Differential DNA Methylation in the Brain as Potential Mediator of the Association between Traffic-related PM2.5 and Neuropathology Markers of Alzheimer’s Disease

INTRODUCTION: Growing evidence indicates fine particulate matter (PM2.5) as risk factor for Alzheimer’s’ disease (AD), but the underlying mechanisms have been insufficiently investigated. We hypothesized differential DNA methylation (DNAm) in brain tissue as potential mediator of this association. METHODS: We assessed genome-wide DNAm (Illumina EPIC BeadChips) in prefrontal cortex tissue and three AD-related neuropathological markers (Braak stage, CERAD, ABC score) for 159 donors, and estimated donors’ residential traffic-related PM2.5 exposure 1, 3 and 5 years prior to death. We used a combination of the Meet-in-the-Middle approach, high-dimensional mediation analysis, and causal mediation analysis to identify potential mediating CpGs. RESULTS: PM2.5 was significantly associated with differential DNAm at cg25433380 and cg10495669. Twenty-six CpG sites were identified as mediators of the association between PM2.5 exposure and neuropathology markers, several located in genes related to neuroinflammation. DISCUSSION: Our findings suggest differential DNAm related to neuroinflammation mediates the association between traffic-related PM2.5 and AD.

DNA methylation (DNAm), which regulates the expression of genes, are associated with indicators of AD as well as PM 2.5 exposure. However, the tissue specificity of DNAm has limited the ability of previous studies to formally investigate mediation. While there is no conclusive evidence of an association between AD and DNAm in blood, 13 DNAm alterations in a number of genes were observed to be associated with AD pathology and neuroinflammation in brain tissues, such as amyloid precursor protein (APP), 14 microtubule-associated protein tau (MAPT), 14 apolipoprotein (APOE) promoter region, 15 homeobox A3 (HOXA3), 16 interleukin-1 beta (IL-1β), 17 and interleukin-6 (IL-6). 17 The association of PM 2.5 with DNAm in blood has been extensively studied. 18 DNAm in interleukin-10 (IL-10), IL-6, tumor necrosis factor (TNF), toll like receptor 2 (TLR2) genes, which play key roles in neuroinflammation, 19 was reported to be significantly altered in response to short-term exposure to PM 2.5 and its species. 20 However, to the best of our knowledge, no human studies have been published on the association between PM 2.5 exposure and DNAm in human brain, which is the most relevant tissue when studying AD. The only evidence to date comes from in-vivo and in-vitro studies.
Tachibana et al. demonstrated with a mouse model that prenatal exposure to diesel exhaust altered DNAm in brain tissues collected from 1-and 21-day-old offspring, and the differentially methylated CpG sites were enriched in the gene ontology (GO) terms related to neuronal development. 21 Wei et al. exposed human neuroblastoma cells to near-road PM 2.5 and found that DNAm was hypermethylated in the promoter regions of genes encoding synaptic neuronal adhesion molecules that mediate essential signaling at synapse. 22 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) 6 The mediating role of DNAm for the association between PM 2.5 and AD pathology has not been well studied, given the limited evidence of an association between PM 2.5 exposure and DNAm in the brain. The current study investigated the relationship among PM 2.5 , DNAm and AD neuropathology in the post-mortem human brain among brain donors of the Emory Goizueta AD Research Center (ADRC) brain bank. We recently showed a significant association between traffic-related PM 2.5 exposure and increased AD neuropathology in this dataset. 23 To elucidate the biological mechanisms for this association, we here investigated whether differential DNAm in the prefrontal cortex tissues mediates the association between long-term exposure to traffic-related PM 2.5 and the levels of AD-related neuropathological markers. This hypothesis was tested using a combination of the Meet-in-the-Middle (MITM) approach and high-dimensional mediation analysis.
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

Study design
The current cross-sectional analysis included study participants recruited by the Emory Goizueta ADRC. The ADRC was founded in 2005 and has maintained a brain bank to facilitate AD research. The study participants were research participants evaluated annually, and others were patients treated by Emory Department of Neurology physicians and diagnosed clinically with AD (biomarker defined) or probable AD. The prefrontal cortex tissues were obtained from the participants who had consented to donate biospecimens to the ADRC brain bank. There were 1011 donors enrolled by the third quarter of 2020. After applying the following inclusion criteria, 264 donors remained eligible for the current study: 1) the availability of residential addresses within Georgia (GA) state; 2) age at death equal to or over 55 years (death earlier than 55 possibly due to competing risks); 3) deceased after 1999 (due to the availability of air quality data) ; 4) no missing values in neuropathology outcomes and key covariates including age at death, race, sex, educational attainment, and APOE genotype. Among these donors, genome-wide DNAm was measured in 161 available samples from the donors deceased after 2007, and after quality control, 159 were included in the current analysis. Written informed consent was provided for all donors, and samples were obtained following research protocols approved by the Emory University Institutional Review Board.

Neuropathology assessment
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted June 30, 2023.  The ADRC performed thorough neuropathologic evaluations on the brains of all donors   using established comprehensive research evaluations and diagnostic criteria. 24 These   neuropathological assessments include a variety of stains and immunohistochemical preparations, as well as semi-quantitative scoring of multiple neuropathologic changes by experienced neuropathologists using published criteria. 25 In this project, AD-related neuropathological changes were evaluated using Braak stage, Consortium to Establish a Registry for AD (CERAD) score, and a combination of Amyloid, Braak stage, and CERAD (ABC) score which were developed based on the Aβ plaques and NFTs. 26 Braak stage is a staging scheme describing NFTs with six stages (Stage I-VI) with a higher stage indicating a wider distribution of NFTs in brain. CERAD score describes the prevalence of Aβ plaques with four levels from no neuritic plaques to frequent. ABC score combines the former two (along with the Thal score for Aβ plaque distribution across various brain regions) 27 and is transformed into one of four levels: not, low, intermediate, or high level of AD neuropathologic changes. The random forest model was trained with the R package randomForest 32 , and two user-defined parameters (i.e., the number of trees and the number of variables randomly tried at each split) were determined by a balance of the efficiency and the outof-bag R 2 value. The final model reached an out-of-bag R 2 of 0.8 and a root-meansquare deviation of 0.2 µg/m 3 . This model was used to predict annual traffic-related PM 2.5 for 2012-2019 with a spatial resolution of 200m. More details can be found elsewhere. 23 Finally, we spatially matched geocoded residential addresses to the centroid of closest grids and calculated the individual long-term exposures as the average of specific exposure windows (1 year, 3 years, and 5 years prior to death).

Genome-wide DNA methylation
DNA was isolated from fresh frozen prefrontal cortex in 161 samples using the QIAGEN GenePure kit. DNAm was assessed with the Illumina Infinium MethylationEPIC BeadChips in batches of 167 prefrontal cortex samples including 6 replicates. The raw intensity files were transformed into a dataset that included beta values for each the . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) 1 0 CpG sites, and these beta values were computed as the ratio of the methylated signal to the sum of the methylated and unmethylated signals, which ranged from 0 to 1 on a continuous scale. Pre-processing and statistics were done using R (v4.2.0). We followed a validated quality control and normalization pipeline as previously published. 33 The detailed data processing and sample quality control can be found in the Supplementary Methods. One hundred and fifty-nine samples passed the quality check, and after excluding SNP probes, XY probes and other low-quality probes, 789,286 CpG sites remained. The final DNAm beta values were further normalized to reduce the probe type differences and corrected by ComBat to remove the batch effect before the downstream analysis. 34 We estimated the cell-type proportions (neuronal vs. nonneuronal cells) for each sample using the most recent prefrontal cortex database and the R package minfi. 35,36

Covariate assessment
The confounding structure was determined according to literature review and our previous studies, which was illustrated by directed acyclic graphs (DAGs) in the Supplement ( Figure S1). Individual-level demographic characteristics [sex, race (Black vs. White), educational attainment (high school or less, college degree, and graduate degree), age at death, APOE ε 4 genotype] were obtained from the medical records.  CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) 1 1 2015. 37 Post-mortem interval (hours) of sample collection was provided by our lab collaborators.

Statistical analysis
Previously, we found higher residential PM 2.5 exposure was associated with increased AD neuropathology in the Emory Goizueta ADRC brain bank. 23 To identify DNAm patterns in brain tissue that potentially mediate the association between PM 2.5 exposure and increased neuropathology markers, we 1) conducted an epigenome-wide association study (EWAS) for the long-term PM 2.5 exposures 1 year, 3 years, and 5 years prior to death and then investigated whether any differentially methylated CpG sites that were significantly associated with PM 2.5 exposure in the EWAS were also associated with increased neuropathology markers; and 2) conducted a combination of Meet-in-the-Middle (MITM) approach and high-dimensional mediation analysis (HDMA) to identify any mediating CpGs that did not reach genome-wide significance in the EWAS of PM 2.5 . The MITM approach and HDMA work complementarily to maximize the detention of potential mediators.
Firstly, we conducted an EWAS to assess associations of long-term PM 2.5 exposures 1 year, 3 years, and 5 years prior to death and methylation levels of CpG sites.
Specifically, we used robust multiple linear regression models as implemented in the R package MASS to identify differentially CpG sites associated with PM 2.5 exposures. 38 To account for measured confounding factors, we included sex, race, educational attainment, age at death, PMI, ADI, and proportion of neuronal cells in the model.
Potential batch effect and other unwanted variation were further corrected using the R . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted June 30, 2023. ; https://doi.org/10.1101/2023.06.30.23292085 doi: medRxiv preprint packages sva 39 (estimating surrogate variables included in the EWAS model as covariates) and Bacon. 40 The sva was used to obtain surrogate variables to be included in the models. To account for multiple testing, the Bonferroni threshold was used for statistical significance (0.05 / 789,286 = 6.33×10 -8 ), while no cut-off was applied on the magnitude of DNA methylation difference. 41 Any CpG sites that were significantly associated with PM 2.5 exposure were then investigated for their associations with neuropathology markers. These associations were extracted from an EWAS of each neuropathology marker (CERAD, Braak stage, ABC score) with methylation levels of all CpG sites, using robust multiple linear regression models with the neuropathology markers converted to continuous outcomes and DNAm beta values of CpG sites as exposures, adjusting for sex, race, educational attainment, age at death, PMI, APOE genotype, and proportion of neuronal cells. We used Bacon 40 to control for unmeasured confounding and bias due to the minor inflation/deflation indicated by raw p-values.
For the MITM, we compared the 1,000 most significant CpGs from the two sets of EWAS on all CpG sites for PM 2.5 exposures and neuropathology markers to identify the differentially methylated CpG sties that were associated with both exposures and outcomes. In other words, the raw p-values of all 789,286 CpG sites were sorted increasingly, which were derived from the two set of EWAS models conducted on PM 2.5 exposure and neuropathology markers, respectively. We selected the CpG sites among the lowest 1000 for both PM 2.5 exposure and neuropathology markers. The MITM approach is widely used in high-dimensional setting to identify intermediate biomarkers. 42 . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted June 30, 2023. ; https://doi.org/10.1101/2023.06.30.23292085 doi: medRxiv preprint Then, we conducted a HDMA using the R packages HIMA and DACT to identify any potential mediating CpG sites between PM 2.5 exposure and neuropathology from all 789,286 CpG sites. HIMA is an R package for estimating and testing high-dimensional mediation effects for omics data, which adopts the multiple mediator model's framework with reducing the dimensionality of omics data via sure independence screening and minimax concave penalty. 43 The divide-aggregate composite null test (DACT) is a more recent method for HDMA, which utilizes the Efron empirical null framework to calculate a weighted sum of p-values obtained from exposure-mediator (EWAS of PM 2.5 exposure as described above) and mediator-outcome (EWAS of neuropathology markers as described above) models for testing the significance of all mediators 44 . We corrected for multiple testing in HIMA and DACT using the Bonferroni method. Lastly, for the mediating CpG sites identified by either HIMA or DACT, we used the R package mediation to conduct a causal mediation analysis obtain their indirect effects. 45-47 The mediation is a frequently used tool which implements the mediation methods and suggestions proposed by Imai et al. 48,49 The average causal mediation effect (i.e., indirect effect) and total effect estimated by mediation were summarized for the CpG sites with positive indirect effects that were in line with the hypothesized adverse effect of traffic-related PM 2.5 on neuropathology markers. In contrast to the MITM approach described earlier, HDMA examine multiple mediators together in a framework of mediation analysis, which allowed us to ascertain the extent to which the particular indirect effects were associated with the mediators.
To aid the interpretation of model results, we conducted a gene ontology analysis using the R package missMethyl based on the top 1000 CpG sites with lowest raw p-. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) 1 4 values 50 . The gene ontology analysis was conducted for the EWAS results of PM 2.5 exposure as well as for the EWAS results of the three neuropathology markers. All CpG sites were annotated using an online annotation data for the 'IlluminaHumanMethylationEPIC'. 51 Additional functional insight on single CpG sites was obtained by searching the corresponding CpG site in publicly available databases, including EWAS catalog 52 .
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

Study population characteristics
A total of 159 donors were included in the current analysis, and their demographic characteristics and neuropathologic markers are described in Table 1

Association between PM 2.5 exposure and DNAm in the brain
After correcting for multiple tests and adjusting for bias and measured and unmeasured confounding, two CpG sites (cg25433380 and cg10495669) were consistently associated with PM 2.5 across different exposure windows ( Figure 2, Table 2; summary statistics for all 789,286 CpG sites are provided as Tables S4-6 in spreadsheets). For . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted June 30, 2023. ; 1 6 example, a 1 µg/m 3 increase in 1-year PM 2.5 exposure was associated with 0.0065 increase in the DNAm beta value of cg25433380 (p = 1.58×10 -8 ). cg25433380 and cg10495669 are on chromosome 9 and 20, respectively, and cg10495669 is assigned to the gene encoding RanBP-type and C3HC4-type zinc finger-containing protein 1 (RBCK1). The two CpG sites were not significantly associated with any neuropathology markers ( Table 2).

Meet-in-the-Middle approach and high-dimensional mediation analysis
For the MITM approach, we explored the overlapping CpG sites among the top 1000 CpG sites for the EWAS of PM 2.5 and the EWAS of neuropathology markers (results presented in Tables S7-S9 in spreadsheets) and identified four overlapping CpG sites (Table S1). Specifically, DNAm in cg01835635 (apolipoprotein A4 gene, APOA4) was associated with CERAD score as well as PM 2.5 exposure for the 1-year and 3-year exposure windows. DNAm in cg09830308 (mixed lineage kinase domain like pseudokinase gene, MLKL) was associated with Braak stage as well as PM 2.5 exposures for the 1-year, 3-year, and 5-year windows; cg16342341 (sorbin and SH3 domain-containing protein 2 gene, SORBS2) was associated with CERAD score as well as 1-year PM 2.5 exposure; and cg27459981 (MLKL gene) was associated with Braak stage and ABC score as well as PM 2.5 exposures for the 3-year and 5-year windows.
The HDMA via HIMA did not identify any CpG sites as significant mediators. In the HDMA using a combination of DACT and causal mediation analysis, we identified twenty-two CpG sites to mediate the positive association between PM 2.5 exposure and ABC score (Table 3), while none were observed for Braak stage and CERAD score.
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted June 30, 2023. ; https://doi.org/10.1101/2023.06.30.23292085 doi: medRxiv preprint One CpG site (cg16342341, SORBS2 gene) was associated with all three exposure windows (1, 3 and 5-years prior to death), and eight with two exposure windows. Of note, cg16342341 (SORBS2) was also identified in the MITM approach described above. The total effect estimated for all mediation analyses was positive but insignificant in this subsample of the cohort (see Christensen et al. 2023 for the significant total effect in the full cohort). 23 The summary statistics for all CpG sites detected by DACT are summarized in the Supplement (Table S2).

Secondary analyses
A gene ontology analysis was conducted for the top 1000 CpG sites associated PM 2.5 and for the top 1000 CpG sites associated with the neuropathology markers. None of the KEGG pathways reached significance after correcting for multiple tests. Therefore, we summarized the top 10 KEGG pathways for each of the PM 2.5 exposures or neuropathology markers in the Supplement (Table S3). One pathway, which is the longevity regulating pathway, was associated with both 3-year exposure to PM 2.5 and CERAD score. Eight genes (HSPA1A, HSPA1L, IRS1, KRAS, NRAS, RPTOR, IRS2, ATG5) in this pathway were enriched by differentially methylated CpG sites that were associated with 3-year PM 2.5 exposure, and ten genes (ADCY3, ADCY5, NFKB1, PRKAG2, RPTOR, TSC2, EHMT1, ULK1, AKT1S1, ATG5) with CERAD score. Of note, AKT1S1 was also among the genes that were identified in the HDMA (DACT and causal mediation analysis).

Discussion
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) In the current study of 159 donors from the Emory Goizueta ADRC brain bank, we identified differential DNAm in prefrontal cortex tissues at two CpG sites to be significantly associated with long-term PM 2.5 exposure. The two CpG sites (cg25433380 and cg10495669) that were associated with PM 2.5 exposure were consistently associated with long-term exposures to traffic-related PM 2.5 1 year, 3 years, and 5 years prior to death, after controlling for measured and unmeasured confounding. While cg25433380 and cg10495669 were not associated with increases in neuropathology markers, we identified 4 CpG sites that overlapped between the top 1000 CpG sites associated with PM 2.5 and neuropathology markers (MITM approach) and 22 CpG sites that mediated the adverse effect of PM 2.5 exposures on AD-related neuropathology markers using HDMA. The longevity regulating pathway was enriched by differentially methylated CpG sites associated with PM 2.5 (3-year exposure window) and CERAD score. This is the first study showing an association between PM 2.5 exposure and differential DNAm in the brain (cg25433380 and cg10495669). Scarce evidence related to air pollution has been reported on cg25433380. Higher DNA methylation levels of cg10495669 in nasal cells have been associated with 1-year ambient PM 2.5 exposure among 503 children in Massachusetts. 54 RBCK1, the gene which cg10495669 is assigned to, is involved in carcinogenesis and inflammation pathways. Yu et al.
suggested that RBCK1 promoted the ubiquitination and degradation of p53. 55 The impairment of p53 expression and activity might participate in neurodegeneration, as p53 can bind to genes that regulate expression of synaptic proteins, neurite outgrowth, and axonal regeneration, which indicated a neuroprotective role against AD . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted June 30, 2023. ; https://doi.org/10.1101/2023.06.30.23292085 doi: medRxiv preprint development. 56 In addition, RBCK1 can regulate the proinflammatory-cytokinesinduced nuclear factor kappa B (NF-kB) activation which serves as a pivotal mediator of inflammatory responses. 57 NF-kB activation is a common feature of many neurodegenerative diseases, 58 and the increased expression and/or activation of NF-kB has been largely observed in post-mortem studies of AD patients. 59 However, the two CpG sites were not found to be associated with any neuropathology markers in the current analysis. More research is warranted on these CpG sites to investigate their potential role in AD development with a larger sample size and participants of more diverse disease stages from preclinical to severe dementia.
We identified four CpG sites (cg01835635, cg09830308, cg16342341, and cg27459981) that overlapped between the top 1000 CpG sites associated with both PM 2.5 and neuropathology markers via MITM approach. Three of them (cg16342341, cg09830308 and cg27459981) or their related genes have been previously associated with AD or PM 2.5 exposure. Cg09830308 and cg27459981, assigned to MLKL, were both associated with Braak stage and PM 2.5 exposure 3 and 5 years prior to death. Collectively, PM 2.5 exposure might induce the TNF-mediated neuroinflammation, resulting in necroptosis, and thus contribute to AD pathogenesis.
Cg16342341, assigned to SORBS2, was also identified as a potential mediator in the HDMA, where it mediated the association of all PM 2.5 exposure windows with ABC score. As SORBS2 is well known for its role in AD and neuroinflammation 63,64 and was associated with PM 2.5 exposure in rats 65 , our findings contribute to the evidence of SORBS2 playing a role in PM 2.5 -induced neuropathologic changes of AD. SORBS2 represses IL-6 and TNF-α expression in the mouse embryonic fibroblasts, 66 and Chen et al. demonstrated that its level was lower in the brains of AD model mice compared to wild type mice, 63 implying a role of SORBS2 in regulating neuroinflammation. In a human study, genetic variation in SORBS2 was associated with age at onset of AD. 64 While evidence on the association between PM 2.5 exposure and SORBS2 is more scarce, Chao et al. reported that prenatal exposure to PM 2.5 induced upregulation of microRNAs targeting SORBS2 gene in fetal rat cortex tissues. 65 We also identified 21 other CpGs as potential mediators of the association between long-term exposure to traffic-related PM 2.5 and ABC score using HDMA, and two of these CpGs (cg07963191 and cg27297993) have been previously reported in association with AD. Cg07963191 was assigned to the dual 3',5'-cyclic-AMP and -GMP phosphodiesterase 11A gene (PDE11A) that participates in neuroplasticity and neuroprotection. 67 Cg27297993 was assigned to the gamma-aminobutyric acid B receptor 1 gene (GABBR1). GABBR1 is the main inhibitory neurotransmitter, which was reported to be downregulated in the brains of AD patients. 68 Iwakiri et al. observed a negative correlation between GABBR1 and NFT formation in the hippocampus of . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted June 30, 2023. ; https://doi.org/10.1101/2023.06.30.23292085 doi: medRxiv preprint 1 seniors, suggesting an increased or stable expression of GABBR1 may contribute to neuronal resistance to AD development. 69 To derive more functional insights, we conducted gene ontology analysis based on KEGG pathway database for the top 1000 CpGs associated with PM 2.5 exposure or neuropathology markers. 50 Proline-rich AKT1 substrate 1 (AKT1S1) was one of the genes enriched in the longevity regulating pathway, an overlapping pathway between PM 2.5 exposure and CERAD score. Of note, differential DNA methylation in cg00633834, which is assigned to AKT1S1, was also identified in HDMA. AKT1S1 can activate mammalian target of rapamycin (mTOR)-mediated signaling pathways when phosphorylated, 70 and mTOR signaling was observed to have higher activity in AD brains, suggesting a role of AKT1S1 in the accumulation of Aβ and tau proteins. 71 The current analysis employed the MITM approach and HDMA simultaneously to maximize the potential of identifying the differentially methylated CpG sites lying on a pathway from PM 2.5 to AD-related neuropathology. The application of the MITM approach was based on the investigation of epigenomics vs. PM 2.5 exposures and ADrelated neuropathology vs. epigenomics, which lent credibility to the association between PM 2.5 exposure and AD-related neuropathology by breaking it down and linking it up with DNAm. 72 Furthermore, while conventional methods of multiple testing correction (e.g., Bonferroni method) may overlook potential relevant CpG sites, especially given a small sample size, the MITM approach serves as a supplement by taking into account the biological relevance regardless of their statistical significance. 72 However, the MITM approach assumes that all intermediate variables are independent, which is not always the case in real-world scenarios. The HDMA focuses more on . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted June 30, 2023. ; https://doi.org/10.1101/2023.06.30.23292085 doi: medRxiv preprint quantifying the indirect effect of the mediator and considers the potential correlation among mediators. 43 Admittedly, we did not observe many consistencies, except for cg16342341 (SORBS2), between the two approaches.
Our study has several strengths. We established for the first time a potential mediation effect of DNAm for the association between PM 2.5 and neuropathological changes of AD. The neuropathological changes of AD were quantified via multiple markers, including Braak stage, CERAD score, and ABC score, which covers the essential components (i.e., NFTs and Aβ plaques) for the neuropathological diagnosis of AD. Further, the neuropathology markers were assessed by experienced neuropathologists at Emory Goizueta ADRC following a standardized protocol, which minimized the misclassification bias of outcomes. Finally, the high-resolution PM 2.5 exposure assessment model enabled the characterization of spatial variation in individual exposure and reduced the potential measurement error. 73 Our study has several limitations. First, the temporal sequence between DNAm changes and AD neuropathology could not be clearly defined because both were assessed post-mortem. Second, traffic-related PM 2.5 exposure was estimated based on the residential address of donors at death. Moving shortly prior to death could have introduced measurement errors in exposure assessment. The selection of exposure windows was arbitrary, as the disease process of AD may start many years before death and vary by patients. Third, the results were from a single brain bank and participants with a high APOE ε 4 carrier rate, so the generalizability should be tested in other brain banks or autopsy cohorts. Fourth, even though most of the study population was White, and we controlled for race, the ancestry effect on DNA methylation might . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted June 30, 2023. ; https://doi.org/10.1101/2023.06.30.23292085 doi: medRxiv preprint persist as residual confounding. Fifth, the current analysis only focused on the effects of PM 2.5 , while other air pollutants such as nitrogen oxides or ozone might also play a role for AD. 74,75 Finally, while the sample size of 159 brain samples was relatively large considering the challenges in collecting such samples, the high dimensionality of the genome-wide DNAm data raises concerns about the reliability of our findings.
Our findings provide important information on the biological mechanisms underlying the PM 2.5 toxicity on AD pathogenesis. Future studies evaluating the mediating role of DNAm on AD-related outcomes should consider: 1) performing the analysis among early-stage AD patients or patients with mild cognition impairment to further illustrate the role of PM 2.5 in AD etiology; 2) performing genome-wide DNAm together with transcriptomics, proteomics, and/or metabolomics to capture a holistic picture of the underlying mechanism.
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

(which was not certified by peer review)
The copyright holder for this preprint this version posted June 30, 2023. ; https://doi.org/10.1101/2023.06.30.23292085 doi: medRxiv preprint . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity.

Consent Statement
Written informed consent was provided for all subjects, and samples were obtained following research protocols approved by the Emory University Institutional Review Board. 3 6 Traffic-related fine particulate matter, Alzheimer's disease, neuropathology, DNA methylation.
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) . CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review)   Abbreviations: PM 2.5 , fine particulate matter; chr, chromosome; ACME, average causal mediated effect (i.e., indirect effect).
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted June 30, 2023. ; https://doi.org/10.1101/2023.06.30.23292085 doi: medRxiv preprint 1 a All CpG sites that were selected by DACT and had a positive ACME were associated with ABC score. No positive associations were found for Braak stage and CERAD score.
b Only the exposure windows were shown for which significant indirect effects were found. c The p-values of mediation effect testing conducted by DACT. d The ACME was associated with one-interquartile-range increase in beta values of CpG sites. e Effect estimates, associated with 1-unit increase, of PM 2.5 exposures on neuropathology markers.
. CC-BY 4.0 International license It is made available under a is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. is the author/funder, who has granted medRxiv a license to display the preprint in perpetuity. (which was not certified by peer review) The copyright holder for this preprint this version posted June 30, 2023. ; https://doi.org/10.1101/2023.06.30.23292085 doi: medRxiv preprint