Longitudinal untargeted metabolomics of cognitive impairment in Parkinson's disease

Abstract

INTRODUCTION

Cognitive impairment is a common non‐motor symptom of Parkinson's disease (PD), yet the mechanism remains unclear.

METHODS

Longitudinal serum samples from PD patients in a community‐based case–control study were profiled, with cognition assessed by the Mini‐Mental State Examination at baseline and follow‐up. We performed three metabolome‐wide association studies: a baseline cross‐sectional analysis (n = 632), a time‐to‐event analysis (n = 412), and a repeated measures analysis (n = 265), followed by “Mummichog” pathway analysis.

RESULTS

Of the 4762 metabolomic features, four reached statistical significance in the baseline analysis, 22 in the time‐to‐event analysis, and none in the repeated measures analysis. Dysregulated pathways included purine metabolism, the pentose phosphate pathway, fatty acid pathways, and C21‐steroid hormone biosynthesis and metabolism.

DISCUSSION

Disruptions in core metabolic pathways may contribute to PD‐related cognitive impairments. Disturbances in pentose and purine pathways may occur before clinical cognitive decline is observed, suggesting early metabolic markers.

Highlights

• We conducted a large metabolomics study with a longitudinal design and used rigorous epidemiological methods like inverse probability weighting to address potential bias.

• Disruptions in several core metabolic pathways are linked to Parkinson's disease (PD) ‐related cognitive impairments.

• Alteration in purine and pentose pathways may precede clinical decline is observed.

1. INTRODUCTION

Cognitive impairment is a common non‐motor symptom of Parkinson's disease (PD) and represents a great disease related burden for patients and their caregivers. Compared to similar age individuals who do not have the disease, PD patients suffer from more rapid decline and higher cumulative risk of developing dementia over time, with as many as 80% of patients eventually developing dementia after 20 years. ¹ , ² , ³ It has also been reported that 25% of PD patients without dementia experience mild cognitive impairment. ⁴ Multiple risk factors have been identified for cognitive impairment in PD, including male sex, older age, lower level of education, duration of PD, severity of other PD symptoms, comorbidity with other diseases, and genetic risk factors. ² , ⁵ We also previously observed that cigarette smoking and exposure to organophosphate pesticides are associated with faster cognitive decline, whereas coffee consumption and a history of competitive sports participation appears to be protective. ⁶ , ⁷

The mechanisms leading to cognitive impairment in PD, however, are not fully understood. ⁸ Previous research has pointed to various pathologies, mainly Lewy body formation in limbic and/or neocortical regions and Alzheimer‐type pathologies such as extracellular beta‐amyloid (Aβ) and intracellular tau deposition. ² Well‐established dementia biomarkers, such as cerebrospinal fluid (CSF) Aβ42 and total tau protein, are also associated with cognitive impairment in PD. ⁹ Furthermore, degeneration of multiple neurotransmitter systems in the brain, including dopamine, noradrenaline, and acetylcholine may also accompany cognitive decline in PD. ² , ⁴ Beyond these brain pathologies, improving our understanding of the systemic metabolism involved in PD dementia may help develop preventive strategies and point to potential treatment targets.

Untargeted metabolomics approaches are increasingly being employed in the search for biomarkers and allow interrogating disease pathology and molecular mechanisms in living patients. To date, however, studies investigating cognitive decline in PD have been generally small in size. For example, a study comparing the serum metabolome of 26 PD patients without to 31 patients with cognitive impairment found inflammation and abnormal lipid metabolism to be more pronounced in the PD patients with cognitive impairment. ¹⁰ Another small study of 41 PD patients reported higher levels of metabolites of fatty acid oxidation in fasting serums from patients with mild cognitive impairment compared to those cognitively normal. ¹¹ Finally, a larger study of 216 PD patients (100 discovery set, 116 validation set) utilized machine learning models to predict cognitive impairment in PD patients and identified the metabolites glycocholic acid and 6‐methylnicotinamide as potential biomarkers. ¹² Targeted metabolomics studies include a study of post‐mortem human brain tissue that described dysregulation in several amino acid pathways and levodopa‐related homocysteine accumulation in the brain of PD patients with dementia. ¹³ An older study focused on monoamine metabolites in CSF and found a deficiency of homovanillic acid and 3‐methoxy‐4‐hydroxyphenylglycol among PD patients with dementia. ¹⁴

Overall, these studies were not only relatively small but also based on cross‐sectional patient data. Here, we conducted a series of untargeted metabolome‐wide association studies (MWAS) of cognitive decline among PD patients using serum‐based data from a large population‐based study: (1) a cross‐sectional analysis of baseline cognitive status; (2) a time‐to‐event analysis; and (3) a repeated measures analysis of cognitive decline.

2. METHODS

2.1. Study population

For analysis, we used PD patients from the UCLA Parkinson's Environment and Genes study (PEG), which is an on‐going community‐based case control study of PD with longitudinal follow‐up of patients, conducted in three agricultural counties from California. Details about the recruitment of study participants have been provided previously. ¹⁵ , ¹⁶ In short, PD patients who are early in the disease course, without serious psychiatric conditions and not being in the later stages of terminal illness, residing at enrollment in three research counties, and within California for ≥ 5 years were eligible for the study. Those who agreed to participate were seen by UCLA movement disorder specialists to confirm the PD diagnosis and collect data on motor and non‐motor features. After initial enrollment, we re‐contacted participants regularly for follow‐up visits. Participants also answered structured interviews and provided blood samples at baseline and during follow‐up visits.

In the current study, we included PD patients from the first two PEG study waves (PEG1: 2001–2007, PEG2: 2011–2018). For the baseline cross‐sectional comparisons, we included all participants who had provided a blood sample and been cognitively tested with the Mini‐Mental State Examination (MMSE) at enrolment (n = 632). For time‐to‐event analysis, we included a subset of patients with serum samples, who were cognitively normal at baseline (MMSE ≥ 25) and had participated in at least one follow‐up visit (n = 412). For the repeated measures analysis, we randomly selected a subset of the follow‐up population to also assess metabolome profiles at follow‐up visits (n = 265), allowing us to assess longitudinal change of the metabolite levels. Figure 1 shows the flow chart of generating the study subjects.

FIGURE 1.

Flowchart demonstrating the process of generating study populations for different analyses.

2.2. Untargeted metabolomics

The serum metabolome was profiled using liquid chromatography–mass spectrometry (LC‐MS); technical details were previously documented. ⁶ , ¹⁷ In brief, serum samples were aliquoted on‐site when patients were examined and kept on ice until they were transferred to a −80°C freezer by the end of the day. Samples were sent to the Emory metabolomics lab in 2019 for metabolomic profiling. After sample preparation, the metabolome was assessed using a dual‐column, dual‐polarity approach, including hydrophilic interaction chromatography with positive electrospray ionization (ESI) and C18 chromatography with negative ESI (hydrophilic interaction liquid chromatography [HILIC]/pos+ and C18/neg‐). Sample extracts were analyzed in triplicate using an Ultimate 3000 liquid chromatography system coupled to a Q‐Exactive high field (HF) mass spectrometer (Thermo Fisher Scientific), with data acquired over an m/z range of 85–1275. The HILIC analyses used a Waters XBridge BEH Amide XP HILIC column (2.1 × 50 mm, 2.5 µm particle size), and the C18 analyses used a Higgins TARGA C18 column (2.1 × 50 mm, 3 µm particle size). We also included a series of quality control (QC) samples to evaluate system performance and applied a series of QC steps. Features were included if Pearson's correlation were > 0.9 and the average coefficient of variation (CV) of feature intensities < 30% within technical replicates.

RESEARCH IN CONTEXT

1. Systematic review: We reviewed the literature using traditional sources such as PubMed. The mechanisms underlying cognitive impairment in Parkinson's disease (PD) are not fully understood, although several studies have utilized metabolomics to investigate the disease‐related metabolic alterations. Those studies are cited and summarized in the introduction.

2. Interpretation: Our findings led to hypotheses regarding molecular mechanisms underlying cognitive decline in PD. Dysregulated pathways including purine metabolism and pentose phosphate pathway may precede clinical decline as they were identified in the time‐to‐event analysis and thus may serve as early markers.

3. Future directions: As an untargeted analysis, this study is primarily hypothesis‐generating. Thus, further work should focus on validating these findings through experimental models (e.g., animal studies) or targeted analyses that more deeply examine the implicated metabolic pathways to better understand their role in disease etiology.

The samples were randomly sorted into runs of 600 samples which were processed in separate lab runs conducted 6 months apart. Longitudinal samples from the same participant (i.e., baseline and follow‐up samples) were included in the same batch of 40 samples to minimize technical variation within person. Metabolite features from the two runs were aligned using the apLCMS R package. As a sensitivity check, we also repeated feature alignment using QC samples only, which yielded an identical set of metabolomics features. This was followed by data pre‐processing that is filtering out features found in less than 50% of the samples, imputation of zeroes using the feature‐wise lowest detected value, log‐transformation, median‐normalization, and controlling for lab batch effects. The final metabolomics matrix included 2046 C18/neg‐ features and 2716 HILIC/pos+ features. Information about data processing success (CV across QCs and before and after data processing, data visualization, etc.) has been previously published. ¹⁸

To annotate  metabolomic features, we used two complementary approaches with different levels of identification confidence. First, we matched features to an Emory in‐house library, a quantitative metabolites database established by the Emory metabolomics lab, ¹⁹ using a mass/charge ratio (m/z) tolerance of < 10 ppm and a retention time (RT) tolerance of < 30 s. These matches were considered higher‐confidence annotations based on accurate mass and RT agreement with the in‐house library. Second, we annotated the remaining features to the Human Metabolome Database (HMDB), Kyoto Encyclopedia of Genes and Genomes (KEGG) compound database, and Lipid Maps using the xMSannotator R package. This last package assigns confidence levels and scores based on isotope/adduct patterns and pathway level associations. ²⁰ We only included annotations with medium or high confidence (confidence level ≥ 2), and for peaks with multiple annotations, we kept those with the highest score. Metabolite names reported here should be interpreted as putative annotations unless confirmed by the Emory in‐house library matching criteria.

2.3. Cognitive function assessment and covariates measurement

Demographic information of patients was collected at baseline in a series of structured interviews that also included lifestyle, medical history, and Geriatric Depression Scale (GDS). Motor symptom severity was assessed by UCLA movement disorder specialists using Unified Parkinson's disease Rating Scale, part III (UPDRS‐III) and cognitive function with the MMSE, an 11‐item measure that tests multiple domains including orientation, registration, attention and calculation, recall, language, and visual construction with a maximum score of 30. We chose a score of < 25 as the threshold suggesting cognitive impairment as this is a widely used cutoff point. A previous review that included 10 studies estimated that among people aged 65 years or older, this MMSE cut‐point has a pooled sensitivity of 0.85 and specificity of 0.90 for cognitive impairment. ²¹ In the time‐to‐event analysis, we defined the event as an MMSE decline ≥ 3, ²² and the time to event as the time difference between the baseline exam and the first follow‐up exam during which the event was observed.

We adjusted for the following variables to control for potential confounders: sex (binary: male, female), race and ethnicity (binary: non‐Hispanic white, others), age at PD diagnosis (continuous, years), PD duration (continuous, years), educational level (3‐categorical: < 12 years, ≥ 12 and < 16 years, ≥ 16 years), smoking history (binary: ever smoked, never smoke), PEG study wave (binary: PEG1, PEG2), and levodopa‐equivalent daily dosage (LEDD, continuous, mg). In sensitivity analyses, we additionally adjusted for baseline GDS and UPDRS‐III scores to account for the influence of other PD features.

2.4. MWAS analysis

2.4.1. Baseline cross‐sectional analysis

We first performed a cross‐sectional MWAS to examine the relationship between baseline metabolomic profiles and cognitive function in PD patients, using empirical Bayes methods included in the limma R package. ²³ Empirical Bayes methods were originally developed for differential expression analyses and have been adopted in metabolomics research as they enhance statistical power and stabilize estimates by borrowing information across features. ¹⁸ We regressed the transformed intensity of metabolomic features on the continuous MMSE score while controlling for potential confounders. To account for multiplicity, we adjusted p‐values using the false discovery rate (FDR) method.

2.4.2. Time‐to‐event analysis

To evaluate how the baseline metabolome predicts cognitive decline over follow‐up, we conducted a time‐to‐event MWAS using weighted Cox proportional hazards (PH) regressions. We included only patients with normal cognitive function (MMSE ≥ 25) at baseline and fit regression models for each metabolomic peak separately. Since we restricted the analysis to patients with at least one follow‐up visit (n = 412), i.e. excluded patients with a baseline visit only (n = 150), selection bias due to informative loss to follow‐up is possible and we accounted for this using inverse‐probability of censoring weights (IPCW). ²⁴ Patients without follow‐up information were considered censored. IPCW creates a pseudo‐population that mimics the population before censoring based on the distribution of predictors for censoring. To calculate IPCW, we selected variables that may affect the probability of censoring using bidirectional stepwise feature selection. The final set of variables used in the IPCW included age at PD diagnosis, smoking history, baseline LEDD, UPDRS‐III, MMSE, and GDS scores. The IPCW was calculated as follows, where C is the indicator of censoring, X is the transformed intensity of the metabolomic feature in each model, and Z is the selected predictors of censoring:

$$IPCW=\frac{P\left(C=0|X\right)}{P(C=0|X,Z)}$$

Metabolomic features with FDR < 0.05 were considered statistically significant in this time‐to‐event MWAS. To test the proportional hazard assumption for Cox PH models, we used the Schoenfeld residual test. The test results and Schoenfeld residual plots are included in Figure S1.

2.4.3. Repeated measures analysis

We fit linear mixed effects models for each metabolomic feature to estimate the relationship of cognitive function with the metabolome longitudinally. We used the transformed intensity of the metabolite as the model's response variable and the fixed effect estimate included the continuous MMSE score, age of PD diagnosis, PD duration at the time of visit, and all other potential confounders described above. We included a random intercept for each patient to account for repeated measures. For some models, the estimated variance of the random effect was close to zero, indicating a singular fit. To ensure reliable estimates, we only retained and interpreted models with a non‐singular fit. Similarly, we reported the FDR to account for multiplicity.

2.4.4. Sensitivity analysis

Several sensitivity analyses were performed. First, to test the robustness of our imputation approach, we repeated the imputation using quantile regression imputation of left‐censored data (QRILC) and repeated the MWAS. ²⁵ Second, as the severity of other PD features may increase the risk of cognitive dysfunction in PD and influence the metabolome, we conducted sensitivity analyses by adjusting baseline UPDRS‐III and GDS scores as additional covariates in our MWAS analyses. We further conducted an additional baseline cross‐sectional MWAS using binary indicator of cognitive function (MMSE < 25 vs. MMSE ≥ 25) as the outcome to check the robustness of our results. To account for the potential sex difference in metabolic changes we observed in the repeated measures analysis, we also did a sex‐stratified analysis following the same analytic approach described before. Additionally, to evaluate whether known PD‐related genetic risk influenced our findings, we conducted a sensitivity analysis excluding participants known to carry adverse glucocerebrosidase beta (GBA) gene mutations (rs409652/G202R, rs78973108/R257Q, rs1064651/D409H, and rs421016/L444P), which we previously assessed in this study population. ²⁶ We compared the results with the original MWAS results by calculating Pearson's correlation coefficients for the regression coefficients from both analyses.

2.5. Pathway analysis

To identify metabolic pathways influenced by cognitive decline, we conducted pathway analyses based on MWAS results using the “Mummichog” software. Mummichog uses a permutation‐base framework and predicts functional network activity directly from untargeted mass spectral data. ²⁷ The input matrix includes the mass/charge ratio, retention time in seconds, polarity (positive or negative), statistics (Z/T‐values), and uncorrected p‐values from MWAS. The significance threshold was set to an uncorrected p‐value < 0.05, the mass tolerance to 5 ppm, and the analytical mode to “mixed”, respectively. We restricted adducts to their main types: M‐H[1‐], M‐2H[2‐], and M‐H2O‐H[1‐] for the negative column and M [1+], M+H [1+], and M+Na [1+] for the positive column. The pathway analyses were done with the MetaboAnalystR R package (version 4.0). ²⁸ Pathways with three or more metabolites detected and a p‐value < 0.1 are included for interpretation.

All statistical analyses were performed in R. Data visualizations were generated using the ggplot2 R package, and schematic illustrations were created with BioRender.com.

3. RESULTS

3.1. Characteristics of the study population and their cognitive function changes

Table 1 shows characteristics of study subjects who contributed to the three different analyses. At study baseline, patients with impaired cognitive function (MMSE < 25) were older, had lower average education, experienced more severe motor symptoms and depression, and were on higher doses of PD medications compared to those with normal cognition. Additionally, a larger proportion of these patients were male and belonged to racial/ethnic groups other than non‐Hispanic White.

TABLE 1.

Characteristics of study population.

	Population for baseline analysis			Population for time‐to‐event analysis	Population for repeated measures analysis
Variable a	Overall n = 632	MMSE ≥ 25 n = 562	MMSE < 25 n = 70	n = 412	n = 265
Baseline MMSE score	27.6 (2.7)	28.4 (1.5)	21.5 (2.9)	28.6 (1.4)	28.2 (2.1)
Age at Parkinson's disease diagnosis	66.9 (10.4)	66.4 (10.5)	71.2 (8.2)	65.5 (10.3)	64.3 (10.4)
Parkinson's disease duration (years) at baseline	3.0 (2.6)	3.0 (2.6)	3.5 (2.4)	3.0 (2.7)	3.5 (2.9)
Sex
Male	403 (64%)	352 (63%)	51 (73%)	251 (61%)	172 (65%)
Female	229 (36%)	210 (37%)	19 (27%)	161 (39%)	93 (35%)
Race and ethnicity
Non‐Hispanic White	475 (75%)	436 (78%)	39 (56%)	323 (78%)	200 (75%)
Others	157 (25%)	126 (23%)	31 (44%)	89 (22%)	65 (25%)
Education level
< 12 years	107 (17%)	71 (13%)	36 (51%)	47 (11%)	32 (12%)
≥ 12 & < 16 years	304 (48%)	278 (49%)	26 (37%)	201 (49%)	124 (47%)
≥ 16 years	221 (35%)	213 (38%)	8 (11%)	164 (40%)	109 (41%)
Smoking history
Ever smoke	293 (46%)	258 (46%)	35 (50%)	172 (42%)	105 (40%)
Never smoke	339 (54%)	304 (54%)	35 (50%)	240 (58%)	160 (60%)
PEG study waives
PEG 1	277 (44%)	251 (45%)	26 (37%)	178 (43%)	36 (14%)
PEG 2	355 (56%)	311 (55%)	44 (63%)	234 (57%)	229 (86%)
Baseline LEDD	417.3 (332.6)	410.7 (331.3)	470.5 (341.3)	396.8 (317.9)	423.9 (336.0)
Baseline GDS score	3.6 (3.2)	3.3 (3.1)	5.7 (3.3)	3.0 (3.1)	3.1 (3.2)
Missing b	2	2		1
Baseline UPDRS‐III score	21.6 (11.4)	20.7 (11.1)	28.1 (11.4)	19.4 (9.6)	19.8 (10.3)
Parkinson's disease duration (years) at following‐up visit	—	—	—	—	7.9 (3.8)
MMSE score at following‐up visit	—	—	—	—	27.5 (3.3)

Abbreviations: GDS, Geriatric Depression Scale; LEDD, levodopa‐equivalent daily dosage; MMSE, Mini‐Mental State Examination; PEG, Parkinson's Environment and Genes study; SD, standard deviation; UPDRS‐III, Unified Parkinson's Disease Rating Scale, part III.

^aData are presented as mean (SD) for continuous variables and count (percentage) for categorical variables.

^bTwo patients missed GDS scores and were therefore excluded from the sensitivity analysis with additional adjustment for baseline GDS.

Figure 2 demonstrates cognitive change patterns for patients included in longitudinal analyses. In the time‐to‐event analysis, the outcome was observed in 106 patients (25.8%), and the median time to cognitive decline was 8.3 years (Figure 2B). Compared to patients who stayed in the study, patients censored after baseline were older, had a higher proportion of smokers, and exhibited more PD‐related symptoms at baseline (Table S1). For the repeated measure analysis, the average time between the baseline and follow‐up visit was 4.4 years (SD = 3.3).

FIGURE 2.

Figures showing the cognitive change patterns for the longitudinal study population. (A) The spaghetti plot showing the MMSE change over follow‐up period for patients included in the time‐to‐event analysis. (B) The Kaplan–Meier survival curve for time‐to‐event analysis. (C) The spaghetti plot showing the repeated MMSE measures for the patients included in the repeated measures analysis. MMSE, Mini‐Mental State Examination.

3.2. MWAS results

Among 4762 metabolomic features included in MWAS, 1689 were we successfully annotated. In the cross‐sectional analysis, empirical Bayes models selected 475 features associated with the MMSE score at baseline with raw p‐values < 0.05 (Figure 3A). After FDR correction, four remained statistically significant, one of which was annotated from xMSannotator; however, the annotation was not unique, corresponding to either phosmet or meclofenamate sodium (m/z = 315.989, RT = 34.8 sec, C18/neg‐). This peak showed a negative association, indicating higher levels were associated with lower cognitive scores (log₂FC = −1.77, FDR = 0.005).

FIGURE 3.

Plots showing and comparing the MWAS results. (A–C) Volcano plots showing the results from the baseline cross‐sectional, time‐to‐event, and repeated measure analyses, respectively. The metabolomic features with FDR < 0.05 and are marked in red and labelled with the annotated compound name. The dashed line indicates the unadjusted alpha threshold of 0.05, and features passed that are considered associated with the outcome in the following pathway analyses. (D–F) Scatter plots comparing coefficients across the different MWAS analyses. Only peaks with a rawp < 0.05 in at least one analysis are shown. Pearson correlation coefficients were calculated for each comparison (showed in the left‐upper corner of each plot). FDR, false discovery rate; MWAS, metabolome‐wide association studies.

The time‐to‐event analyses identified a different set of features associated with incident cognitive decline. Of the 441 features with a raw p‐value < 0.05, 22 met the FDR threshold of 0.05, and six were annotated at high confidence. These included three metabolites that were uniquely annotated from xMSannotator: dibenzyl ether (m/z = 197.097, RT = 177.6 sec, C18/neg‐ column, logHR = −0.17, FDR = 0.002), 12‐dichloroethane (m/z = 140.002, RT = 294.7 sec, HILIC/pos+ column, logHR = −0.26, FDR = 0.010), and 11‐aminoundecanoic acid (m/z = 202.18, RT = 28.2 sec, HILIC/pos+ column, logHR = 0.242, FDR = 0.038).

In the repeated measures analysis, 4199 of the 4762 (88.2%) linear mixed‐effects models achieved a non‐singular fit and were retained for further evaluation. Among these, 353 features had a raw p‐value < 0.05; however, none remained statistically significant after adjustment for multiple testing. The information on ,   annotated and statistically significant metabolomic features is presented in Table 2. The upper panels of Figure 3 present volcano plots for three MWAS analyses, highlighting features that remained statistically significant after FDR correction. Comprehensive MWAS results, including annotations for all features, can be found in Table S2.

TABLE 2.

Significant and annotated metabolites from metabolome‐wide association study results.

Analysis	m/z	RT (sec)	Column	Annotation	Database ID	Database	Multiple annotation	Regression coefficient a	p‐value	FDR q‐value
Baseline cross‐sectional analysis	315.989	34.8	C18/neg‐	Phosmet; Meclofenamate sodium	C18756; C02996	KEGG	Yes	−1.769	2.5E–06	0.005
Time‐to‐event analysis	197.097	177.6	C18/neg‐	Dibenzyl ether	HMDB32078	HMDB		−0.174	1.9E–06	0.002
	140.002	294.7	HILIC/pos+	12‐Dichloroethane	HMDB29571	HMDB		−0.258	3.7E–05	0.010
	202.180	38.2	HILIC/pos+	11‐Aminoundecanoic acid	C19325	KEGG		0.242	2.7E–04	0.038
	387.275	37.9	HILIC/pos+	Cryptomeridiol 11‐rhamnoside; Mangalkanyl glucoside; Glycerol trihexanoate	HMDB38018; HMDB36015; HMDB31125	HMDB	Yes	−0.201	5.2E–05	0.013
	405.300	38.1	HILIC/pos+	11‐Carboxy‐gamma‐chromanol; MG(22:5(7Z10Z13Z16Z19Z)0:00:0); MG(22:5(4Z7Z10Z13Z16Z)0:00:0); MG(0:022:5(7Z10Z13Z16Z19Z)0:0); MG(0:022:5(4Z7Z10Z13Z16Z)0:0)	HMDB12517; HMDB11586; HMDB11585; HMDB11556; HMDB11555	HMDB	Yes	−0.282	2.1E–05	0.008
	429.325	39.6	HILIC/pos+	Glyceryl lactooleate; Sorbitan oleate	HMDB32298; HMDB29886	HMDB	Yes	−0.161	1.1E–05	0.005

Abbreviations: FDR, false discovery rate; m/z, mass/charge ratio; PH, proportional hazards; HILIC, hydrophilic interaction liquid chromatography; HMBD, Human Metabolome Database; KEGG, Kyoto Encyclopedia of Genes and Genomes.

^aRegression coefficients were from empirical Bayes models for the baseline cross‐sectional analysis and Cox PH models for the time‐to‐event analysis.

We conducted pairwise comparisons among the three MWAS, as shown in Figure 3D–F. The regression coefficients from the baseline and repeated‐measures analyses were moderately correlated (Pearson's r = 0.31, Figure 3D). In contrast, the results from the time‐to‐event analysis are notably different from the other two MWAS (Figure 3E and 3F). Beyond differences in model setup, this discrepancy highlights that each analysis captures distinct aspects of cognitive decline in PD patients.

3.3. Pathway analysis results

The Mummichog pathway analysis revealed different sets of dysregulated metabolic pathways for the three different analyses. In the baseline analyses, we identified the TCA cycle (p‐value = 0.05), purine metabolism (p‐value = 0.08), ascorbate (Vitamin C) and aldarate metabolism (p‐value = 0.09), and tyrosine metabolism (p‐value = 0.10). The time‐to‐event analysis linked two pentose‐related pathways to cognitive decline: the pentose phosphate pathway (PPP, p‐value = 0.007) and pentose and glucuronate interconversions (p‐value = 0.04). Additionally, purine metabolism was identified in the time‐to‐event analysis with a p‐value = 0.09. In the repeated measures analysis, most identified pathways were related to lipid/lipid‐like molecules, including fatty acid activation, C21‐steroid hormone biosynthesis and metabolism, and fatty acid metabolism. Vitamin K metabolism was also found to be dysregulated in this analysis. The pathway analysis results are presented in Figure 4. Information on the peaks contributing to the dysregulated pathways is provided in Table S3.

FIGURE 4.

Pathway analysis results using the Mummichog algorithm. The three panels represent results from the baseline cross‐sectional, time‐to‐event, and repeated measures analyses, respectively. The pathway p‐values were calculated using Fisher exact tests. The dashed line shows the alpha threshold of 0.1. The enrichment factors were calculated by dividing the number of significant peaks by the expected number of peaks within that pathway. Pathways with three or more peaks associated with the outcome (significant hits with rawp < 0.1) are marked as red in the figures.

3.4. Sensitivity analysis results

The QRILC‐based sensitivity analysis produced results that were highly consistent with the primary analysis. The MWAS coefficients from the two approaches were strongly correlated (r > 0.94, Figure S2), and similar sets of features were identified, supporting the robustness of the primary imputation strategy.

To account for any potential confounding caused by other PD features, we additionally adjusted for the UPDRS‐III or the GDS scores. Correlations between results from the sensitivity analyses and the original analyses are shown in Figure S3. Pearson's correlation coefficients are all above 0.97, indicating a very strong correlation between the sensitivity analysis results and the main results. Thus, we concluded that adjusting for UPDRS‐III or GDS did not change the original results.

We also conducted sensitivity analysis using a binary indicator of cognitive impairment at baseline. The correlation between these analyses was high (Figure S4, Pearson's r = 0.81). Two pathways were successfully replicated in this analysis, that is, purine metabolism (p‐value = 0.01) and the TCA cycle (p‐value = 0.04), while the remaining pathways were not statistically significant (Table S4).

To further investigate sex‐related pathways (i.e., C21‐steroid hormone biosynthesis and metabolism) identified in the repeated measures analysis, we stratified by sex (the correlation pattern is shown in Figure S5). In the sex‐stratified MWAS, one peak from the C18 column (m/z = 750.565, RT = 234.8 sec) and one peak from the HILIC column (m/z = 209.073, RT = 75.1 sec) were found to be positively associated with MMSE in women, yet neither of them was annotated. None of the remaining features achieved formal statistical significance after FDR correction. Pearson's correlation coefficients were 0.73 and 0.74 respectively for results in women and men compared to the overall sample. The C21‐steroid hormone biosynthesis and metabolism were not identified in the sex‐stratified analyses. Nevertheless, most peaks within this pathway demonstrated similar regression coefficients in the combined and stratified analysis (Table S5). The top pathways picked in the pathway analysis can be found in Table S4.

In our study population, only a small proportion of participants carried known pathogenic GBA mutations (7/632, 1.1%), suggesting that PD cases in the PEG study were predominantly sporadic. After excluding participants known to carry these GBA mutations, the MWAS remained largely unchanged (Figure S6); thus, we concluded that our results were not driven by these pathogenic mutation carriers.

4. DISCUSSION

The community‐based population of PD patients from the PEG study enabled us to investigate cognitive function changes in PD with metabolomic tools in a longitudinal manner. Relying on three different MWAS analyses, we identified seven annotated and 19 unidentified metabolites associated with either prevalent cognitive impairment or cognitive decline of initially unimpaired PD patients during follow‐up. We also found several different pathways to be disturbed either before the cognitive decline occurred or during the progression of the condition. Specifically, differences in tyrosine metabolism, the TCA cycle, and ascorbate and aldarate metabolism were observed among PD patients early in the disease (< 5 years of PD duration) comparing those with normal versus low cognitive function at baseline. Purine metabolism differed both at baseline and in those who experienced faster cognitive decline during follow‐up. Dysregulation of pentose‐related pathways was also detected among patients at enrolment who then declined faster, while alterations in fatty acid metabolism and C21‐steroid hormone metabolism occurred alongside cognitive changes over time. Although the metabolomic patterns from the three analyses appear distinct, each targets a different mechanism: the baseline analysis examines cognitive impairment in early PD, the time‐to‐event analysis investigates the speed of cognitive decline, and the repeated measures analysis assesses metabolic changes occurring alongside the decline.

Lipid metabolism plays a multifaceted role in brain health and is closely related to cognitive function. ²⁹ In the central nervous system, lipids are major structural components of myelin, and maintenance of myelin integrity is essential for efficient axonal conduction, circuit connectivity, and cognitive processing. ³⁰ Disruption of lipid homeostasis may therefore contribute to cognitive decline not only through oxidative stress, blood–brain barrier disruption, neuroinflammation, mitochondrial dysfunction, and amyloidogenesis, ³¹ but also through altered myelin turnover, repair, or stability. Previous MWAS and lipidomics studies have found alterations in lipid composition and lipid pathways in PD and Alzheimer's disease (AD), including studies of PD patients with cognitive impairment. ¹⁰ , ¹¹ Among lipid‐related findings reported to date, structural lipid classes such as glycerophospholipids and sphingolipids have been most consistently associated with both PD and AD. ³² , ³³ In PD, dysregulation of monounsaturated fatty acid metabolism has also been linked to exacerbation of alpha‐synuclein neurotoxicity. ³⁴ In our study, lipid pathways repeated measures analysis highlighted several fatty acid pathways as dysregulated among patients whose MMSE scores dropped during progression. These findings align with prior AD and PD studies that reported perturbations in fatty acid biosynthesis and metabolism, supporting the hypothesis that lipid metabolic changes may accompany or potentially contribute to neurodegenerative processes relevant to cognitive decline in PD. Future lipidomics and multi‐omics studies are needed to clarify how these pathway‐level signals relate to underlying disease mechanisms. Repeated measures analysis also implicated the metabolism of C21‐steroid hormones, which includes the family of progestogens, in the development of cognitive impairment. The role of progesterone in cognition and memory has been investigated for a long time, ³⁵ , ³⁶ yet research on progesterone's effect on cognitive function among menopausal women has been contradictory. ³⁷ , ³⁸ In our sex‐stratified analysis, we observed comparable regression coefficients for the associated peaks within this pathway for both men and women, but they did not reach formal statistical significance, likely due to limited statistical power. This certainly warrants further investigation into the role that steroid hormones may play in cognitive decline in PD.

Our time‐to‐event analysis highlighted two pentose‐related pathways, including the PPP, and both the time‐to‐event and cross‐sectional analyses consistently found purine metabolism to be altered among patients with either low cognitive scores at baseline or faster cognitive decline during follow‐up. Purine metabolism and PPP are involved in a wide range of fundamental cellular processes, including DNA repair, energy transfer, redox balance, and intracellular signaling. And emerging evidence supports that a deficit or deregulation of the enzymes involved in purine metabolism may contribute to multiple neurological conditions including cognitive impairment. ³⁹ Additionally, purine nucleosides (adenosine and guanosine) are involved in neuroprotective mechanisms in the brain, ⁴⁰ and urate, the end product of purine metabolism, is a circulating antioxidant that may have protective effects or reflect oxidative and metabolic processes relevant to diseases such as PD and AD. ⁴¹ , ⁴² Notably, although purine metabolism was identified at the pathway level, we did not observe a significant association between serum uric acid level and cognitive decline among individuals with PD in our population. Nicotinamide adenine dinucleotide phosphate (NADPH), another product of PPP, is a cofactor in anabolic reactions (including nucleic acid synthesis) and glutathione regeneration and plays an essential role in reducing oxidative stress and neuroinflammation in the brain. ⁴³ It has been reported that PPP dysfunction and overacting NADPH oxidase contribute to dopaminergic neurodegeneration, and they have been suggested as potential targets for PD treatment. ⁴⁴ Our time‐to‐event analysis suggests that the dysregulation of PPP and purine metabolism precedes faster cognitive decline; thus, it is possibly an early marker as well as a valuable target for interventions aiming to slow cognitive decline in PD patients.

Our study is to date the largest metabolomic study of PD‐associated cognitive impairment and decline. The longitudinal design of the PEG study provided us with the unique opportunity to study different aspects of cognitive impairment and decline. We identified metabolic disturbances associated cross‐sectionally with the presence of cognitive dysfunction early in the PD course, and disturbances preceding the clinical manifestation of the condition, potentially providing us with predictors of cognitive decline in PD. However, we were only able to annotate just over one third of all features. Unannotated metabolite features are a limitation of current technologies and remain the biggest challenge of any untargeted metabolomics study, as there is a sizeable proportion of unannotated features as well as multiple annotations tied to the same features. Specifically, annotations assigned to MWAS‐highlighted features are putative. Some features may reflect environmental, medication‐related, or other exposure‐associated signals rather than endogenous metabolites. We avoided over‐interpreting these findings without further evaluation based on targeted metabolite confirmation and relevant exposure information. Additionally, we were only able to assess cognitive functions using the MMSE and there are reports that the MMSE is less sensitive in screening for mild cognitive impairment. ⁴⁵ Thus, non‐differential misclassification of the outcome is a concern. In the time‐to‐event analysis, we used IPCW to account for informative censoring after baseline. However, we cannot exclude the possibility of residual selection bias. Finally, by design, our analyses are somewhat exploratory, with the goal of generating plausible hypotheses for future investigations of this important condition.

In conclusion, by utilizing LC‐MS based metabolomics we identified metabolites and pathways that are related to cognitive decline in PD including the dysregulation of fundamental metabolic processes, such as pentose and purine metabolism pathways, and the metabolism of certain lipid/lipid‐like molecules including different fatty acids, and possibly progestogens. These findings may help nominate biological processes for future mechanistic and intervention‐focused studies of cognitive decline in PD. Purine metabolism and the PPP may reflect altered nucleotide metabolism and redox balance, and fatty acid pathways indicate lipid homeostasis. However, this observational study used an untargeted LC‐MS and should be considered hypothesis‐generating. Targeted metabolomic and experimental studies as well as comparisons with non‐PD related cognitive impairment are needed to inform possible intervention strategies.

CONFLICT OF INTEREST STATEMENT

The authors declare no conflicts of interest. Author disclosures are available in the Supporting Information.

CONSENT STATEMENT

This study was approved by the Institutional Review Boards of the University of California Los Angeles. All participants were informed of the procedures and their rights and were provided with written informed consent.

Supporting information

Supporting Information

Supporting Information

Supporting Information

Lin Y, Yu Y, Zhang K, et al. Longitudinal untargeted metabolomics of cognitive impairment in Parkinson's disease. Alzheimer's Dement. 2026;22:e71890. 10.1002/alz.71890