Immune Cell Type‐Specific DNA Methylation Regions Associate With 24‐Hour Blood Pressure Regulation in Black People

Abstract

Background

DNA methylation and immune cells have been linked to blood pressure (BP) regulation and the development of hypertension. However, the immune cell profiles and the cell type‐specific DNA methylation associated with BPs remain unclear.

Methods

This study evaluates the 19 cell type deconvolution algorithms using reduced representation bisulfite sequencing data, comparing them to in silico mixtures derived from whole‐genome bisulfite sequencing. The top‐performing algorithm, Epigenetic Dissection of Intra‐Sample Heterogeneity (EpiDISH)‐Robust Partial Correlations, was applied to 281 Black inpatients with 24‐hour BP monitoring. The immune cell profiles and cell type‐specific DNA methylation regions associated with these BP phenotypes were further investigated using regression analysis.

Results

In patients with hypertension, B‐cell and CD4 effector memory T‐cell abundances were significantly elevated. Monocyte and CD8 effector memory T‐cell fractions positively correlated with nighttime BP, and CD3 T cells were inversely associated with office BP. These associations remained robust after covariate adjustments and were partially validated in the Medical Information Mart for Intensive Care‐IV cohort. For the first time, we identified several cell type‐specific DNA methylation regions as being associated with BP phenotypes and patterns across 13 immune cells, with approximately one third predominantly found in effector CD8 T cells.

Conclusions

These findings provide novel insights into the epigenetically regulated immune mechanisms underlying BP regulation and identify potential targets for hypertension management.

Nonstandard Abbreviations and Acronyms

CpG

cytosine‐phospho‐guanine dinucleotide

CTS‐MRs

cell type‐specific DNA methylation regions

JSD

Jensen–Shannon divergence

NK

natural killer

PP

pulse pressure

RMSE

root mean squared error

RRBS

reduced representation bisulfite sequencing

Clinical Perspective.

What Is New?

• We provide a thorough evaluation of cell type deconvolution algorithms using reduced representation bisulfite sequencing data and demonstrate significant associations between immune cell composition and 24‐hour blood pressure phenotypes and patterns.

• We report one of the first studies of cell type‐specific DNA methylation regions associated with 24‐hour blood pressure phenotypes and patterns at the cell type‐specific level, with approximately one third of cell type‐specific DNA methylation regions present in effector CD8 T cells.

What Are the Clinical Implications?

• Our study supports the involvement of epigenetics‐regulated immune dysregulation in blood pressure variability and highlights potential biomarkers for future hypertension management strategies.

High blood pressure (BP) is one of the leading risk factors for cardiovascular and renal diseases, contributing to substantial morbidity and mortality worldwide. ¹ , ² , ³ This burden is particularly pronounced among Black people, who have a higher prevalence of hypertension, an earlier onset, and an increased risk of resistant hypertension compared with other populations. ⁴ , ⁵ Extensive studies in animal models and human studies have linked various immune cell types with BP levels, underscoring the potential role of the immune system in BP control. ⁶ , ⁷ , ⁸ , ⁹ , ¹⁰ , ¹¹

DNA methylation is a well‐established epigenetic marker that has been reported to be associated with the regulation of BP and the development of hypertension. ¹² , ¹³ , ¹⁴ , ¹⁵ However, cell type heterogeneity is one of the major sources of variation in DNA methylation that can complicate the identification of DNA methylation markers associated with BP regulation. Therefore, accounting for cellular heterogeneity is critical in DNA methylation analysis. ¹⁶

Although immune mechanisms have been implicated in the pathogenesis of hypertension, the precise underlying mechanisms remain poorly understood. ¹⁷ Although immune cell composition can be determined through methods such as immunohistological analysis, cell sorting, or single‐cell RNA‐sequencing, these approaches are often costly and time‐consuming and require fresh blood samples. A valuable alternative is cell type deconvolution, which decomposes bulk tissue signals into their constituent cell types' signals. ¹⁸ , ¹⁹ , ²⁰ Importantly, distinct DNA methylation profiles across different cell types enables cell type deconvolution from bulk samples, facilitating the identification of cell type‐specific DNA methylation markers associated with various diseases. A recent study estimated the circulating proportions of 12 leukocyte subsets from both lymphoid and myeloid lineages by deconvolving cell‐type‐specific DNA methylation data (450K and Estimate the Proportion of Immune and Cancer cells arrays) in 4124 women and examined their relationship with the development of elevated BP. ²¹ However, to date, no study has identified cell type‐specific DNA methylation markers associated with BP regulation.

In contrast to single‐time‐point BP measurements during office or research visits, 24‐hour BP monitoring provides a more comprehensive view, capturing short‐term BP variability and diurnal variation. ²² This method allows for the detection of masked, white coat, or nocturnal hypertension—conditions that may otherwise go unnoticed, leading to inappropriate treatments and poor clinical outcomes. ¹²

To address these gaps, we aim to systematically evaluate cell type deconvolution algorithms using data generated from reduced representation bisulfite sequencing (RRBS), one of the most commonly used and efficient method for profiling genome‐wide DNA methylation. ²³ , ²⁴ Our objectives are to identify the top‐performing deconvolution algorithms for our real RRBS data, investigate the relationship between immune cell profiles and 24‐hour BP phenotypes, and identify the cell type‐specific DNA methylation regions (CTS‐MRs) associated with these BP phenotypes in Black people.

METHODS

The DNA methylation data from whole‐genome bisulfite sequencing is available in the Gene Expression Omnibus database (accession no. GSE186458). The Medical Information Mart for Intensive Care IV (MIMIC‐IV) database is available through the PhysioNet repository (https://physionet.org/content/mimiciv/1.0/). Relevant code for simulation of deconvolution benchmarking on a demo data set is available at https://github.com/CORRS‐LAB/devtEp. All the other data that support the findings of this study are available from the corresponding author upon reasonable request.

Study Participants and Phenotypes

Participants from 2 cohorts were included in this study. The first cohort was the 24‐hour BP study, which was used to estimate peripheral immune cell composition and identify the CTS‐MRs. This cohort comprised 281 Black participants from our recent studies. ¹² , ²⁵ These participants determined as being Black based on self‐identification, birth in the United States by Black parents, and English as the native language. They were inpatients with a 3‐day stay at the Clinical Research Center and were drawn from a larger Milwaukee‐area cohort of 3943 normotensive and hypertensive Black volunteers, aged between 18 and 55, recruited between 1994 and 2006. Our exclusion criteria for this study included pregnancy, diabetes, serum creatinine >2.2 mg/dL, body mass index (BMI) >36 kg/m², active malignancy, a history of stroke or myocardial infarction within 6 months, and diastolic BP (DBP) >110 mm Hg despite being on antihypertensive therapy. Additionally, antihypertensive medications were discontinued for at least 1 week before participation, and lipid‐lowering medications were stopped for 1 month. The Milwaukee‐area cohort was approved by the Medical College of Wisconsin (Milwaukee, WI) Institutional Review Board and participants provided informed consent. For further details, please refer to our previous work. ¹²

In this study, hypertensive individuals were defined as having a systolic BP (SBP) >140 mm Hg or DBP >90 mm Hg, or on antihypertensive medication. The 24‐hour BP of 281 inpatients was recorded using an Accutracker monitor (Sun Tech Medical Instruments, Inc, Raleigh, NC, USA), with measurements taken every 20 minutes during the daytime (5 a.m.–11 p.m.) and every 45 minutes during the nighttime (11 p.m.–5 a.m.). BP phenotypes included in this study were 24‐hour average, daytime, nighttime, and office BP measurement of SBP, DBP, mean arterial pressure (MAP), and pulse pressure (PP).

Participants were further classified based their nighttime SBP decline into the following dipping patterns: dippers (a nocturnal SBP decline of at least 10% but <20%), nondippers (a nocturnal SBP decline of at least 0% but <10%), and reverse dippers (a nocturnal SBP increase). ²⁶ , ²⁷ , ²⁸

The second cohort was from the publicly available MIMIC‐IV database, ²⁹ specifically focusing on patients in intensive care units. It was used to validate the association between estimated peripheral immune cell composition and 24‐hour BP. Of the 383 220 inpatients recorded in MIMIC‐IV, 80526 were Black. To ensure data quality, immune cell proportion observations flagged as “abnormal” were excluded before the downstream validation analysis. Due to the limited availability of data on other immune cell types in the MIMIC‐IV data set, this study focuses on validating the association between monocytes and 24‐hour BPs. Monocyte values were defined as the absolute number of monocytes per unit volume of blood (eg, cells/μL), measured using Logical Observation Identifiers Names and Codes 742‐7. In order to avoid data duplication, multiple hospitalizations or multiple records from the same patient were removed, retaining only the first hospitalization and the first record of monocyte value for each inpatient. BPs within 24 hours of the chart blood collection time (used to measure monocyte fractions) were categorized into daytime and nighttime readings. In total, 2729 inpatients with recorded monocyte value were identified. After excluding those missing 24‐hour BP data or BMI, the final cohort consisted of 1017 inpatients.

Deconvolution Methods

Reference‐based cell type deconvolution algorithms typically model the observed bulk DNA methylation profile as a linear combination of a given set of reference data profiles that represent the underlying cell types. Let Y be a vector representing methylation levels at m cytosine‐phospho‐guanine dinucleotides (m CpGs; or regions) from a bulk DNA methylation profile in a sample, C a vector representing proportions of the K underlying cell types, and X the reference signature, a methylation data matrix with dimensions m by k, containing unique DNA methylation characteristics for each specific cell type. The deconvolution model can be expressed as $\mathrm{Y}=\mathrm{XC}+\mathrm{ϵ}$, where $\mathrm{ϵ}$ is the error term accounting for measurement noise and other unmodeled factors. This model interprets the observed bulk DNA methylation profile as the sum of the DNA methylation profiles from each contributing cell population. As the most basic model, the cell type proportions are estimated from the weight coefficients by the ordinary least square regression. For our benchmarking, we included 16 cell type deconvolution algorithms from a recent study ³⁰ and added 3 more algorithms: Estimate the Proportion of Immune and Cancer cells, EpiDISH‐CIBERSORT and EpiDISH‐Constrained Projection (Table 1). ³¹ , ³² , ³³ , ³⁴ , ³⁵ , ³⁶ , ³⁷ , ³⁸ , ³⁹ , ⁴⁰ , ⁴¹ , ⁴²

Table 1.

Overview of Cell Type Deconvolution Algorithms Included in the Study

Method	Description	Software
BVLS	Bounded‐variable least squares	R package: bvls 31
Digital cell quantifier	Elastic net regularization	R package: ADAPTS 32
eMeth‐Binomial	Expectation maximization with binominal likelihood function	R package: Emeth 33
eMeth‐Normal	Expectation maximization with normal likelihood function	R package: Emeth 33
eMeth‐Laplace	Expectation maximization with normal likelihood function	R package: Emeth 33
EPIC	Regression models	R package:EPIC 34
EpiDISH‐CBS	CIBERSORT using support vector regression	R package: EpiDISH 35
EpiDISH‐CP	Linear constrained projection	R package: EpiDISH 35
EpiDISH‐RPC	Robust partial correlation	R package: EpiDISH 35
Elastic net	Elastic net regularization	R package: glmnet 36
FARDEEP	Least trimmed squares	R package: FARDEEP 37
Immune Cell Deconvolution in Tumor Tissues	Expectation maximization	R package: iCeDT 38
Lasso	Lasso regression	R package: glmnet 36
Meth atlas	Non‐negative least squares	Pyhton: Meth atlas 39
MethylResolver	Least trimmed squares	R package: MethylResolver 40
Minfi	Linear constrained projection	R package: Minfi 41
NNLS	Nonnegative least squares	R package: nnls 42
OLS	Ordinary least square regression	base R package
Ridge	Ridge regression	R package: glmnet 36

EPIC indicates Estimate the Proportion of Immune and Cancer cells; EpiDISH, Epigenetic Dissection of Intra‐Sample Heterogeneity; FARDEEP, Fast and Robust Deconvolution of Expression Profiles; and Lasso, least absolute shrinkage and selection operator.

Data Normalizations in Deconvolution

Seven normalization techniques were applied to the DNA methylation profiles before deconvolution: column‐wise Z score, column‐wise min‐max normalization, quantile normalization, log normalization, and no normalization. Specially, quantile normalization was performed using the ‘normalize.quantiles’ function of the R‐package ‘preprocessCore.’ ³⁰ The log normalization was implemented as ln(x + 1).

Metrics for Accessing Deconvolution Accuracy

To assess the accuracy of cell type deconvolution algorithms, three metrics were employed: R‐squared, root mean squared error (RMSE), and Jensen–Shannon divergence (JSD). These metrics were used to quantify the agreement between true and estimated cell proportions.

DNA Methylation Data

Two sets of DNA methylation profiles were analyzed in this study. The first data set comprised 28 217 448 CpGs obtained from whole‐genome bisulfite sequencing (WGBS), with150‐bp‐long paired‐end reads to a mean depth of at least 30 coverage. This WGBS data set was downloaded from the Gene Expression Omnibus database (accession no. GSE186458), ⁴³ which includes 35 blood samples across 13 cell types, including CD3, CD4, CD8, CD4 central memory T cells, CD8 effector T cells, CD4 effective memory T cells, CD8 effective memory T cells, CD8 naïve T cells, natural killer cells, monocytes, granulocytes, B cells, and memory B cells (Table S1). These profiles were used to generate reference signatures and in silico mixtures for benchmarking the 19 cell type deconvolution algorithms.

The second data set comprised DNA methylation profiles from 281 inpatients, obtained using RRBS. This RRBS data set contained 1 549 368 CpG sites with at least 5× coverage in at least 95% of the samples. It was used to investigate the association between peripheral immune cell composition and 24‐hour BP phenotypes and the association between the cell type‐specific DNA methylation makers and 24‐hour BP phenotypes.

Reference Signatures and In Silico Mixtures in Deconvolution

Reference signatures and in silico mixtures were generated using the WGBS data set, which contained DNA methylation profiles from 35 blood samples across 13 cell types. To assess the performance of deconvolution methods in RRBS, we retained only the CpG sites detected in both the WGBS and RRBS data sets. Given the high correlation between CpGs in adjacent genomic regions, ⁴⁴ CpGs were merged into nonoverlapping genomic regions spanning 100 bp.

Regions were then selected as deconvolution makers using the algorithm described by Luo et al. ³⁰ , ⁴⁵ To increase the computational efficiency, we optimized the original marker selection algorithm using C/C++ language and incorporated it into our R package “devtEp.” For each region, 2‐sample t tests were applied to assess methylation difference between the target cell type and all other cell types. The top 200 regions with the lowest P values were selected for each cell type, and from these, the 100 regions with the highest mean methylation differences were retained. After removing duplicated regions across cell types, the final reference signature consisted of 1143 regions.

For each of 35 blood samples, 3 repeats were randomly generated by the binomial distribution, with parameters as total count and methylation rate at each CpG site within the reference region. This produced 105 simulated RRBS samples across 13 cell types.

Two types of ground‐truth cell proportions were generated for in silico mixtures: homogeneous and heterogeneous. Homogeneous cell proportions were generated from Dirichlet distribution using R package ‘LaplacesDemon.’ Heterogeneous cell proportions were generated from uniform distributions centered around the estimated proportion for each cell type, with the values rescaled to sum to 1. Finally, simulated bulk samples were generated by summing the methylation levels at genomic regions for each simulated cell type, weighted by the corresponding cell proportion. A total of 100 in silico mixtures were generated in each scenario.

Statistical Analysis

Age was described using median and interquartile range. The Mann–Whitney test was used to evaluate differences in median age between hypertensive and normotensive subjects. Continuous variables such as BMI, BP phenotypes, and cell proportions were described using mean±SD. Two‐sample t test was performed to compare differences between 2 groups. Categorical variables, including sex, hypertension status, dipper, nondipper, and reverse dipper, were described as frequencies and percentages. Fisher's exact tests were used to examine proportions for categorical variables.

Regression analyses, with or without covariates adjustments, were performed to investigate the association between peripheral immune cell composition and blood pressure phenotypes. In univariable regression analyses, linear regression was used for continuous dependent variables, and logistic regression was used for binary dependent variables. In multivariable regression analyses, covariates including age, sex, and BMI were further adjusted in linear and logistic regressions. P value <0.05 was considered statistically significant.

CTS‐MRs associated with BP phenotypes were identified using regression analyses conducted with the “CellDMC” tool. ⁴⁶ In these analyses, the bulk DNA methylation level at each region served as the dependent variable, immune cell fractions and the interaction with BP phenotypes were treated as independent variables. Models were performed with or without covariate adjustments. Multiple comparisons were adjusted using Bonferroni–Hochberg correction. All statistical analyses were performed using R Statistical Software (version 4.4.1) (https://www.r‐project.org/).

RESULTS

Study Overview

This study involved 3 key parts: benchmarking, identifying immune cells and CTS‐MRs associated with BP phenotype, and patterns (Figure 1).

Figure 1. Schematic representation of the study.

A, Reference data set (WGBS) containing DNA methylation profiles of cell types across all CpG sites. B, Bulk data set (RRBS) containing DNA methylation profiles across all CpG sites. C, Reference data set containing DNA methylation profiles of cell types in selected regions. D, Bulk data set containing DNA methylation profiles in selected regions. E, Estimated proportion of cell subtypes for each bulk sample using the deconvolution algorithm. F, Identify BPs related cell subtypes in inpatients using RRBS. G, Identify BPs related CTS‐MRs using RRBS. BP indicates blood pressure; CpG, cytosine‐phospho‐guanine dinucleotide; CTS‐MR, cell type‐specific DNA methylation regions; HT, subjects with hypertension; ICU, intensive care unit; JSD, Jensen–Shannon divergence; NT, subjects without hypertension; RMSE, root mean squared error; RRBS, reduced representation bisulfite sequencing; and WGBS, whole‐genome bisulfite sequencing.

(1) During the benchmarking, we systematically evaluated 19 commonly used or recently developed cell type deconvolution algorithms, listed in Table 1. These included bounded variable least‐squares, Elastic net regression, Emeth‐Binomial, Emeth‐Laplace, Emeth‐Normal, EpiDISH‐Robust Partial Correlations (RPC), EpiDISH‐CIBERSORT, EpiDISH‐Constrained Projection, least absolute shrinkage and selection operator, Meth atlas, MethylResolver, Minfi, nonnegative least squares, ordinary least squares, Ridge regression, Estimate the Proportion of Immune and Cancer cells, Fast and Robust Deconvolution of Expression Profiles, Immune Cell Deconvolution in Tumor Tissues, and digital cell quantifier. Additionally, we investigated 7 data normalization approaches as selecting appropriate normalization approach can improve deconvolution accuracy. This resulted in 133 total algorithm‐normalization combinations. However, because Immune Cell Deconvolution in Tumor Tissues and Estimate the Proportion of Immune and Cancer cells cannot deal with negative values, they are incompatible with (column) Z score and Z‐score normalization, reducing the total of 129 combinations.

(1a) CpG site retention: To assess deconvolution methods in RRBS data sets, we retained CpG sites that were consistently detected in both WGBS and RRBS data sets.

(1b) Regional merging: Given the high correlation between adjacent CpG sites, ³¹ we merged these sites into nonoverlapping 100 bp regions. Methylation marker regions were selected using the WGBS data set and an optimized algorithm from our R package, “devtEp.” For each cell type, we selected the top 200 regions with the lowest P values and further refined this to retain the 100 regions with the highest mean methylation differences between that cell type and the other cell types.

(1c) Simulated RRBS data generation: For each of the 35 blood samples in the WGBS data set, we generated 3 random RRBS repeats using the binomial distribution. Two types of ground‐truth cell proportions were created for in silico mixtures: homogeneous proportions (drawn from a Dirichlet distribution using the “LaplacesDemon” R package) and heterogeneous proportions (from uniform distributions centered around estimated proportions, rescaled to sum to 1). We generated 100 in silico mixtures for each scenario by summing the methylation levels at genomic regions for each cell type, weighted by their respective proportions.

(1d) Deconvolution analysis: The different deconvolution and normalization methods were then applied to in silico mixtures to estimate cell type proportions. Performance was evaluated by comparing the estimated proportions with the ground‐true cell proportions using R‐squared, RMSE, and JSD.

(2) Association analysis between immune cell composition and BP phenotypes: The best‐performing algorithm from the benchmark analysis was subsequently applied to real RRBS data from the 24‐hour BP study to estimate immune cell abundance. These abundance estimates were then examined for associations with 24‐hour blood pressure phenotypes and patterns. The findings are partially validated from MIMIC‐IV data sets.

(3) Association analysis between CTS‐MRs and BP phenotypes: The estimated immune cell abundance as well as the DNA methylation profiles were further analyzed to identify CTS‐MRs associated with 24‐hour blood pressure phenotypes and patterns.

Clinical Characteristics of Study Participants

In the 24‐hour BP study, 281 inpatients were included, with a median age of 44 years, and nearly half were female. Among these, 142 subjects had hypertension, with a median age 46 years, significantly older than subjects without hypertension. BMI and BP phenotypes, including office, 24‐hour, day, and night BPs (SBP, DBP, MAP, PP) were significantly higher in subjects with hypertension than subjects without hypertension. Although a slight difference in the dipper and reverse‐dipper patterns was observed between subjects with and without hypertension, this did not reach statistical difference. No significant difference was observed in sex distribution or the nondipper pattern between subjects with and without hypertension (Table 2).

Table 2.

Clinical Characteristics of Study Participants

Characteristic	Discovery				Validation
	Total (N=281)*	Hypertension (n=142)	Normotension (n=139)	P value	Total (N=1017)
Age, y, median (interquartile range)	44 (40–49)	46 (42–51)	43 (39–47)	6.6E‐04	62 (49–72)
Sex
Male	140 (49.8%)	72 (50.7%)	68 (48.9%)	0.8	533 (52.4%)
Female	141 (50.2%)	70 (49.3%)	71 (51.1%)		484 (47.6%)
Body mass index, kg/m2	28.2 (4.5)	29.2 (4.6)	27.1 (4.2)	7.1E‐05	28.5 (7.1)
SBP office, mm Hg	131.9 (22.4)	149.0 (18.5)	114.4 (7.8)	<2.2E‐16	…
DBP office, mm Hg	86.2 (15.0)	97.6 (12.2)	74.7 (6.1)	<2.2E‐16	…
MAP office, mm Hg	101.5 (16.8)	114.7 (13.1)	87.9 (5.9)	<2.2E‐16	…
PP office, mm Hg	45.6 (12.4)	51.4 (13.8)	39.7 (6.9)	<2.2E‐16	…
SBP 24 h, mm Hg	130.4 (19.0)	144.5 (15.2)	115.9 (9.2)	<2.2E‐16	120.3 (17.3)
DBP 24 h, mm Hg	78.6 (12.5)	87.6 (9.7)	69.3 (6.9)	<2.2E‐16	66.1 (11.8)
MAP 24 h, mm Hg	95.9 (14.3)	106.6 (10.9)	84.8 (7.3)	<2.2E‐16	80.5 (11.8)
PP 24 h, mm Hg	51.8 (9.2)	56.9 (9.4)	46.6 (5.3)	<2.2E‐16	54.2 (15.1)
SBP day, mm Hg	131.6 (19.0)	145.5 (15.2)	117.3 (9.5)	<2.2E‐16	120.6 (18.0)
DBP day, mm Hg	79.6 (12.5)	88.4 (9.8)	70.5 (7.3)	<2.2E‐16	66.3 (12.3)
MAP day, mm Hg	96.9 (14.3)	107.5 (11.1)	86.1 (7.7)	<2.2E‐16	80.7 (12.3)
PP day, mm Hg	52.0 (9.1)	57.0 (9.1)	46.9 (5.4)	<2.2E‐16	54.3 (15.8)
SBP night, mm Hg	125.2 (20.1)	139.3 (17.4)	111.2 (10.6)	<2.2E‐16	120.0 (19.3)
DBP night, mm Hg	74.6 (13.3)	84.0 (11.5)	65.4 (7.2)	<2.2E‐16	65.8 (13.1)
MAP night, mm Hg	91.5 (15.2)	102.4 (12.9)	80.7 (7.8)	<2.2E‐16	80.1 (13.4)
PP night, mm Hg	50.4 (9.9)	55.2 (10.3)	45.8 (7.0)	5.4E‐15	54.2 (16.2)
Dipper	43 (17.5%)	16 (13.0%)	27 (22.0%)	0.09	145 (14.3%)
Nondipper	159 (64.6%)	79 (64.2%)	80 (65.0%)	1.0	371 (36.5%)
Reverse dipper	43 (17.5%)	27 (22.0%)	16 (13.0%)	0.09	479 (47.1%)

DBP indicates diastolic blood pressure; MAP, mean arterial pressure; PP, pulse pressure; and SBP, systolic blood pressure.

^*24‐hour SBPs were not available for 35 of 281 subjects.

Evaluation of Cell Type Deconvolution Algorithms

We evaluated the performance of 19 cell type deconvolution algorithms by applying 7 different transformations on in silico mixture samples composed of either homogeneous or heterogeneous cell type proportions. For each transformation on each sample, cell type proportions were estimated by various deconvolution algorithms. The accuracy between the ground‐truth cell proportion and the estimated cell proportion was quantified by R‐squared, RMSE and JSD (Figures 2 and 3).

Figure 2. Evaluating cell type deconvolution algorithms using R‐squared.

A, Homogeneous cell proportions used for generating in silico mixtures. B, R‐squared between estimated and true homogeneous cell proportions across deconvolution algorithms with different transformations. C, Heterogeneous cell proportions used for generating in silico mixtures. D, R‐squared between estimated and true heterogeneous cell proportions across deconvolution algorithms with different transformations. E, Estimated cell proportions of 281 inpatients. Color intensities indicate the magnitude of the R‐squared (ranging from 0 to 1). BVLS indicates bounded‐variable least squares; B_Mem, B memory; Col Z score, column‐wise Z score; Col min‐max: column‐wise minimum‐maximum; DCQ, digital cell quantifier; EPIC, Estimate the Proportion of Immune and Cancer cells; EpiDISH‐CBS, Epigenetic Dissection of Intra‐Sample Heterogeneity‐CIBERSORT; EpiDISH‐CP, Epigenetic Dissection of Intra‐Sample Heterogeneity‐Constrained Projection; EpiDISH‐RPC, Epigenetic Dissection of Intra‐Sample Heterogeneity‐Robust Partial Correlations; FARDEEP, Fast and Robust Deconvolution of Expression Profiles; IceDT, Immune Cell Deconvolution in Tumor Tissue; Lasso, least absolute shrinkage and selection operator; NK, natural killer; NNLS, nonnegative least squares; None, no normalization; OLS, ordinary least square; QN, quantile normalization; T_Eff_CD8, T effector CD8; T_EffMem_CD8, T effector memory CD8; T_EffMem_CD4, T effector memory CD4; and T_CenMem_CD4, T central memory CD4.

Figure 3. Evaluating cell type deconvolution algorithms using RMSE and JSD.

A, RMSE between estimated and true homogeneous cell proportions across deconvolution algorithms with different transformations. B, JSD between estimated and true homogeneous cell proportions across deconvolution algorithms with different transformations. C, RMSE between estimated and true heterogeneous cell proportions across deconvolution algorithms with different transformations. D, JSD between estimated and true heterogeneous cell proportions across deconvolution algorithms with different transformations. Color intensities, ranging from blue to red, represent the magnitude of RMSE and JSD, with blue indicating smaller values and red indicating larger values. BVLS indicates bounded‐variable least squares; Col Z score, column‐wise Z score; Col min‐max, column‐wise minimum‐maximum; DCQ, digital cell quantifier; EPIC, Estimate the Proportion of Immune and Cancer cells; EpiDISH‐CBS, Epigenetic Dissection of Intra‐Sample Heterogeneity‐CIBERSORT; EpiDISH‐CP, Epigenetic Dissection of Intra‐Sample Heterogeneity‐Constrained Projection; EpiDISH‐RPC, Epigenetic Dissection of Intra‐Sample Heterogeneity‐Robust Partial Correlations; FARDEEP, Fast and Robust Deconvolution of Expression Profiles; IceDT, Immune Cell Deconvolution in Tumor Tissue; JSD, Jensen–Shannon divergence; Lasso, least absolute shrinkage and selection operator; NK, natural killer; NNLS, nonnegative least squares; None, no normalization; OLS, ordinary least square; QN, quantile normalization; and RMSE, root mean squared error.

On average, R‐squared for homogeneous cell type proportions was 0.784 with quantile normalization and 0.866 without normalization (Figure 2A and 2B). For heterogeneous cell type proportions, the R‐squared improved to 0.864 in quantile normalization and 0.955 without normalization (Figure 2C and 2D). Higher accuracy in heterogeneous cell proportions was also observed in terms of RMSE and JSD, particularly without normalization (Figure 3).

We then compared the performance of the 19 cell type deconvolution algorithms. On average, the R‐squared across transformation methods for homogeneous cell type proportions was 0.925 using the EpiDISH‐RPC algorithm and 0.143 using the digital cell quantifier algorithm (Figure 3A and 3B). For heterogeneous cell type proportions, the R‐squared values were 0.974 and 0.572 for EpiDISH‐RPC and digital cell quantifier, respectively (Figure 3C and 3D). Notably, the R‐squared reached up to 0.99 with the EpiDISH‐RPC algorithm and up to 0.579 with the digital cell quantifier algorithm without normalization. In conclusion, the EpiDISH‐RPC algorithm without transformation generated the best deconvolution performance when applied to benchmarking using RRBS data and thus was used in our real RRBS data analysis.

Deconvolution of Immune Cells in 281 Inpatients

Using the Epidish‐RPC algorithm, we estimated the fractions of 13 immune cell types in 281 inpatients (Figure 2E). On average, nearly half of the cells were granulocytes, and 10.8% were monocytes. T central memory CD4, T effector memory CD4, B cells, and natural killer cells each constituted around 5% of the total. B‐memory, T CD8, T effector CD8, T effector memory CD8, and T CD3 cells were each less than 2%.

We conducted a 2‐sample t test to compare the differences in immune cell fraction between subjects with and without hypertension, with about half of the 281 inpatients having hypertension. Compared with subjects without hypertension, patients with hypertension had 1.5% higher fractions of B cells and T effector memory CD4 cells, reaching statistical significance (Table 3).

Table 3.

Estimated Cell Proportions in 281 Inpatients

Cell type (%), mean±SD	Total (N=281)	Hypertension (n=142)	Normotension (n=139)	P value
T CD8	1.2 (2.0)	1.0 (1.6)	1.4 (2.4)	0.131
B‐memory	1.3 (1.7)	1.2 (1.7)	1.5 (1.7)	0.084
T effector CD8	1.6 (2.3)	1.5 (2.0)	1.7 (2.6)	0.597
T effector memory CD8	1.8 (1.5)	1.8 (1.5)	1.7 (1.5)	0.555
T CD3	1.9 (2.6)	1.7 (2.6)	2.0 (2.6)	0.256
T CD4	3.8 (4.2)	3.6 (3.8)	4.0 (4.6)	0.449
T naïve CD8	3.9 (2.6)	3.9 (2.4)	4.0 (2.8)	0.689
Natural killer	4.7 (2.3)	4.7 (2.2)	4.8 (2.3)	0.878
B cells	5.0 (3.9)	5.8 (4.2)	4.3 (3.5)	0.001*
T effector memory CD4	6.4 (4.1)	6.9 (4.3)	5.8 (3.8)	0.017*
T central memory CD4	8.1 (4.9)	7.9 (5.2)	8.3 (4.6)	0.444
Monocytes	10.8 (3.5)	10.9 (3.3)	10.6 (3.6)	0.576
Granulocytes	49.6 (9.0)	49.2 (9.9)	50.0 (8.2)	0.470

^* P <0.05 was considered statistical significance.

BPs Related to Immune Cells

We further investigated the association between estimated immune cell fractions and BP phenotypes or patterns in 281 inpatients. Linear regression was used for continuous dependent BP phenotypes, including 24‐hour average, daytime, nighttime, and office BP phenotypes (SBP, DBP, MAP and PP). Additionally, logistic regression was used for binary dependent BP phenotypes and patterns, including hypertension, dipper, nondipper, and reverse dipper (Figure 4, Tables S2 and S3).

Figure 4. Immune cell types associated with BP phenotypes and patterns.

A, Results from regression models analyzing the association between immune cell abundance and BP phenotypes/patterns without covariate adjustment. B, Results from regression models analyzing the association between immune cell abundance and BP phenotypes/patterns with covariate adjustment. The numbers in the cells represent P values from the regression models, and color intensities indicate the magnitude of the regression coefficients. B_Mem indicates B memory; BMI, body mass index; BP, blood pressure; DBP, diastolic blood pressure; MAP, mean arterial pressure; NK, natural killer; PP, pulse pressure; SBP, systolic blood pressure; T_Eff_CD8, T effector CD8; T_EffMem_CD8, T effector memory CD8; T_EffMem_CD4, T effector memory CD4; and T_CenMem_CD4, T central memory CD4.

Using the univariate linear and logistic regression models, we found that office SBP, DBP, and MAP were negatively associated with the fraction of CD3 T cells. In contrast, 24‐hour DBP and MAP, daytime DBP and MAP, and nighttime DBP and MAP were positively associated with the fraction of CD8 effective memory T cells. Furthermore, nighttime SBP, DBP, and MAP were associated with the fraction of monocytes. Office SBP, MAP, and PP; 24‐hour SBP, MAP, and PP; daytime BP, and nighttime PP were positively associated with the fraction of B cells. Moreover, the fractions of CD4 effective memory T cells and B cells were significantly higher in hypertensives than normotensives (Figure 4A and Table S2).

After adjusting covariates of age, sex, and BMI, office SBP, DBP, and MAP and daytime SBP and PP remained negatively associated with CD3 T cells. Additionally, 24‐hour and daytime DBP and MAP and nighttime DBP were positively associated with CD8 effective memory T cells. Office, 24‐hour, daytime, and nighttime PP were negatively associated with natural killer cells. Nighttime SBP, DBP, and MAP were positively associated with monocytes. Office, 24‐hour, and daytime SBP and PP were positively associated with B cells. Furthermore, the fractions of CD4 effective memory T cells and B cells were significantly higher in hypertensives than normotensives (Figure 4B and Table S3).

Validation of BPs Related to Immune Cells

To validate the association between immune cells and BP phenotypes, we analyzed 1017 intensive care unit inpatients from a publicly available database MIMIC‐IV. In this validation cohort, the median age was 62, 52.4% were female, 14.3% were classified as dipper, 36.5% as nondipper, and 47.1% as reverse dipper (Table 2).

After adjusting covariates of age, sex, and BMI, we found that nighttime SBP, DBP, and MAP were positively associated with monocyte fractions, reaching statistical significance. Although lower monocyte fractions were observed in inpatients with dipper BP pattern in both discovery and validation cohorts, the P value in validation cohort did not reach statistical significance (Table S4).

BPs Related to CTS‐MRs

To identify CTS‐MRs associated with office and 24‐hour BP phenotypes and patterns, we applied linear regression models incorporating the interaction between BP phenotype and estimated cell fraction, with or without covariates adjustments (age, sex, and BMI), for each methylation region. A substantial number of CTS‐MRs were found to be associated with office and 24‐hour BP phenotypes and patterns across 13 immune cell subtypes (Figure 5A and 5B).

Figure 5. Number of CTS‐MRs associated with BP phenotypes and patterns.

A, Number of CTS‐MRs identified to be associated with BP phenotypes/patterns without covariate adjustment. B, Number of CTS‐MRs identified to be associated with BP phenotypes/patterns with covariate adjustment. The color intensities indicate the number of CTS‐MRs identified. C, Distribution of CTS‐MRs from different number of cell types. D, Number of overlapped CTS‐MRs identified to be associated with office, 24‐h, daytime, and nighttime BPs. B_Mem indicates B memory; BP, blood pressure; CTS‐MR, cell type‐specific DNA methylation regions; DBP, diastolic blood pressure; MAP, mean arterial pressure; NK, natural killer; PP, pulse pressure; SBP, systolic blood pressure; T_Eff_CD8, T effector CD8; T_EffMem_CD8, T effector memory CD8; T_EffMem_CD4, T effector memory CD4; and T_CenMem_CD4, T central memory CD4.

After adjusting for covariates, 692 CTS‐MRs (Tables S5 and S6) were identified as being associated with various BP phenotypes and patterns across various immune cells. Among these, 76 CTS‐MRs were linked to office BPs, significantly fewer than those associated with 24‐hour BPs. Rather than uniformly distributed across the 13 immune cell types, CTS‐MRs were primarily derived from effector CD8 T cells (32.2%), CD4 T cells (28.2%), and monocytes (17.3%). Approximately 5% of the CTS‐MRs were associated with BPs from CD8 T cells, effector memory CD8 T cells, and memory B cells. Less than 2% of CTS‐MRs were associated with BPs from CD3 T cells, central memory CD4 T cells, effector memory CD4 T cells, naïve CD8 T cells, natural killer cells, granulocytes, and B cells (Tables S7 and S8).

To further examine the immune cell specificity of CTS‐MR, we analyzed whether a given CTS‐MR was identified from multiple immune cell types. Out of 304 unique methylation regions, 84.5% were identified from a single immune cell type, 13.8% were in 2 immune cell types, and the remaining regions were found across 3 immune cell types. Among CTS‐MRs from a single cell immune cell type, effector CD8 T cells, CD4 T cells, CD8 T cells, and monocytes are the most common sources. For CTS‐MRs identified in multiple cell types, CD4 T cells and monocytes are predominant, rather than effector CD8 T cells (Figure 5C).

We also investigated whether a CTS‐MR was associated with multiple BP phenotypes and patterns. Among 304 unique regions, almost half were associated with multiple phenotypes (ranging from 2 to 12). Of the 132 CTS‐MRs associated 24‐hour BPs, approximately 90% were also linked to other BP phenotypes including daytime, nighttime, and office BPs. However, approximately 58.5% CTS‐MRs associated with office BPs were not validated through associations with 24‐hour BPs, underscoring the relevance of CTS‐MRs related to 24‐hour BPs (Figure 5D).

To explore potential overlap with previously identified methylation regions, we compared the 692 CTS‐MRs with regions identified in our previous study, where dozens of methylation regions were linked to 24‐hour BP phenotypes from the bulk DNA methylation profiles. After adjusting for covariates (age, sex, and BMI), none of the 692 CTS‐MRs overlapped with these previously identified regions. However, 2 genes among the 692 CTS‐MRs, POU3F1 and PRDM16, were associated with BP phenotypes in the absence of covariate adjustments (Table 4). Specifically, POU3F1 was identified as being negatively associated with daytime SBP in monocytes after covariates adjustment, consistent with bulk DNA methylation profiles without covariates adjustments. Conversely, PRDM16 was identified as being positively associated with 24‐hour PP in CD4 T cells after covariates adjustments while showing a negative association with 24‐hour PP in bulk DNA methylation profiles without covariates adjustments. Overall, these findings underscore the importance of accounting for cell heterogeneity in identifying DNA methylation regions associated with BP phenotypes and patterns.

Table 4.

Genes Identified From CTS‐MRs and Bulk DNA Methylation Profiles

Gene	BPs	Cell type	Chromosome	Start	End	α	P value	FDR
POU3F1	Systolic blood pressure day	Monocytes	1	38 513 201	38 513 300	−0.004	3.06E‐05	0.0451
		Bulk 12	1	38 513 269	38 513 669	−19.159	4.97E‐05	0.0499
PRDM16	Pulse pressure 24h	T CD4	1	2 990 101	2 990 200	0.014	5.68E‐05	0.0163
		Bulk 12	1	3 102 619	3 102 727	−9.050	2.76E‐05	0.0479

BPs indicates blood pressures; CTS‐MR, cell type‐specific DNA methylation regions; FDR, false discovery rate; PP, and SBP.

DISCUSSION

In this study, we uncovered extensive immune cell type heterogeneity in DNA methylation, revealed the relevant immune cell subtypes in BP regulation and hypertension, and identified CTS‐MRs associated with 24‐hour BP phenotypes and patterns. To our knowledge, this is the first investigation revealing the BP‐related DNA methylation regions as epigenetic markers at the cell type‐specific level. These findings support the involvement of epigenetics‐regulated immune dysregulation in BP variability and highlight potential biomarkers for future hypertension management strategies.

We adopted a comprehensive approach to assess BP phenotypes, extending beyond traditional office BP measurements to include 24‐hour BP phenotypes and patterns in this study. Notably, the majority of the identified CTS‐MRs were associated with 24‐hour BP phenotypes, emphasizing the value of continuous BP monitoring. This methodology captures critical BP variability and nocturnal dipping, enhancing our understanding of masked and nocturnal hypertension. The positive association between monocyte fractions and nighttime BP observed in both our discovery and validation cohorts aligns with earlier evidence implicating immune cells in hypertension. ⁶ Importantly, our study provides more robust evidence by using continuous BP monitoring rather than relying solely on office‐based measurements.

To derive reliable estimation of immune cell fractions, we systematically evaluated the performance of 19 cell type deconvolution algorithms using RRBS‐derived DNA methylation profiles. Our findings demonstrate the feasibility of immune cell deconvolution from RRBS data, with the EpiDISH‐RPC algorithm achieving the highest accuracy, particularly in heterogeneous cell type mixtures. Additionally, we identified significant associations between immune cell fractions (eg, monocytes and effector CD8 T cells) and various BP phenotypes, offering new insights into the immune system's role in BP regulation. These associations remained robust after adjusting for covariates and were partially validated in the MIMIC‐IV cohort, further supporting the reliability of our immune cell fraction estimation and subsequent CTS‐MR identification.

The identification of hundreds of CTS‐MRs associated with office and 24‐hour BP phenotypes underscores the importance of conducting cell type‐specific analyses. Among these, 2 genes, POU3F1 and PRDM16 stand out. POU3F1, a transcription factor implicated in neural development, has been linked to BP regulation, potentially through its role in mediating inflammatory responses. ⁴⁷ , ⁴⁸ Notably, POU3F1 can bind to and regulate Nav2, a gene essential for cranial nerve development and BP regulation. ⁴⁹ , ⁵⁰ In addition, polymorphisms in the PRDM16 gene have been associated with essential hypertension in the Xinjiang Uygur population. ⁵¹ Prdm16 is crucial for arterial flow recovery in pathological condition, primarily by maintaining endothelial function. ⁵² Furthermore, mice with cardiac‐specific Prdm16 deficiency exhibit with hypotension and cardiac hypertrophy. ⁵³ These findings suggest a complex interplay between immune function and BP regulation, emphasizing the need for further exploration of these pathways.

The significant association of monocytes with nighttime BP phenotypes contributes to a growing body of literature on the role of innate immune cells in hypertension. ⁷ Monocytes are known to play a critical role in inflammatory processes that promote vascular dysfunction, potentially explaining their link with elevated nighttime BP. Notably, monocyte depletion in mice has been shown to attenuate BP increases in angiotensin II‐induced hypertension models, underscoring the importance of innate immune activation in vascular dysfunction. ⁵⁴ Our cohort study in humans further supports the hypothesis that inflammation and immune dysregulation contribute to nondipping and reverse‐dipping BP patterns, both of these patterns are associated with worse cardiovascular outcomes. ²⁷ Additionally, the elevated fractions of B cells and CD4 effector memory T cells in patients with hypertension compared with individuals without hypertension align with previous studies implicating adaptive immune responses in hypertension. ⁹ , ¹¹ B cells play a role in antibody‐mediated inflammatory responses, ⁵⁵ and their increased presence in individuals with hypertension may reflect heightened immune activation. These findings suggest that targeting specific immune pathways may have therapeutic potential for managing hypertension.

Notably, our CTS‐MR analysis revealed that 32.2% of the identified CTS‐MRs linked to BP‐related phenotypes are derived from effector CD8 T cells, underscoring their critical role in BP regulation. Effector CD8 T cells are key components of the adaptive immune response, known for their ability to directly target and eliminate infected or transformed cells. ⁵⁶ , ⁵⁷ Recent studies suggest that they also modulate cardiovascular function, particularly through cytokine secretion that influences vascular tone and inflammation. ⁵⁸ , ⁵⁹ The association between effector CD8 T cells and hypertension may reflect their involvement in inflammatory processes contributing to vascular dysfunction and altered BP regulation. Furthermore, the significant presence of CTS‐MRs derived from effector CD8 T cells points to potential epigenetic mechanisms through which these immune cells could affect BP. Understanding the specific DNA methylation patterns associated with effector CD8 T cells may unveil novel pathways linking immune function and hypertension, shedding light on the complex interplay between the immune system and cardiovascular health.

However, several limitations should be noted. First, due to the limited availability of single cell DNA methylation samples of immune cell types, we presented data from only 13 immune cell types, all derived from healthy tissues. These reference profiles may not fully capture the heterogeneity of cell states present in clinical or disease‐specific context, potentially introducing bias into estimated cell‐type proportions. With the rapid development of single cell DNA methylation techniques, future studies will be able to incorporate a broader range of immune cell types under diverse clinical conditions to better capture phenotypic variation in complex diseases. Second, our cohort consisted exclusively of Black patients. Although this enhances the representation of an understudied population, it may limit the generalizability of our findings to other racial or ethnic groups and outpatient populations. Third, our validation cohort from the MIMIC‐IV database included only monocyte fractions, limiting the ability to fully validate findings related to other immune cell types. Future studies with larger and more diverse cohorts, expanded clinical variables, a broader spectrum of immune cell types under various clinical settings, and ambulatory BP monitoring will be important for further validation and generalization of our findings. Fourth, this is an association study. Causal inference to investigate the role of immune cells or methylation in BP regulation still needs further investigation. The cross‐sectional and observational nature of this study precludes causal inference regarding the role of immune cells or DNA methylation in BP regulation. Longitudinal and mechanistic studies will be needed to clarify these causal relationships. Finally, the lack of integrated transcriptomic profiles alongside the methylation data constrained our ability to identify potential therapeutic targets, as it limited our understanding of the direct functional relationships between epigenetic modifications and their downstream molecular effects. Future studies incorporating transcriptomic profiles will be essential to build upon the methylation findings and provide a more comprehensive understanding of these mechanisms. Despite these limitations, our analytical pipeline remains broadly applicable for identifying robust immune cell types and CTS‐MRs associated with complex disease phenotypes in diverse populations.

Our study highlights the potential of using DNA methylation‐based deconvolution to assess immune cell compositions and their associations with BP phenotypes at cell type‐specific levels. The identification of monocytes and B cells as correlates of nighttime BP opens new avenues for personalized hypertension management, particularly in addressing nondipping and reverse‐dipping patterns. Future research should focus on validating these associations in larger and more diverse cohorts and exploring the mechanisms by which immune cells influence BP regulation. Additionally, enhancing the accuracy and scalability of deconvolution algorithms, particularly for use with RRBS data, could improve the applicability of these methods in clinical and research settings. Integrating multiomics approaches—combining methylation data with transcriptomic or proteomic profiles—may also yield a more comprehensive understanding of the immune mechanisms underlying hypertension.

Conclusions

In summary, this study provides a thorough evaluation of cell type deconvolution algorithms using RRBS data and demonstrates significant associations between immune cell composition and 24‐hour BP phenotypes. We revealed extensive cell‐type heterogeneity in DNA methylation and identified CTS‐MRs associated with 24‐hour BP phenotypes and patterns at the cell type‐specific level. Notably, approximately one third of CTS‐MRs were present in effector CD8 T cells. This suggests that effector CD8 T cells are important in BP regulation, most likely through epigenetic regulation. These findings support the involvement of epigenetically regulated immune dysregulation in the molecular mechanisms of blood pressure variability and highlight potential biomarkers for future hypertension management strategies.

Sources of Funding

This work has been supported by the Natural Science Foundation of Shanghai (24ZR1456400) and the National Institutes of Health (HL149620).

Disclosures

None.

Supporting information

Tables S1–S8