Epigenetic age acceleration is associated with oligodendrocyte proportions in MSA and control brain tissue

Abstract Aims Epigenetic clocks are widely applied as surrogates for biological age in different tissues and/or diseases, including several neurodegenerative diseases. Despite white matter (WM) changes often being observed in neurodegenerative diseases, no study has investigated epigenetic ageing in white matter. Methods We analysed the performances of two DNA methylation‐based clocks, DNAmClockMulti and DNAmClockCortical, in post‐mortem WM tissue from multiple subcortical regions and the cerebellum, and in oligodendrocyte‐enriched nuclei. We also examined epigenetic ageing in control and multiple system atrophy (MSA) (WM and mixed WM and grey matter), as MSA is a neurodegenerative disease comprising pronounced WM changes and α‐synuclein aggregates in oligodendrocytes. Results Estimated DNA methylation (DNAm) ages showed strong correlations with chronological ages, even in WM (e.g., DNAmClockCortical, r = [0.80–0.97], p < 0.05). However, performances and DNAm age estimates differed between clocks and brain regions. DNAmClockMulti significantly underestimated ages in all cohorts except in the MSA prefrontal cortex mixed tissue, whereas DNAmClockCortical tended towards age overestimations. Pronounced age overestimations in the oligodendrocyte‐enriched cohorts (e.g., oligodendrocyte‐enriched nuclei, p = 6.1 × 10−5) suggested that this cell type ages faster. Indeed, significant positive correlations were observed between estimated oligodendrocyte proportions and DNAm age acceleration estimated by DNAmClockCortical (r > 0.31, p < 0.05), and similar trends were obtained with DNAmClockMulti. Although increased age acceleration was observed in MSA compared with controls, no significant differences were detected upon adjustment for possible confounders (e.g., cell‐type proportions). Conclusions Our findings show that oligodendrocyte proportions positively influence epigenetic age acceleration across brain regions and highlight the need to further investigate this in ageing and neurodegeneration.


INTRODUCTION
DNA methylation-based estimators of chronological age, termed 'epigenetic clocks', make use of penalised regression models such as elastic net regression to select signature CpG sites (CpGs) as biomarkers that correlate with chronological age [1,2]. Multiple clocks have been developed for use as molecular indicators of DNA methylation (DNAm) ages in different human tissues such as peripheral blood [3], multi-tissues (DNAmClock Multi ) [1] and the brain cortex (DNAmClock Cortical ) [2]. These epigenetic clocks are being increasingly used to determine the biological age of tissues and organs in various contexts; epigenetic age acceleration has been reported in multiple tissues, including the peripheral blood and post-mortem brain tissues, in neurodegenerative disorders such as Alzheimer's disease (AD), Huntington's disease (HD) and Parkinson's disease (PD) [4][5][6].
The DNAmClock Multi has been applied in tissues from several brain regions such as the dorsolateral prefrontal cortex (DLPFC) [6], frontal, occipital, temporal and parietal lobes, cerebellum, hippocampus, midbrain, caudate nucleus and cingulate gyrus [4,7]; similarly, the DNAmClock Cortical has specifically been tested in the dorsolateral prefrontal cortex and posterior cingulate cortex [8,9]. However, the sensitivity of the different clocks to tissue-specific changes (e.g., different cell-type proportions) varies depending on the tissues used in the reference training data, with a higher sensitivity likely to be observed in tissue-specific clocks compared with pan/multi-tissue clocks [10].
Moreover, when clocks were trained using data from one tissue at a time, their performances in predicting age in another tissue worsened [11]. However, the performance of these clocks in tissues that are particularly enriched for specific brain cell types, such as oligodendrocytes in the case of white matter, remains unexplored.
About half of the human brain is made up of white matter, which primarily comprises myelinated axons and glial cell types such as astrocytes, microglia and oligodendrocytes, with oligodendrocytes being by far the most abundant cell type [12]. White matter has been shown to play prominent roles in the normal functioning of the brain including cognitive [13] and motor functions [14]. Many neurodegenerative diseases show white matter changes, including AD, PD and multiple system atrophy (MSA) [15][16][17]. MSA is a rare adult-onset progressive neurodegenerative disorder with prominent white matter involvement [15]. MSA belongs to the group of α-synucleinopathies along with PD and Dementia with Lewy bodies (DLB) [18]. However, pathological hallmarks of MSA include the abnormal accumulation of misfolded α-synuclein primarily in the oligodendrocytes rather than in neurons, and these aggregates are termed glial cytoplasmic inclusions [19]. MSA is a predominantly sporadic disease, with a complex aetiology involving the interplay of genetic, epigenetic and environmental factors [20]. Recent epigenome-wide association studies have revealed aberrant DNA methylation changes in the MSA post-mortem brain tissue [21][22][23].
We, therefore, investigated the performances of DNAmClock Multi and DNAmClock Cortical in white matter from multiple brain regions and compared them with tissues comprising a mixture of grey and white matter in neurologically healthy controls as well as in a disease context. As MSA is a neurodegenerative disease model involving white matter pathological hallmarks, we investigated if patients with MSA exhibited accelerated ageing in the white matter of different brain regions such as the cerebellum and frontal and occipital lobes, as well as prefrontal cortex grey and white matter mixed tissue.

Cohort description
The first cohort (Cohort 1) consisted of white matter tissue carefully dissected from the cerebellum, and frontal and occipital lobes of postmortem brain tissues obtained from brains donated to the Queen Square Brain Bank. Cohort 1 comprised a total of 127 samples from MSA cases (N = 77) and neurologically healthy controls (N = 50), which included 93 samples from a previously published study [21] and 34 newly profiled samples obtained from the same brain bank. Cohort 2 consisted of samples from a publicly available dataset (GSE143157) comprising grey and white matter tissue mix from the prefrontal cortex of MSA cases (N = 40) and controls (N = 37) [22]. We further included a white matter cohort-Cohort 3, which consisted of control samples (N = 9) from the corpus callosum, the biggest white matter structure of the brain, obtained from a publicly available dataset (GSE109381) [24]. As MSA is characterised by oligodendroglial pathology, we included an additional Cohort 4, which consists of sorted SOX10+ (oligodendrocyte-enriched) immunolabelled populations from bulk DLPFC healthy control tissue (N = 15) [25]. The demographic details of all cohorts including sample numbers and chronological age are described in Table S1.

Key points
• DNAmClock Cortical and DNAmClock Multi are applicable to white matter from different brain regions.
• In white matter tissues of cortical and subcortical regions, DNAmClock Cortical performs better than DNAmClock Multi .
• DNAmClock Cortical shows lower accuracy (performs suboptimally) in cerebellar white matter compared with DNAmClock Multi .
• Using these two clocks, epigenetic age acceleration was not observed in MSA compared with controls upon adjustment for possible confounders such as cellular proportions, suggesting their limited applicability in diseases affecting specific cell types.
• Oligodendrocyte proportions are positively correlated with epigenetic age acceleration across brain regions.

DNA methylation profiling and data preprocessing
For the newly profiled samples from Cohort 1 (N = 34 samples), genomic DNA was extracted from frozen brain tissues using a previously established phenol-chloroform-isoamyl alcohol extraction method.
Bisulphite conversion was then carried out with 750 ng of DNA using the EZ DNA Methylation Kit (Zymo Research, Irvine, USA), followed by genome-wide methylation profiling using the Infinium Human-MethylationEPIC Bead Chip (Illumina). Bisulphite conversion and methylation profiling were carried out at UCL genomics, and the raw intensity files (.idat) were generated. Methylation profiles of 97 samples (a subset of Cohort 1) generated by Bettencourt et al. [21], 77 methylation profiles from the publicly available Cohort 2 [22] and 15 SOX10+ methylation profiles of Cohort 4 [25] had been previously generated using the Infinium HumanMethylationEPIC BeadChip (Illumina), whereas nine methylation profiles of Cohort 3 [24] had been previously generated using the Infinium HumanMethyla-tion450K BeadChip (Illumina). All cohorts were subjected to harmonised quality control checks and preprocessing.
Raw intensity files for all cohorts were imported into R using the WateRmelon package [26]. The following preprocessing and quality control checks were performed: (i) Raw intensities were visualised to identify and filter out atypical and failed samples, (ii) outlier detection to identify and remove outliers, (iii) bisulphite conversion assessment, where intensities from the probes were converted into percentages and samples with <80% bisulphite conversion were removed, and (iv) probe filtering to identify poorly performing probes by assessing the bead counts and detection-values. Samples with >1% probes above the 0.05 detection p-value threshold, probes with a beadcount of <3 in 5% of the samples and sites for which over 1% of samples showed a detection p-value >0.05 were discarded. Samples were also discarded if sex predictions did not match with phenotypic sex. Following the preprocessing, dasen normalisation was carried out. Celltype proportions of neuron-enriched (NeuN+), oligodendrocyteenriched (SOX10+) and other glial brain cell type (NeuNÀ/SOX10À) populations in the bulk tissues were estimated using the CETYGO package (https://github.com/ds420/CETYGO) and a sorted cell-type reference dataset as described by Shireby et al. [25]. CETYGO estimates cell proportions and quantifies the CEll TYpe deconvolution GOodness (CETYGO) score of a set of cellular heterogeneity variables derived from a genome-wide DNA methylation profile for an individual sample by capturing the deviation between the DNAm profile of a sample and its expected profile.

DNA methylation age estimation
Normalised beta values were used for the estimation of DNA methylation (DNAm) ages using two established epigenetic clocks trained to predict chronologic age, the DNAmClock Multi [1], which employs 353 CpGs curated from analysis of DNA methylation from 51 different tissues and cell types, and the DNAmClock Cortical [2] designed using 347 CpGs to specifically predict age in the human cortex. DNAm ages for DNAmClock Multi were calculated using the advanced analysis with normalisation, on the online calculator (http://dnamage.genetics.ucla. edu/). The DNAm ages for DNAmClock Cortical were calculated as described by Shireby et al. [2].

Statistics
All statistical analyses were performed in R (ver. 4.1.1). For both clocks, DNAm age acceleration was calculated as (1) the difference between the predicted DNAm age and chronological age and (2) residuals obtained by linear regression of DNAm age on chronological age and adjusting for possible confounders such as cell proportions (either neuronal or oligodendroglial) and duplicated individuals with multiple data points. Cohort 1 consisted of a subset of tissues for multiple brain regions obtained from the same individuals, therefore, mixed effects linear regression models were implemented using the lme4 package (https://github.com/lme4/lme4/) [27], where DNAm age was regressed against chronological age, estimated cell proportions and individual. For the mixed-effect regression models, chronological age and cellular proportions derived using CETYGO [25] were considered as fixed effects, and the individual's ID was included as a random effect. For Cohorts 2, 3 and 4, as there was only one sample per individual, we employed standard linear regression models where DNAm age was regressed with both age and cell proportion estimates. Furthermore, as post-mortem interval (PMI) was available for Cohort 1, on a separate analysis, we also included PMI as a fixed effect in addition to chronological age and cell proportion estimates and included individual as a random effect in the mixed effects linear regression model. Correlations between chronological and epigenetic ages for the different groups were calculated using Pearson's coefficient. Comparisons across groups/brain regions were performed using the Kruskal-Wallis test, and pairwise comparisons between the groups (MSA vs. controls) within each brain region were performed using pairwise Wilcoxon test with the Benjamini-Hochberg multiple testing correction. Correlations between epigenetic age acceleration (i.e., residuals obtained by linear regression of DNAm age on chronological age) and cell-type proportions (e.g., estimates of oligodendrocyte proportions) were calculated using Pearson's coefficient.

RESULTS
Comparison of the performances of DNAmClock Multi and DNAmClock Cortical in white matter from different brain regions and correlations with chronological age  was observed in white matter, but with higher error values ( Figure 1, Table 1). The findings from white matter and oligodendroglial- Oligodendrocyte proportions are associated with epigenetic age acceleration As DNA methylation patterns often are cell-type specific, it is reasonable to hypothesise that differences in specific cell-type proportions in a tissue sample will influence the estimates of epigenetic age based on DNA methylation levels and consequently epigenetic age acceleration. Previous work has demonstrated that a significant negative correlation exists between epigenetic age acceleration and the proportion of neurons in the prefrontal cortex [29]. We found the same direction of effect in our prefrontal cortex grey and white matter mix tissues with the DNAmClock Cortical , with a significant negative correlation observed between epigenetic age acceleration and neuronal proportions (r = À0.47, p = 1.6 Â 10 À5 ) ( Figure S1). As expected, in our white matter cohorts, the estimated neuronal proportions were minimal in frontal (mean 1.8 ± 0.6) and occipital lobes (mean 1.98 ± 1.6) as well as in the corpus callosum (mean 5.7 ± 0.8). Furthermore, there was no relationship between the residual neuronal proportions and epigenetic age acceleration in the white matter for these regions.
Lower neuronal proportions (i.e., higher glial proportions) have been associated with higher estimates of epigenetic ages [2] Figure S4). Altogether, our data support a role for oligodendrocytes pushing towards older epigenetic age estimates, whereas it would be the other way around for neurons and the other brain cell types.

Epigenetic age acceleration in the different brain regions in MSA
As increased epigenetic age acceleration has been reported to occur in neurodegenerative diseases [4][5][6], we investigated the presence of biological age acceleration in MSA when compared with controls based on results from the two clocks. For Cohort 1, with both DNAm-Clock Multi and DNAmClock Cortical , increased age acceleration was observed in the MSA white matter of all brain regions when compared with the controls (Figure 3). Differences ranged between 2 years in the frontal lobe for DNAmClock Cortical and 7 years in the occipital lobe for both DNAmClock Multi and DNAmClock Cortical (Figure 3A,C). Pairwise comparisons revealed that these differences were statistically significant only in the occipital lobe for DNAmClock Multi (p = 0.042).
In the prefrontal cortex mixed tissue (Cohort 2), $2 years of acceleration was observed in the MSA samples with DNAmClock Multi (p = 0.038, Figure 3B). A smaller difference was found in the same direction with the DNAmClock Cortical ( Figure 3D).
We then proceeded to verify whether the age acceleration could be influenced by possible confounding factors. Chronological age is known to have an impact on the DNAm age estimations [32]. As the chronological ages of the control group were generally higher than that of the MSA group for all brain regions, this had to be accounted for. In addition, for Cohort 1, although white matter tissue was carefully hand-dissected and enriched for oligodendrocytes, neuronal contaminants could still be present, and therefore, the neuronal proportions needed to be accounted for. Also, Cohort 1 included multiple brain regions per individual. Therefore, the second measure of age acceleration corresponded to the residuals obtained by regressing unchanged, and no significant age acceleration was observed in the MSA group in any of the studied brain regions ( Figure S6).

DISCUSSION
This is the first study, to our knowledge, that comprehensively ana- Therefore, in addition to tissue specificity, proportions of different cell types should also be taken into consideration to increase the accuracy of epigenetic clocks. Our results on the varied performances of the two clocks in different tissue types and regions within the brain further support the need for the development of specific clocks for different cell types and tissues as well as clocks specifically tailored in the context of brain ageing and disease as previously suggested [2,9].
It is of note that DNAm states have been shown to play a part in driving cell division and the differentiation of oligodendrocyte precursor cells into mature oligodendrocytes [34]. Therefore, defective DNAm changes could lead to unbalances in the number and/or function of oligodendrocytes in ageing and neurodegenerative diseases.
As the differentiation of oligodendrocyte precursor cells into mature myelinating oligodendrocytes depends on high metabolic and mitochondrial demand, these cells are particularly susceptible to oxidative stress, which affects oligodendrocyte precursor cell plasticity and differentiation capacity in ageing and neurodegenerative diseases [35].
Increased numbers of oligodendrocyte precursor cells have indeed been reported in several neurodegenerative diseases, including amyotrophic lateral sclerosis [36] and MSA [37,38]. Given the observed positive correlations between oligodendrocyte proportions and DNAm age acceleration, in both MSA and ageing controls, one may hypothesise that disruptions in the oligodendrocyte differentiation process, due to oxidative stress and other factors, are involved in the observed relationship with DNAm age acceleration and oligodendrocyte sensitivity to ageing. As the algorithm used to predict cell-type proportions cannot distinguish cell types within the oligodendrocyte lineage (e.g., oligodendrocyte precursor cells from mature oligodendrocytes), further studies would be necessary to dissect whether specific cells within the lineage underpin the observed correlations as well as to explore the potential applicability of oligodendrocyte proportions as an ageing biomarker.
Epigenetic age acceleration in specific brain regions has been reported in other neurodegenerative diseases such as AD and HD [4,39]; however, no studies exist on epigenetic age acceleration in the brain tissues for any of the α-synucleinopathies. Our rigorous multi-step analysis of the epigenetic clocks in the context of MSA demonstrates that although increased DNAm age acceleration differences were observed in multiple brain regions in the MSA group compared with controls with both epigenetic clocks, no statistical significance was observed upon adjustment for chronological age, neuronal/oligodendrocyte proportions and duplicated individuals with multiple data points. Our results highlight the importance of accounting for confounders during the estimation of age acceleration and emphasise the need for researchers to interpret the results in ageing and disease models with caution.
Our study has several limitations, the sample sizes for each of the brain regions are relatively small, and future studies with larger sample sizes would be important to further validate our observations; in addition, the chronological age for the controls was considerably higher than that for the MSA cases, and although these differences were accounted for during the statistical analysis, different age ranges may have influenced the performance of both clocks differently for MSA cases and controls; finally, both clocks used in this study have not been specifically designed for white matter tissues, and follow-up studies with more specialised epigenetic clocks are warranted. Moreover, we cannot exclude the possibility that other factors, including variable PMI or variable number of missing probes across cohorts, could have had an impact on DNAm age predictions. However, it is of note that when accounting for PMI in the regression models used for Cohort 1, it did not change our overall results. Additionally, after following the methodology described by McEwen et al. [40], it is unlikely that the missing probes would have considerably affected the precision of the DNAm age predictions.
Notwithstanding, our data point towards important contributions of oligodendrocytes in brain tissue biological ageing and emphasise the need for additional studies on brain ageing and neurodegeneration focusing on this cell type.
In conclusion, this study demonstrates that both DNAmClock-Cortical and DNAmClock Multi showed a high correlation between DNAm age and chronological age, even in white matter, expanding the applicability of these clocks. As expected, performances and DNAm age estimates varied considerably between clocks and tissue types. With these available tools, we could not detect changes in epigenetic age acceleration in MSA compared with controls when accounting for known confounding factors. Additional important findings from this study consist of the positive correlations observed between epigenetic age acceleration and oligodendrocyte proportions across brain regions and tissue types, with the opposite being observed for all other brain cell types. This, together with previous reports of older DNAm ages in glial cells compared with neurons from the same individual, support differential biological ageing in different brain cell types, and the possibility of accelerated ageing in oligodendrocytes. Our findings highlight the need for cell type and tissue-specific clocks and clocks that include shared markers of common aberrant epigenomic patterns underlying neurodegeneration to accurately dissect disease-related DNAm age acceleration from normal ageing.