Cardiac sarcoidosis: new insights beyond the granuloma using spatial proteomics

Abstract

Background and Aims

Cardiac sarcoidosis (CS) is poorly understood, challenging to diagnose, and portends a poor prognosis. A lack of animal models necessitates the use of residual human samples to study sarcoidosis, which in turn necessitates the use of analytical tools compatible with archival, fixed tissue. This study employed high-plex spatial protein analysis to characterize the immunologic, fibrotic, and metabolic landscape of CS.

Methods

Cardiac tissues from 64 patients—39 with CS and 25 without CS (Controls)—were analysed using the GeoMx Digital Spatial Profiler. Regions of interest (ROIs) with and without granulomatous inflammation were selected to examine protein abundance. ROIs were further divided into areas of illumination to separately analyse distinct protein profiles in myocytes, stroma, and vessels. A 79-protein panel was used to characterize immune and parenchymal cell populations. Differential protein abundance was evaluated using mixed-effects modeling. A predictive model was developed using a subset of seven proteins to distinguish inflammation-free CS ROIs from control ROIs.

Results

Spatial profiling demonstrated marked heterogeneity within and between sarcoidosis granulomas. Inflammation-free ROIs from CS hearts exhibited a ‘distance gradient’, with protein abundance varying by proximity to granulomatous inflammation. The seven-protein predictive model achieved 95.9% accuracy and an area under the receiver operating characteristic curve of 0.993 in differentiating inflammation-free CS ROIs from Controls.

Conclusions

Spatial proteomics revealed novel features of the CS microenvironment and identified a distinct immune signature in tissue lacking overt inflammation. These findings may improve diagnostic accuracy in clinical biopsies from patients with suspected CS.

Structured Graphical Abstract

Structured Graphical Abstract.

GeoMx spatial proteomic profiling was performed on 64 human cardiac tissue samples-39 from cardiac sarcoidosis (CS) patients and 25 from non-sarcoidosis controls. After 4-plex immunofluorescence staining, sections were incubated with oligonucleotide-tagged antibodies and processed on the GeoMx Digital Spatial Profiler. Regions of interest (ROIs) were selected to capture both granuloma-containing tissue and a spectrum of non-granuloma parenchymal regions, including peri-granuloma, granuloma-remote, and vascular-adjacent parenchyma. Within each parenchymal ROI, digital pathology enabled compartment-specific segmentation into Myocyte, Stroma, and Vascular domains for high-resolution proteomic analysis. Comparisons were performed (1) across granuloma ROIs, (2) between peri-granuloma and granuloma-remote ROIs within CS hearts, and (3) between granuloma-remote CS ROIs and control ROIs. These analyses revealed granuloma marked heterogeneity, spatial immune-activation gradients in the cardiac parenchyma linked to granuloma proximity, and distinct sarcoidosis-associated protein signatures distributed throughout the myocardium which differentiate CS hearts from Control hearts.

See the editorial comment for this article 'Resolution of spatiotemporal building blocks in the diagnosis of cardiac sarcoidosis deconvolutes immunopathogenic mechanisms', by J.M. Keane et al., https://doi.org/10.1093/eurheartj/ehaf1000.

Translational perspective.

This study demonstrates the utility of spatial proteomics in uncovering novel insights into cardiac sarcoidosis, a disease that remains difficult to diagnose and treat. Analysis of archival tissue revealed distinct immunologic and protein abundance profiles at various disease stages and in different cardiac compartments. Predictive modeling using just seven in situ biomarkers accurately distinguished sarcoidosis tissue from diverse controls, even in the absence of nearby histologic evidence of sarcoidosis. These findings highlight the potential of tissue-based biomarkers to support a new diagnostic strategy—one that could improve the diagnostic yield of endomyocardial biopsies in patients with suspected cardiac sarcoidosis.

Introduction

Sarcoidosis is a multi-system inflammatory disease of uncertain cause and with widely ranging clinical manifestations. Cardiac involvement is considered the disease’s most serious manifestation, conferring a high risk of heart failure, life-threatening arrhythmias, and all-cause mortality.^1–3 The exact prevalence of cardiac sarcoidosis (CS) is difficult to assess due to a significant burden of occult disease and missed diagnoses,^2,4,5 with many cases only diagnosed after sudden cardiac death⁶ or at the time of heart transplantation.^7,8

Though incurable, CS is a treatable disease. The mainstay of therapy is systemic immunosuppression, initially using high-dose glucocorticoids, followed by a transition to systemic steroid-sparing agents.^2,3 However, because of the potential risks of systemic immunosuppression, current guidelines only recommend initiating treatment in patients with definitive, active CS. Owing to the particularly high risk of life-threatening arrhythmias in the CS population,⁹ guidelines for primary prevention of implantable cardiac defibrillator (ICD) make special mention of CS, recommending ICDs for many CS patients even when left ventricular ejection fraction (LVEF) is relatively preserved. However, because ICD therapy has known risks,¹⁰ these recommendations are only applicable to patients with definitive CS diagnoses.

Given the high morbidity of CS, making a timely and accurate diagnosis is of paramount importance. In general terms, diagnosis relies on clinical characteristics, tissue sampling, and exclusion of alternative causes. More specifically, the diagnostic reference standard for CS is histologic identification of non-caseating granulomas on endomyocardial biopsy (EMB) tissue.^1,3,4 Unfortunately, due to the heterogeneous and sparse distribution of these pathognomonic granulomas, the true diagnostic yield of EMB is estimated at 25%.^2,11,12 Advanced imaging techniques such as cardiac magnetic resonance imaging and fluorodeoxyglucose positron emission tomography have value in CS as both screening tests and treatment-surveillance tests, but remain most useful as ancillary diagnostics due to limitations of their sensitivity and specificity.¹

The high incidence of occult disease, risky and untargeted treatment regimens, and insensitive diagnostic gold standard each highlight key areas of unmet need in CS. Without a known cause for sarcoidosis, the development of animal models to study this disease has not been possible. Thus, researchers are forced to rely on human biosamples as the primary source for diagnostic, mechanistic, and therapeutic investigation. Until recently, tools for examining these residual or archival human tissue samples have been limited, which in turn has limited breakthroughs in translational CS research. In this manuscript, we describe the application of high-plex spatial protein analysis within a large cohort of archival human CS tissue samples. The present study differs from prior CS research¹³ by leveraging the full potential of spatial-omics technology to individually study the myocytes, stromal cells, and vascular cells beyond the pathognomonic granulomas that define the disease. This approach is motivated by the fact that although granulomas are the hallmark of CS, they occupy only a small fraction of the total cardiac tissue area. We therefore hypothesize that key CS patient outcomes like cardiac pump failure and arrhythmias may reflect CS-specific, tissue-level changes that can be detected in regions far-removed from overt, active CS inflammation. Using this novel conceptual framework and cutting-edge spatial protein profiling technology, we describe novel CS biology and uncover a concise set of tissue-level biomarkers that provide high diagnostic accuracy even in tissue without pathognomonic granulomas.

Methods

The study cohort consisted of n = 64 formalin-fixed, paraffin-embedded cardiac tissue samples from CS patients and non-CS controls. N = 39 samples were derived from patients with tissue-confirmed CS, including n = 14 EMBs obtained early in CS disease course and n = 25 ‘advanced-stage’ samples obtained at cardiac transplant. Although all CS patients contributing samples had tissue-confirmed CS, among the specific tissue blocks used for this study, n = 24 contained pathologist-documented granulomatous inflammation, n = 7 had inflammatory cell infiltrates without granulomas, and n = 8 had no identified inflammation. The remaining n = 25 study tissues comprised a diverse control cohort, including: (i) n = 9 ‘heart failure controls’ with advanced non-inflammatory cardiomyopathy, (ii) n = 8 ‘non-failing controls’ sourced from heart donors who had suffered cardiac arrests, and (iii) n = 8 ‘inflammatory controls’, consisting of n = 6 myocarditis patients and n = 2 heart transplant recipients with chronic allograft dysfunction. Patient-level demographics and comorbidities for the study cohorts are summarized in Table 1.

Table 1.

Cohort details with patient- and tissue-level summary statistics

Tissue type (number patients)	Self-reported ethnicity (number of patients)	Age (years)	Male sex (number of patients)	Left ventricular ejection fraction	Hyper-tension (number of patients)	Diabetes (number of patients)	Hyper-lipidemia (number of patients)	CKDa (number of patients)	Granuloma present (number of patient samples)b	Inflammation present (number of patient samples)b	On ISxc at tissue sampling (number of patients)
Sarcoid (n = 39)	White: 53.8% (n = 21) Black: 43.6% (n = 17) Not Disclosed: 2.6% (n = 1)	53.8% (n = 21)	87.2% (n = 34)	26.10%	41% (n = 16)	33.3% (n = 13)	28.2% (n = 11)	7.7% (n = 3)	64.1% (n = 25)	53.8% (n = 21)	15.4% (n = 6)
Biopsy (n = 14)	White: 78.6% (n = 11) Black: 21.4% (n = 3)	50.1	85.7% (n = 12)	30.30%	42.9% (n = 6)	28.6% (n = 4)	42.9% (n = 6)	0% (n = 0)	57.1% (n = 8)	57.1% (n = 8)	14.3% (n = 2)
Explant (n = 25)	White: 40% (n = 10) Black: 56.0% (n = 14) Not Disclosed: 4.0% (n = 1)	51.6	88.0% (n = 22)	28.60%	40.0% (n = 10)	36.0% (n = 9)	20.0% (n = 5)	12.0% (n = 3)	68.0% (n = 17)	52.0% (n = 13)	16.0% (n = 4)
Control (n = 25)	White: 68.0% (n = 17) Black: 32.0% (n = 8)	55.6% (n = 5)	52.0% (n = 13)	39.90%	64.0% (n = 16)	12.0% (n = 3)	28.0% (n = 7)	32.0% (n = 8)	0% (n = 0)	24.0% (n = 6)	12.0% (n = 3)
Total (n = 64)	White: 59.4% (n = 38) Black: 39.0% (n = 25) Not Disclosed: 1.6% (n = 1)	54.2% (n = 26)	73.5% (n = 47)	31.40%	50.0% (n = 32)	25.0% (n = 16)	28.0% (n = 18)	17.2% (n = 11)	39.1% (n = 25)	42.2% (n = 27)	14.1% (n = 9)
Sarcoid vs Control (P)		.926	.002	.002	.075	.056	.986	.011	<.001	.0183	.704

^aCKD, chronic kidney disease.

^bPresence in tissue sample is based on histologic diagnosis by pathologist review of tissue sample used for spatial protein expression profiling. Note that all patients contributing sarcoid samples had a tissue-confirmed histologic diagnosis of cardiac sarcoidosis, though the sample used for digital spatial profiling did not necessarily demonstrate diagnostic histologic features.

^cISx: Immunosuppression. Defined as active treatment of the patient with systemic corticosteroids, mycophenolate, azathioprine, methotrexate, leflunomide, hydroxychloroquine, tacrolimus, sirolimus, or tumor-necrosis factor inhibitor.

Using the GeoMx Digital Spatial Profiler instrument, tissue ‘regions of interest’ (ROIs) were selected to assess differentially abundant proteins (DAPs) across a variety of conditions, both within and beyond granulomas. Importantly, every study ROI was manually screened by study investigators to exclude significant tissue artifacts (folds, tears, staining/processing defects) as well as non-relevant tissue types (e.g. large areas of epicardium or endocardium). This manual screening during ROI selection helped to ensure a degree of consistency in gross tissue quality and content. The panel for GeoMx spatial protein abundance analysis consisted of n = 79 protein targets which provide detailed descriptions of immune and parenchymal cell populations and their activation and differentiation states. Refer to the Supplementary methods and Supplementary data online, Table S1 for details. Within the study ROIs containing cardiac parenchymal tissue (as opposed to those drawn solely around granulomas), we conducted ‘tissue-compartment-specific’ analyses of protein abundance via digital image segmentation of ROIs into distinct ‘areas of illumination’ (AOIs). These AOIs were labeled according to the parenchymal tissue-types contained within them: cardiomyocyte AOIs (containing cardiomymocytes), interstitial stroma AOIs (inclusive of fibroblasts, capillaries/small vessels, and immune cells), and the vascular compartment AOIs (containing the walls and perivascular stroma of larger vessels). Figure 1 provides a summary of the study workflow with visual examples of ROIs and AOIs.

Figure 1.

GeoMx workflow. (Panel A) after 4-plex immunofluorescence staining with CD31 for vessels, Troponin I for myocytes, Collagen VI for collagenous stroma, and a pan-nuclei stain, study slides are bathed in nucleotide-tagged antibodies for each of the protein panel markers and loaded into the GeoMx digital spatial profiling instrument. (Panel B) Region of interest (ROI) selection as per study experimental design, focusing on granuloma ROIs and a series of cardiac parenchyma ROIs corresponding to ‘peri-granuloma- regions, ‘granuloma-remote’ regions, and large vascular bed regions. (Panel C) Prior to expression data collection, digital pathology image segmentation of parenchyma ROIs is performed to divide ROIs into tissue-compartment-specific ‘areas of Illumination’ (AOIs). AOIs representing the Myocyte compartment, the Stroma compartment, and the vascular compartment are digitally segmented using pixel intensity thresholds, hole-filling, and dilation operations. Finally, focused ultraviolet light from the GeoMx instrument is used to cleave nucleotide barcodes and sequentially extract protein expression data from each AOI in each ROI.

The research focused on four major experiments: (i) an analysis of intra- and inter-sample heterogeneity in the protein abundance profiles of CS granulomas, (ii) an analysis of protein abundance across CS disease stages, comparing the cardiac parenchyma (e.g. non-granuloma-containing ROIs) from ‘earlier-stage’ EMBs to ‘later-stage’ explant samples, (iii) an analysis of ‘distance-dependent’ protein abundance patterns in CS parenchyma, comparing inflammation-and-granuloma-free ‘peri-granuloma’ ROIs (<300 µm from granulomas) to ‘granuloma-remote’ ROIs (>500 µm away), and (iv) an analysis comparing granuloma-remote, inflammation-free ROIs of CS tissues to ROIs from the diverse control cohort. The goal of this final analysis was to uncover protein abundance patterns that are specific to CS but which are more homogenously distributed than classical granulomatous inflammation and thus capable of improving CS diagnostic accuracy.

The GeoMx workflow generated 363 ROIs (totaling 713 AOIs after tissue compartment segmentation). The CS cohort consisted of n = 39 ‘granuloma ROIs’, n = 177 granuloma-free ‘cardiac parenchyma ROIs’, and n = 29 large ‘vascular bed ROIs’. Refer to Figure 1 for an overview of the GeoMx and ROI/AOI selection process. Among granuloma-free ‘parenchyma ROIs’ within the CS cohort, there were n = 47 from EMB tissues (n = 6 ‘peri-granuloma’, n = 54 ‘granuloma-remote’), and there was n = 130 from advanced-stage explant tissues (n = 19 peri-granuloma, n = 111 granuloma-remote). Notably, n = 33 CS parenchyma ROIs were designated as ‘inflammation(+)’ due to having overt histologic immune cell infiltration despite the lack of any observable granulomas, while the rest are designated ‘inflammation(−)’. Finally, the diverse control cohort consisted of n = 118 ROIs, including n = 92 parenchymal ROIs and n = 26 large vascular bed ROIs.

For data analysis, principal component analysis (PCA) and t-distributed Stochastic Neighbor Embedding (t-SNE) were used for whole-panel data visualization. Variance analysis was performed to evaluate protein abundance heterogeneity in granuloma-containing regions. Differential abundance testing for study experiments was conducted via a linear mixed model on quantile-normalized data. All statistical analyses of protein abundance were performed in an exploratory framework, with raw (unadjusted) P-values reported throughout to minimize type II error. Given the novelty of the cohort and methodology, and the absence of prior deep tissue phenotyping in CS, these analyses were designed to generate hypotheses and highlight signals that warrant further investigation. Adjusted P-values via the Benjamini–Hochberg method are available in the Supplementary data online, Tables S4–S9 for readers interested in a more conservative approach to hypothesis testing.

As a proof-of-principle for using spatial tissue analysis to enhance the diagnostic yield of tissue sampling in CS, prediction modeling was performed to differentiate granuloma-remote CS from Controls. Only protein markers suitable for immunofluorescence (IF) were eligible for the model (n = 122 total), as future diagnostic applications would require an IF-based approach. LASSO feature selection and stepwise logistic regression were employed to generate ROI-level and patient-level models. See Supplementary Methods for detailed cohort, GeoMx, tissue processing/staining, and data analysis methods.

Results

Spatial protein abundance results demonstrate substantial CS granuloma heterogeneity, both within and between patients

The first analyses focused on examining heterogeneity in protein abundance within CS granulomas. Variance comparisons across all 74 protein biomarkers revealed that for 62% (46/74) of proteins, the population variance in abundance across granuloma-containing ROIs was greater than that observed in non-granuloma parenchymal ROIs. To statistically assess the significance of this pattern, we performed F-tests for each protein, comparing variances between the two tissue groups. Among the 74 proteins, 33 exhibited significantly different variances (P < .05), and of these, 25 (76%) had significantly greater variance in granuloma ROIs (two-sided proportion z-test P = .0032), supporting the interpretation that granuloma ROIs exhibit systematically higher heterogeneity in protein expression. This increased heterogeneity is particularly notable given that the parenchymal ROIs also represent biologically diverse regions—encompassing varying degrees of stromal and myocyte content, LVEFs, and the presence or absence of interstitial inflammatory cell infiltrates. The 10 most variably expressed proteins in granuloma ROIs were predominantly immune-related markers associated with macrophages, T cells, granulocytes, and antigen-presenting cells, including: CD68, HLA-DR, CD11c, CD45, CD3, IDO1, CD44, GAPDH, CD66b, and OX40L (see Supplementary data online, Table S2).

There were only modest differences in the granuloma protein abundance between ‘advanced-stage disease’ tissues and EMB tissues obtained earlier in the disease course. Immune checkpoint molecule Tim-3 and activated fibroblast marker FAP-alpha have significantly increased abundance in advanced-stage CS hearts (P = .02 and P = .017, respectively). In contrast, immune checkpoint molecule VISTA, nuclear/proliferation marker Histone-H3, and activated MEK1 (part of the RAF/MEK/ERK pathway known to be involved in inflammation and linked to granuloma formation when inhibited)^14,15 each had significantly increased abundance in earlier/active stage CS (P = .007, P = .013, and P = .017, respectively).

We performed a sub-analysis of the nine CS samples which contributed multiple granuloma ROIs to the dataset. The PCA biplot in Supplementary data online, Figure S1 suggests substantial intra-sample heterogeneity in CS granulomas, with only modest within-sample groupings. Nearest-neighbor analysis of the PCA plot demonstrates that for cases contributing multiple granuloma ROIs, the closest-clustering granuloma ROI is more likely to be from a different sample than from the same sample (13/23, 56.5%). At the individual protein level, the intra-sample coefficient of variation (COV) exceeded inter-sample COV for an average of 11 protein markers in each multi-granuloma sample, with the most substantial intra-sample variation seen in T-cell and cytotoxic cell markers (CD27, CD8, GZMB), apoptotic markers (CD95/FAS, Cleaved-Caspase-9), and checkpoint molecules (PDL2) (Supplementary data online, Table S3).

Analysis of the cardiac parenchyma highlights the protein abundance profiles of active and ‘burnt out’ CS

Analysis of intrinsic cardiac parenchymal ROIs (e.g. non-granuloma ROIs) demonstrates substantial differences in the abundance of immunologic, cell survival, and cell death pathways between EMB samples obtained during the active workup/management phase of CS and advanced-stage disease samples obtained at cardiectomy. These differences are readily apparent via unsupervised data visualization with PCA and t-SNE plots in Figure 2A. When analysed as non-compartmentalized ROIs, n = 33 proteins show significant differential abundance based on disease stage (Figure 2B and Supplementary data online, Table S4). When segmented as compartment-specific AOIs, there are n = 24 significant DAPs in the myocyte compartment, and n = 27 DAPs in the stroma compartment.

Figure 2.

Cardiac sarcoidosis (CS) disease stage analyses, comparing spatial protein abundance between biopsy tissue samples obtained early during disease management and advanced-stage disease samples obtained at cardiac explant. (A) Principal component analysis (PCA, on left) and t-SNE (on right) plots of panel-wide protein abundance in the study regions of interest (ROIs). ROIs from biopsy samples obtained earlier in disease (purple circles in PCA and in t-SNE plots) predominantly group together at one edge of the dataset, while ROIs from advanced-stage samples largely occupy the rest of the field (pink circles in PCA and t-SNE plots). (B) Tabular results for differential protein abundance analyses comparing samples by CS disease stage. ROI-level protein abundance for biopsy samples obtained earlier in disease are compared to explant samples obtained in advanced-stage disease. In addition to ‘full-ROI-level’ differential abundance results, digital segmentation of study ROIs enables tissue-compartment-specific analysis of protein abundance in discrete ‘areas of illumination’ (AOIs). Via this method, cardiomyocyte-specific and stromal-tissue-specific differential protein abundance comparisons between biopsy and explant samples are also presented. Differentially expressed proteins with P-values <.05 are shown for full-ROI level data and for myocyte-specific and stroma-specific data

While substantial overlap exists between DAPs in the ROI- and AOI-level analyses, AOI analyses add important context. We observed a significant shift towards pro-apoptotic conditions (increased BIM, decreased BXLCL) as well as decreased MAP-kinase and PI3K/AKT pathway activity in the myocyte compartment AOIs of advanced-stage CS hearts—a finding consistent with prior research on end-stage cardiomyopathy more generally.^16,17 In stroma compartment AOIs of advanced-stage hearts, we observed significant increases in markers of activated/differentiated fibroblasts (FAP-alpha and SMA). Again, this finding is largely consistent with known advanced-stage cardiomyopathy biology.¹⁸ However, we also observed numerous significant shifts in abundance of immune-related protein markers in both the stroma and myocyte AOIs when comparing tissues acquired earlier vs later in CS, which are not as easily explained by the common final pathways of end-stage cardiomyopathy.

In advanced-stage CS, there was a significant decrease in several major lineage markers for activated macrophages and effector T cells (CD3, CD4, CD163), which coincided with significant increases in markers of longer-lasting regulatory T cell (Treg), memory T cell, and B-cell populations (ICOS, FOXP3, CD45RO, CD127, CD20). Interestingly, though CD163 macrophages are significantly less abundant in advanced-stage CS (Figure 2B), a modest increase in CD14 abundance was noted, suggesting a shift in macrophage phenotype in advanced-stage CS without a significant change in overall macrophage abundance (as assessed by pan-macrophage marker CD68, which was unchanged, refer to Supplementary data online, Table S4). An increased abundance of checkpoint receptors PD-1 and Tim-3 was also observed in late-stage CS—a protein expression pattern strongly associated with T-cell exhaustion following prolonged antigen exposure.¹⁹ The myocytes of advanced-stage CS manifest a less ‘immune primed’ state relative to earlier-stage cases, with decreased abundance of major histocompatibility (MHC) molecule HLA-DR, checkpoint molecules PD-L1/PD-L2, and interferon-producing STING. This combination of protein abundance patterns is consistent with the theoretical biology of the ‘late fibrous phase’ of CS (sometimes called ‘burnt-out’ CS),²⁰ which is thought to involve increased fibrosis along with a shift from active inflammatory mediators to a more chronic, indolent state.²⁰ Overall, our findings provide material support to long-standing theories about the biology of CS, highlighting numerous key immune cell types and effectors that change as CS progresses.

Examining the ‘distance gradient’ of cardiac parenchymal biology in CS

Spatial analysis of inflammation(−) CS parenchyma reveals a previously unreported ‘distance gradient’ in protein abundance patterns. As shown in Figure 3 and Supplementary data online, Table S5, when comparing histologically bland, inflammation(−) ROIs from study-defined ‘peri-granuloma’ regions vs those from ‘granuloma-remote’ regions, we found that numerous panel proteins exhibited differential abundance based on an ROI’s distance from granulomas.

Figure 3.

Cardiac sarcoidosis (CS) ‘distance gradient’ analyses, highlighting the impact of proximity to granulomatous inflammation on cardiac parenchyma protein abundance. (A) Principal component analysis (PCA, on left) and t-SNE (on right) plots of panel-wide protein abundance in the study regions of interest (ROIs). ‘Peri-granuloma’ ROIs located within 300 µm of a granuloma (purple circles in PCA and in t-SNE plots) predominantly group together on the left edge of the dataset, while ‘granuloma-remote’ ROIs located >500 µm from any granuloma largely occupy the rest of the field (pink circles in PCA and t-SNE plots). (B) Tabular results for differential protein abundance analyses comparing samples by relative proximity to granulomatous inflammation. ROI-level protein abundance from peri-granuloma cardiac parenchyma is compared to ROI-level abundance from granuloma-remote cardiac parenchyma. In addition to ‘full-ROI-level’ differential abundance results, digital segmentation of study ROIs enables tissue-compartment-specific analysis of protein abundance in discrete ‘areas of illumination’ (AOIs). Via this method, cardiomyocyte-specific and stromal-tissue-specific differential protein abundance comparisons between peri-granuloma and granuloma-remote parenchymal regions are also presented. Differentially expressed proteins with P-values <.05 are shown for full-R OI level data and for myocyte-specific and stroma-specific data

For additional context, we also examined ‘inflammation(+)’ ROIs which do have discrete (though non-granulomatous) interstitial inflammatory infiltrates, comparing these extreme examples of parenchymal proximity to inflammation to the ‘inflammation(−)’ ROIs. Unsurprisingly, there were many significant DAPs between these groups, including increased abundance of numerous immune cell lineage markers: CD3, CD4, CD8, CD68, CD163, GZMA, CD14, and CD45 (Supplementary data online, Table S6). Interestingly, at the ROI level, 70% (12/17) of the significant DAPs which were observed during comparison of peri-granuloma to granuloma-remote ROIs are also significant DAPs when comparing ‘inflammation(+)’ to ‘inflammation(−)’ ROIs. However, these overlapping DAPs are largely not lineage markers of acute, pro-inflammatory, immune cell-types, instead highlighting subtler patterns of distance-dependent immune biology within inflammation(−) CS parenchyma.

Of the major lineage markers listed above, only general leukocyte marker CD45 exhibited a distance-dependent abundance between inflammation(−) peri-granuloma and granuloma-remote ROIs. Upon closer examination, this difference is largely due to significant increases in long-lived immune ‘sentinels’ CD11c+ dendritic cells and CD45RO+ memory T-cells. Notably, these are relatively low-abundance cell-types, which may explain why their differential abundance does not manifest as an obvious histologic cellular infiltrate. Beyond specific immune cell lineage markers, in the stromal compartment of inflammation(−) peri-granuloma ROIs, we observed increased abundance of class I and II MHC molecules (beta-2-microglobulin and HLAD-DR)—a histologically imperceptible phenomenon documented previously in inflammatory myocardial conditions.^21–25 Peri-granuloma stroma also exhibits increased inflammation-associated pro-fibrotic mediators IDO-1, arginase-1,^26,27 fibronectin,²⁸ and CD44,^29,30 while manifesting decreased anti-inflammatory immune checkpoint receptor Tim-3.^19,31 Within the cardiomyocyte compartment of peri-granuloma ROIs, we observed a similar ‘stress-activated’ state,³² with increases in inflammation-associated class I/II MHC molecules, CD40,³³ and immune checkpoint PD-L1³⁴ along with increased injury repair, fibrosis, and stiffness-associated CD44^35–37 and fibronectin (each of which likely co-localizes with cardiomyocytes rather than being expressed by them).²⁸

The protein abundance profile of ‘granuloma remote’ CS parenchymal tissue differs from that of controls

A fundamental question at the outset of this research was whether tissue biomarkers of CS exist which can be detected even when no histologic evidence of CS is present. As shown in Figure 4A and Supplementary data online, Table S7, we identified numerous significant DAPs between granuloma-remote, inflammation(−), CS tissue and tissue from non-CS controls. Compared to controls, granuloma-remote CS cardiac parenchyma is characterized by increased lymphocyte and lymphocyte co-stimulatory markers (CD127, CD20, 4-1BB), MHC-II molecule HLA-DR, checkpoint receptor Tim-3, inflammation-induced immune-modulator IDO-1, cell-adhesion molecule CD44, endothelial marker CD34, and global nuclei/proliferation marker Histone-H3. CS parenchyma manifests decreased abundance of NK-cell marker CD56, inhibitory checkpoint ligand PDL2, stromal remodeling/fibrosis marker fibronectin, pro-apoptosis markers BIM, CD95/Fas, and Cleaved-Caspase-9, and an array of PI3K and MAPK pathway markers.

Figure 4.

Comparison of cardiac sarcoidosis (CS) cardiac parenchyma and non-CS control tissue parenchyma. (A) List of differentially expressed proteins (DAPs) between granuloma-remote, histologic inflammation-free regions-of-interest (ROIs) in cardiac sarcoidosis tissue samples as compared to ROIs from a diverse Control tissue population of failing and non-failing hearts. In addition to ‘full-ROI-level’ differential abundance results, digital segmentation of study ROIs enables tissue-compartment-specific analysis of protein abundance in discrete ‘areas of illumination’ (AOIs). Via this method, cardiomyocyte-specific and stromal-tissue-specific differential protein abundance comparisons between inflammation-free/granuloma-remote parenchymal regions and Control parenchymal regions are also presented. Differentially expressed proteins with P-values <.1 are shown for full-ROI level data and for myocyte-specific and stroma-specific data. (B) Spatial abundance trends for selected categories of protein markers. Abundance is shown as a function of relative distance from granulomatous inflammation: peri-granuloma regions of CS tissue, granuloma-remote regions, and non-CS Control tissue regions. As expected, effector immune cell activity declines with increasing distance, while unexpected variation is observed in checkpoint molecules and regulatory T-cell markers. In plots 4B.iv and 4B.v, immune marker abundance is further stratified by Control subgroup (inflammatory vs non-inflammatory cardiomyopathies). Error bars represent standard errors. Statistical comparisons corresponding to these trends are provided in the Results text and in Figures 3B and 4A; FDR-adjusted P-values are available in the Supplement. Superscript notation is used to identify statistically significant trends to aid in figure interpretation as follows: * = significant distance trend for peri-granuloma CS vs granuloma-remote CS ROI comparisons, Ŧ = significant distance trend for granuloma-remote vs Control ROI comparisons, § = significant distance trend across ordinal distance gradient.’

To place this finding in the context of the ‘distance gradient’ described in the previous section, we performed ordinal logistic regression, treating peri-granuloma, granuloma-remote, and Control regions as ordinal classes of increasing distance from granulomas. Interestingly, 66.6% of panel proteins (n = 50) demonstrated differential abundance with increasing distance (Supplementary data online, Table S8). Specifically, histologically bland, inflammation(−) cardiac parenchyma located relatively closer to granulomas is characterized by higher levels of immune-modulators such as FOXP3, CD25, Tim-3, VISTA, and IDO-1, along with immune ‘activated’ myocytes expressing co-stimulatory molecules HLA-DR, CD40, CD44, and 4-1BB. On the other hand, cardiac parenchyma located closer to granulomas had significantly lower abundance macrophage marker CD163, NK-cell marker CD56 and NK-cell-associated enzyme GZMA, lower abundance of DNA damage (PARP) and pro-apoptotic markers (BIM, BAD, CD95/Fas), and lower PI3K and MAPK pathway markers. Figure 4B provides a visual summary of many of these distance-related phenomena, showing changes among several key groups of proteins when moving along this ordinal distance scale. Taken together, these results reinforce the presence of a ‘distance gradient’ within inflammation(−) CS tissue for some markers, while also highlighting more diffuse, organ-wide changes in CS which distinguish CS ROIs from Controls regardless of granuloma proximity.

It is worth noting that the Control cohort deliberately contained tissues from hearts with both definitive diagnoses of inflammatory and non-inflammatory cardiomyopathies. While the abundance of major immune markers like CD3, CD68, and CD45, and CD8 are not significantly different between granuloma-remote CS ROIs and Control ROIs overall, Supplementary data online, Figures S4B.iv and S4B.v highlight clear differences between inflammatory Controls and other study populations. Inflammatory Controls have significantly more CD3 (P < .001), CD68 (P = .0358), and CD45 (P = .0012) than non-inflammatory Controls, and significantly more CD3 (P = .002), CD68 (P < .001), and CD8 (P < .001) than granuloma-remote CS parenchyma. Overall, visual inspection of Supplementary data online, Figures S4B.iv and S4B.v suggest that inflammatory controls more closely resemble peri-granuloma CS regions with respect to many immune cell markers. The differences between inflammatory and non-inflammatory Controls are subtler among the other thematic groups of markers plotted in Figure 4B (plots-by-control subgroup not shown due to uniform results).

A limited, pre-specified sub-analysis of ROIs derived from larger vascular beds was performed to assess whether the vasculature in CS differs from Controls. Overall, among n = 29 from CS cases, n = 26 from controls, the differential abundance testing an immunologically active environment in CS vascular beds relative to controls, with significantly increased abundance CD127, CD20, Tim-3, 4-1BB (suggesting immune peri-vascular immune cell presence) and endothelial immune response protein STING. In addition, significant increases in CD34 and SMA suggest alterations in endothelial and vascular smooth muscle cell biology may be present in CS vasculature (Supplementary data online, Table S9).

Spatial protein abundance biomarkers enable accurate prediction of occult CS

To assess whether the observed protein abundance differences between areas of inflammation(−) CS parenchyma and Controls could have diagnostic value, we developed a binary prediction model to classify parenchyma tissue ROIs as originating from CS vs Control patients. To maximize clinical utility as a tool capable of improving the diagnostic yield of tissue sampling in CS, the model was specifically developed using only data from CS ROIs which were ‘granuloma-remote’ and ‘inflammation(−)’. After optimization via leave-one-out cross validation, our final 7-variable logistic ROI-level model was comprised of MHC molecule HLADR, Treg markers FOXP3, CD25, memory T-cell marker CD127, cytotoxic immune cell markers CD8 and CD56, and endothelial cell marker CD34 (Figure 5A). Figure 5B and Supplementary data online, Figure S2 provide representative IF images highlighting the relative differences in Treg and cytotoxic T-cell marker abundance between CS and Controls in granuloma-remote tissue.

Figure 5.

(A) Predictive modeling results for differentiating granuloma-remote, inflammation-free tissue from cardiac sarcoidosis (CS) patients vs controls. The modeling workflow for generating both a region-of-interest (ROI)-level prediction model, and a subsequent patient-level prediction model are provided on the left. Despite being far removed from any conventional histologic evidence of sarcoidosis or active inflammation, the seven tissue biomarkers of CS listed in the figure achieve excellent discrimination between CS and a diverse cohort of non-CS Control tissues. Confusion matrices and receiver operator curves (ROC) represent performance in the held-out test set of unseen cases for the ROI-level model and the subsequent patient-level model. (B) representative multi-marker immunofluorescence images from a CS and non-CS Control case (in this example, a tissue sample from a patient with myocarditis). Antibodies for HLA-DR, CD25, FOXP3, and CD8 are shown, each of which are predictive biomarkers in the prediction model summarized in 5A. Troponin stain used to identify cardiomyocytes, which appear uniform-gray in the micrographs. DAPI staining in blue identifies nuclei. Note the relative increased abundance of HLA-DR and regulatory T-cell markers CD25 and FOXP3 in the CS sample compared to the Control sample, along with relatively reduced abundance of cytotoxic T-cell marker CD8 in CS. HLA-DR imaged at 8× magnification to highlight broad parenchymal staining, all other image tiles imaged at 40× magnification

As shown in Figure 5, the ROI-level model achieved excellent performance in the ‘held-out’ validation set of n = 49 ROIs, with an accuracy of 95.9% (47/49 correct), AUROC of 0.99, sensitivity of 100%, and specificity of 92.3%. ROI-level predictions were also consistent at the patient level, with confident but discordant ROI-level predictions being exceedingly rare. Specifically, there were only n = 2 patients for which ROIs had an assigned probability of both >0.95 and <0.05. In both instances, the discrepancy was driven by a single ROI rather than a broader pattern across the tissue. In other words, incorrect, highly confident, and highly discordant predictions occur in just 2/217 (0.92%) of ROI-level predictions in our experiment. Given that it is also possible to sample peri-granuloma regions during a clinical EMB procedure while still ‘missing’ a granuloma, we also assessed performance of the final model on inflammation(−), peri-granuloma regions. Performance was excellent on these as well, achieving accuracy of 100% (20/20). As described in the methods and Figure 5, we next aggregated and averaged the ROI-level prediction results on a per-patient basis to generate a simple, single-variable, patient-level model for identifying CS. This patient-level model achieved perfect discrimination (100% accuracy, AUROC: 1.0) between CS and Control tissue on n = 13 patients (n = 6 CS patients, n = 7 Controls, including n = 4 inflammatory cardiomyopathy controls). Refer to Supplementary data online, Table  S10 and S11 for per-ROI and per-patient details on model classification performance.

Discussion

In this manuscript, we utilized advanced spatial protein abundance profiling, meticulous digital pathology methods, and a unique experimental design to comprehensively study cardiac tissue from CS patients. Our findings provide numerous insights into CS biology, not only within characteristic granulomas, but also within the cardiac parenchymal tissue near granulomas and spatially far removed from granulomas. Particularly notable is the discovery of a small set of protein markers that are differentially expressed in CS tissue as compared to a diverse group of controls, even when no inflammatory process or granuloma is evident in the surrounding area. This finding has translational value that extends beyond descriptive biology, introducing the possibility of improving the diagnostic yield of clinical biopsies for patients with suspected CS. We believe this report represents an important contribution to mechanistic research in CS, lays the groundwork for a novel precision diagnostic tool, and highlights the translational potential of spatial-omics methods when paired with thoughtful, ‘spatially conscious’ experimental designs.

Molecular heterogeneity of CS granulomas—both within and between samples—is a novel finding in our analysis. Enabled by a cohort with six times as many CS tissues and 30 times as many AOIs as the next-largest spatial profiling study,¹³ our experiments were better equipped to characterize the substantial variability of the CS immune response than any prior investigation. As a retrospective cohort study, confounders such as varied disease stages and treatment regimens could have contributed to the diversity of protein abundance patterns we observed in this experiment. However, neither of those confounders would explain the marked intra-sample heterogeneity we observed, which must necessarily arise from an intrinsic biological mechanism within each patient. It is conceivable that CS granulomas develop at different times, and that the relative ‘age’ of a granuloma may dictate the local immune profile. Granulomas may also go through periods of relative activation and quiescence, and thus may present different profiles at different times. There is precedent for granuloma heterogeneity and granuloma ‘aging’ in published tuberculosis research, though no prior CS study has been able to investigate this phenomenon.³⁸ In the absence of animal models for CS, prospective cohorts with serial biosampling might be required to provide further mechanistic insights into the causes of granuloma heterogeneity.

We also observed ‘temporal’ heterogeneity in CS, with differential protein profiles arising from differences in ‘disease stage’ when examining granuloma-free parenchyma tissue from CS patients. An examination of potential confounders does not allow for easy dismissal of these findings. In our study, the ‘earlier-stage’ EMB cases and ‘advanced-stage’ explant subgroups had similar proportions of patients on active treatment at tissue acquisition (14.2% vs 15.3%). In addition, major differences in intrinsic tissue content between endocardial EMB samples and transmural explant samples (e.g. large vascular beds, epicardial tissue, fibro-fatty scar, etc.)—a sampling bias which frequently confounds traditional, homogenate-based, ‘omics’ assays—is less relevant in the present experiment. This is because the spatial platform utilized in this research permits direct visual selection of ROIs, ensuring that unwanted/extraneous tissue areas do not contaminate the dataset. We further confirmed this assumption via quantitative, whole-slide, morphology analysis of all CS parenchymal ROIs across both EMB and explant groups in our study cohort, measuring cardiomyocyte, stromal, and vascular areas as well as nuclear densities (see Supplementary methods). This analysis revealed no statistically significant differences in between early-stage EMBs and late-stage explants (all P-values >.2, see Supplementary data online, Table  S12). Thus, we conclude that even though local conditions at sampling sites may introduce some confounding, it is the progression from active CS to ‘burnt out’ disease that is the primary process underlying the large number of DAPs we observed. While our results demonstrate that advanced-stage CS is indeed less immunologically ‘active’ than earlier disease with decreases in several immune lineage markers and an increase T-cell exhaustion markers, our results also demonstrate that numerous pro-fibrotic pathways remain quite active in late-stage CS. Cell-therapy clinical studies targeting activated cardiac fibroblasts have received significant attention in recent years,¹⁸ and our findings suggest that CS may be another potential population for these novel therapeutics.

Utilizing the unique opportunity spatial biology platforms provide to directly visualize and analyse distinct tissue regions, we were able to uncover remarkable ‘spatial’ heterogeneity within CS tissues as well. Numerous protein markers were differentially expressed based on an ROI’s relative distance from CS inflammation. Unsurprisingly, inflammation(+) cardiac parenchyma, which contains overt histologic cellular infiltrates, has increased abundance of nearly two-thirds of the study panel proteins relative to inflammation(−) regions. More interestingly, when comparing inflammation(−) ‘peri-granuloma’ tissue to ‘granuloma-remote’ regions, we continued to find significant differential abundance of many immune markers and pathways. This suggests that immune activity in CS extends beyond the sites of overt inflammation, and does so in a graded, distance-dependent fashion. Whether this observed ‘distance gradient’ is a consequence of recent inflammation, of paracrine cytokine effects from nearby granulomas, or both, cannot be ascertained without animal models or serial tissue sampling. Nevertheless, it is clear that relative distance from inflammation has biological effects on the surrounding parenchyma, even in the absence of histologic inflammation.

Adding to this narrative is the discovery that even granuloma-remote, inflammation(−) regions in CS hearts manifest protein abundance profiles which distinguish them from a diverse population of Controls. While this finding has obvious diagnostic utility (as demonstrated in our predictive modeling efforts), when interpreted in the context of the ‘distance gradient’ experiments, it also highlights the diffuse and specific nature of parenchymal immune activation in CS. As an example, HLA-DR manifests significant, graded, differential abundance in cardiomyocytes across our experiments. HLA-DR is increased in inflammation(+) CS tissue relative to inflammation(−) tissue, is increased in peri-granuloma tissue relative to granuloma-remote tissue, and is increased in granuloma-remote tissue relative to Controls. In fact, two-thirds of our study markers manifest a statistically significant ‘distance gradient’ when peri-granuloma, granuloma-remote, and Control tissue regions are treated as ordinal classes of ‘distance’ from overt CS inflammation. Taken together, these observations lead to the conclusion that while parenchymal immune activation in CS is indeed distance-dependent, it is also sufficiently widespread to enable differentiation of CS tissue from non-CS tissue.

The abundance patterns of Treg markers in our experiments represent a particularly interesting manifestation of both widespread and distance-dependent protein abundance. Treg and Treg-associated markers FOXP3 and CD25 are significantly increased relative to controls in DAP testing and are incorporated into our final CS predictive model. Historically, there have been conflicting reports on both the abundance and functional abilities of Tregs in sarcoidosis.^39,40 It has been speculated that Tregs are functionally deficient in sarcoidosis,^39,40 and given the serious cardiomyopathy suffered by all cohort patients in spite of having elevated Treg-associated protein abundance, it is tempting to agree with this theory. Further supporting this position, we detected significantly increased CD127 in CS tissue relative to Controls, with CD127 also appearing as a strong predictor in our CS predictive model. CD127, an IL-7 receptor, is associated with memory T cells and has been shown to promote long-lasting Treg survival,^41,42 but has also been implicated as a key marker of Tregs which have lost effective immunosuppressive abilities.^41,43 This finding, along with increased T-cell exhaustion marker Tim-3 (which closely tracks CD127 abundance in our distance gradient framework—refer to Figure 4B.iv), suggests a unique population of long-lasting but potentially ineffective Tregs in CS.

Importantly, not all the study data support this dysfunctional Treg theory. Within CS tissue, we only observe increased Treg abundance in regions without overt inflammation. In fact, Treg abundance is significantly reduced in inflammation(+) CS parenchyma relative to inflammation(−) parenchyma. Given that nearly every other T-cell marker—except FOXP3, CD25, and CD127—is significantly increased in inflammation(+) ROIs, we must consider the possibility that a relative insufficiency in Treg number, rather than a deficiency in Treg function, is contributing to inflammatory injury in areas with active CS inflammation. There is precedence for such a hypothesis: several publications have reported a decrease in the ratio of Tregs to effector T cells within tissue experiencing pathologic immune activity.^44–46

The observation that CD8 levels are increased in inflammation(+) CS tissue, while Treg markers are decreased, might suggest a pathological shift in the balance between pro-inflammatory and anti-inflammatory factors in these regions without implicating any dysfunction in the Tregs themselves. In contrast, the inverse relationship between CD8 and inflammation(−) CS (as indicated by the negative regression coefficient in the CS predictive model), along with the positive relationship between Treg markers and inflammation(−) CS, may reflect a protective, anti-inflammatory, milieu these regions. In this vein, even the unexpected, positive relationship between CD34 and inflammation(−) CS might be explained. CD34, a protein expressed on cardiac vascular and lymphatic endothelial cells, is known to induce Treg differentiation in T cells entering the tissue compartment.⁴⁷ Therefore, CD34 may be playing a key role in promoting this observed shift in Treg abundance within inflammation(−) CS regions.

Our multi-marker IF findings lend orthogonal support and credibility to this theory. While we were not able to validate every marker in our GeoMx panel or diagnostic model due to technical limitations, the presence of ‘immune-primed’ parenchyma expressing diffuse HLA-DR—along with a persistent CD25⁺/FOXP3⁺ cell population and an absence of CD8⁺ cells—suggests that Tregs may be serving a role in maintaining tissue quiescence in histologically bland regions of CS. This observation aligns with emerging models of tissue-resident Tregs maintaining immune homeostasis in other settings, including transplant allografts.^48,49 Although the use of archival tissue in this study limits our ability to assess dynamic Treg function directly, the consistent spatial distribution of Treg-associated markers across multiple analytic contexts, combined with targeted IF confirmation, supports a potential anti-inflammatory role. Future prospective studies—including serial sampling and broader IF validation—will be necessary to clarify whether Tregs are functionally impaired in CS, or whether they serve a compensatory, protective role.

From a translational perspective, the CS predictive model developed in this manuscript proves the existence of broadly expressed tissue biomarkers that can be leveraged to discriminate between CS patients and non-CS patients, even in the absence of nearby granulomas. Though our single-center cohort was modestly sized, the compelling modeling results represent an important step towards establishing a diagnostic paradigm for CS that is more resilient to the sampling error than conventional histologic diagnosis. While our current workflow relies on manual ROI selection to ensure tissue quality, we acknowledge that this introduces operator-dependent bias and may limit scalability. To address this, future implementations will likely benefit from automated or semi-automated ROI selection strategies guided by predefined quality thresholds—an increasingly common solution in digital pathology workflows.

Although clinical translation of the GeoMx spatial technology used in this report is impractical due to cost and technical complexity, there are alternative diagnostic avenues worthy of consideration. Quantitative digital pathology using multi-marker immunostaining and whole-slide image analysis has been piloted in oncologic^50–52 and cardiovascular research,^48,49 and represents a more clinically viable methodology. With appropriate external validation, development of a quantitative immunopathology scoring system incorporating the full set of diagnostic model biomarkers identified in this study could substantially improve the diagnostic yield of clinical biopsies. Such a system would support earlier intervention for CS patients and more accurate cohort identification for future therapeutic trials.

In conclusion, this study represents the largest application of high-plex spatial protein profiling to date in human cardiac tissue one of the largest investigations of in situ sarcoidosis biology ever conducted. Leveraging the unique capabilities of spatial biology platforms, we reveal striking heterogeneity in granuloma composition, disease stage-associated transitions from immune activation to chronic immune exhaustion, a spatial ‘distance gradient’ of subtle immune and fibrotic signaling around granulomas, and widespread protein abundance changes that distinguish even histologically bland CS tissue from controls. Beyond these specific findings, our approaches to cohort design, ROI selection, and digital pathology-based AOI segmentation demonstrate the discovery and translational potential of emerging spatial profiling technologies. While limitations of retrospective sampling and the GeoMx platform constrain some mechanistic conclusions, our multi-layered internal validation and limited orthogonal confirmation support the robustness and reproducibility of the reported findings. Importantly, these analyses should be viewed within an exploratory, hypothesis-generating framework; definitive confirmation will require replication in independent cohorts and with complementary approaches. Future studies leveraging prospective sarcoidosis cohorts, deeper spatial profiling, single-cell assays, and systems biology approaches will be essential to build upon these findings and further elucidate the pathobiology of CS.

Supplementary data

Supplementary data are available at European Heart Journal online.

Declarations

Disclosure of Interest

Nothing to declare.

Data Availability

The data that support the findings of this study are presented in the Manuscript and Supplementary Data sections, with additional datasets available on Figshare. Unprocessed raw data and patient-level data corresponding to study tissues can be requested from the corresponding author with legitimate scientific justification and provided appropriate inter-institutional data sharing agreements are in place.

Funding

Research reported in this publication was supported by the National Heart, Lung and Blood Institute: R01HL15807101A1, K08HL159344, by the W.W. Smith Charitable Trust, by the Foundation for Sarcoidosis Research, and by the Bogle Family Foundation.

Ethical Approval

Procurement of human myocardial tissue was performed under protocols and ethical regulations approved by Institutional Review Boards at the University of Pennsylvania and the Gift-of-Life Donor Program (Philadelphia, Pennsylvania, United States, protocol 830123), and complies with the Declaration of Helsinki. Access to clinical data and archival clinical tissue was approved by the University of Pennsylvania Institutional Review Board (protocol 848421), with waiver of consent authorized by 45 CFR 46.116(d) and 45 CFR 164.512(i).

Pre-registered Clinical Trial Number

Not applicable.