Multimodal peripheral blood immunophenotyping and single cell epigenetic analysis of knee osteoarthritis reveals expansion and epigenetic dysregulation of regulatory T-cell and nonclassical monocyte populations

Abstract

Objective:

Osteoarthritis (OA) involves systemic inflammation, yet peripheral blood immunophenotypes and their underlying epigenetic landscapes remain poorly defined. We characterized cellular and chromatin accessibility profiles in knee OA and radiographic progressors (RP).

Methods:

We performed multimodal immunophenotyping using mass cytometry (CyTOF) on peripheral blood from 21 knee OA patients and 11 healthy controls (HC). To confirm findings, we developed DNA methylation-based imputation models and applied them to two independent validation cohorts (n=723). Finally single-cell ATAC-seq (scATAC-seq) was performed to interrogate chromatin accessibility landscapes in a patient subset.

Results

CyTOF identified significant expansions of CD8+ central memory T-cells, CD4+ regulatory T-cells (T-regs), and nonclassical monocytes in OA vs. HC (all P≤0.05). These expansions were robustly confirmed in the large-scale validation cohorts (all P<0.0001). Conversely, RPs displayed validated reductions in circulating CD4+ and CD8+ central memory pools. scATAC-seq revealed extensive epigenetic remodeling, particularly within monocytes and Tregs. Paradoxically, unsupervised clustering indicated that epigenetically defined pro-inflammatory clusters were depleted in OA blood. Pathway analysis revealed that these depleted clusters possess activated, migratory phenotypes, whereas the expanded circulating cells among OA patients display quiescent, non-migratory epigenetic signatures.

Conclusions:

Knee OA is characterized by validated systemic expansions of specific T-reg and monocyte subsets. Single-cell epigenetic profiling suggests epigenetically distinct subpopulations driven by age, OA, comorbid conditions, or migration from circulation to the periphery.

Introduction:

Osteoarthritis (OA) is a chronic, debilitating musculoskeletal disease characterized by progressive loss of function of joints. One of the most common age-associated diseases, it is also the leading cause of chronic disability in the US 1 and is expected to affect more than one billion individuals worldwide by 2050 ². Furthermore, the rate of severe OA is increasing beyond what would be expected from societal increases in obesity and aging 3. Despite its importance, there remain no anti-osteoarthritic drugs available for OA therapy, in stark contrast to the multitude of disease-modifying therapies available for autoimmune arthritis.

It is now well understood that chronic inflammation plays a key role in the development and progression of OA. Several inflammatory processes have been implicated, including innate immune activation, macrophage inflammatory responses, toll-like receptor (TLR) activation, etc. However, most OA inflammation research, including recent deep immunophenotyping analyses ⁴, have focused on the joint itself, rather than the systemic inflammatory context, with the exception of a recent study which identified an expansion of CD25^hi switched memory B cells in OA ⁵.

Indeed, previous efforts aimed at deep immunophenotyping of immune cell populations in peripheral blood specimens have been limited by the large sample volume required for multiplex flow cytometry. The recent development of mass cytometry (cytometric time of flight, CyTOF) has allowed for the detection of many parameters simultaneously, opening the door to deeper immunophenotyping from a relatively small sample volume. In the present study, we applied this technique to peripheral blood specimens in knee OA and OA progressor phenotypes. To confirm our findings, we then went on to develop a set of peripheral blood immunophenotyping imputation models based on genome-wide DNA methylation arrays in these same patients and applied these models to a large set of data from two independent cohorts. Finally, we took a deeper look at epigenetic regulation within cell subsets using single-cell chromatin accessibility analysis (scATAC-Seq) to gain novel insights into epigenetic dysregulation of individual peripheral blood cellular subtypes in knee OA.

Methods:

Power calculations

We used the scPower package ⁶ to estimate group sizes under the following conditions: “blood tissue” type, min cell frequency 0.01, sample size ratio 1, reference study Blueprint CLL. A group size of 15 was estimated give a detection power of 0.936 to detect differentially expressed/eQTL genes (used as a surrogate for our CyTOF analysis) with a target cell number (passing QC filters) per sample of ~8000. To confirm the likely number of cells required to hit our target, we performed a second calculation using SCOPIT v1.1.4 ⁷, a tool for estimating the number of cells to be sequenced to observe cell types in a single-cell sequencing experiment. Using CD4+ regulatory T cells as an example, at a sequencing depth of 200 cells per cell subtype, a population frequency of 5% and a detection probability of 0.90, a minimum of 4300 cells passing QC was estimated to be required per sample.

Blood sample sources

Peripheral blood samples for mass cytometry (CyTOF) immunophenotyping were obtained from our longitudinal Systematic Oklahoma Osteoarthritis iNflammation and Epigenetics Research (SOONER) cohort and Oklahoma Immune Cohort (OIC), both housed at the Oklahoma Medical Research Foundation (OMRF). All participants provided written informed consent; the study was approved by the IRB at OMRF (IRB# 23–10). Knee OA patients were recruited with a preexisting physician diagnosis of knee OA and/or with a self-report of chronic knee pain on most days for ≥6 months. Patients returned to the clinic every 6 months for at least 2 years for radiographic, clinical, and biospecimen data collection. Demographic information, medication indices, and OA pain scores (WOMAC) were collected. Semiquantitative Kellgren-Lawrence radiographic grades (KLG) and quantitative minimum medial joint space width (mmJSW) were calculated using ImageBiopsyLab’s knee osteoarthritis labeling assistant (KOALA) tool ⁸. Of note, radiographic progressors and pain progressors were mostly distinct groups: two of the PPs were also RPs, whereas the remaining 6 PPs were non-radiographic progressors. OA patient progression definitions are presented in the Supplementary Methods.

CyTOF sample processing and analysis

Previously fresh-frozen leukocytes were thawed, counted, washed, resuspended in cell staining buffer, Fc-blocked (Human Trustain FcX, Biolegend), then fixed in 1.6% formaldehyde for 10 minutes. Cells were intercalated then added to Maxpar Direct Immune Profiling Assay (MDIPA) tubes. The following day, cells were profiled using a Standard BioTools Helios system. A multistep workflow for CyTOF data cleaning and analysis was constructed within the OMIQ software environment (Dotmatics). First, each feature underwent inverse hyperbolic sine scaling (arcsinh) with a cofactor value of 5. Next, outliers were removed using flowCut, using default settings, and a QC gating strategy was implemented using parameters recommended by BioTools, including multiple cleanup steps (residual, center, offset, width, event length, alive/dead, and two intercalating DNA stains). All samples were subsampled to 10,000 events and t-SNE parameter selection was performed using opt-SNE ⁹. Canonical cellular composition group assignments were made using manual gates as recommended by BioTools. Group differences in cellular composition were calculated using a generalized linear model-likelihood ratio test as implemented using the edgeR package within OMIQ and P-values calculated. Given the limited sample number and focus on discovery of our CyTOF analysis, we defined group statistical significance as P≤0.05.

Methylation confirmation cohort

Methylation data for our confirmation studies were obtained from the Johnston County Osteoarthritis Project (JoCoOA) and the OABC ¹⁰, a subset of the Osteoarthritis Initiative (OAI). and Samples included healthy controls without a history of knee OA (n=237), established knee OA patients who would go on to experience radiographic progression within the subsequent 24–96 months (n=84 from JoCoOA and n=81 from OABC, 165 total), established knee OA patients who would go on to experience pain progression within the subsequent 24–96 months (n=77, all from OABC), and established knee OA patients with no evidence for future pain or radiographic progression (n=56 from JoCoOA and n=189 from OABC, n=245 total). See supplementary methods section for details of each cohort.

Peripheral blood cellular composition models based on DNA methylation microarrays

Peripheral blood was then used to develop whole-genome DNA methylation-based immunophenotyping models. Specifically, 500ng of genomic DNA extracted from peripheral blood buffy coat was treated with sodium bisulfite (Zymo EZ DNA methylation kit), then loaded onto Illumina Infinium MethylationEPIC v2.0 beadchips and imaged at the Clinical Genomics Center at OMRF. Statistical analysis and imputation model details in the supplementary methods section.

Single-cell ATAC-Seq analysis

Frozen PBMCs from 12 OA patients (6 RP and 6NP) and 6 HC were obtained from the SOONER and OIC as above and used for single cell ATAC-Seq analysis. Initial nuclei extraction and sample processing were done according to the 10X Genomics protocol for primary cells. Nuclei were resuspended at 3500/μl and submitted to the Clinical Genomics Center at the OMRF, where chromatin was extracted, treated with Tn5 transposase, and single cell ATAC-Seq libraries prepared using the Chromium Next GEM Single Cell ATAC Kit according to the manufacturer’s protocol. The quality of the libraries was assessed by a DNA-based fluorometric assay and automated capillary electrophoresis (Agilent Technologies, Inc.). Libraries were pooled and sequenced on a NovaSeq X Plus instrument (Illumina) using a 50bp paired end read approach. Statistical analysis details in the supplementary methods section.

Genome Set Enrichment Analysis (GSEA)

Gene set enrichment analysis (GSEA) was performed using the Ingenuity Pathway Analysis (IPA) suite (Qiagen). Transcriptional accessibility from ATAC-Seq analyses was mapped to gene expression log ratio for putative epigenetic ‘activation’ of involved pathways. IPA also defines “upstream regulators”, which are any molecule (for example, a transcription factor, kinase, cytokine, drug, or small molecule) that is known from prior experimental studies to influence the expression or activity of other genes or proteins and is therefore inferred to lie causally upstream of the differentially expressed genes in the dataset. For analysis, only canonical pathways including 5 or more target genes were considered significant.

Results

Deep immunophenotyping of peripheral blood in OA patients reveals significant differences in cellular composition

Peripheral blood was drawn from symptomatic knee OA patients (OA, n=21) and healthy controls (HC, n=11). OA patients included 6 individuals who would progress radiographically within the subsequent 24 months (radiographic progressors, RP) and 15 radiographic nonprogressors (NP). OA patients were also divided by future pain progressors (PP, n=6) and non-pain progressors (NP, n=15). Patient age, sex, BMI, and race/ethnicity were similar between groups (Table 1). We then stained whole blood with a 30-marker standardized human antibody panel (Maxpar Direct Immune Profiling Assay, MDIPA, Fluidigm) and performed CyTOF. After data cleanup, we assigned cell subset data via traditional manual gating recommended by Fluidigm, defining n=33 cellular subsets. Individual patient characteristics including age, BMI, and baseline WOMAC pain did not correlate significantly with cellular subsets (Supplementary Table 1).

Table 1:

Baseline participant characteristics and imputed peripheral blood cell composition, discovery (CyTOF) cohort.

	Age (mean±SD), years	Sex (%female)	BMI (mean±SD), kg/m2	Race / Ethnicity
				Caucasian	African-American	Multi-racial	Hispanic/Latino	Baseline radiographic KLG	Baseline WOMAC pain
Discovery (CyTOF) cohort
Healthy controls (HC) (n=12)	52.8±9.7	64%	35±6	100%	0%	0	1 (9%)	n/a	n/a
All OA (n=21)	59.9±11	67%	30±7	95%	0%	0	1 (5%)	2±1	4.6±3.2
OA radiographic progressors (n=6)	59.2±10	17%	29±5	83%	0%	17%	0	2±2	4.2±2.9
OA non-radiographic progressors (n=15)	60.1±11	87%	31±8	100%	0%	0	0	2±2	4.7±3.3
OA pain progressors (n=6)	59.9±2.5	33%	31±2	100%	0%	0	1 (7%)	2±2	4.6±0.9
OA non-pain progressors (n=15)	59.8±6.0	73%	30±2	93%	0%	7%	0	2±1	4.5±1.3
Single cell ATAC-Seq cohort
HC (n=6)	50±13	17%	31±6	83%	0%	13%	14%	n/a	n/a
All OA (n=12)	53±13	59%	30±6	92%	0%	8%	8%	2±2	4.3±3.6
RP (n=6)	55±13	17%	29±5	83%	0%	17%	17%	2±1	4.3±2.8
Non-RP (n=6)	52±14	100%	31±6	100%	0%	0%	0%	2±1	4.5±4.5
Confirmation cohort (DNA methylation imputation)
HC n=234	61±9	68%	29±5	86%	13%	1%	0.4%	n/a	n/a
All OA n=489	62±9	64%	32±6	73%	25%	2%	0.4%	2±1	3.5±3.9
RP n=158	64±8	62%	33±7	71%	28%	1%	1%	2±1	4.4±4.4
PP n=91	59±9	63%	31±5	70%	30%	0%	1%	2±1	2.8±2.9
Non-RP/PP n=240	62±9	66%	31±6	76%	22%	2%	0%	2±1	3.2±3.8

Values presented in mean±SD. P values by Student t-test (continuous data, normally distributed), Chi square test + Fisher’s exact (categorical data). KLG = -Kellgren Lawrence grade, JSW = joint space width (mm), WOMAC = Western Ontario and McMaster Universities Osteoarthritis Index subscores.

Next, we identified six differences between OA and HC (Table 2, Figure 1A, 1D, 1E), including expansions of CD8+ central memory cells (OA:0.007±0.007 vs. HC:0.002±0.001, mean±SD fraction total cells, P=0.0001), CD4+ regulatory T cells (OA:0.01±0.01 vs. HC:0.005±0.002, P=0.0007), nonclassical monocytes (OA:0.008±0.005 vs. HC:0.004±0.003, P=0.007), basophils (OA:0.01±0.007 vs. HC:0.005±0.003, P=0.01), MAIT/NKT cells (OA:0.02±0.01 vs. HC:0.01±0.01, P=0.04), and plasmablasts (OA:0.0009±0.0008 vs. HC:0.0004±0.0005, P=0.04). We also identified a nearly-statistically-significant increase in eosinophils in OA (OA:0.001±0.002 vs. HC:0.0005±0.0005, P=0.053).

Table 2:

Immunophenotype differences between various OA groups and controls in CyTOF discovery and DNA methylation-imputed validation cohorts.

Cellular subset	OA/RP/PP frequency: CyTOF discovery cohort	HC/NP frequency: CyTOF discovery cohort	OA/RP/PP frequency: imputed validation cohort	HC/NP frequency: imputed validation cohort	OA/RP/PP: HC/NP ratio, discovery cohort	OA/RP/PP: HC/NP ratio, validation cohort	Validated difference
OA vs. HC
CD8+ central memory cells	0.007±0.007	0.002±0.001	0.01±0.004	0.004±0.004	3.5	2.5	Yes
CD4+ regulatory T cells	0.01±0.01	0.005±0.002	0.02±0.005	0.006±0.004	2.0	3.3	Yes
Nonclassical monocytes	0.008±0.005	0.004±0.003	0.01±0.005	0.004±0.0004	2.0	2.5	Yes
Basophils	0.01±0.007	0.005±0.003	0.02±0.004	0.006±0.004	2.0	3.3	Yes
Plasmablasts	0.0009±0.0008	0.0004±0.0005	0.007±0.005	0.002±0.004	2.3	3.5	Yes
Eosinophils	0.001±0.002	0.0005±0.0005	0.007±0.005	0.002±0.004	2.0	3.5	Yes
MAIT/NKTs	0.02±0.01	0.01±0.01	0.007±0.006	0.008±0.004	2.0	0.9	No
RP vs. NP
CD4+ central memory cells	0.05±0.01	0.08±0.03	0.03±0.008	0.04±0.007	0.6	0.8	Yes
CD8+ central memory cells	0.004±0.003	0.01±0.007	0.01±0.004	0.011±0.004	0.4	0.9	Yes
CD8+ terminal memory cells	0.008±0.008	0.02±0.02	0.005±0.007	0.005±0.006	0.4	1.0 (n/s)	No
PP vs. NP
CD4+ terminal effector cells	0.02±0.02	0.01±0.01	0.01±0.005	0.01±0.004	2.0	1.0 (n/s)	No
CD8+ terminal memory cells	0.02±0.03	0.01±0.01	0.006±0.008	0.005±0.006	2.0	1.2 (n/s)	No

HC = healthy control, RP = radiographic progressor, PP = pain progressor, NP = nonprogressor, n/s = not significant. Frequency represents mean±SD fraction total cells.

Figure 1:

CyTOF analysis. (A) Volcano plot of cellular subset differences among OA patients vs. healthy controls (HC). (B) Volcano plot of cellular subset differences within OA patient radiographic progressors (RP) vs. nonprogressors (NP). (C) Volcano plot of cellular subset differences within OA patient pain progressors (PP) vs. nonprogressors (NP). (D) Uniform manifold approximation and projection (UMAP) plot of all cellular populations using optsne algorithm. (E) UMAP of cellular populations different between OA and HC.

In radiographic progression we found three differences (Table 2, Figure 1B), including reductions in CD4+ central memory cells (RP:0.05±0.005 vs. NP:0.08±0.008, P=0.01), CD8+ central memory cells (RP:0.004±0.001 vs. NP:0.008±0.002, P=0.04), and CD8+ terminal memory cells (RP:0.008±0.003 vs. NP:0.02±0.005, P=0.04). Finally, we identified two differences among pain progressors (Table 2, Figure 1C): increases in CD4+ terminal effector cells (PP:0.02±0.007 vs. NP:0.01±0.002, P=0.03) and CD8+ terminal memory cells (OA:0.02±0.01 vs. NP:0.01±0.002, P=0.05).

DNA methylation-based immunophenotyping models as confirmation of peripheral blood cell subset differences in large, independent datasets

We then sought to validate these findings in an independent dataset. A large-scale validation using the same CyTOF method is not practical; however, previous studies have shown that DNA methylation-based deconvolution models can accurately predict cellular composition in peripheral blood ^11,12. Indeed, most DNA methylation analysis packages (e.g. minfi) include an estimateCellCounts function which allows for deconvolution of a small number of general cell types. Thus, we next generated genome-wide DNA methylation microarray data on our CyTOF samples and employed a cross-validated, parsimonious, elastic net-regularized (using the Lasso penalty) generalized linear model approach to produce predictive models ¹³. Then, we applied these models to our previous large set of peripheral blood DNA methylation data from OA patients and controls from the Osteoarthritis Initiative (OAI) and Johnston County Osteoarthritis Project (JoCoOA) ^14,15, which consisted of 723 individuals: 489 knee OA patients (158 RP, 91 PP, and 240 NP) and 234 HC.

Using these imputed cellular composition data (Table 2, Figure 2A), we confirmed differences in CD8+ central memory cells (OA:0.011±0.004 vs. HC:0.0037±0.004, mean±SD imputed fraction total cells, P<0.0001), CD4+ regulatory T cells (OA:0.015±0.005 vs. HC:0.0055±0.004, P<0.0001), nonclassical monocytes (OA:0.010±0.005 vs. HC:0.0044±0.0004, P<0.0001), plasmablasts (OA:0.0071±0.005 vs. HC:0.002±0.004, P<0.0001), and basophils (OA:0.016±0.004 vs. HC:0.0062±0.004, P<0.0001). Furthermore, we found a significant increase in eosinophils in our imputation cohort (OA:0.0072±0.005 vs. HC:0.0017±0.004, P<0.0001). We did not confirm differences in MAIT/NKT cells (OA:0.0071±0.006 vs. HC:0.0077±0.004, P=0.006).

Figure 2:

Comparison of CyTOF cellular subset findings and imputed cellular compositions using peripheral blood DNA methylation-derived machine learning models in a confirmation cohort (OAI + JoCoOA). (A) OA patients vs. HC. (B) RP vs. NP, (C) PP vs. NP.

We also confirmed two of the three cell subset differences in RP vs. NP from our discovery cohort (Table 2, Figure 2B): CD4+ central memory cells (RP:0.031±0.008 vs. NP:0.036±0.007, P<0.0001) and CD8+ central memory cells (RP:0.010±0.004 vs. NP:0.011±0.004, P=<0.0001). We did not confirm either of the two cell subset differences in PP vs. NP from our discovery (Table 2, Figure 2C), as neither were significantly different in our validation data.

Epigenomic differences in peripheral blood cell subsets revealed by single-cell transcriptional accessibility assays (ATAC-Seq)

Finally, we sought to determine whether epigenetic patterns within these cellular subsets were altered. To that end, we performed single-cell ATAC-Seq, which identifies regions within the genome of differential epigenetic/chromatin accessibility, on a subset of patients including 12 OA (6 NP and 6 RP) and 6 HC. Defining cellular subtypes in ATAC-Seq is challenging given a paucity of deconvolution reference data; thus, we analyzed the cell clusters from available ATAC-Seq reference datasets that most closely aligned with those we identified and validated in the above analyses. First, we identified and annotated regions of differential chromatin accessibility within each cell population in OA patients vs. HC (Table 3). We found the most differential chromatin accessibility among monocytes (n=1805 differentially accessible peaks at P≤0.01, n=346 at q≤0.05, Figure 3A) with fewer differentially accessible peaks among CD4+ T-regs (n=1174 peaks at P≤0.01, 162 at q≤0.05, Figure 3C). We identified few differentially accessible chromatin peaks in the remaining cell types (Table 3). Comparing radiographic progressors vs. nonprogressors, we identified few epigenetic differences (CD4+ central memory cells: n=387 at P≤0.01, n=6 at q≤0.05; CD8+ central memory cells: n=222 at P≤0.01, 0 at q≤0.05; CD8+ terminal memory cells: n=188 at P≤0.01, n=5 at q≤0.05).

Table 3:

Epigenetic transcriptional accessibility (ATAC-Seq) differences between various OA and control groups

Cellular subset	#ATAC peaks different at P≤0.01	#ATAC peaks different at q≤0.05	#ATAC peaks more accessible in OA	Top 5 genes more accessible in OA	#ATAC peaks less accessible in OA	Top 5 genes less accessible in OA
OA vs. HC
CD8+ memory cells	516	0	n/a	n/a	n/a	n/a
CD4+ regulatory T cells	1174	162	6	ASMT, CCND3, PRAG1, CROCC, PRKX	156	LMNL, CHAF1B, FSIP1, GALNT1, KLF6
Monocytes	1805	346	272	CNYP1, GAB2, ADAMTS4, MDM1, MYH9	74	CHAF1B, BCAS1, IFNGR2, LDHA, SCD
Basophils	50	0	n/a	n/a	n/a	n/a
Plasmablasts	329	3	0	n/a	3	LMNL, AC067752.1, SORL1
Eosinophils	551	0	n/a	n/a	n/a	n/a
MAIT/NKTs	736	3	0	n/a	3	CHAF1B, FGD1, TBKBP1
RP vs. NP
CD4+ central memory cells	387	6	4	PRKY, TXLNGY, EIF1AY, DDX3Y	2	RP11, ZRSR2
CD8+ central memory cells	222	0	n/a	n/a	n/a	n/a
CD8+ terminal memory cells	188	5	5	PRKY, EIF1AY, DDX3Y, TXLNGY, UTY	0	n/a

Figure 3:

ATAC analysis. (A) Top genes differentially accessible in all monocytes among OA patients vs. HC. (B) UMAP of unsupervised clustering of monocytes via chromatin accessibility; clusters with asterisk (*) significantly different between OA and HC. (C) Top genes differentially accessible in CD4+ regulatory T cells among OA patients vs. HC. (D) UMAP of unsupervised clustering of T-regs via chromatin accessibility; clusters with asterisk (*) significantly different between OA and HC.

Gene set enrichment analysis (GSEA) of ATAC-Seq data in monocytes

We then performed GSEA on the monocyte cell population identified above and found 209 pathways significantly enriched among differentially accessible genes with 152 predicted inhibited, 22 activated, and 35 of indeterminate status (Supplementary Table 3). The top inhibited pathways included PI3K/AKT signaling (P=5E-10, activation Z-score −1.1), Sertoli cell-germ cell signaling pathways (P=3E-8, Z=−0.5), TGF-β signaling (P=3E-7, Z=−1.3), integrin signaling (P=3E-7, Z=−1.9), and PTEN signaling (P=3E-7, Z=+0.30). Ingenuity Pathway Analysis (IPA), the software package we used for GSEA, also evaluates upstream regulators, which are defined as molecules (for example, a transcription factor, kinase, cytokine, drug, or small molecule) that are known from prior experimental studies to influence the expression or activity of other genes or proteins. Key upstream regulators among differentially accessible genes in monocytes included lipopolysaccharide (P=3E-15, Z=−3.8), TNF (P=5E-15, Z=−3.7), IL-4 (P=1E-13, Z=−3.2), tretinoin (P=6E-13, Z=−2.4), and IL-1B (P=5E-11, Z=−3.1).

Unsupervised clustering of monocyte ATAC-Seq data

Given the substantial number of ATAC-Seq (or chromatin accessibility) differences in monocytes, we next wondered whether these differential epigenetic patterns might reflect differences in monocyte (and, later, T-regs) subclusters, as defined by epigenetic accessibility, as previous studies have shown high-resolution subclustering is possible leveraging epigenetic patterns, particularly for T cell subclasses ¹⁶. Thus, we then performed unsupervised clustering of these data, optimizing the number of clusters using the Leiden algorithm across a resolution grid and summarized at a representative/optimized resolution. Five clusters were identified, with HC cells enriched in one cluster and OA cells enriched in two clusters (cluster 3: 95% HC vs. 5% OA, q=9E-130, cluster 2: 81% OA vs. 19% HC, q=2E-33, cluster 4: 75% OA vs. 25% HC, q=1E-8, Figure 3B), the remaining two clusters not being significantly different. We then found 146 genes within regions of differential chromatin accessibility that characterized clusters enriched in OA (2, 3) vs. the cluster enriched in HC (4). All had reduced accessibility among OA clusters. Pathway analysis of these genes revealed similar results compared to the previous all-monocyte analysis.

GSEA of CD4+ regulatory T cell ATAC-Seq data

In the T-reg population, we identified 23 pathways as significantly enriched, all estimated to be activated (Supplementary Table 4). The top non-oncologic pathways included HIF1a signaling (P=6E-9, Z=+0.04), RAR activation (P=8E-5, Z=+0.04), HMGB1 signaling (P=2E-4, Z=+0.03), and germ cell-sertoli cell junction signaling (P=2E-4, Z=0.03). Upstream regulators were similar to monocytes and included TGFB1 (P=9E-13, Z=−4.0), IL-4 (P=8E-11, Z=−2.2), IL-2 (P=1E-10, Z=−2.6), LMO2 (P=1E-10, Z=−0.26), TNF (P=2E-10, Z=−3.9), and lipopolysaccharide (P=5E-10, Z=−4.3).

Additionally, we analyzed differential chromatin accessibility within loci governing migration ¹⁷ and exhaustion ¹⁸ among T-regs. Specifically, we evaluated classical markers of T-cell exhaustion (PDCD1 [PD-1], CTLA4, TIGIT, HAVCR2 [TIM-3], LAG3, TOX, CD244, PRDM1 (Blimp-1), and IL2RA. We also evaluated markers of cellular migration; specifically, genes critical for cell motility and adhesion, including MMP9, RAC1, CDC42, VIM (Vimentin), CSCR4, CCr7, ITGB1 (Integrain Beta 1), ITGA4, ITGAL, IRGAM, and SELPLG. Migration markers including CXCR4, CX3CR1, and ITGB1 and markers of activation and metabolism (IL2RA, encoding CD25, and LDHA), and exhaustion markers TIGIT and TGFB1 all demonstrated reduced accessibility in OA, suggesting reduced capacity of these cells for migration and reduced likelihood of an exhaustion phenotype.

Unsupervised clustering of CD4+ regulatory T cell ATAC-Seq data

As in the monocyte population, we also performed unsupervised clustering of T-reg ATAC-Seq data, although many more clusters were identified. Among 13 total clusters, 12 were significantly different between OA patients and HC (Figure 3D) (cluster 7: 98% HC vs. 2% OA, q=2E-190, cluster 8: 99% HC vs. 1% OA, q=2E-162, cluster 11: 99% HC vs. 1% OA, q=7E-90, cluster 12: 97% HC vs. 3% OA, q=6E-74, cluster 2: 80% OA vs. 20% HC, q=1E-66, cluster 1: 77% OA vs. 23% HC, q=8E-52, cluster 4: 80% OA vs. 20% HC, q=4E-31, cluster 6 78% OA vs. 22% HC, q=2E-15, cluster 10: 79% OA vs. 21% HC, q=5E-14, cluster 9: 78% OA vs. 22% HC, q=6E-13, cluster 3: 51% HC vs. 49% OA, q=2.9E-12, cluster 13: 56% HC, 44% OA, q=2E-3. Notably, most clusters were lost in OA patients vs. HC. As in the monocyte population, pathway analysis of OA-associated vs. HC-associated clusters was similar to the analysis of the total monocyte population (Supplementary Table 4).

Finally, given the above GSEA results indicating a reduced (negative Z score) for many inflammatory pathways, we then performed an analysis specifically of the clusters that were absent among OA patients (specifically, clusters 7, 8, and 11). GSEA of these clusters showed pathway activation in IL8 signaling, Th1/Th2 activation, neutrophil degranulation, with enriched upstream regulators including lipopolysaccharide, IL4, TNF, immunoglobulin, IL2, and interferon-gamma, all suggesting that these clusters, absent in OA patients, were indeed pro-inflammatory cellular subsets (Supplementary Table 5).

Discussion

In the present study, we demonstrate that baseline peripheral blood immune cell subsets are altered in patients with primary knee OA compared to healthy controls and among OA patient phenotypes, particularly future radiographic progressors. We went on to confirm our findings by generating novel peripheral blood DNA methylation array-based imputation models and applying these models to previously generated peripheral blood DNA methylation data from independent (OAI and JoCoOA) cohorts. We validated nearly all the differences in OA vs. HC and validated two of three cell subset differences in RP vs. NP, although we did not validate either of the two cell subset differences in PP vs. NP. Finally, we delved deeper into the epigenetic landscape of these immune cell subsets by performing single-cell transcriptional accessibility assays (ATAC-Seq) on a subset of OA and HC peripheral blood samples. Epigenetic differences were most pronounced in monocytes and T-regs with an anti-inflammatory signature noted in OA. We found few epigenetic differences in RP vs. NP, although these secondary analyses were likely underpowered.

Previously published data have been mixed regarding the T-reg population in OA; a 2016 study showed a ~30% increase in T-regs in OA patients vs. HC ¹⁹, although these cells were characterized by reduced IL10 secretion and were thus likely dysfunctional. However, a 2015 study found a slight reduction in peripheral blood T-reg levels among OA patients ²⁰. A 2025 deep immunophenotyping study of T cells specifically identified expansion and activation of Tregs and an overrepresentation of innate immune response genes in OA patients with high pain intensity ²¹, which our study was not designed to address. Our finding of increased T-reg frequency in peripheral blood both directly (via CyTOF) and indirectly on a large validation set (via epigenetic imputation) would suggest an increase in T-reg frequency across the disease spectrum. Furthermore, we identified widespread epigenetic differences among T-regs, including many imputed T-reg subpopulations, many of which were depleted in OA patients. As discussed above, the majority of ATAC-Seq differences in OA vs. HC were inflammatory in nature, although we did see evidence for decreased signaling of GADD45, a marker of chondrocyte senescence ²² and previously shown to be protective against OA ²³; indeed, the canonical senescence pathway was predicted to be inhibited in this population. We also saw evidence for HIF-1alpha pathway epigenetic inhibition and predicted inhibition of the retinoid X receptor (RXR) activation pathway, both previously described in OA ^24,25.

Our paradoxical finding of an expanded circulating T-reg pool alongside diminished subcluster diversity suggests a specific remodeling of the immune compartment in OA. One potential explanation is that the epigenetically pro-inflammatory cellular subsets that would ordinarily be present within peripheral blood have migrated out of circulation and into target joint tissues as part of the disease process, leaving epigenetically less inflammatory cells behind. This interpretation is supported by our data: GSEA of T-reg clusters absent in OA samples (7, 8, and 11) showed inflammatory pathway activation. Furthermore, we went on to confirm that cells remaining in peripheral blood among OA patients had a reduced migratory and exhaustion phenotype. Taken together, these findings suggest that the T-regs in the clusters that remain within peripheral blood of OA patients are quiescent, non-migratory, and non-exhausted; thus, the clusters that are absent in the periphery of OA patients had a phenotype of activation, migration, and potential exhaustion. Future studies should be conducted with paired samples from both peripheral blood and synovium to confirm this migration hypothesis.

We also identified and validated an increase in nonclassical monocytes in OA patients. A number of publications have highlighted the key role these cells play in OA progression; indeed, monocytes are the most frequent cell type found in OA synovitis ²⁶. A recent key publication by Philpott et al. performed a single cell transcriptomic analysis of OA synovium and demonstrated an increase in intermediate monocytes within the synovial lining and found these to be correlated with OA pain ⁴. Other studies have found that peripheral blood monocytes display an activated phenotype in women with OA ²⁷, as well as demonstrate an increase in osteoclastogenic differentiation capacity ²⁸. Furthermore, soluble macrophage and monocyte biomarkers correlate positively with clinical knee pain severity, osteophyte progression, and physical disability among OA patients ^29,30. In our epigenetic analysis, we saw significant inhibition of the macrophage alternative activation signaling pathway, which would correlate with reduced production of M2 anti-inflammatory macrophages, consistent with what has been reported in OA ³¹. Also of note, we found the wound healing pathway to be epigenetically inhibited within monocytes. Nonclassical monocytes have been shown to be rapidly recruited to sites of skin injury, where they generate wound healing macrophages ³²; disruption of this trafficking and/or phenotype change via epigenetic inhibition may be important in OA progression, frequently described as a chronic wound ³³. Similar to T-regs above, GSEA of clusters absent in OA showed were characterized by increased accessibility of inflammatory markers: monocyte cluster 3, for example, demonstrated activation of Th1, Th2, viral entry, and Fc-gamma-mediated phagocytosis pathways, along with enrichment in TCR, immunoglobulin, and lipopolysaccharide-associated genes (Supplementary Table 6).

Comparing the GSEA analysis of OA-enriched clusters in monocytes and T-regs, we find an expected discordance in canonical pathways given their substantially different roles in the immune system, although some overlap exists (macrophage alternative activation signaling, S100 family, and serotonin receptor signaling are all inhibited in both, for example). Intriguingly, the upstream regulators enriched in transcriptionally accessible regions are far more concordant, with OA-associated clusters in both cell types showing predicted inhibition of inflammatory and growth-associated regulators including TGFB1, IL4, TNF, lipopolysaacharide, KLF6, IL1B, NFkB, PDGF, and predicted activation of the miRNA-regulation complex DICER1, miR-21 (previously associated with OA ³⁴ and specifically with macrophage phenotype in OA ³⁵), and the estrogen receptor transcription factor ESR1, among others (Supplementary Table 7). This suggests that a core set of shared epigenetic pathways may be affected independent of cell type during OA development.

Among other cell subsets, CD8+ central memory cells have not previously been directly evaluated in the context of OA, although studies have shown a positive correlation of pan-CD8+ T cells with OA and correlated with OA disease severity ^20,36. Another small study of 30 OA patients found no difference in overall CD8+ number in OA, although they did find an increase in CD8+CD45RA- T cells, which are likely indicative of memory CD8+ cells, agreeing with our findings ³⁷. The aforementioned Philpott study identified a trend towards a lower CD8+ memory T cell within the subintima and lower CD8+ memory T cells within microvessels which correlated with worse OA pain ⁴, although no disease-free tissue was used for comparison. We did not find any significant differences in transcriptional accessibility among CD8+ memory T cells in OA patients vs. HC, suggesting no significant epigenetic immunophenotype differences within this population.

We also found and validated an increase in plasmablasts in OA. Few studies have focused in the past on the B-cell/plasmablast/plasma cell axis; a 2023 study found circulating B cells in OA had a more mature phenotype and found that transcriptomic profiles of synovial B cells from OA patients resembled those of plasma cells ³⁸, agreeing with our findings of a shift towards maturation in this axis. Further, an important study published in 2025 by Sahu and colleagues presented an analysis like our own, employing CyTOF profiling of peripheral blood from OA patients, healthy controls, and degenerative meniscal tears with follow-up single cell RNA sequencing on a subset of samples. They identified depletion of naïve B cells in OA but an increase in a unique subset of B cells, called switched memory B cells ⁵. Although we did not identify a similar pattern in our samples, this study nonetheless highlights the potential importance of B cells and their terminal differentiated state (plasmablasts / plasma cells) in the pathogenesis of OA. Future studies should further evaluate this lineage which has thus far received little attention in the OA literature.

Our study has several strengths. Firstly, we performed one of the only comprehensive deep immunophenotyping investigations of peripheral blood in OA patients using standardized CyTOF, and ours is the first such study to validate our findings using an imputation from two independent OA cohorts. Our investigation of single cell peripheral blood epigenetics is novel and reflects immune cell priming and memory, offering a glimpse into long-term durable cellular programming compared to transcriptomic data ³⁹. Furthermore, our study has implications for future biomarker development. If specific differences in OA endotypes (e.g. future radiographic progressors) are confirmed in larger studies, these data would be invaluable in further classifying OA and enriching clinical trials in specific OA endotypes. However, our study does have several weaknesses. The most prominent of these is a lack of matched joint tissue which would have allowed us to validate our cell migration hypothesis, and a lack of our ability to test other hypotheses for the reduction in pro-inflammatory immune cells in circulation (e.g. OA+comorbid conditions). Buffy coat DNA samples used in this study for confirmation (from the OAI cohort) were not optimized for single cell analysis and low frequency cell populations may not be well represented; thus, our ability to impute these populations may be limited. The current study did not include end-stage OA patients where synovium would be readily accessible; rather, they represented moderate disease. Another weakness is our inability to evaluate single-cell epigenetic differences in PP vs. NP due to a lack of power; future studies should be conducted with appropriate power to detect changes in these additional patient phenotypes. Finally, our study only evaluated a single timepoint for a given patient (baseline); thus, we are unable to draw conclusions about the stability of our epigenetic findings, or of the overall cellular subpopulation differences that might precede OA during preclinical disease.

In summary, we defined and validated peripheral blood immune cell subset differences in knee OA patients compared to healthy controls and performed the first single cell epigenetic analysis of OA peripheral blood, which identified novel T-reg and monocyte cellular subsets associated with OA. Our study highlights key cellular subpopulations for future study and offers evidence that dysregulation of systemic immunity in OA is not only confined to the innate immune system but may also be present in certain lymphoid lineages (most prominently CD4+ T-regs) which deserve additional scrutiny. Finally, our ATAC-Seq data, along with our laboratory’s previous publications demonstrating differences in peripheral blood DNA methylation ^14,40 add to the growing evidence of systemic immuno-epigenetic dysregulation in OA.

Data availability statement:

The datasets and models generated during the current study are available upon reasonable request.