Plasma Proteomic Signatures of Physical Activity Provide Insights into Biological Impacts and its Protective Role against Dementia

Abstract

Purpose:

Physical activity (PA) and sedentary behavior (SB) are associated with many diseases, including Alzheimer disease and all-cause dementia. However, the specific biological mechanisms through which PA protects against disease are not entirely understood. This study aims to address this gap, with a specific focus on all-cause dementia.

Methods:

We first assessed the conventional observational associations of three self-reported and three device-based PA/SB measures with circulating levels of 2911 plasma proteins measured in the UK Biobank (n_max = 39,160) and assessed functional enrichment of identified proteins. We then used bidirectional Mendelian randomization to further evaluate the evidence for causal relationships of PA/SB with protein levels. Finally, we performed mediation analyses to identify proteins that may mediate the relationship of PA with incident all-cause dementia.

Results:

Our findings revealed 41 proteins consistently associated with all PA measures and 1027 proteins associated with at least one PA measure. Both conventional observational and Mendelian randomization study designs converged on proteins that appear to increase as a result of PA, including integrins such as ITGAV and ITGAM, as well as MXRA8, CLEC4A, CLEC4M, LPL, and ADGRG2; and on proteins that appear to decrease as a result of PA such as LEP, INHBC, CLMP, PTGDS, ADM, OGN, and PI3; and on proteins that are more responsive to high-intensity PA, such as CA14, CA6, CA4, KIT, and ANGPT2. Functional enrichment analyses revealed processes such as cell-matrix adhesion, integrin-mediated signaling, and collagen binding. Finally, GDF15, ITGAV, ITGAM, ITGA11, HPGDS, GFAP, ADM, AHNAK, and DPP4 were among 21 unique proteins found to mediate the relationship of PA with all-cause dementia, implicating processes such as synaptic plasticity, neurogenesis, and inflammation.

Conclusions:

Our results provide insights into how PA affects biological processes and protects against dementia, and provide avenues for future research into the health-promoting effects of PA.

Physical activity (PA) and sedentary behavior (SB) are associated with health outcomes such as cardiovascular function, weight management, cognitive function, and mental health (1). There are well-established associations between PA and risk of several chronic diseases such as heart disease, type 2 diabetes, depression, anxiety, osteoporosis, and cancer (2,3). Similarly, SB has been linked to adverse cardiometabolic outcomes (4,5). PA and SB are also increasingly recognized for their potential role in the risk and progression of dementia (6–10).

Regular PA may enhance neuroplasticity and cognitive functioning through improved cerebral blood circulation, stimulation of neural tissue development, and the release of other immunological and endocrine factors (11–13). Additionally, PA improves metabolic functioning, in part, by reducing systemic inflammation and oxidative stress, which are pathological mechanisms also implicated in all-cause dementia (14–17). SB has also been linked to adverse cognitive outcomes such as reduced cerebral flow (9,18). Although the neurocognitive benefits of PA are often observed, the precise mechanisms and biological underpinnings of this protection are not well understood (9,14,18,19). Investigating the biological pathways that underlie the association between PA/SB and all-cause dementia can increase opportunities for the precise and effective prevention and management of dementia (12).

Investigating circulating plasma protein levels offers a promising avenue to uncover the biological mechanisms underlying the health-promoting effects of PA and limiting SB. These insights could pave the way for innovative applications in clinical practice, enabling more targeted and effective interventions (19,20). However, previous human studies examining the relationship between PA/SB and plasma protein levels have been based on relatively small sample sizes or small sets of measured proteins (21–25). For example, Stattin et al. (21) examined 184 proteins among 6500 individuals. Moreover, only self-reported PA measures are typically examined (21,22). While substantial progress has been made in understanding the proteomic changes associated with cardiovascular health and PA, there remains a critical gap in linking these molecular changes to cognitive outcomes such as dementia (20–27). For instance, a study by Robbins et al. (28) demonstrated how plasma proteomic changes in response to exercise training are associated with cardiorespiratory fitness adaptations, highlighting the diverse physiological pathways through which PA can confer cardiorespiratory benefits. Furthermore, while previous research has attempted to find associations between PA and plasma protein levels, conventional observational studies may be biased by unmeasured and residual confounding factors and reverse causation (29). As a result, alternative study designs such as Mendelian randomization (MR) can help to assess causal relationships (29–31).

We performed the most extensive study to date testing the associations of PA and SB with the levels of 2911 plasma proteins, using self-reported and device-based measures (n_max = 39,160) from the UK Biobank. We compared the evidence for causality across the different types of PA/SB measures and two study designs: a conventional observational study and a bidirectional MR study. We then assessed the functional enrichment of identified proteins and assessed whether any specific proteins mediate the relationship between PA/SB and all-cause dementia in the UK Biobank (Fig. 1).

FIGURE 1.

Study design and analytical framework for investigating the relationship of physical activity and sedentary behaviors with plasma proteins and incident all-cause dementia. The framework consists of four main components: 1) Phenotypic (conventional observational) associations: Analysis of associations between PA/SB measures and 2911 plasma proteins from the UK Biobank, with sample sizes ranging from n = 9496 to n = 39,160. 2) Pathway enrichment analyses: Identification of protein signaling pathways, Gene Ontology (GO) biological processes, and GO molecular functions enriched among the associated proteins. 3) Bidirectional Mendelian randomization: Examination of causal relationships between PA and proteins using genetic instruments (SNPs) for both PA and protein levels, employing Mendelian randomization methods such as Inverse Variance Weighted (IVW), Wald Ratio, and MR-Egger. 4) Mediation and interaction analyses: Assessment of the direct, indirect, mediated interaction, and interaction effects of PA-associated proteins on the risk of clinically diagnosed incident all-cause dementia. (Created with BioRender.com).

METHODS

UK Biobank Proteomics

The UK Biobank is a large prospective study with over 500,000 participants aged 39–73 yr at the time of recruitment between 2006 and 2010 (32). Between November 2020 and March 2021, the Pharma Proteomics Project (PPP) consortium selected 54,219 representative participants through age, sex, and study center stratification for the proteomics substudy (33). Blood samples were obtained at each participant’s baseline visit. The time of day of blood sample collection varied for each individual (33). UK Biobank has approval from the North West Multi-Center Research Ethics Committee as a Research Tissue Bank.

Between May 2021 and November 2022, plasma profiling was performed using Olink Explore 3072 (Uppsala, Sweden) technology, in which a matched pair of antibodies labeled with unique complementary oligonucleotides (proximity probes) binds to the respective target proteins in each sample (33). The oligonucleotides come into proximity and hybridize with each other to enable DNA amplification of the protein signal, which is quantified on a next-generation sequencing readout, as detailed elsewhere (33). Antibodies for 2923 unique plasma proteins were available in the UK Biobank PPP (33). The protein panel covers major biological pathways in inflammation, oncology, cardiometabolic, and neurological domains (33). The performance of each protein assay was validated for specificity, sensitivity, dynamic range, precision, scalability, detectability, and endogenous interference (33). Additional details on sample batching, data preprocessing, and quality checking were previously published (33). Quality control of the proteomic data for the current study was performed in R (version 4.3.3; R Foundation for Statistical Computing, Vienna, Austria) (34) by excluding proteins with more than 20% missing values (n = 12) and samples with more than 50% missing values, leaving a set of 2911 proteins for analyses.

Conventional Observational Associations of PA With Protein Levels

We used linear regression to examine the cross-sectional associations of six PA-related variables with each plasma protein. Three PA variables were self-reported: moderate to vigorous PA (MVPA), vigorous PA (VPA), and strenuous sports and other exercises (SSOE). As previously described, these were derived from questionnaire responses obtained during the UK Biobank baseline assessment (2006–2010) (35). Additionally, three device-based PA/SB measures were assessed in 2014 for associations with plasma protein levels measured 4–8 yr earlier at the baseline visit (35). A subset of 103,684 participants agreed to wear a 3-axis logging accelerometer (Activity) for 24 h·d⁻¹ for 7 d on their dominant wrist, from 2013 to 2015. These device-based measures were: average acceleration (AvgAcc), the fraction of accelerations >425 milligravities (FracAcc > 425), and SB (AccSed) derived from a machine-learning model (36) as previously applied (9). The questions for self-reported PA measurements, participant device-wear information, and variable construction were created and transformed as previously described (37). We applied consistent exclusion criteria and data quality control measures to the accelerometry-based measures. Individuals with at least 3 d of data were included in the study (9,37).

The numbers of participants with proteomic data and self-reported PA/SB measures were as follows: self-reported measures (MVPA: n = 40,787; VPA n = 10,502 physically active and 18,092 controls; SSOE: n = 13,179 physically active and 25,770 controls) and device-based measures (AvgAcc: n = 9851; FracAcc > 425: n = 9496; AccSed: n = 9806). We conducted linear regression for the independent relationship between each PA/SB variable as the exposure and each protein as the outcome, resulting in a total of 6 × 2911 regressions. All proteins were rank-based inverse normal transformed (RNT) before analysis to ensure normality and unit comparability. Of the six PA/SB variables, VPA and SSOE were treated as binary measures, MVPA and FracAcc >425 as RNT continuous measures, and AvgAcc and AccSed as continuous variables in the models. Data cleaning and variable derivation steps were applied as previously described (9,35,36). Each model assessing the PA-protein association was adjusted for age (at baseline or device use), sex, season (questionnaire or device use), race and ethnicity, body fat percentage (measured by a Tanita BC418MA bioimpedance scale), alcohol intake (drinks per week), smoking status (never, former, and current), education (college or higher vs less), and the presence of major health conditions (cancer, type 2 diabetes, and hypertension). All statistical analyses of conventional observational associations were performed using R software (version 4.3.3). We selected all PA-protein pairs with a P < 1.7 × 10⁻⁵ as statistically significant, based on a Bonferroni correction for 2911 tests. Although six PA/SB measures were tested, we consider this a conservative correction given the substantial correlations among proteins and among PA/SB measures.

Pathway and Functional Enrichment Analyses of Conventional Observational Results

We conducted gene enrichment analyses for the gene set corresponding to the proteins associated with each PA/SB measure and the set of proteins associated with all six PA/SB measures. We analyzed DAVID gene enrichment and functional annotation using the complete set of 2911 Olink 3072 proteins as the background (38–40). We derived enrichment terms for Gene Ontology biological process and Gene Ontology molecular function. A Benjamini–Hochberg correction was used to assess whether our identified genes are enriched for specific functional pathways.

Bidirectional MR

We performed bidirectional MR for PA/SB and protein levels for all pairs with a conventional observational association (P < 1.7 × 10⁻⁵). This approach relies on three core assumptions: 1) the relevance assumption requires that the selected genetic variants are robustly associated with the exposure; 2) the independence assumption assumes that there are no confounders of the instrument-outcome association; and 3) the exclusion restriction assumption requires that the genetic variants influence the outcome only through their effect on the exposure. Specifically, our MR analyses focused on the proteins (n = 1027) that were associated with at least one PA or SB measure in the conventional observational analyses.

We assessed PA/SB exposure SNPs selected at the P < 5 × 10⁻⁸ in the primary analyses and SNPs selected at P < 5 × 10⁻⁶ for further exploratory analyses. We used GWAS of the same set of PA variables described in Klimentidis et al. (35), using the same derivation and transformation steps for the PA variables. Additional MR analyses included publicly available GWAS summary statistics of two self-reported measures based on larger sample sizes (41): leisure screen time (n = 526,725) and MVPA during leisure time (n = 608,595). We used publicly available GWAS summary statistics reported in Sun et al. (33) to genetically instrument protein levels. Specifically, we used the GWAS based on the European ancestry subset in the UK Biobank PPP (n = 34,557) since this corresponds to the ancestry used in the PA GWAS. Additionally, we performed MR with each protein as the exposure, limiting our exposure SNPs to cis-pQTLs within 1 Mb upstream and downstream of the transcription start site for the respective protein-coding gene to reduce the potential for pleiotropic variants outside the protein-encoding gene (42). All MR statistical analyses were performed in R (version 4.2; R Foundation for Statistical Computing, Vienna, Austria) using the “TwoSampleMR” package (43). Clumping for MR was performed with Plink 1.9 using the “EUR” subpopulation from 1000 Genomes as a reference. We first assessed results from the inverse variance weighted (IVW) method. Data was harmonized using the harmonise_data() function from the “TwoSampleMR” R package with the default option that tries to infer positive strand alleles using allele frequencies, and excludes palindromic SNPs with intermediate allele frequencies to minimize the risk of strand ambiguity. When only one SNP exposure instrument was available, we applied the Wald ratio method and reported those results. To strengthen our findings, we examined the results of sensitivity analyses, including MR-Egger regression to account for directional pleiotropy, the weighted median method to account for the presence of invalid SNPs, and a mode-based estimator that grouped similar SNPs to minimize the impact of outliers. F-statistics for PA exposure instruments are all >10, as presented elsewhere (44,45). We selected all PA-protein pairs that showed evidence from MR for a causal effect in either direction, using a preset significance criterion of P < 4.9 × 10⁻⁵. This criterion was chosen based on a Bonferroni correction for the number of proteins tested. The protein-PA pairs identified in MR were forwarded to the colocalization analysis. We used MR to further examine the evidence for causality and directionality for the set of proteins associated with all PA and SB measures.

We created four composite categories to classify the strength of evidence for PA-protein MR associations based on P value thresholds and consistency of effect direction across methods. To further assess the consistency of the effect estimates of PA/SB on protein levels across conventional observational and MR analyses, we plotted the beta coefficient estimates from the conventional observational analyses that included only age and sex as covariates for self-reported PA and age, sex, and season as covariates for device-based PA/SB against the beta coefficient estimates from the corresponding MR-IVW analyses for each of the 1027 proteins. When evaluating our results across the two study designs, we emphasized proteins exhibiting directionally consistent effects and statistically significant at P < 1 × 10⁻³ in both conventional observational and MR analyses. We used a more liberal threshold for statistical significance, considering the lower null likelihood of directionally consistent results in both designs.

Colocalization

We performed colocalization analyses for PA/SB-protein pairs that showed associations in MR analyses to rule out the potential for linkage disequilibrium (LD) confounding that could bias the MR results. We used the “coloc” package in R (46). We applied the “coloc.abf” function using a genomic region of ± 250 kb around the variants to assess the colocalization of PA/SB measures and proteins (47). We used the default threshold for colocalization evidence using the posterior probability (PP) of hypothesis 4 (H_4: PA/SB and protein have one shared common associated/causal variant for both measures). A PP of H₄ > 0.8 was considered as nominal evidence of colocalization.

Conventional Observational Associations of PA, Plasma Proteins, and All-Cause Dementia

To establish the foundation for mediation analyses, we first confirmed the associations between PA/SB and incident all-cause dementia in the UK Biobank, and then evaluated the associations between individual plasma proteins and incident all-cause dementia. We then focused mediation analyses on the proteins that were associated with at least one PA/SB measure and with incident all-cause dementia.

We used Cox proportional hazards regression models to examine the relationships of each of the six PA/SB measures and RNT proteins with the incidence of all-cause dementia. Follow-up time and dementia incidence data were obtained from the UK Biobank. Dementia cases were identified through inpatient hospital records and death registry data, as previously described. The outcome encompassed all dementia subtypes, including Alzheimer disease, vascular dementia, and other forms (9). Follow-up began at the baseline assessment of self-reported PA (2006–2010) or at the device-wear period (2013–2015), depending on the analysis, and continued until the earliest of dementia diagnosis, last health record, or the year 2021 (9). All models were adjusted for the same set of covariates listed above. For both sets of analyses, only participants 55 yr or older, without a prior diagnosis of dementia at baseline, were included. Proteins meeting the Bonferroni-corrected significance threshold (P < 1.7 × 10⁻⁵) were considered associated. Proteins showing associations with both a given PA/SB measure and dementia were carried forward to mediation analyses for the respective PA/SB-protein pair. All Cox models were implemented in R (version 4.3.3).

Conventional Observational Mediation and Interaction Analyses (Four-Way Decomposition)

Plasma proteins associated with at least one PA/SB measure (n = 1027) and with all-cause dementia were analyzed using a four-way decomposition method to estimate the extent to which specific protein(s) mediate the PA/SB-dementia relationship. Two primary models were employed to identify protein mediators of the PA-dementia relationship. For the exposure–mediator relationship, we used linear regression to estimate the association between each PA/SB measure (independent variable) and the selected protein mediator (dependent variable), as described above. For the mediator–outcome relationship of a protein on dementia, we used Cox proportional hazards regression models, as described above. The four-way approach decomposes the total effect between PA/SB and dementia into four main parameters: 1) Controlled direct effect: This parameter estimates the direct effect of a PA/SB measure on the incidence of dementia, independent of the protein mediator (48). The protein is set to a fixed value for the entire study population to mimic the absence of the mediator. In this study, the fixed value of each protein was set to the mean of that protein (2). Reference interaction: This parameter indicates the effect of PA/SB on the incidence of dementia due to only an interactive effect between the PA/SB and the protein on dementia (3). Mediation interaction: This parameter refers to the effect of a PA/SB measure on the incidence of dementia from both mediation by the protein and the interaction between PA/SB and the protein (4). Pure indirect effect: Also referred to as “mediated main effect,” this parameter estimates how much of the relationship between PA/SB and dementia is explained by the influence of a given protein, which in turn influences dementia (48). A Bonferroni correction was applied to assess the statistical significance of the pure indirect effect P values based on the number of proteins tested per PA/SB measure. Only proteins showing directions of effect consistent with a protective (negative) effect of PA on dementia—those increased by higher PA and associated with lower dementia risk, or decreased by higher PA and associated with higher dementia risk—were retained as the final set of mediators. Similarly, the opposite directional effects were considered for SB. We used the med4way command in Stata 17 (StataCorp, College Station, TX) to perform the four-way decomposition analysis (49).

RESULTS

Our comprehensive findings are available through an interactive web platform (https://github.com/klimentidis-lab/ProteomicsofPhysicalActivity2024.git), where it is also possible to visually browse the complete set of results with interactive volcano plots.

Conventional Observational Associations of PA With Proteins

Participants were 54% female and had a mean age of 56.8 yr (Supplemental Table 1, Supplemental Digital Content 1, http://links.lww.com/MSS/D385). PA/SB measures were correlated with each other with r/V values ranging from −0.50 to 0.76 (Supplemental Fig. 1, Supplemental Digital Content 2, https://links.lww.com/MSS/D364). Across all three self-reported (n_max= 39,160) and three device-based (n_max=9,496) PA measures (Supplemental Table 2, Supplemental Digital Content 1, http://links.lww.com/MSS/D385), there were a total of 1027 proteins (Supplemental Table 3, Supplemental Digital Content 1, http://links.lww.com/MSS/D385) that were associated with at least one measure, and 41 proteins (Supplemental Table 4, Supplemental Digital Content 1, http://links.lww.com/MSS/D385) that were associated with all six measures (Supplemental Tables 5–10, Supplemental Digital Content 1, http://links.lww.com/MSS/D385). Some of the strongest and most consistently associated proteins across self-report and device measures included increased levels of ITGAV, ITGAM, ITGA11, MXRA8, CA14, CLEC4A, MEGF10, MYOM3, MYL3, ADGRG2, ADAMTS8, and PON3, and decreased levels of LEP, GDF15, and IL-6, among others (Fig. 2). Overall, most proteins showed broadly consistent directions and magnitudes of association across measures. A small subset of proteins was consistently more strongly associated with high-intensity PA compared to overall-intensity, including CA14, CA4, ANGPT2, ENPP5, KIT, and MEGF10 (Supplemental Fig. 2, Supplemental Digital Content 2, https://links.lww.com/MSS/D364). Similarly, PA was strongly associated with proteins such as ITGAV, CLEC4A, MEGF10, and CA14 compared to device-based SB (Supplemental Fig. 3, Supplemental Digital Content 2, https://links.lww.com/MSS/D364).

FIGURE 2.

Volcano plots showing the associations of physical activity (PA) and sedentary behavior (SB) measures with protein levels. The six panels illustrate different PA/SB measures, including self-reported measures (moderate to vigorous PA [MVPA], vigorous PA [VPA], strenuous sport or other exercise [SSOE]) on the top row, and device-wear measures (average acceleration, fraction acceleration milligravities, and sedentary behavior) on the bottom row. For continuous PA measures, the X-axis shows the β coefficient per 1 rank-normalized standard deviation (RNT SD) increase in PA. In contrast, for binary measures (VPA and SSOE), it represents the standardized mean difference (in RNT SD units) in protein level between active and inactive groups. Each point corresponds to a protein, with positive associations (P < 1.7 × 10⁻⁵, denoted by the gray horizontal dashed line) highlighted in purple (self-reported measures) and blue (device-wear measures). Negative associations are depicted in orange (self-reported measures) and red (device-wear measures).

Pathway and Functional Enrichment Analyses of Conventional Observational Results

The set of genes coding for the 41 proteins we found to be associated with every PA/SB measure was enriched for several biological processes and molecular functions, including integrin-mediated signaling, cell-matrix adhesion, and collagen binding (Table 1; Supplemental Table 11, Supplemental Digital Content 1, http://links.lww.com/MSS/D385). Carbohydrate binding and serine-type endopeptidase activity were enriched in protein-coding genes specific to MVPA and VPA and to the set of 1027 proteins associated with at least one PA/SB measure (Supplemental Tables 12–13, Supplemental Digital Content 1, http://links.lww.com/MSS/D385). Cell adhesion processes emerged as most prominently enriched for VPA, SSOE, and all device-based measures (Supplemental Table 12, Supplemental Digital Content 1, http://links.lww.com/MSS/D385). The pathway enrichment analysis also revealed that cell adhesion mediated by integrins was consistently observed across four measures, including self-reported and device-based measures (Supplemental Table 12, Supplemental Digital Content 1, http://links.lww.com/MSS/D385). Furthermore, proteins identified for AvgAcc and AccSed were found to participate in integrin-mediated signaling pathways (Supplemental Table 12, Supplemental Digital Content 1, http://links.lww.com/MSS/D385). Integrin binding was recognized as a common pathway across the various device-based measures examined (Supplemental Table 11, Supplemental Digital Content 1, http://links.lww.com/MSS/D385).

TABLE 1.

Enrichment of Gene Ontology biological process and molecular function terms (with Benjamini FDR q < 0.05) among the genes coding for the set of 41 proteins that were associated with all PA/SB measures.

Enriched Term	Gene (n)	Benjamini FDR q Value	Gene Names
Integrin binding	12	8.8E-07	ADAMTS8, GFRA1, COMP, EGFR, FAP, ITGA11, ITGA2, ITGAM, ITGAV, ITGB1, ITGB2, ITGB5
Cell adhesion mediated by integrin	7	3.3E-05	ITGA11, ITGA2, ITGAM, ITGAV, ITGB1, ITGB2, ITGB5
Cell-matrix adhesion	9	5.1E-05	CD34, L1CAM, ITGA11, ITGA2, ITGAM, ITGAV, ITGB1, ITGB2, ITGB5
Virus receptor activity	8	2.5E-04	CLEC4M, CD209, DPP4, EGFR, ITGA2, ITGAV, ITGB1, ITGB5
Integrin-mediated signaling pathway	8	5.7E-04	APOA1, ITGA11, ITGA2, ITGAM, ITGAV, ITGB1, ITGB2, ITGB5
Symbiont entry into host cell	8	1.8E-03	CLEC4M, CD209, DPP4, EGFR, ITGA2, ITGAV, ITGB1, ITGB5
Cell-cell adhesion	9	5.1E-03	CD34, EGFR, ITGA11, ITGA2, ITGAM, ITGAV, ITGB1, ITGB2, ITGB5
Cell adhesion	12	1.2E-02	CLEC4A, L1CAM, COMP, DPP4, FAP, ITGA11, ITGA2, ITGAM, ITGAV, ITGB1, ITGB2, MXRA8
Collagen binding involved in cell-matrix adhesion	3	2.5E-02	ITGA11, ITGA2, ITGB1
D-mannose binding	4	2.7E-02	CLEC10A, CLEC4A, CLEC4M, CD209
Negative regulation of vasoconstriction	3	4.7E-02	ADM, ITGB1, LEP

Benjamini FDR q value: This represents the false discovery rate (FDR) adjusted q according to the Benjamini–Hochberg procedure, presenting here results after correcting for the multiple testing. This ensures that the reported enriched terms are statistically within an acceptable rate of false positives. Enriched term: The enriched Gene Ontology (GO) biological processes and molecular function terms that were identified as overrepresented among the genes coding for the proteins associated with PA/SB measures.

Mendelian Randomization

Of 1027 proteins tested based on showing at least one PA/SB association (P < 1.7 × 10^–5) in the above conventional observational analyses, we found 191 MR-IVW associations (P < 4.9 × 10⁻⁵) between PA/SB exposure and protein level (with exposure SNPs at P < 5 × 10⁻⁸) and 282 proteins (IL-6, LEP, ITGAV, MXRA8, and TNF were associated with four or more measures) with the PA/SB exposure SNPs at the P < 5 × 10⁻⁶ threshold (Fig. 3; Supplemental Fig. 5, Supplemental Digital Content 2, https://links.lww.com/MSS/D364 Supplemental Tables 14–19, Supplemental Digital Content 1, http://links.lww.com/MSS/D385). Using only cis-pQTL SNPs as protein exposure instruments for analyses in the reverse direction revealed two (SLC9A3R2 and CCL11) and three (SLC9A3R2, AP3B1, and CCL11) proteins at the P < 5 × 10⁻⁸ and P < 5 × 10⁻⁶ thresholds, respectively. (Supplemental Fig. 5, Supplemental Digital Content 2, https://links.lww.com/MSS/D364 and Supplemental Tables 20–26, Supplemental Digital Content 1, http://links.lww.com/MSS/D385) MR sensitivity analyses generally revealed directionally consistent estimates for most IVW top hits (Supplemental Tables 27–28, Supplemental Digital Content 1, http://links.lww.com/MSS/D386). Based on our colocalization analyses, 110 distinct proteins and PA-protein SNP pairs were observed to have a PP of colocalization greater than 0.80, supporting a shared causal variant for both traits. At the level of PA-protein pairs, up to 16% of SNPs had PPH4 > 0.8. These results suggest that the association signal is likely driven by the same underlying variant rather than distinct variants in linkage disequilibrium, which helps to rule out LD-driven confounding of the MR analyses. (see Supplemental Tables 29–30, Supplemental Digital Content 1, http://links.lww.com/MSS/D386).

FIGURE 3.

A, Volcano plots depicting Mendelian randomization results of genetically-predicted physical activity (PA) and sedentary behavior (SB) measures on protein levels, using IVW or Wald ratio methods. The top row shows results using exposure SNPs with P values below the stringent genome-wide significance threshold (P < 5 × 10⁻⁸). The figures in the panels highlight each genetically-predicted PA and SB measure related to protein levels, with a statistical significance threshold of P < 4.9 × 10⁻⁵, indicated by the horizontal gray dashed line. Effect estimates (β) are expressed per one standard deviation (SD) unit increase in the PA and SB exposure, corresponding to the associated change in protein level measured in inverse rank-normalized standard deviation (RNT SD) units. For continuous exposures that were inverse rank-normalized in the exposure GWAS (MVPA and Fraction Acceleration >425 mg), β represents the change in protein level (RNT SD units) per one RNT SD increase in the exposure. For continuous exposures analyzed on their native scale (average acceleration and acceleration sedentary behavior), β represents the change in protein level (RNT SD units) per one SD increase in the exposure. For binary exposures (leisure-time MVPA, VPA, and SSOE), estimates represent the standardized mean difference in protein level (RNT SD units) between active and inactive groups. Proteins with positive associations are shown in purple for self-reported measures and blue for device-wear measures, while proteins with negative associations are shown in orange and red. B, Heatmap of Mendelian randomization results for PA/SB exposures on protein levels of the top 41 proteins identified in the conventional observational analysis. Each tile represents the direction and strength of evidence for a causal effect of a PA/SB trait (Y-axis) on a protein level (X-axis), using exposure SNPs with P < 5 × 10⁻⁸. Colors reflect the level of evidence based on P value thresholds and consistency of effect direction across four other MR sensitivity analyses. Strong evidence corresponds to an IVW P < 4.9 × 10⁻⁵ and a consistent direction of the beta coefficients across at least two of the MR sensitivity analyses. Moderate evidence corresponds to an IVW P value < 0.005 and a consistent direction of the beta coefficients across at least two of the MR sensitivity analyses. Suggestive evidence corresponds to an IVW P value < 0.05 and a consistent direction of the beta coefficient across all of the MR sensitivity analyses. White tiles indicate no evidence. Proteins are ordered from left to right based on the total evidence across all PA exposures.

Across MR analyses, we identified several proteins for which MR supported a causal relationship of PA and SB exposures with protein levels (Fig. 3; Supplemental Tables 31–38, Supplemental Digital Content 1, http://links.lww.com/MSS/D386). Notably, ITGAV, INHBC, LEP, and MXRA8 continued to emerge as strongly or moderately supported across MR analyses (Fig. 3, Supplemental Fig. 5, Supplemental Digital Content 2, https://links.lww.com/MSS/D364). Additional results using the genome-wide threshold of P < 5 × 10⁻⁶ provide further support for the above proteins as well as ADM, CLEC4M, and PON3 (Supplemental Fig. 6, Supplemental Digital Content 2, https://links.lww.com/MSS/D364). In the overall comparison of consistency of conventional observational results with MR, we identified ITGAV, ITGAM, MXRA8, LEP, and CXCL10 as exhibiting consistent associations across study designs and PA measures. (Fig. 4, Supplemental Figs. 7–8, Supplemental Digital Content 2, https://links.lww.com/MSS/D364).

FIGURE 4.

Comparison of observational and Mendelian randomization effect estimates for the associations between physical activity (PA) exposures and protein levels. The figures demonstrate the relationship between beta coefficients derived from conventional observational associations (Y-axis) and MR analyses (X-axis) for self-reported PA measures (moderate to vigorous PA, vigorous PA, strenuous sports or other exercises) and device-wear measures (acceleration average, fraction acceleration > 425 mg). Each colored data point corresponds to a protein, with associations in both the observational and MR analyses (P < 1 × 10⁻³) highlighted. These plots offer a visual comparison to evaluate the consistency in effect estimates across both observational and MR analyses. (Created in part with BioRender.com).

Conventional Observational Associations of PA and Plasma Proteins With All-Cause Dementia

Across all three self-reported and three device-based measures. Higher levels of PA were consistently associated with a lower risk of incident all-cause dementia (Supplemental Table 39, Supplemental Digital Content 1, http://links.lww.com/MSS/D386). MVPA, VPA, and SSOE showed protective associations, with HR of 0.88 (95% confidence interval [CI]: 0.80–0.97), 0.78 (95% CI: 0.64–0.93), and 0.62 (95% CI: 0.52–0.74), respectively. Among device-based measures, both AvgAcc (HR = 0.94, 95% CI: 0.90–0.98) and FracAcc > 425 mg (HR = 0.26, 95% CI: 0.13–0.51) were inversely associated with dementia risk, whereas AccSed was not significantly associated (HR = 1.14, 95% CI: 0.99–1.32).

In conventional observational analyses of 2911 plasma proteins on incident all-cause dementia, we identified 57 proteins that were associated after multiple testing correction (Supplemental Table 40, Supplemental Digital Content 1, http://links.lww.com/MSS/D386). Notably, higher levels of GFAP, NEFL, GDF15, SYT1, and VGF were observed among individuals who developed dementia, whereas APOE, APOA1, FURIN, HPGDS, ITGAV, and ITGAM showed inverse associations, as previously identified (50).

Conventional Observational Mediation and Interaction Analyses

There were up to 859 incident dementia cases identified among participants with both proteomic and PA data (Supplemental Table 41, Supplemental Digital Content 1, http://links.lww.com/MSS/D386). Our investigation employing the four-way decomposition method identified 21 unique proteins that mediated the relationship between PA/SB and incident dementia (P < 4.9 × 10⁻⁵). Sixteen of these proteins were found to mediate the relationship between MVPA and dementia; 10 proteins mediated the relationship between VPA and dementia, and 10 proteins mediated the relationship between SSOE and dementia (Table 2; Supplemental Tables 42–45, Supplemental Digital Content 1, http://links.lww.com/MSS/D386). We did not identify any protein as a mediator of the effect of device-based measures on dementia (Supplemental Tables 46–48, Supplemental Digital Content 1, http://links.lww.com/MSS/D386). GDF15, HPGDS, ITGAV, and ITGA11 mediated the relationship between all three PA measures and dementia. ITGAM mediated the effects of MVPA and SSOE, and GFAP mediated those of MVPA and VPA. Additional mediators identified for at least two PA measures included AHNAK, ADM, and DPP4 for both MVPA and SSOE, and PI3 and GFAP for MVPA and VPA.

TABLE 2.

Proteins that mediated the relationship between PA and all-cause dementia.

PA measure	Protein	Total Excess Risk Ratio (TERERI)	TEREEI_P	Controlled Direct Effect (CDE)	CDE_P	Pure Indirect Effect (PIE)	PIE_P
MVPA
	GDF15;Growth/differentiation factor 15	−0.04 (−0.11, 0.02)	1.66E-01	−0.03 (−0.10, 0.03)	3.44E-01	−0.01 (−0.01, −0.01)	1.09E-08
	HPGDS;Hematopoietic prostaglandin D synthase	−0.04 (−0.09, 0.01)	1.33E-01	−0.02 (−0.07, 0.03)	4.44E-01	−0.01 (−0.02, −0.01)	1.85E-08
	ITGAV;Integrin alpha-V	−0.05 (−0.10, −0.00)	3.62E-02	−0.03 (−0.08, 0.02)	2.28E-01	−0.02 (−0.02, −0.01)	1.39E-07
	ITGAM;Integrin alpha-M	−0.05 (−0.10, 0.00)	6.12E-02	−0.03 (−0.08, 0.02)	2.22E-01	−0.01 (−0.02, −0.01)	4.63E-06
	PEPD;Xaa-Pro dipeptidase	−0.03 (−0.09, 0.02)	2.26E-01	−0.02 (−0.07, 0.03)	4.63E-01	−0.01 (−0.02, −0.01)	4.60E-05
	TFF3;Trefoil factor 3	−0.05 (−0.10, −0.00)	4.14E-02	−0.05 (−0.10, 0.01)	7.95E-02	−0.01 (−0.01, −0.00)	1.60E-05
	DPP4;Dipeptidyl peptidase 4	−0.05 (−0.10, 0.00)	6.06E-02	−0.04 (−0.09, 0.01)	1.33E-01	−0.01 (−0.01, −0.00)	3.53E-04
	BAG3;BAG family molecular chaperone regulator 3	−0.05 (−0.10, 0.00)	5.68E-02	−0.04 (−0.09, 0.01)	1.49E-01	−0.01 (−0.01, −0.00)	8.70E-05
	ITGA11;Integrin alpha-11	−0.04 (−0.09, 0.01)	1.56E-01	−0.02 (−0.07, 0.03)	3.64E-01	−0.01 (−0.02, −0.01)	5.21E-05
	PI3;Elafin	−0.04 (−0.09, 0.01)	9.90E-02	−0.03 (−0.08, 0.02)	2.13E-01	−0.01 (−0.01, −0.00)	3.15E-05
	AHNAK;Neuroblast differentiation-associated protein AHNAK	−0.03 (−0.09, 0.02)	2.64E-01	−0.02 (−0.08, 0.03)	3.92E-01	−0.01 (−0.01, −0.00)	1.35E-04
	GFAP;Glial fibrillary acidic protein	−0.06 (−0.11, −0.01)	2.09E-02	−0.06 (−0.12, −0.01)	3.07E-02	−0.01 (−0.02, −0.01)	2.50E-06
	EGFR;Epidermal growth factor receptor	−0.05 (−0.10, 0.00)	5.30E-02	−0.04 (−0.10, 0.01)	9.11E-02	−0.01 (−0.01, −0.00)	8.92E-04
	TGFA;Protransforming growth factor alpha	−0.05 (−0.10, 0.00)	6.34E-02	−0.04 (−0.09, 0.01)	9.43E-02	−0.01 (−0.01, −0.00)	1.86E-04
	ADM;Pro-adrenomedullin	−0.06 (−0.11, −0.01)	1.66E-02	−0.06 (−0.11, −0.01)	2.91E-02	−0.01 (−0.01, −0.00)	2.68E-04
	AREG;Amphiregulin	−0.05 (−0.10, −0.00)	3.78E-02	−0.05 (−0.10, 0.00)	6.30E-02	−0.01 (−0.01, −0.00)	8.47E-05
VPA
	ITGAV;Integrin alpha-V	−0.27 (−0.41, −0.13)	2.04E-04	−0.21 (−0.36, −0.07)	4.37E-03	−0.06 (−0.09, −0.04)	3.37E-07
	HPGDS;Hematopoietic prostaglandin D synthase	−0.23 (−0.38, −0.08)	2.60E-03	−0.18 (−0.33, −0.03)	2.24E-02	−0.04 (−0.05, −0.02)	9.17E-07
	GDF15;Growth/differentiation factor 15	−0.16 (−0.34, 0.02)	8.54E-02	−0.10 (−0.30, 0.10)	3.32E-01	−0.05 (−0.06, −0.03)	1.85E-07
	NEFL;Neurofilament light polypeptide	−0.33 (−0.47, −0.18)	1.48E-05	−0.39 (−0.56, −0.22)	7.28E-06	−0.02 (−0.04, −0.01)	1.05E-04
	GFAP;Glial fibrillary acidic protein	−0.28 (−0.43, −0.13)	1.64E-04	−0.29 (−0.46, −0.12)	7.06E-04	−0.04 (−0.05, −0.02)	1.79E-06
	CLEC3B;Tetranectin	−0.12 (−0.30, 0.06)	1.90E-01	−0.09 (−0.27, 0.09)	3.45E-01	−0.02 (−0.03, −0.01)	4.30E-04
	ADAMTS8;A disintegrin and metalloproteinase with thrombospondin motifs 8	−0.22 (−0.37, −0.08)	2.84E-03	−0.19 (−0.34, −0.04)	1.36E-02	−0.04 (−0.05, −0.02)	9.24E-05
	PI3;Elafin	−0.23 (−0.38, −0.08)	3.06E-03	−0.19 (−0.35, −0.04)	1.43E-02	−0.03 (−0.05, −0.01)	1.36E-04
	ITGA11;Integrin alpha-11	−0.22 (−0.37, −0.07)	4.02E-03	−0.19 (−0.34, −0.03)	1.69E-02	−0.04 (−0.06, −0.02)	2.49E-04
	TFF3;Trefoil factor 3	−0.23 (−0.38, −0.08)	2.51E-03	−0.22 (−0.38, −0.05)	8.54E-03	−0.02 (−0.04, −0.01)	1.94E-04
SSOE
	GDF15;Growth/differentiation factor 15	−0.38 (−0.50, −0.26)	1.07E-09	−0.33 (−0.47, −0.19)	3.21E-06	−0.05 (−0.07, −0.04)	2.82E-10
	HPGDS;Hematopoietic prostaglandin D synthase	−0.36 (−0.48, −0.25)	5.48E-10	−0.32 (−0.44, −0.21)	5.66E-08	−0.02 (−0.03, −0.02)	2.07E-07
	ITGAV;Integrin alpha-V	−0.41 (−0.52, −0.30)	5.09E-14	−0.39 (−0.50, −0.28)	7.24E-12	−0.04 (−0.05, −0.02)	7.94E-07
	ITGAM;Integrin alpha-M	−0.38 (−0.49, −0.27)	1.60E-11	−0.35 (−0.46, −0.24)	5.98E-10	−0.03 (−0.04, −0.01)	3.93E-06
	BCAN;Brevican core protein	−0.36 (−0.48, −0.25)	4.04E-10	−0.32 (−0.44, −0.20)	8.28E-08	−0.02 (−0.02, −0.01)	1.60E-05
	DPP4;Dipeptidyl peptidase 4	−0.38 (−0.49, −0.27)	1.62E-11	−0.36 (−0.47, −0.25)	3.85E-10	−0.01 (−0.02, −0.01)	2.43E-04
	EGFR;Epidermal growth factor receptor	−0.35 (−0.47, −0.23)	4.30E-09	−0.33 (−0.45, −0.21)	4.28E-08	−0.02 (−0.03, −0.01)	1.63E-04
	AHNAK;Neuroblast differentiation-associated protein AHNAK	−0.33 (−0.46, −0.20)	4.56E-07	−0.32 (−0.44, −0.19)	1.39E-06	−0.01 (−0.02, −0.01)	7.86E-04
	ITGA11;Integrin alpha-11	−0.37 (−0.48, −0.26)	7.05E-11	−0.35 (−0.46, −0.24)	1.17E-09	−0.02 (−0.03, −0.01)	1.28E-04
	ADM;Pro-adrenomedullin	−0.37 (−0.49, −0.26)	1.77E-10	−0.35 (−0.47, −0.24)	4.89E-09	−0.02 (−0.03, −0.01)	3.15E-04

Total excess relative risk (TERERI) represents the overall excess risk (or protection) of dementia associated with PA through protein mediation/interaction pathways and direct pathways, adjusting for other potential confounders. Controlled direct effect (CDE) represents the direct effect of PA on dementia risk when the mediator protein is held constant at a fixed mean level. PIE (Pure indirect effect) represents the portion of the association between PA and all-cause dementia that is through a protein mediator, independent of any interaction between PA and protein. Presented here are the protein mediators that are associated after correction for multiple testing (PIE P value threshold was 2.17E-03, 2.50E-03, 2.17E-03 for MVPA, VPA, and SSOE, respectively). Effect estimates are presented with 95% confidence intervals in parentheses.

DISCUSSION

In this study, we aimed to explore how PA and SB influence the plasma proteomic landscape and to identify proteins that mediate PA’s protective effect on dementia. We conducted thorough analyses across 2911 proteins to identify consistent findings across different PA and SB measures and study designs. Conventional observational analyses revealed 1027 proteins associated with at least one of the six PA/SB measures, and 41 proteins associated with all six PA measures. We leveraged bidirectional MR analyses to validate and enhance the robustness of our findings regarding the putative causal effects of PA on protein levels. Pathway enrichment analyses revealed several biological pathways that may be upregulated in response to PA, and mediation analyses revealed specific proteins that may mediate the relationship between PA and dementia.

Our study extends previous findings by confirming several proteins associated with habitual PA and identified in post-intervention studies, such as FAP, LEP, CD209, CD248, MXRA8, WFIKKN2, ADM, APOA1, and IL-6 (21,24,28). We also identified many novel proteins and evaluated the evidence for causality and direction of effect between PA/SB and each protein using a genetically-informed study design. The identified proteins fall into several broad functional categories, including cardiovascular and vascular regulation (e.g., ADM and ITGB5), lipid metabolism (e.g., APOA1 and LPL), inflammatory and metabolic regulation (e.g., DPP4, GDF15, and HPGDS), and tissue remodeling and fibrosis (e.g., FAP), highlighting the diverse physiological mechanisms through which PA/SB influences critical cellular functions (51). The top proteins emerging from the leisure time screen time MR analysis were implicated in pathways such as adipokine signaling (e.g., LEP), low-grade inflammation (e.g., IL-6, TNF, and CCL3), and vascular remodeling (e.g., HGF and ANGPT2). Many of these proteins were also observed in the dementia-associated protein signatures (e.g., IL-6 and GFAP), suggesting a shared immuno-metabolic association between SB and neurodegenerative pathology.

Across conventional observational, MR, mediation, and pathway enrichment analyses, we consistently identified the integrin family of proteins (e.g., ITGAV, ITGAM, and ITGA11) as increasing in response to PA, decreasing in response to SB, and mediating the effect of PA/SB on all-cause dementia. Integrins are transmembrane receptors involved in cell adhesion and signaling and have been shown to influence the extracellular matrix (ECM), critical for maintaining tissue integrity and facilitating cellular interactions (52). Integrins such as ITGAV, ITGAM, and ITGA11 were consistently associated across multiple PA/SB measures, underscoring the role of PA/SB in ECM remodeling and cell-ECM adhesion. Another protein that we consistently identified was MXRA8, which is also involved in ECM processes (53). MXRA8 may interact with integrins such as ITGAV, playing a role in the maturation and maintenance of the blood-brain barrier (53). Integrins could also explain the potential roles of PA in reducing the risk of all-cause dementia and Alzheimer disease via facilitating cell-ECM adhesion functions that impact neuronal signaling pathways, synaptic plasticity, neuroinflammation, and blood-brain-barrier integrity (54). Additionally, integrin-mediated pathways, such as those involving ITGA2β1 and ITGAVβ1, have been found to reduce amyloid-beta toxicity, further linking the positive effects of PA with brain health and reduced neurodegenerative risk (55). Finally, integrins are strongly linked to irisin, a previously identified exerkine (56,57). Irisin facilitates neurogenesis and amyloid-beta degradation through integrin receptors, such as ITGAVβ5, on astrocytes and may also be involved with BDNF activation via integrin receptor complexes (56,57).

We also found that proteins such as IL-6, HPGDS, and GDF15 were consistently associated with PA and SB, implicating the modifications of inflammatory, metabolic, and vascular processes (58,59). Previous studies have indicated that IL-6 and GDF15 levels may transiently rise after acute PA, but tend to decrease with sustained long-term PA, suggesting that our measurements may capture these lasting benefits effectively (60–64). HPGDS was previously found to have a role in dementia via inflammatory pathways (59). Notably, HPGDS and GDF15 also emerged as mediators between PA and the risk of all-cause dementia, suggesting that PA protects against dementia by reducing inflammation. The findings reinforce the broader evidence supporting the involvement of these proteins in neurodegenerative pathways (65). The increase of HPGDS via PA could be related to prostaglandin-mediated inflammatory processes implicated in neurodegeneration (59). Elevated GDF15 has been previously associated with increased risks of all-cause dementia, vascular dementia, and Alzheimer disease, particularly in the context of cerebrovascular damage (50,58,65). The reduction in GDF15 that we observed with long-term PA suggests an anti-inflammatory compensatory response in the brain, which may minimize damage and promote recovery (50,65). These results underscore the need to explore further associations between PA and these inflammatory markers and their role in vascular health, reducing neuroinflammation, and promoting neuronal resilience.

Other proteins that were consistently associated with PA included LPL, DPP4, FAP, LEP, and IGFBP1, which may play roles in cardiometabolic diseases and cancer. LPL, a protein that breaks down triglycerides into free fatty acids for energy, is increased with PA and decreased with SB, affecting the risk of hyperlipidemia and cardiovascular diseases (66–70). Another protein, FAP, which is primarily involved in tissue remodeling and fibrosis, was found to increase with PA. In previous research, elevated FAP expression after exercise correlated with improved cardiovascular metrics, such as VO₂ max, suggesting FAP’s possible role in musculoskeletal tissue repair and cardiovascular health (28,71–74). Moreover, we also identified LEP, a well-known metabolic hormone that regulates energy balance and satiety (75). It consistently decreased with PA after controlling for percent body fat. This supports the previously reported role of PA in regulating appetite, possibly via LEP (76,77). IGFBP1, which increases with PA, regulates insulin-like growth factors essential for glucose metabolism and cell growth (78). Elevated IGFBP1 levels aid in better insulin regulation and metabolic stability, which may be crucial for preventing type 2 diabetes and supporting tissue repair (78,79). Finally, we found that EGFR, another protein known for its role in cell growth and differentiation, increases with PA, which may help reduce cancer risk by preventing aberrant cellular proliferation (80). Although PA has been studied as a protective factor for cancer, the relationship between PA and EGFR is not well understood, warranting further study (81).

Our study is notable because it is the largest and most comprehensive study of PA/SB and plasma protein levels. Our approach included various self-reported and device-based PA measures, which allowed us to identify proteins with the most consistent associations across these measurement types, and proteins that exhibited associations with only a subset of measures. Employing a genetically-informed study design enabled us to examine the existence and direction of causal links between proteins and PA. Thus, specific proteins with consistent magnitudes and directions of association across study designs and PA measures could be identified. To assess the mediation of the PA-dementia relationship through specific proteins, we employed the four-way decomposition method to disentangle the distinct contributions of PA’s direct, indirect, and interactive effects on dementia through a given protein. This approach facilitated the identification of specific pathways and mechanisms that could underlie the protective effect of PA.

Analyses of device-based measures in our conventional observational design may be vulnerable to temporal ambiguity and potential reverse causation since there is a 4–8-yr time gap between baseline blood sample collection and the device-based measure. Bidirectional MR analyses may partly address this limitation but they rely on assumptions that are difficult to fully test. In colocalization analyses, most signals showed limited evidence for shared causal variants, indicating that we cannot fully rule out the possibility of LD confounding. MR analyses using a relaxed instrument threshold of P < 5 × 10⁻⁶ were considered exploratory and provided an opportunity to further prioritize protein signals for future investigation. However, these findings should be interpreted with caution. Furthermore, overlap between exposure and outcome GWAS samples may bias MR estimates toward observational results, particularly for weaker instruments. Finally, the gene-environment equivalence assumption posits that the downstream effects of genetic variants on a given exposure are like those of the environmental exposure itself, which may not always hold. Mediation analyses might be prone to residual confounding and temporal ambiguity for the device-based measures, as mentioned above. Moreover, in mediation and interaction analyses, the total effect is only represented by the exposure and mediator, making strong assumptions about unaccounted confounding and the independence of protein mediation. The study population primarily consists of middle-aged individuals of European descent, which limits the generalizability of the findings to other ancestries and age groups.

We identified several proteins influenced by PA/SB that span functions such as cell adhesion, inflammation, and cardiometabolic and neuroprotective roles. By integrating proteomic data with detailed epidemiological information and various study designs, our study provides novel insights into how PA/SB affects health at the molecular level, potentially guiding precision public health interventions. These proteins reveal biological mechanisms that warrant closer investigation, especially in their role as molecular transducers of PA and SB. The findings could inform strategies for tracking and enhancing PA behaviors and highlight biological pathways linked to dementia, which may be prioritized for therapeutic intervention.

Supplemental tables can be found in our GitHub repository: https://github.com/klimentidis-lab/ProteomicsofPhysicalActivity2024.git and Zenodo repository: https://doi.org/10.5281/zenodo.16923652. Individual-level data from the UK Biobank can be requested at https://www.ukbiobank.ac.uk/enable-your-research/apply-for-access. Summary statistics for all GWAS used in this study are publicly available from the GWAS catalog (64) and UKB-PPP at https://doi.org/10.7303/syn51364943. Our Supplemental tables, analysis codes, and results, including interactive visualizations of results, are available in a GitHub repository: https://github.com/klimentidis-lab/ProteomicsofPhysicalActivity2024.git. This research has been conducted using the UK Biobank Resource under Application Number 21259. It uses data provided by patients and collected by the NHS as part of their care and support. The authors thank the UK Biobank’s organizers and participants. The authors acknowledge that this study’s results are presented clearly, honestly, and without fabrication, falsification, or inappropriate data manipulation. The results of the present study do not constitute endorsement by the American College of Sports Medicine. The authors would also like to acknowledge funding from NIH R01AG072445. No conflicts of interest were disclosed.

Supplementary Material

msse-58-1193-s001.pdf ^{(1.8MB, pdf)}

msse-58-1193-s002.xlsx ^{(8.9MB, xlsx)}

msse-58-1193-s003.xlsx ^{(6.9MB, xlsx)}