across the two cohorts, we identified 19 PD-enriched bacterial species, which effectively distinguished between patients from controls and were associated with more severe clinical indicators of PD progression.

Implications of all the available evidence

We detected more abundant microbial signatures in the blood from patients with PD compared to control subjects (CS). The potential origin of these signals, such as translocation from human body sites, remains speculative and requires functional validation. This study provides a preliminary exploration of blood-derived microbial signatures in the context of host–microbe interactions. While the analytical workflow used here enabled the detection of low-biomass signals, further optimisation and orthogonal validation are required before it can be broadly applied to other complex diseases.

Introduction

The human microbiome, particularly at external and mucosal interfaces such as the skin, oral cavity, gastrointestinal tract, and respiratory system, has been extensively studied and has revealed critical roles in maintaining physiological homoeostasis and influencing human health.1, 2, 3 While blood was traditionally considered as a sterile environment, modern sequencing technologies have uncovered traces of microbial DNA, albeit low-biomass.4, 5, 6 The notion of blood microbiome remains controversial,⁷ yet the microbial signals identified in a large-scale whole-genome sequencing (WGS) study on healthy population were indeed observed and attributed to the transient and sporadic translocation from other human body sites.^8,9 Also, scanning electron microscopy analysis visually confirmed the presence of coccus- and bacillus-shaped bacteria alongside erythrocytes in patients with Parkinson's disease (PD).¹⁰ As blood circulates throughout the body, it carries integrated microbial signals derived from various tissues and organs, thereby offering an informative window into systemic physiological and pathological states.^9,11 Therefore, profiling microbial signatures in the blood may hold significance for investigating a previously underexplored aspect of human disease.

PD is a chronic and progressive neurodegenerative disorder that predominantly affects middle-aged and older adults, and its global burden has risen steadily over the past three decades.^12,13 Gut microbiota dysbiosis has been implicated in the pathogenesis, progression, and treatment response of PD.^14,15 Actually, microbial dysbiosis and impaired mucosal integrity, commonly observed in conditions like PD, may facilitate translocation of intestinal microbiota and their components into the bloodstream, potentially promoting systemic inflammation and exacerbating central neuroinflammation.^16,17

Recent studies on investigating blood-derived microbial signals in patients with PD have generated valuable data, suggesting a potential link between blood microbial signatures and the disease.18, 19, 20 However, the scope of these studies has been constrained by limited sample sizes and a lack of validation cohorts. That is, these characteristics remain to be elucidated in large-scale, multi-centre datasets.

In this study, to identify those blood microbial signatures potentially associated with PD whilst minimising the impact of potential contaminants and background noise, we leveraged large-scale WGS datasets from two independent cohorts and further examined these signatures in two re-sequenced datasets. Our findings indicated a potential association between blood microbial signatures and PD, providing a preliminary framework for similar investigations in other diseases.

Methods

Data source and participants

The study was approved by the Ethics Committee of Xiangya Hospital of Central South University (ethical approval no: 202308749), and written informed consent was obtained from all participants or their legal guardians. A total of 4018 WGS datasets derived from blood samples were acquired from two independent cohorts, i.e. Parkinson's Disease & Movement Disorders Multicenter Database and Collaborative Network in China (PD-MDCNC, http://pd-mdcnc.com) and Parkinson's Progression Markers Initiative (PPMI, http://www.ppmi-info.org/data, January 27, 2023).^21,22

In PD-MDCNC cohort, we obtained 3241 WGS bam files, including 1962 individuals with PD and 1279 control subjects (CS), which have been utilised and published in previous studies.^23,24 Therein, each patient was diagnosed by at least two neurologists according to either the Movement Disorder Society Clinical Diagnostic Criteria or the United Kingdom Parkinson's Disease Society Brain Bank Clinical Diagnostic Criteria.^25,26 For the present study, we selected individuals aged >50 years, including patients with PD and age-matched CS without neurological disorders. Peripheral blood sample of 10 mL was collected from each participant. Genomic DNA was extracted by the phenol-chloroform method, assessed for quality by agarose gel electrophoresis, and quantified using a Qubit fluorometer. Sequencing libraries were constructed from 300 ng of high-quality genomic DNA (OD_260/280 ratio: 1.8–2.0) using Sureselect^XT reagent kit (Agilent) according to the manufacturer's instructions. Briefly, DNA was sheared, followed by end-repair, A-tailing, adaptor ligation, and purification using AMPure XP beads. After PCR amplification, the libraries were sequenced on an Illumina NovaSeq 6000 platform to generate 150-bp paired-end reads (WuXi AppTec Co., Ltd.). The WGS data achieved an average depth of 12× (range: 8×–18×). All participants in PD-MDCNC cohort are of Chinese descent, and the available information from electronic medical records were collected, including demographic traits and clinical indicators (Table S1). Sex information was self-reported by the participants. To handle missing data, a complete-case analysis was performed; participants with missing data for any of the variables of interest were excluded from the analysis. In addition, to verify the detectability of blood microbial signatures in the PD-MDCNC cohort, 35 extracted DNA samples were selected and re-sequenced by WGS (Validation dataset 1).

In addition, a total of 777 WGS bam files at baseline, including 586 PD and 191 CS files, were downloaded from PPMI. These WGS files were sequenced with 150-bp pair-end reads and achieved an average depth of 38× (range: 30×–80×). The majority of participants in PPMI cohort are white, and their available demographic traits and clinical characteristics at baseline were obtained as well (Table S2). Also, a complete-case analysis was performed to handle the missing data. In addition, to verify the detectability of blood microbial signatures in those WGS files from PPMI cohort, we also downloaded RNA-seq files corresponding to 631 subjects at baseline, matched by participant ID (Validation dataset 2).

Taxonomic annotation of blood microbial signatures

In this study, referring to a prior study,⁸ a workflow was developed for taxonomic annotation of microbial sequences in the blood (Fig. 1A). The WGS files were separately aligned to the human reference genome (hg38) by using BWA mem (version 0.7.17-r1188) with default parameters; non-human reads were subsequently extracted from each alignment file by using Samtools (version 1.9). To obtain high-quality non-human reads, BBDuk (version 39.01) was applied for base trimming and reads filtering, followed by Kneaddata (version 0.12.0) which incorporates Trimmomatic (version 0.33) and Bowtie 2 (version 2.2.3) to remove residual low-quality sequences and further eliminate human-derived reads. Meanwhile, Fastp (version 0.23.4) was employed to assess sequence quality, and those samples with high quality scores (Q20 > 0.95 and Q30 > 0.85) were retained. Eventually, the resulting high-quality non-human reads in each sample were individually applied for taxonomic annotation by using Kraken 2 (version 2.1.3) and Bracken (version 2.9) in combination with the PlusPF database (https://benlangmead. github.io/aws-indexes/k2, October 9, 2023).27, 28, 29 In order to minimise false-positive microbial assignments originating from host sequences,³⁰ the PlusPF database integrates human genome with a comprehensive microbial genome repository encompassing Bacterial, Viral, Archaeal, and Fungal domains. Similarly, the two validation datasets (i.e. WGS files, RNA-seq files) were also subjected to microbial taxonomic annotation as described above.

Fig. 1.

Comparisons of non-human reads in the sequencing data from blood samples of PD-MDCNC or PPMI cohort. (A) Microbial annotation and filtration process used in this study. (B, D) Comparisons between PD and CS at the levels of non-human reads, microbes, and bacteria in PD-MDCNC (B) or PPMI (D) cohort. Boxplots show median (line) and mean (point), respectively. Significance was determined by two-tailed Wilcoxon rank-sum test. (C, E) Comparisons of the proportions of microbes to non-human reads (left) and the proportions of bacteria to microbes (right) between PD and CS in PD-MDCNC (C) or PPMI (E) cohort. Boxplots show median (line) and mean (point), respectively. Significance was determined by two-tailed Wilcoxon rank-sum test. Those circles in the dashed box (right) represent those subjects with the proportions of bacteria to microbes of less than 0.5. (F, H) Comparisons of bacterial reads between PD and CS at different ages in PD-MDCNC (F) or PPMI (H) cohort. The points in red and in blue represent the PD and CS subjects, respectively. The lines in red and in blue refer to the fitted lines in PD and CS, respectively. Spearman's correlations between bacterial reads and subjects' ages in PD and CS were calculated, respectively. (G, I) Comparisons of bacterial reads between PD and CS at different sex in PD-MDCNC (G) or PPMI (I) cohort. Boxplots show median (line) and mean (point), respectively. Significance was determined by two-tailed Wilcoxon rank-sum test.

Filtering processes of contaminants and noise

To identify robust microbial signals in blood, we implemented a population-based and cross-cohort filtering strategy. Firstly, to remove background noise, a threshold of <3 was counted as absent (set to zero counts). Subsequently, since it was reported that a list of 93 bacterial genera acting as common reagent contaminants,³¹ thus those species that belong to any of the 93 genera were filtered as contaminant species. Next, those remaining species that exhibited a strong positive Spearman's correlation with any of the excluded contaminant species were also removed as contaminants (ρ > 0.7 and p < 0.05).³² Spearman's correlation between two species was calculated based on the centred log-ratio-transformed microbial abundance using the clr function of R package compositions (version 2.0–8).³³

Prevalence filtering was also conducted. Species with a prevalence of <0.1% in both cohorts were treated as noise or contaminants and were excluded. Then, species with a prevalence of >1% in at least one cohort were retained for subsequent analyses. Meanwhile, in either the PD or CS group, when the prevalence of a given species exceeded >30% in a certain group of one cohort, the inter-cohort difference in prevalence was constrained to within a 1.5-fold range.^8,32 In addition, the abundance matrices of two validation datasets were similarly subjected to de-noising operation (≥3) and prevalence filtering (>1%).

Normalisation of bacterial species abundance

We speculated that the number of non-human reads in WGS data was proportional to the total reads. To ensure comparability across samples and mitigate biases introduced by the varying sequencing depth, the number of non-human reads in each sample was normalised by the sequencing depth of the human genome, as following:

$${Normalisedread}_i=\frac{{read}_i}{\raisebox{1ex}{${depth}_i$}\!\left/ \!\raisebox{-1ex}{$Depth$}\right.}$$ (1)

in which read_i and Normalised read_i separately represent the number of non-human reads before and after normalisation in the i-th sample; depth_i stands for the sequencing depth of human genome in the i-th sample, while Depth denotes the average depth across the cohort (i.e. PD-MDCNC or PPMI).

To mitigate quantification bias introduced by the genome sizes of different bacterial species,³⁴ we further normalised the abundance matrices of those key blood species, as following:

$${Normalisedabundance}_j=\frac{2\times \frac{{read}_j}{{genome}_j}\times 150}{{depth}_i}=\frac{2\times \frac{{read}_j}{\raisebox{1ex}{${depth}_i$}\!\left/ \!\raisebox{-1ex}{$Depth$}\right.}\times 150}{{genome}_j\times Depth}$$

$${Normalisedabundance}_j=\frac{2\times {Normalisedread}_j\times 150}{{genome}_j\times Depth}$$ (2)

in which read_j and Normalised read_j separately represent the number of non-human reads before and after normalisation the j-th species in i-th sample; Normalised abundance_j and genome_j respectively stand for the normalised abundance and the genome size of j-th species; depth_i and Depth refer to the sequencing depth of the i-th sample and the average depth in the cohort (i.e. PD-MDCNC or PPMI). The constant 150 corresponds to the read length of 150-bp. And, the constant 2 reflects the assumption that bacterial species are haploid in contrast to the diploid human genome. For ease of interpretation, the normalised abundance reflects the copy number of microbial genome per unit sequencing depth of human genome. In Equ. 2, the genome sizes of those key blood species were sourced from NCBI Genome Database (https://www.ncbi.nlm.nih.gov/datasets/genome/).

Diversity analysis at the bacterial species level

To evaluate the sample sizes of the PD or CS groups in the two cohorts, rarefaction analyses were performed by the R package vegan (version 2.6-4) on the R platform (version 4.1.2). Based on those samples that have detected at least one species in the cohort, the Bray–Curtis dissimilarity indices between samples were estimated by R package vegan, after performing logarithm transformation of the normalised abundance matrix. Then, principal coordinate analysis (PCoA) and permutational multivariate analysis of variance (PERMANOVA) were performed to visualise the beta-diversities between samples and determine the significance of differences between groups, respectively. To evaluate the bacterial alpha-diversity within each sample, the species richness, evenness, and Shannon indexes were separately calculated by R package vegan.

Construction of classification models

The normalised abundance matrices of those key blood species in PD-MDCNC and PPMI cohorts was divided into training dataset and test dataset in a 7:3 ratio, respectively. The XGBoost algorithm was employed to construct classification models by R package xgboost (version 1.6.0.1). Therein, three distinct feature sets were used to construct classification models: (i) demographic characteristics (sex and age in PD-MDCNC cohort; sex, age, BMI, education years, and race in PPMI cohor), (ii) the 19 differential species identified in both PD-MDCNC and PPMI cohorts, and (iii) a combination of both demographic characteristics and the 19 differential species. The AUC values of the classification models in the test dataset were computed and visualised by R package pROC (version 1.18.5).

Also, we treated PPMI and PD-MDCNC dataset as training dataset and test dataset to build the classification models. The three distinct feature sets included: (i) demographic characteristics (age and sex), (ii) the 19 differential species, and (iii) a combination of both demographic characteristics and the 19 differential species.

Differential analysis

Based on the normalised abundance of those key species, we performed differential analysis with adjustment for those available covariates (sex and age in PD-MDCNC cohort; sex, age, BMI, education years, and race in PPMI cohort). The analysis was conducted using R package MaAsLin2 (version 1.15.1) with TSS normalisation and logarithm transformation,³⁵ a recommended best practice for microbiome differential abundance analysis.³⁶ Statistical significance was assessed using Benjamini–Hochberg (BH) correction for multiple testing. In addition, the detection rates (prevalence) of those species were compared using Fisher's exact test with BH-correction. To further evaluate the associations between microbial presence and PD diagnosis, we fitted generalised linear models (GLMs) and took those available covariates into adjustment. The use of GLMs to estimate odds ratios (OR) for microbial presence, following the approach of prior microbes–disease association studies.³⁷ Statistical significance was assessed using a BH-corrected p-value of <0.05 in the PD-MDCNC cohort, whereas a nominal p-value of <0.05 was used for the smaller PPMI validation cohort.

Correlation analysis and visualisation

We conducted correlation analyses between blood microbial signatures and clinical indicators of PD progression, with adjustment for relevant covariates in the PD-MDCNC and PPMI cohorts, respectively. The categorical indicators were analysed by GLMs, whilst the continuous indicators were analysed by partial correlation analysis using R package ppcor (version 1.1). Covariate adjustment varied by cohort: in the PD-MDCNC cohort, we adjusted for age, sex, disease duration, living environment, hypertension, diabetes, and hyperlipidaemia; in the PPMI cohort, adjustment included age, sex, BMI, years of education, and race. The correlations of categorical variables were visualised by forest plot using R package see (version 0.12.0), whilst the correlations of continuous variables were displayed as heatmap produced by R package corrplot (version 0.95).

Statistical analysis

No sensitivity analyses were performed, as all analyses were based on complete cases. Statistical analyses were conducted using the two-tailed Wilcoxon rank-sum test, and the corresponding p-value are displayed in each plot. Partial correlation analysis was conducted to evaluate the associations between blood microbial signatures and continuous indicators of PD progression, and statistical significance was defined as ∗ p < 0.1; ∗∗p < 0.05; ∗∗∗p < 0.01; ∗∗∗∗p < 0.001.

Role of funders

The funders of the study had no role in study design, data collection, data analysis and interpretation, or writing of the report.

Results

Significantly elevated microbial signatures in the blood of patients with PD compared to controls

In this study, a total of 4018 WGS datasets of blood samples from two independent cohorts (i.e. PD-MDCNC and PPMI) were acquired to extract high-quality non-human reads and explore the microbial sequences in the blood (Fig. 1A, left). We found that the quantity of detected non-human reads increased with the sequencing depth of the human genome (Fig. S1A, C). This trend was consistently observed in both PD and CS groups within both cohorts, possibly attributable to the proportional generation of non-human reads to human reads during WGS. To minimise the bias in subsequent comparisons, thus we normalised the non-human reads (Equ. 1), and observed that the number of non-human reads in PD was stably higher than those in CS in both cohorts (Fig. S1B, D).

Further to taxonomically determine which microorganisms high-quality non-human reads in blood belong to, we performed taxonomic annotation using Kraken 2 and Bracken in combination with the PlusPF database (Fig. 1A, left).27, 28, 29 In both cohorts, we consistently detected significantly elevated levels of non-human reads, microbes, and bacteria in patients with PD compared to CS (Fig. 1B, D). In addition, patients with PD also showed increased proportion of microbial reads to non-human reads (Fig. 1C, E, left) and elevated proportion of bacteria to microbes (Fig. 1C, E, right). Therein, over 98% of the classified microbial reads were bacterial sequences in both PD and CS groups. Consequently, we focused on bacteria and selected those samples with a bacterial proportion exceeding 50% for subsequent investigations. Further, stratified statistical analysis also revealed consistently elevated bacterial signatures in patients with PD compared to CS, regardless of age (Fig. 1F, H) and sex (Fig. 1G, I).

Our analysis detected microbial DNA in the sequencing data from blood samples of both PD and CS groups across the two cohorts, with the majority deriving from bacterial sources. And, patients with PD exhibited higher microbial signatures compared to CS, independent of sequencing depth and demographic variables such as sex and age.

Identification of key bacterial species in the blood across PD-MDCNC and PPMI cohorts

A total of 10,885 and 8956 bacterial species were taxonomically annotated in PD-MDCNC and PPMI cohorts, respectively (Fig. 1A, left), and a series of filtering processes were carried out to remove low-abundance noise and putative reagent contaminants (Fig. 1A, right).^31,32 Consequently, 2120 and 2842 species were retained in PD-MDCNC and PPMI cohorts, respectively. To minimise the impact of cohort-specific artifacts, we intersected species lists of the two cohorts and performed prevalence filtering, yielding 562 shared bacterial species (Fig. 1A, right).

Further to verify the detectability of the 562 bacterial species in the blood, two re-sequenced datasets from PD-MDCNC and PPMI cohorts were obtained and utilised (Fig. 1A, right). That is, among those samples (participants) analysed in PD-MDCNC and PPMI cohorts above, we selected 35 DNA samples from the PD-MDCNC cohort for WGS again as Validation dataset 1 (Table S3), and downloaded 631 RNA-seq files at baseline from the PPMI cohort as Validation dataset 2 (Table S4). Among the 562 species retained above, a subset of 126 species was detected in the 35 WGS files from PD-MDCNC cohort, and 61 (48.4%) were confirmed by those 35 re-sequenced WGS files. And, 420 (74.8%) species were verified in those 631 RNA-seq files from PPMI cohort.

Ultimately, by intersecting the discovery and validation datasets, we defined a key subset consisting of 126 bacterial species for downstream analysis (Table S5). The abundance matrices of these key species in PD-MDCNC and PPMI cohorts were normalised, respectively (Equ. 2). Considering the distinct ethnicities of participants in the two cohorts, thus these key species in the blood may be independent of race.

Characterisation of the key bacterial species in the blood of patients with PD

Based on the prevalence of these 126 key species in PD-MDCNC and PPMI cohorts, we observed a significant positive correlation, indicating the consistency of these bacteria in both cohorts (Fig. 2A). Therein, 18 bacterial species detected in the oral or gut samples of patients with PD have previously been reported in associations with this disease, i.e. Actinomyces oris, Brucella intermedia, Brucella melitensis, Brucella sp. 458, Citrobacter braakii, Citrobacter freundii, Enterococcus faecium, Halomonas sp. M4R1S46, Halomonas sp. NyZ770, Halomonas sp. PGE1, Klebsiella grimontii, Klebsiella michiganensis, Klebsiella oxytoca, Klebsiella pneumoniae, Klebsiella variicola, Salmonella enterica, Sphingobacterium multivorum, and Veillonella parvula (Table S6).38, 39, 40, 41, 42, 43, 44, 45

Fig. 2.

Characteristics of the 126 key species identified across the two cohorts. (A) Pearson's correlation of prevalence of the key species in the PD-MDCNC and PPMI cohorts. (B) Oxygen requirement classification of the key species. (C) Annotation profiling of the key species, showing the proportions of species with blood culture records (left), known as human pathogen (middle), and documented human association (right). (D) Histogram of human-associated species detected in body sites, including gut, genitourinary tract, respiratory tract, skin, oral, eye, and blood. (E) Rarefaction curves for PD and CS samples in the PD-MDCNC (left) or PPMI (right) cohorts. (F) Histogram of samples with a gradient species in PD-MDCNC (left) or PPMI cohort (right). The parts in red and in blue represent PD and CS samples, respectively. (G) PCoA based on the Bray–Curtis dissimilarity indexes shows beta-diversity between PD and CS groups of the PD-MDCNC (top) or PPMI (bottom) cohort. Significance was determined by PERMANOVA. (H–I) The cumulative abundance of the key species between PD and CS groups in the PD-MDCNC (H) or PPMI (I) cohort. Boxplots show median (line) and mean (point), respectively. Significance was determined by two-tailed Wilcoxon rank-sum test. (J–K) Species richness (left), Evenness (middle), and Shannon index (right) between PD and CS groups in the PD-MDCNC (J) or PPMI (K) cohort. Boxplots show median (line) and mean (point), respectively. Significance was determined by two-tailed Wilcoxon rank-sum test.

These key blood species are predominantly aerobic or facultatively anaerobic (Fig. 2B). In comparison with several previous studies (Table S6),46, 47, 48 we found that only 40 (31.7%) of these species were cultivable by inoculating blood samples (Fig. 2C, left), that 42 (33.3%) species were pathogenic in human (Fig. 2C, middle), and that 81 (64.3%) species were associated with human diseases (Fig. 2C, right). Notably, these 81 human-associated species can be detected at various human body sites, especially in the gut and genitourinary tract (Fig. 2D). This indicated that nearly two-thirds of these species may originate from the translocation of human microbes, either microbial DNA fragments or the entire viable bacteria.

Based on the normalised abundance matrices of these key species in both cohorts, rarefaction analyses showed the saturation curves, confirming the adequate sample sizes in PD and CS groups of the two cohorts (Fig. 2E). We found that at least one species was present in 92.29% of PD-MDCNC samples and 88.54% of PPMI samples (Fig. 2F), showing an underestimated detectability of microbial signatures in the blood. Subsequently, a subset of samples positive for at least one species was selected for diversity analysis. A significant difference in beta-diversity between PD and CS groups was both observed in the two cohorts (Fig. 2G). The cumulative normalised abundance remained stably higher in PD compared to CS in the two cohorts (Fig. 2H–I). And, the alpha-diversities were evaluated in the two cohorts (Fig. 2J–K), in which comparing with CS group, we observed that higher species richness and lower evenness in PD group, while Shannon index that incorporates both species richness and evenness was inconsistent. These situations were also observed when samples were stratified by age and sex in the two cohorts, respectively (Fig. S2A–L). These findings suggested that patients with PD harboured a larger species richness, but the community may be dominated by a few species, leading to a less balanced distribution.

Identification of characteristics species with differential abundance and prevalence in patients with PD

Drawing upon the findings described above, we performed differential analysis based on both normalised abundance and prevalence of the key species adjusting for covariates, respectively. The PD-MDCNC cohort was treated as a discovery dataset, in which we identified 45 species with differential abundance and 60 species with differential prevalence in patients with PD. To focus on those robust candidates, we selected those species showing consistent directional differences in both abundance and prevalence, yielding a total of 38 species: 25 PD-enriched and 13 PD-depleted (Table S7). These candidates were further validated in the independent PPMI cohort using the same analytical approach. As a result, a total of 19 PD-enriched species were successfully demonstrated, including Achromobacter pestifer, Bordetella avium, Bordetella bronchialis, Bordetella hinzii, Bordetella parapertussis, Bordetella pertussis, B. intermedia, C. freundii, Diaphorobacter nitroreducens, Diaphorobacter sp. JS3051, Dickeya zeae, Hydrogenophaga crassostreae, Halomonas sp. NyZ770, Orrella dioscoreae, Pectobacterium carotovorum, Pusillimonas sp. M17, Raoultella ornithinolytica, S. enterica, and Serratia liquefaciens. Notably, all 19 species belong to the phylum Pseudomonadota (formerly known as Proteobacteria) and exhibited significantly higher abundance and prevalence in patients with PD compared to CS across both cohorts (Fig. 3A–B). And, the odds ratio (OR) for these species was all greater than 1 (Tables 1 and 2), underscoring their potential as detrimental biomarkers for PD.

Fig. 3.

Differential species identified between PD and CS groups across the two cohorts. (A–B) Characterisation of the 19 differential species in the PD-MDCNC (A) and PPMI (B) cohorts. For each cohort, the prevalence (left) and abundance (middle) of the 19 differential species between PD and CS groups. Effect size of the 19 differential species with PD (right) produced by using MaAsLin2. (C–D) Classification performance of prediction models in the PD-MDCNC (C) and PPMI (D) cohorts. Receiver operating characteristic curves for the models fitted by (i) demographic characteristics, (ii) the 19 differential species, and (iii) a combination of both demographic characteristics and the differential species, respectively. (E) Cross-cohort validation of the models fitted by (i) demographic characteristics, (ii) the 19 differential species, and (iii) a combination of both demographic characteristics and the differential species, respectively.

Table 1.

Characteristics of the nineteen validated differential species in PD-MDCNC cohort.

Differential species	Prevalence (Fisher's test)			Prevalence (GLM)			Abundance (MaAsLin2)
	HC (n = 1201)	PD (n = 1717)	p-value (BH-correction)	Estimate	OR	p (<|z|)	Effect size	p-value (BH-correction)
Achromobacter pestifer	52 (4.33%)	340 (19.80%)	2.50E-36	1.73	5.61	1.76E-24	1.62	1.08E-32
Bordetella avium	1 (0.08%)	39 (2.27%)	1.40E-07	3.07	21.55	2.96E-03	0.13	4.58E-06
Bordetella bronchialis	2 (0.17%)	46 (2.68%)	3.50E-08	2.47	11.82	1.01E-03	0.12	1.46E-06
Bordetella hinzii	22 (1.83%)	73 (4.25%)	9.50E-04	1.00	2.72	5.02E-04	0.17	2.53E-04
Bordetella parapertussis	1 (0.08%)	38 (2.21%)	2.50E-07	3.05	21.09	3.17E-03	0.16	1.04E-05
Bordetella pertussis	2 (0.17%)	43 (2.25%)	1.80E-07	2.50	12.20	8.59E-04	0.14	1.60E-05
Brucella intermedia	231 (19.23%)	474 (27.61%)	7.70E-07	0.45	1.56	4.29E-05	0.72	2.12E-04
Citrobacter freundii	48 (4.00%)	455 (26.50%)	1.20E-63	2.02	7.54	9.26E-32	1.84	8.04E-57
Diaphorobacter nitroreducens	7 (0.58%)	65 (3.79%)	2.10E-08	1.80	6.04	2.37E-05	0.23	1.40E-06
Diaphorobacter sp. JS3051	11 (0.92%)	45 (2.62%)	2.60E-03	1.14	3.14	3.14E-03	0.17	2.06E-03
Dickeya zeae	3 (0.25%)	27 (1.57%)	9.50E-04	1.57	4.78	1.80E-02	0.12	2.67E-03
Halomonas sp. NyZ770	17 (1.42%)	121 (7.05%)	5.30E-13	1.85	6.34	3.30E-09	0.44	5.82E-10
Hydrogenophaga crassostreae	7 (0.58%)	55 (3.20%)	1.40E-06	1.75	5.74	4.04E-05	0.24	1.82E-05
Orrella dioscoreae	5 (0.42%)	97 (5.65%)	2.30E-16	2.46	11.67	2.82E-07	0.50	1.49E-12
Pectobacterium carotovorum	0 (0%)	30 (1.75%)	6.80E-07	17.11	2.69E+07	9.85E-01	0.12	4.71E-05
Pusillimonas sp. M17	23 (1.92%)	467 (27.20%)	2.50E-87	2.83	16.86	1.56E-35	2.87	9.71E-69
Raoultella ornithinolytica	5 (0.42%)	60 (3.49%)	1.20E-08	2.04	7.71	4.47E-05	0.10	7.73E-04
Salmonella enterica	90 (7.49%)	439 (25.57%)	1.90E-37	1.32	3.74	3.26E-21	1.34	3.53E-27
Serratia liquefaciens	75 (6.24%)	318 (18.52%)	4.90E-22	1.06	2.89	2.30E-12	1.62	3.13E-25

Table 2.

Characteristics of the nineteen validated differential species in PPMI cohort.

Differential species	Prevalence (Fisher's test)			Prevalence (GLM)			Abundance (MaAsLin2)
	HC (n = 188)	PD (n = 579)	p-value	Estimate	OR	p (<|z|)	Effect size	p-value
Achromobacter pestifer	6 (3.19%)	74 (12.76%)	5.63E-05	1.53	4.61	5.00E-04	0.74	1.31E-03
Bordetella avium	1 (0.53%)	57 (9.83%)	1.59E-06	3.10	22.23	2.30E-03	0.46	7.25E-05
Bordetella bronchialis	6 (3.19%)	83 (14.31%)	9.85E-06	1.68	5.35	1.00E-04	0.60	1.34E-04
Bordetella hinzii	9 (4.79%)	86 (14.83%)	1.16E-04	1.26	3.53	6.00E-04	0.68	2.75E-03
Bordetella parapertussis	0 (0%)	55 (9.48%)	1.30E-07	16.61	1.64E+07	9.75E-01	0.30	3.45E-05
Bordetella pertussis	2 (1.06%)	40 (6.90%)	1.29E-03	1.96	7.11	7.60E-03	0.24	2.36E-02
Brucella intermedia	20 (10.64%)	135 (23.28%)	1.02E-04	0.95	2.58	3.00E-04	1.16	1.33E-05
Citrobacter freundii	15 (7.98%)	110 (18.97%)	2.47E-04	1.02	2.76	6.00E-04	0.59	3.93E-02
Diaphorobacter nitroreducens	2 (1.06%)	43 (7.41%)	5.06E-04	2.02	7.51	5.90E-03	0.27	3.49E-02
Diaphorobacter sp. JS3051	3 (1.60%)	60 (10.34%)	2.59E-05	1.97	7.19	1.00E-03	0.51	8.92E-03
Dickeya zeae	1 (0.53%)	59 (10.17%)	9.77E-07	3.12	22.71	2.10E-03	0.61	5.41E-05
Halomonas sp. NyZ770	0 (0%)	17 (2.93%)	1.79E-02	15.53	5.58E+06	9.78E-01	0.12	2.76E-02
Hydrogenophaga crassostreae	4 (2.13%)	83 (14.31%)	4.56E-07	2.08	7.99	1.00E-04	0.56	1.02E-04
Orrella dioscoreae	7 (3.75%)	78 (13.45%)	8.42E-05	1.39	4.02	7.00E-04	0.76	9.60E-04
Pectobacterium carotovorum	0 (0%)	17 (2.93%)	1.79E-02	15.64	6.18E+06	9.78E-01	0.17	3.36E-02
Pusillimonas sp. M17	4 (2.13%)	52 (8.97%)	1.02E-03	1.54	4.67	3.70E-03	0.32	2.44E-02
Raoultella ornithinolytica	1 (0.53%)	41 (7.07%)	1.37E-04	2.58	13.19	1.13E-02	0.39	4.73E-03
Salmonella enterica	9 (4.79%)	111 (19.14%)	4.12E-07	1.58	4.84	0.00E+00	0.70	3.92E-05
Serratia liquefaciens	5 (2.66%)	84 (14.48%)	1.38E-06	1.76	5.81	2.00E-04	0.86	7.85E-04

To evaluate the diagnostic utility of these differential species relative to demographic characteristics, we constructed multiple classification models. Models based on these 19 species outperformed those using only demographic features, as evidenced by higher AUCs in both cohorts (Fig. 3C–D). Furthermore, a combined model integrating both microbial and demographic features achieved the highest AUCs (Fig. 3C–D). The generalisability of this finding was confirmed through cross-cohort validation, in which models trained on the PPMI cohort and tested on the PD-MDCNC cohort yielded consistent results (Fig. 3E). That is, the 19 characteristics species can be associated with PD status.

Positive associations between blood microbial signatures and PD progression

We next examined the associations between blood microbial signatures and clinical indicators of PD progression. As shown in Fig. 4A–B, several associations all achieved OR of greater than 1 in each of the two cohorts. H. crassostreae was consistently linked to the progressive Hoehn-Yahr stages across the two cohorts (Fig. 4A–B), and several cohort-specific associations were also observed in PPMI cohort, such as non-human reads, Shannon index, S. liquefaciens, B. avium, C. freundii, and B. parapertussis (Fig. 4B). In addition, in PD-MDCNC cohort, several microbial signatures were observed to facilitate dyskinesia and/or freezing gait of patients with PD, such as non-human reads, Shannon index, species richness, D. nitroreducens, and O. dioscoreae (Fig. 4A). Similarly, in PPMI cohort, non-human reads, Shannon index, and species richness were also associated with postural instability and gait disorder (PIGD) subtype (Fig. 4B). These findings suggested that blood microbial signatures may be associated with PD progression, especially in relation to motor dysfunction.

Fig. 4.

Associations of microbial characteristics with PD progression in the PD-MDCNC or PPMI cohort. (A–B) Partial correlation analysis between microbial characteristics and categorical indicators in PD-MDCNC (A) or PPMI (B) cohort. Only correlations with p < 0.1 are shown. (C–D) Partial correlation analysis between microbial characteristics and continuous indicators in PD-MDCNC (C) or PPMI (D) cohort. Significance levels: ∗p < 0.1; ∗∗p < 0.05; ∗∗∗p < 0.01; ∗∗∗∗p < 0.001.

In addition, we found that the microbial signatures (reflected by non-human reads, microbes sum and Bacteria) and alpha-diversity (represented by Shannon index and species richness) were negatively associated with the age of PD onset in both cohorts (Fig. 4C–D). That indicated that higher blood microbial signatures and alpha-diversity may be associated with earlier disease onset.

Regarding the 19 differential species identified above, most exhibited positive associations with Unified Parkinson's Disease Rating Scale (UPDRS) scores in PPMI cohort, with several replicating in the PD-MDCNC cohort (Fig. 4C–D). Specifically, B. bronchialis was consistently associated with UPDRS-I (non-motor aspects of experiences of daily living), while D. nitroreducens and O. dioscoreae were consistently linked to UPDRS-II (motor aspects of experiences of daily living). Similarly, in PD-MDCNC cohort, six species were positively associated with the score of 39-item Parkinson's Disease Questionnaire (PDQ-39, a self-reported instrument assessing mobility, activities of daily living, emotional well-being, and stigma), i.e. A. pestifer, B. bronchialis, B. hinzii, D. nitroreducens, H. crassostreae, and Pusillimonas sp. M17 (Fig. 4C). Moreover, in both cohorts, B. intermedia, D. nitroreducens, Halomonas sp. NyZ770, and H. crassostreae were consistently correlated with higher PIGD score (Fig. 4C and D). Inverse correlations with cognitive score were respectively observed for multiple species (MMSE in PD-MDCNC cohort; MoCA in PPMI cohort), with four species showing consistent associations across both cohorts, i.e. A. pestifer, B. avium, B. hinzii, and D. nitroreducens (Fig. 4C–D). Additionally, we also found in both cohorts that microbes sum, Bacteria, and four species (C. freundii, P. carotovorum, R. ornithinolytica, S. enterica) were positively correlated with the levodopa equivalent daily dose (LEDD) of patients with PD (Fig. 4C–D).

Taken together, despite some heterogeneous associations in the specific species observed between cohorts, our findings suggested that the blood microbial signatures may be associated with adverse outcomes in patients with PD.

Discussion

In this study, we used PD as a model to explore the potential clinical implications of microbial signatures detected in the blood. By analysing large-scale WGS datasets of two independent cohorts, we found that microbial sequences were widely detected in the sequencing data derived from blood samples, with over 98% of the microbial signals taxonomically assigned to bacterial species. And, significantly elevated microbial signatures were observed in the blood of patients with PD compared to CS. Here, through a population-based and cross-cohort filtration process with verification of re-sequenced data, a key set consisting of 126 bacterial species was identified, in which nearly two-thirds could be detectable in other body sites, especially in the gut and genitourinary tract. Meanwhile, 19 bacterial species defined by higher both abundance and prevalence in PD, effectively discriminated patients from controls. And, several microbial signatures were found to correlate with more severe PD progression. Collectively, these findings support the potential of blood microbial signatures as biomarkers for PD progression and point to their underlying pathophysiologic associations.

Gastrointestinal dysfunction is one of the most common non-motor symptoms in patients with PD, which cause persistent imbalance of gut microbiota and increased permeability of intestinal mucosa.^49,50 These alterations were reported to link with both development and treatment efficacy of PD.¹⁴ Although our study detected stronger microbial DNA signals in the blood of patients with PD, it remains speculative whether this reflect increased mucosal permeability and translocation from the gut, as previously hypothesised.⁸ In addition, across the two cohorts, we consistently observed 19 PD-enriched bacterial species. These species are all Gram-negative members of the phylum Proteobacteria, which has previously been reported to be enriched in the gut microbiota of individuals with various disorders, including PD.^51,52 Beyond blood samples, A. pestifer has also been found in skin and genitourinary tract samples,^53,54 whereas S. liquefaciens has been detected in gut, respiratory, and genitourinary specimens.55, 56, 57 In particular, oral B. intermedia, and gut C. freundii and S. enterica have been reported to be associated with PD progression.^39,42,45

Our exploration of blood microbial signatures, based on WGS data of genomic DNA extracted from peripheral blood leukocytes, suggested that the detected non-human reads may have originated from phagocytosed microorganisms.⁵⁸ In addition, circulating bacterial-derived DNA level was reported to correlate with the degree of systemic inflammatory state among patients with peritoneal dialysis.⁵⁹ Also, in patients with PD, this inflammatory response can be driven by lipopolysaccharide, an endotoxin derived from Gram-negative species.⁶⁰ Such persistent inflammation may be implicated in the disruption of the blood–brain barrier and the degeneration of nigral dopaminergic neurons, thereby potentially accelerating PD progression, such as motor dysfunction and cognitive impairment.60, 61, 62 Therefore, the detection of microbial signatures in the blood raises the hypothesis that exposure to microbial antigens, and subsequent inflammation, may not be confined to the gut but could also originate directly within the circulation.

Several limitations of this study should be acknowledged. First, our filtering workflow, which relied on population detectability and resequencing verification, prioritised reproducibility at the expense of less prevalent yet genuine blood microbial signatures. Second, we quantified microbial signatures using a custom normalisation method adapted from a previous study.³⁴ Given that this method relies on unvalidated assumptions and lacks benchmarking against known inputs, future studies should compare alternative normalisation strategies. Third, the inconsistent or missing recording of some clinical indices and confounders across cohorts may affect the interpretability of certain associations, although we have adjusted for the available confounders. Fourth, given the cross-sectional design of this study, no definitive causal inferences can be drawn. Moreover, the underlying roles of those identified blood microbial signatures in PD pathogenesis warrant further validation through dedicated in vitro and in vivo experiments.

In summary, this study established an analytical workflow to explore microbial signatures in the blood, and observed that those in patients with PD may serve as potential biomarkers for PD progression. Our findings shed light on a previously underexplored dimension of host–microbe interactions. This approach may also be applicable to other complex diseases.

Contributors

Study concept and design: Jinchen Li and Xinging Jian. Methodology and investigation: Xingxing Jian, Pei Yu, Keman Wu, and Yi Zhang. Data curation: Hongxu Pan, Yuwen Zhao, Yige Wang, Jifeng Guo, Yuanfeng Huang, and Beisha Tang. Computing platform: Yijing Wang and Qiao Zhou. Visualisation: Xingxing Jian. Interpretation: Guihu Zhao, Bin Li, Xiaotuan Zhang, Hongxi Zhang and Hao Zhang. Writing—original draft: Xingxing Jian and Pei Yu. Writing—review & editing: Jinchen Li. Supervision: Kun Xia and Beisha Tang. Xingxing Jian and Pei Yu accessed and verified the underlying data. All authors read and approved the final version of the manuscript.

Data sharing statement

No original sequencing data was generated in this study. The outcomes of microbial annotation produced in this study were stored in a public database National Omics Data Encyclopedia (NODE), with access number OEZ00021768 (https://www.biosino.org/node/review/detail/OEV00000669?code=UWBXEHCQ) and OEZ00021769 (https://www.biosino.org/node/review/detail/OEV00000668?code=VYR6RPBY).

Declaration of interests

The authors declared no competing interests.