Whole blood epigenomic and transcriptomic characterization identifies vulnerable molecular subtypes of chronic coronary disease

Abstract

Chronic coronary disease (CCD) remains a leading cause of morbidity and mortality worldwide. However, current clinical assessments, including tests of inducible ischemia or coronary artery disease severity poorly discriminate risk for future cardiovascular (CV) disease events among this population with established CCD. To address this gap, our study leverages high-dimensional molecular data from the ISCHEMIA (International Study of Comparative Health Effectiveness with Medical and Invasive Approaches) Trials biorepository to molecularly characterize patients with CCD. By integrating transcriptomic (N = 646) and methylomic (N = 732) data with core-lab confirmed clinical phenotyping, we describe molecular signatures associated with disease severity and identify distinct whole-blood molecular subtypes of CCD. These subtypes demonstrate differential risks of CV events, independent of traditional clinical risk scores, and have distinct molecular and immune profiles. Validation of the transcriptomic and methylomic subtypes in two independent external cohorts confirms the clinical relevance and generalizability of our findings. These findings underscore the potential of blood-based multi-omic approaches to refine risk stratification, improve personalized treatment strategies and advance secondary prevention in CCD. Clinical Trial Registration: ClinicalTrials.gov identifier: NCT01471522; https://clinicaltrials.gov/ct2/show/NCT01471522.

Blood-based RNA and DNA methylation profiles reveal chronic coronary disease subtypes with distinct risk of cardiovascular events, improving cardiovascular risk assessment with validation in independent cohorts.

Introduction

Chronic coronary disease (CCD) is the leading cause of morbidity and mortality worldwide, affecting nearly 20 million Americans and resulting in 365,000 deaths annually¹. While the basic mechanisms and contributing factors of CCD are well characterized^2,3, the heterogeneity of risk for patients with CCD has made risk stratification and prognostication of future cardiovascular (CV) events a challenge. As such, the positive predictive value across CAD is poor⁴. Current assessment of risk for patients with CCD includes evaluation of achieving risk factor modification and non-invasive and invasive testing, but does not take into account molecular markers for essential processes underlying atherosclerosis, angina, and thrombosis. Assessment of inducible ischemia is commonly performed using imaging and non-imaging modalities, but contemporary data demonstrate uncertainty regarding the risk of CV outcomes based on severity of ischemia in those with established CCD^5–8. While coronary artery disease severity is associated with adverse CV outcomes, most patients with CCD do not experience a CV event. This gap in clinical assessments in CCD demonstrates a critical need for improved prognostic and predictive assays that may be developed using blood-based multi-omic approaches.

In this study, we used patients from the ISCHEMIA (International Study of Comparative Health Effectiveness with Medical and Invasive Approaches) and ISCHEMIA Chronic Kidney Disease (CKD) Trials biorepository (N = 1064) with at least moderate inducible ischemia and obstructive coronary artery disease to molecularly characterize patients with CCD^9–13.

Our analytic approach combines transcriptomic, methylomic, and core-lab confirmed clinical phenotyping to identify a multimodal molecular signature of CCD and CV event risk. A particular strength of our approach is the ability to link core-lab confirmed severity of ischemia and coronary artery disease with adjudicated CV events, providing a unique opportunity to use whole blood (WB) molecular markers for association and prediction of CV phenotypes and events. Using a combination of supervised and unsupervised approaches, we identified distinct WB transcriptomic and methylomic subtypes of CCD. Together, these findings advance risk stratification, prognostication, and our understanding of CCD disease state beyond state-of-the-art clinical testing.

Results

Description of the ISCHEMIA omics cohort

The ISCHEMIA Trials biorepository includes biospecimens from 1064 participants enrolled in the ISCHEMIA and ISCHEMIA-CKD trials, which recruited patients with chronic coronary disease and moderate-to-severe ischemia on stress testing. Of the 1064 participants, 779 patients provided whole blood (WB) samples of RNA-seq or methylation data (646 samples for WB RNA-seq and 732 samples for WB methylation)^13,14. Overall, 599 participants had both RNA-seq and methylation data available (Fig. 1A–B). Demographic information and clinical variables including severity of inducible ischemia and coronary artery disease are noted in Table 1 and Supplementary Fig. 1. Among participants with RNA-seq or methylation data available (N = 779; referred to as the biorepository cohort), the median (25^th, 75^th percentile) age was 66.9 (60.6, 72.1), 19% were female, 79% were non-Hispanic White, 11.2% Black or African American, and 5% Hispanic or Latino (Table 1).

Fig. 1. Schematic overview of molecular subtype discovery in ISCHEMIA Biorepository and validation cohorts.

A Schematic of the ISCHEMIA biorepository multi-omic workflow. B Venn diagram of the overlap of the samples with RNA-seq data available and samples with methylation data available. C Schematic overview of multi-omic molecular subtype discovery in ISCHEMIA biorepository and external validation in PACE and YFS cohorts. Panel A and C were created in https://BioRender.com/ys0a3l1, licensed under CC BY 4.0.

Table 1.

Description of the ISCHEMIA biorepository and key clinical measures

Demographics	Feature	RNA or Methylation (N = 779)	RNA (N = 646)	Methylation (N = 732)	Both RNA and Meth (N = 599)
Age	median (x±iqr)	66.98 (60.57–72.1)	67.17 (61.14–72.09)	67 (60.56–72.01)	67.27 (61.15–72)
Sex	Female, N (%)	148 (19)	126 (19.5)	140 (19.1)	118 (19.7)
Race	American Indian or Alaska Native, N (%)	6 (0.8)	4 (0.6)	6 (0.8)	4 (0.7)
	Asian, N (%)	18 (2.3)	18 (2.8)	17 (2.3)	17 (2.8)
	Black or African American, N (%)	87 (11.2)	69 (10.7)	81 (11.1)	63 (10.5)
	Multiple Races, N (%)	4 (0.5)	4 (0.6)	4 (0.5)	4 (0.7)
	Native Hawaiian or Other Pacific Islander, N (%)	10 (1.3)	9 (1.4)	9 (1.2)	8 (1.3)
	White, N (%)	651 (83.6)	539 (83.4)	612 (83.6)	500 (83.5)
	Unknown, N (%)	3 (0.4)	3 (0.5)	3 (0.4)	3 (0.5)
Ethnicity	Hispanic or Latino, N (%)	39 (5)	37 (5.7)	36 (4.9)	34 (5.7)
	Not Hispanic or Latino, N (%)	730 (93.7)	599 (92.7)	686 (93.7)	555 (92.7)

Clinical Characteristics	Feature	RNA or Methylation (N = 779)	RNA (N = 646)	Methylation (N = 732)	Both RNA and Meth (N = 599)
BMI	median (x±iqr)	29.4 (26.4–33.4)	29.4 (26.5–33.2)	29.4 (26.3–33.4)	29.4 (26.5–33.3)
Hypertension	N (%)	655 (84.1)	536 (83)	621 (84.8)	502 (83.8)
Diabetes	N (%)	343 (44)	274 (42.4)	325 (44.4)	256 (42.7)
CKD	N (%)	89 (11.4)	66 (10.2)	84 (11.5)	61 (10.2)
History of Dialysis	N (%)	43 (5.5)	26 (4)	40 (5.5)	23 (3.8)
History of Prior MI	N (%)	181 (23.2)	139 (21.5)	169 (23.1)	127 (21.2)
Smoking Status	Current, N (%)	85 (10.9)	73 (11.3)	81 (11.1)	69 (11.5)
	Former, N (%)	426 (54.7)	355 (55)	398 (54.4)	327 (54.6)
	Never, N (%)	268 (34.4)	218 (33.7)	253 (34.6)	203 (33.9)

Baseline Labs	Feature	RNA or Methylation (N = 779)	RNA (N = 646)	Methylation (N = 732)	Both RNA and Meth (N = 599)
LDL-C, mg/dL	median (x±iqr)	78 (60–105)	78 (60–104)	78 (60–105)	77 (60–103)
HDL-C, mg/dL	median (x±iqr)	43 (36–52.98)	42.92 (35.19–52.2)	43 (36–52.59)	42.92 (35.19–52)
Triglycerides, mg/dL	median (x±iqr)	127.72 (95–185.84)	126 (96–184.04)	127.43 (94–184.07)	124.39 (95.43–182.52)
Total Cholesterol, mg/dL	median (x±iqr)	150.91 (130–180.59)	150.81 (129.74–179)	150.81 (129.93–179.81)	150.62 (129.16–178.27)
SBP, mmHg	median (x±iqr)	130 (120–144)	130 (120–143)	130 (120–143)	130 (120–142)
DBP, mmHg	median (x±iqr)	75 (68–80)	74 (68–80)	75 (68–80)	74 (68–80)
HbA1c, %	median (x±iqr)	6.27 (5.7–7.3)	6.2 (5.7–7.26)	6.3 (5.7–7.3)	6.25 (5.7–7.22)
LVEF	median (x±iqr)	60 (54–65)	60 (54.5–65)	60 (54–65)	60 (54.75–65)
eGFR, ml	median (x±iqr)	75 (59–91)	75 (60–91)	75 (59–91)	75 (60–91)

Baseline Medications	Feature	RNA or Methylation (N = 779)	RNA (N = 646)	Methylation (N = 732)	Both RNA and Meth (N = 599)
Statin	No, N (%)	57 (7.3)	37 (5.7)	50 (6.8)	30 (5)
	Yes, N (%)	722 (92.7)	609 (94.3)	682 (93.2)	569 (95)
Aspirin	No, N (%)	42 (5.4)	32 (5)	40 (5.5)	30 (5)
	Yes, N (%)	737 (94.6)	614 (95)	692 (94.5)	569 (95)

Assessment of Ischemia	Feature	RNA or Methylation (N = 779)	RNA (N = 646)	Methylation (N = 732)	Both RNA and Meth (N = 599)
Imaging Degree of Ischemia	Severe, N (%)	302 (38.8)	243 (37.6)	284 (38.8)	225 (37.6)
	Moderate, N (%)	354 (45.4)	289 (44.7)	338 (46.2)	273 (45.6)
	Mild / None, N (%)	72 (9.3)	65 (10.0)	62 (8.5)	55 (9.1)
	Uninterpretable, N (%)	51 (6.5)	49 (7.6)	48 (6.6)	46 (7.7)
Number of Diseased Vessels (70%)	3 v, N (%)	63 (8.1)	48 (7.4)	61 (8.3)	46 (7.7)
	2 v, N (%)	88 (11.3)	74 (11.5)	80 (10.9)	66 (11)
	1 v, N (%)	139 (17.8)	114 (17.6)	132 (18)	107 (17.9)
	0 v or Not Evaluable, N (%)	238 (30.5)	212 (32.8)	226 (30.8)	200 (33.4)
	Not Available, N (%)	251 (32.2)	198 (30.7)	233 (31.8)	180 (30.1)

Events	Feature	RNA or Methylation (N = 779)	RNA (N = 646)	Methylation (N = 732)	Both RNA and Meth (N = 599)
Maximum Follow-up	o	3.58 (2.41–4.85)	3.39 (2.26–4.38)	3.67 (2.49–4.89)	3.43 (2.43–4.43)
5-Point Composite Outcome	0, N (%)	622 (79.8)	519 (80.3)	588 (80.3)	485 (81)
	1, N (%)	142 (18.2)	117 (18.1)	129 (17.6)	104 (17.4)
	2, N (%)	15 (1.9)	10 (1.5)	15 (2)	10 (1.7)
CVD / MI	0, N (%)	641 (82.3)	535 (82.8)	605 (82.7)	499 (83.3)
	1, N (%)	122 (15.7)	100 (15.5)	111 (15.2)	89 (14.9)
	2, N (%)	16 (2.1)	11 (1.7)	16 (2.2)	11 (1.8)
ACD	0, N (%)	714 (91.7)	596 (92.3)	670 (91.5)	552 (92.2)
	1, N (%)	65 (8.3)	50 (7.7)	62 (8.5)	47 (7.8)
CVD	0, N (%)	714 (91.7)	596 (92.3)	670 (91.5)	552 (92.2)
	1, N (%)	47 (6)	38 (5.9)	44 (6)	35 (5.8)
	2, N (%)	18 (2.3)	12 (1.9)	18 (2.5)	12 (2)
ACD / MI	0, N (%)	640 (82.2)	535 (82.8)	604 (82.5)	499 (83.3)
	1, N (%)	139 (17.8)	111 (17.2)	128 (17.5)	100 (16.7)
MI	0, N (%)	644 (82.7)	538 (83.3)	608 (83.1)	502 (83.8)
	1, N (%)	93 (11.9)	77 (11.9)	82 (11.2)	66 (11)
	2, N (%)	42 (5.4)	31 (4.8)	42 (5.7)	31 (5.2)

BMI body mass index, CKD chronic kidney disease, MI myocardial infarction, LDLC low-density lipoprotein cholesterol, HDLC high-density lipoprotein cholesterol, LVEF left-ventricular ejection-fraction, EGFR estimated glomerular filtration rate, CVD cardiovascular death, ACD all cause death.

All participants in the ISCHEMIA biorepository had site determined moderate to severe ischemia and CCD and underwent core-lab assessment to quantitate inducible ischemia and extent of obstructive coronary artery disease^9,11. Inducible ischemia by stress imaging was severe in 302 (39%), moderate in 354 (45%) and mild/none for 72 (9%) of participants. Among the 779 participants with methylation or RNA data available, 251 (32%) were not required to have CCTA performed due to kidney impairment¹⁵; 64 (8%) had nonobstructive CAD < 70% stenosis and 174 (22%) were non-evaluable for determination of number of diseased vessels by CCTA due to technical limitations including motion artifact, calcification, or image quality¹⁵. Of the 290 participants with a CCTA performed and evaluable for number of diseased vessels, 139 (48%) had 1-vessel coronary artery disease (CAD), 88 (30%) 2-vessel CAD, and 63 (22%) had 3-vessel CAD (Table 1)^13,15.

Molecular drivers of ischemia severity in ISCHEMIA

Assessment of severity of inducible ischemia and coronary atherosclerosis remains a cornerstone for the diagnosis, risk stratification, and management of patients with CCD. Using our omics dataset, we first sought to identify molecular features that differ by ischemia and coronary atherosclerosis severity. To do this, we completed two sets of differential expression for ischemia severity and coronary atherosclerosis severity using transcriptomics and methylation data, independently (Supplementary Fig. 2). Baseline characteristics by groups are described in Supplementary Data 1a.

Comparing the transcriptome of patients with severe vs mild/no ischemia by stress imaging, we found 1,933 differentially expressed genes between groups. (p-adj <0.05; log2fc > 0.25: 1,129; log2fc < −0.25: 804, Supplementary Fig. 2a, Supplementary Data 1a). Based on Gene Set Enrichment Analysis (GSEA), patients with severe ischemia compared with those with mild/no ischemia, had pathways enriched in immunoglobulin complex (NES = 3.7; p-adj <0.001), antigen binding (NES = 3.4; p-adj <0.001), B cell mediated immunity (NES = 2.8; p-adj <0.001), and innate immune response (NES = 2.5; p-adj <0.001, Supplementary Fig. 2b, Supplementary Data 2a).

Investigating the WB methylome, patients with severe compared with mild/no ischemia, differential analysis revealed 643 differentially methylated probes, (p-adj <0.05; deltaBeta > 0.03: 55; deltaBeta < −0.03: 588, Supplementary Fig. 2c, Supplementary Data 1b). Probe Set Enrichment Analysis (PSEA) identified differentially methylated pathways (p-adj <0.01) including immunoglobulin receptor binding (NES = −1.89), B cell differentiation (NES = −1.5), and platelet/thrombotic response pathways including icosanoid binding (NES = −1.82), platelet aggregation (NES = −1.4); and blood coagulation (NES = 1.56, Supplementary Fig. 2d, Supplementary Data 2b).

Next, we investigated association of severity of ischemia with gene expression and CpG site methylation. 666 pathways were significantly altered at both the GSEA and PSEA level (p-adj <0.05) (Fig. 2A). Of these, 633 gene sets were found to be both hypomethylated and increased transcriptionally, consistent with traditional paradigms of hypermethylation and transcriptional inhibition¹⁶. These included pathways of immune system development, apoptotic signaling, and B cell differentiation (Fig. 2A).

Fig. 2. Pathway-level integration of differential gene expression and methylation analyses for ischemia and CAD severity.

A Scatterplot of all GO terms with their respective RNA-seq GSEA NES values and methylation PSEA NES values for ischemia severity (severe vs mild/none); significant pathways (padj <0.05) are highlighted. B Scatterplot of RNA-seq log-fold changes and methylation deltaBeta values of gene-probe pairs from the B Cell Differentiation pathway, highlighting three genes of interest. Correlation between the methylation and expression values is shown by color and size. C Scatterplot of all GO terms with their respective RNA-seq GSEA NES values and methylation PSEA NES values for CAD severity (3-vessel vs 1-vessel). D Scatterplot of RNA-seq log-fold changes and methylation deltaBeta values of gene-probe pairs from the Leukocyte Mediated Immunity pathway, highlighting three genes of interest. Correlation between the methylation and expression values is shown by color and size. GSEA and PSEA were performed with a two-sided weighted Kolmogorov–Smirnov-like statistic. P-values were adjusted for multiple comparisons using the Benjamini–Hochberg method.

Consistent with this finding, many studies have highlighted a role of specific B cell subtypes in atherosclerosis^17,18 and identified B cells as a key immune mediator in atherosclerotic lesion development acting through cytokines and Ig production¹⁹. Based on our data and the supporting findings, we isolated all annotated gene-probe pairs for genes annotated in the B-cell differentiation gene set (Fig. 2B) and compared the log2 fold change from differential expression with delta beta differential methylation values for patients with severe compared with mild/no ischemia. Overall, this revealed hypomethylation and increased transcript expression in B cell genes including SPI1, NFAM1 and ADGRG3—all of which have been implicated in B cell activation (Fig. 2B)^20,21.

Molecular drivers of coronary artery disease severity in ISCHEMIA

Following a similar analytic approach, we then assessed molecular drivers of CAD severity, comparing patients with triple-vessel (3 V) vs single-vessel (1 V) CAD. Baseline characteristics by groups are described in Supplementary Data 1b. This analysis found no significant differences (p-adj <0.05) in the differential analysis of WB transcriptomic data (Supplementary Data 1c); however, GSEA identified similar pathways as those found in our ischemia severity analysis, with a few notable additions centering around neutrophilic and granulocytic response. Significant pathways (p-adj <0.01) included antigen binding (NES = 2.62), specific granule lumen (NES = 2.04), B cell mediated immunity (NES = 2.29) and complement activation (NES = 2.49) (Supplementary Fig. 2f, Supplementary Data 2c).

Differential analysis of methylation of CAD severity also did not identify significantly differential probes (p-adj <0.05, Supplementary Fig. 2g). PSEA of genes associated with our differentially methylated probes did identify significantly downregulated pathways (p-adj <0.001), however, including neutrophil mediated cytotoxicity (NES = −2.13), immunoglobulin receptor binding (NES = −2.09) and thrombin activated receptor response (NES = −1.88; Fig. 2H, Supplementary Data 2d).

RNA-methylation integration analysis for CAD severity was performed as was done for ischemia severity, to identify coordinated pathway-level changes across the transcriptome and methylome. 35 pathways were significantly hypomethylated and transcriptionally increased (p-adj <0.05; Fig. 2C), including innate immune response and leukocyte mediated immunity. Within innate immune response, we found profound changes in several pathways related to phagocytosis (phagocytosis NES = 1.81; phagocytosis recognition NES = 2.60; Phagocytic Vesicle NES = 1.92) and complement activation (NES = 2.48) (Supplementary Data 2c). Focusing on leukocyte mediated immunity, we isolated all annotated gene-probe pairs from this pathway, comparing the log2 fold change from differential expression with the delta beta from differential methylation for patients with 3-vessel vs 1-vessel disease. We identified decreased pathway methylation and corresponding increased transcript expression of granulocytic cell response genes, including AZU1, ELANE, and CD177 (Fig. 2D).

Transcriptomic molecular subtyping of the ISCHEMIA cohort

Next, we applied an unsupervised clustering approach to identify RNA- and methylation-based subtypes to elucidate the molecular underpinnings of CCD. In our transcriptomics data, patient samples were clustered using non-negative matrix factorization (NMF)²², and gene features were clustered using weighted gene correlation network analysis (WGCNA)²³ to create co-regulated gene modules. This resulted in 4 RNA-seq subtypes (RS1, RS2, RS3, RS4) and 15 gene modules (annotated by their respective colors) (Fig. 3A).

Fig. 3. RNA-seq based subtyping of ISCHEMIA patients.

A Heatmap of ISCHEMIA samples with RNA-seq and their respective gene expression. The heatmap is pre-separated by molecular subtype and gene module based off unsupervised clustering methods (B) Heatmap of the RNA-seq cohort reducing genes down to single composite eigengene values representative of expression of the gene module. The side annotation indicates the rank of the expression of the module for each subtype with 4 as the highest and 1 as the lowest. Gene modules are annotated using enrichment tests of the genes that make up each module. C Heatmap of enrichment tests for each RS vs the rest of the population. All non-gray cells indicate statistically significant associations (Two-sided Fisher’s exact test p < 0.05), with the color scale representative of the respective odds ratio. D Heatmap of Kaplan–Meier (KM) for each event comparing each RS vs the rest of the population. All non-gray values have a KM log-rank p < 0.05 with the color representative of the p-value and directionality of the survival test result. E KM event-free survival analysis for all RSs comparing the major adverse cardiovascular events (MACE) composite outcome calculated using a log rank p-value. F KM plot for RS1 vs not RS1 and (G) RS2 vs not RS2 for the MACE outcome. H KM plot for RS4 vs not RS4 for the outcome of cardiovascular (CV) death. Survival curves were compared using a two-sided log-rank test. Adjusted p-values were derived from Cox proportional hazards regression adjusted for age, sex, race, and ethnicity.

We then investigated enrichment of clinical characteristics within each molecular subtype. Transcriptomic subtype RS1 was found to have over-representation of female participants, higher HDL-C, and severe ischemia. Subtype RS2 was enriched with older participants, White race, higher HDL-C concentration while subtype RS3 was enriched for a higher LDL-C and a lower HDL-C. RS4 showed enrichment for a lower eGFR (indicative of impaired kidney function), hypertension and diabetes. Molecular clustering did not identify a clear pattern of enrichment by severity of ischemia and CAD severity (Fig. 3C), suggesting that a transcriptomic-based approach may identify molecular signatures underlying clinical risk that are complimentary to other testing modalities.

We next assessed differences in known markers of risk and probability of CV events in our cohort based on transcriptomic subtypes (see Supplementary Table 1 for definitions of CV events). Here, we observed significant differences in transcriptomic subtypes with different CV events (log-rank p < 0.05) including composite outcomes of CV death (CVD) or myocardial infarction (MI), and all cause death (ACD) or MI (Fig. 3D). Patients with RS1 subtype had fewer events (Fig. 3E–F) while patients in subtypes RS2 and RS4 had greater risk of events, Fig. 3E, G, H). More specifically, we found that subtype RS2 was at higher risk of Major Adverse Cardiovascular Event (MACE), CV death or MI, ACD or MI, and MI while patients with RS4 were at higher risk of death when compared to the rest of the population (Fig. 3D). Transcriptomic subtypes RS2 and RS4 had a greater risk of CV events uncaptured by measures of ischemia or CAD severity. (Fig. 3C–D). Interestingly, established biomarkers of CV risk, including hsTNT, GDF-15, NT-proBNP, hsCRP and Cystatin C were all found to be elevated in the high risk RS4 but not in RS2, indicating that the molecular signal identified in this transcriptomic subtype provides information beyond previously established molecular blood risk markers (Supplementary Fig. 3).

To better understand molecular signatures associated with each molecular subtype differential analysis of each subtype was completed (Supplementary Fig. 4a–c) and weighted gene co-expression network analysis (WGCNA)²³ was utilized to find gene modules enriched for each subtype (Fig. 3B). Gene modules were annotated using GSEA and GO terms encompassing the strongest overall signature for each module are shown. A full list of enriched GO terms can be found in Supplementary Data 3a. RS1 and RS2 were enriched in immune processes, most notably the adaptive immune system, including B cell and T cell response (RS1), and innate immune response (RS2). RS4 was enriched for a hypercoagulable and prothrombotic state with upregulation of pathways associated with wound healing, coagulation, and heme metabolism (Fig. 3B).

Cell type deconvolution within subtypes identified a significant enrichment of regulatory Tregs, activated NK cells, and plasma B cells in RS1, aligning with the adaptive immune pathways identified by WGCNA. In contrast, RS2 showed increased neutrophil, M0 and M1 macrophage populations, indicative of heightened innate immunity, again consistent with the gene module enrichment (Supplementary Fig. 3). These findings highlight distinct immune signatures within our subtypes and underscores key differences in immune regulation that likely contributes to subtype-specific cardiovascular risk and therapeutic strategies.

Methylation based molecular subtyping of the ISCHEMIA biorepository

A similar molecular subtyping framework was applied in our methylation cohort. Unsupervised clustering yielded 3 Methylation Subtypes (MS) and WGCNA analysis resulted in 13 co-regulated probe modules (Fig. 4A, Supplementary Data 3b. Clinical features (Fig. 4C) and known CV risk biomarkers (Supplementary Fig. 3) did not differ significantly across methylation subgroups. Calculating the incidence of CV events by methylation subtype, we found that MACE differed significantly between methylation subtypes (Fig. 4D–E), with fewer events in MS1 and MS2, and more events in MS3 (Fig. 4F–H). Specifically, patients with subtype MS2 had significantly fewer CV deaths or MIs, fewer all-cause deaths or MIs, and fewer MIs alone when compared to non-MS2 patients. Patients with subtype MS3 were at higher risk for all events (Fig. 4E).

Fig. 4. Methylation based subtyping of ISCHEMIA patients.

A Heatmap of methylation probes for ISCHEMIA biorepository samples. The heatmap is pre-separated on columns by molecular subtype based on unsupervised clustering and (B) rows show probe based WGCNA modules as composite eigengene values representative of expression of each gene module. Side annotation indicates the rank of the methylation expression of the module for each subtype with 3 as the highest and 1 as the lowest. Probe modules are annotated using enrichment tests of the genes closest to the probes that make up each module. C Heatmap of enrichment tests for each MS vs the rest of the population. All non-gray cells indicate statistically significant associations (Two-sided Fisher’s exact test p < 0.05), with the color scale representative of the respective odds ratio. D Heatmap of Kaplan–Meier (KM) results for each event comparing each MS vs the rest of the population. All non-gray values have a KM log-rank p < 0.05 with the color representative of the p-value and direction of the survival test result. E KM event-free survival analysis for all MSs comparing the major adverse cardiovascular events (MACE) composite outcome calculated using a log rank p-value. F KM plot for MS1 vs not MS1 for the MACE outcome. G KM plot for MS2 vs not MS2 and (H) MS3 vs not MS3 for cardiovascular (CV) death or myocardial infarction (MI). Survival curves were compared using a two-sided log-rank test. Adjusted p-values were derived from Cox proportional hazards regression adjusted for age, sex, race, and ethnicity.

Molecular characterization of subtypes found that MS1 was characterized by significant hypermethylation of pathways related to cell signaling, inflammation and immune regulation. In contrast, MS3 showed hypermethylation of immune processes, specifically B cell response, T-cell differentiation/activation, and platelet activation while MS2 demonstrated hypermethylation of pathways related apoptosis and cell cycle processes (Fig. 4B). Methylation probes significantly associated with each subtype are shown in Supplementary Fig. 4.

Overall, cell deconvolution of methylation subtypes showed fewer differences in immune cell populations compared to transcriptomic subtypes. Specifically, M3 showed slight reductions in CD8 and CD4 memory resting T cells, corresponding to hypermethylation of T-cell pathways in this subtype (Supplementary Fig. 3). These finding indicate that while there are detectable changes in immune cell proportions by methylation subtypes, these changes are less pronounced when compared to transcriptomic subtypes.

Combining molecular subtypes

Given our findings that molecular features at the transcript and methylation level may independently drive clinically relevant subtypes, we hypothesized that integration of transcriptome and methylation subtypes could identify CCD patients at differential risk of cardiovascular events. We observed that a patient’s RS and MS subtypes are not independent, suggesting a potential concordance of signatures (chi square p = 9.21e-11, Fig. 5A). Testing for clinical enrichment on combinations of RS + MS we found many small but significant changes, particularly in race, ethnicity, hemoglobin count, HDL, and CAD severity (Fig. 5B).

Fig. 5. Combinatorial analysis of ISCHEMIA RNA-seq and methylation subtypes.

A Alluvial plot depicting the overlap of RS and MS for the ISCHEMIA omic repository. The accompanying table shows the raw number of patients that have each combination of subtypes. B Heatmap of enrichment tests for each combination of molecular subtypes vs the rest of the population. All non-gray cells indicate statistically significant associations (Fisher’s exact test p-value < 0.05), with the color scale representative of the respective odds ratio. C Heatmap depicting the KM results for each event comparing each RS vs the rest of the population. All non-gray values have a KM log-rank p < 0.05 with the color representative of the p and direction of the survival test result. D KM plot for CS1 vs not CS1 for the MACE composite outcome. E KM plot for CS2 vs not CS2 for cardiovascular death. Outlier analysis on (F) CS1 vs. not-CS1 at the methylomics level (FDR < 0.0001). and (G) CS2 and not-CS2 at the transcriptomics level (FDR < 0.05). Survival curves were compared using a two-sided log-rank test. Adjusted p-values were derived from Cox proportional hazards regression adjusted for age, sex, race, and ethnicity.

Focusing on transcriptomic-methylomic subtype-specific differences in CV outcomes, we identified two distinct high-risk combination subtypes. The first, RS2 + MS3 (hereafter combination subtype 1, CS1) demonstrated significantly increased risk across multiple events, including MACE (Fig. 5D, defined in Supplementary Data 1), CV death/MI, all-cause death/MI, MI, and all-cause death (Fig. 5C). The second high-risk group, RS4 + MS3 (hereafter, combination subtype 2, CS2) showed increased risk specifically for CV death and all-cause death (Fig. 5C, E). In contrast, the group of participants belonging to both low risk subtypes, RS1 + MS2, had a lower risk of MI (Fig. 5C).

Despite their elevated risk profiles, neither CS1 nor CS2 strongly associated with traditional clinical risk scores like the Framingham Risk Score²⁴ or the 10-year ASCVD pooled cohort equations²⁵, indicating an orthogonal axis of risk uncovered using this molecular approach (Supplementary Fig. 3). However, the two high-risk subtypes showed distinct biomarker profiles, as CS2 showed significant elevations in established CV risk biomarkers including hsTNT, GDF-15, NT-proBNP, hsCRP and Cystatin-C, whereas CS1 showed elevation only in myeloperoxidase (MPO) (Supplementary Fig. 3).

To better understand potential molecular drivers of these subgroups, we investigated immune cell proportions and gene modules most enriched in combination subtypes CS1 and CS2, revealing distinct biological processes underlying each subtype. Combination subtype CS1 was highest in the RNA-based yellow and red modules, which are enriched in leukocyte and neutrophil pathways and myeloid secretion, respectively (Supplementary Fig. 5). Consistent with this transcriptomic signature and selective MPO elevation, deconvolution analyses of CS1 revealed a significant enrichment in neutrophil and macrophage populations, the primary cellular sources of MPO, alongside reductions in B cell and CD4+ activated, and CD8 + T cell abundances (Supplementary Fig. 3). In contrast, CS2 exhibited high expression of RNA modules green-yellow (corresponding to wound healing, coagulation, and platelet activity), and black (corresponding to ribosomal and mitochondrial activity) (Supplementary Fig. 5). Consistent with this, we saw very few changes to deconvoluted immune populations in CS2 beyond a modest increase in M2 macrophages and reduction in CD4⁺ resting T cells (Supplementary Fig. 3). Together, these findings suggest that CS1 represents immune-mediated risk driven by myeloid activation, whereas CS2 reflects molecular changes relating to wound healing and traditional biomarkers of CV risk.

Combined subtype molecular characterization

To better understand molecular alterations occurring in these at risk subgroups, we investigated transcripts and methylated regions with extreme high or low expression in the CS1 and CS2 subtypes using outlier analysis, a highly sensitive approach to identify features with altered expression in small groups of interest compared with a larger comparison cohort²⁶. Following false discovery rate (FDR) filtering, combination subtype CS1 had significant outliers only at the methylation level (Fig. 5f, FDR < 0.0001) and CS2 only at the RNA level (Fig. 5G, FDR < 0.05). Combination subtype CS1 had 5 genes with significantly increased outlier methylated probes (PDCD1, LCK, CD247, IKZF1 and ARID5B) and 40 genes with significantly decreased probes (Fig. 5F).

Of the 40 significantly hypomethylated genes, 13 are involved in transcriptional regulation (SMAD3, NR1H3, SMARCA2, ARID1B, PML, ETV6, WT1, ZMIZ1, RARA, TRIM27, JARID2, RREB1, BRD4) and 4 in innate immunity (FGR, NOD2, CLU, PML) (Fig. 5F). The most hypomethylated gene identified by this method was ETV6 (FDR = 3e-22), a transcription factor essential for bone marrow hematopoiesis and megakaryocyte development. Interestingly, autosomal dominant variants in ETV6 may lead to hematologic malignancies, inherited thrombocytopenia, and alterations in hemostasis and thrombosis²⁷. Combination subtype CS2 transcriptome outliers included 3 genes that were increased (CTNNBIP1, CHID1, ITGA7) and 5 that were decreased (IL18, PTGDR2, RIF1, MED17, POT1) (Fig. 5G). Of these, CTNNBIP1 and IL18, have established roles in atherosclerosis and cardiac risk through Wnt signaling impacts on cardiac function²⁸ and MI-associated proinflammatory cytokine production²⁹, respectively.

To explore the translatability of these subtypes, we investigated FDA approved gene targets differentially expressed in high-risk RSs, MSs, and combination subtypes CS1 and CS2. Based on the most differentially abundant genes and probes by subtype we identified FDA-approved gene targets using the Drug Gene Interaction Database (DGIdb, v5.0) (Supplementary Fig. 6). Notably, individuals in high risk RS4 showed upregulation of CA1, encoding carbonic anhydrase 1. CA1 is inhibited by acetazolamide, a carbonic anhydrase inhibitor, that has been shown to decrease expression of vascular calcification-related genes, suggesting CA1 inhibition as a potential therapeutic for patients in this group^30,31.

High risk combination subtype CS1 was characterized by hypermethylation of PDCD1, which encodes the immune checkpoint PD-1 (Supplementary Fig. 6). Evidence in preclinical studies indicates that inhibition of PD-1 may protect against CV risk³² and we found that PDCD1 transcription was downregulated in all RS2 combination subtypes, including CS1, consistent with its hypermethylation (Supplementary Fig. 4). We also found hypermethylation of CD247, coding for the T-cell receptor CD3 zeta chain, a gene which has been previously reported as having reduced expression in patients with acute coronary syndrome and several other chronic inflammatory diseases^33,34. Together, these findings provide evidence for PDCD1 and CD247 as promising therapeutic targets, particularly for patients in the high-risk CS1 subtype.

Multiomic subtypes offer improved risk prediction

To explore the contribution of multiomic subtypes to risk prediction, we compared 3 models of increasing complexity: (1) a base demographic model (age, sex, race, and ethnicity), (2) a clinical risk factor model (age, sex, race, ethnicity, eGFR, dialysis, and LVEF), (3) and an -omic subtype model (age, sex, race, ethnicity, eGFR, dialysis, LVEF, and -omic subtype) for cardiovascular events (Fig. 6). A notable increase in AUC was noted from the base demographics to the risk factor model for all outcomes. We then tested the prognostic value of the multiomic subtypes by comparing the risk factor model to the -omic subtype model. While we found negligible differences according to single -omic subtypes (Supplementary Fig. 7, Supplementary Data 4), we found stark increases in AUC integrating composite subtypes (Fig. 6A). Further, we noted an increase in prediction of MACE from 0.659 [0.598–0.720] to 0.698 [0.643–0.755] (p = 0.025) with the addition of composite subtypes (Fig. 6B). Similarly, the prediction of CV Death increased from 0.705 [0.609–0.794] to 0.764 [0.675–0.846] (p = 0.09) (Fig. 6C). Overall, we found our composite subtypes increased the AUC of all outcomes by an average of 0.05 over a model comprised of traditional risk factors.

Fig. 6. Additive risk prediction of molecular subtypes for cardiovascular outcomes.

A Forest plot comparing the area under the receiver operating characteristic (ROC) curve (AUC) performance of the three models across major adverse cardiovascular events (MACE), cardiovascular death (CVD) or myocardial infarction (MI), all cause death (ACD) or MI, MI, CVD, or ACD for the composite subtypes, with a bar plot of the -log10(p-value) for the comparison of the risk factor model versus the multiomic subtype model colored by significance. B ROC curves for MACE prediction across the three models using the composite subtypes. The subtype model achieves the highest predictive performance (AUC: 0.698). C ROC curves for CV Death prediction across the three models using the composite subtypes. The subtype model achieves the highest predictive performance (AUC: 0.764). Curves were compared using a two-sided bootstrap method for two correlated ROC curves. Data are presented as AUC ± 95% confidence interval.

Validation of RNA subtypes in a separate high risk cardiovascular cohort

Since, to our knowledge, there are no well-phenotyped cohorts with established CCD that include both transcriptomic and methylomic data, we conducted separate validation analyses for each data type. Starting with transcriptomic validation of our models, to determine if our molecular subtypes have prognostic relevance in other cohorts of high-risk cardiovascular disease patients, we applied the ISCHEMIA derived RNA-subtypes to an external cohort 106 patients with symptomatic peripheral artery disease (PAD) in the PACE (Platelet Activity in Vascular Surgery and Cardiovascular Events) study³⁵. PACE recruited patients with lower extremity atherosclerosis and collected whole blood for RNA-seq prior to lower extremity revascularization, and were followed longitudinally for a median follow-up of 18 months³⁵.

We then built a subtype classification model using the ISCHEMIA cohort based on representative SingScore values derived for each transcriptomic subtype³⁶. SingScore values were calculated from the top 50 differentially expressed genes for each subtype, resulting in 4 total scores for each ISCHEMIA patient representative of their enrichment to each respective RNA Subtype. We then trained a lasso model on 70% of our ISCHEMIA patients to predict the RNA subtypes based off the 4 RNA subtype scores with ten 5-fold cross validations. The resulting model was tested on the remaining 30% of ISCHEMIA biorepository participants to determine if the SingScore model could accurately recapture the transcriptomic subtypes. This multi-classifier model achieved pairwise AUC values ranging from 0.933 to 0.998 in the 30% hold-out validation cohort (Fig. 7A, Supplementary Fig. 8).

Fig. 7. Validation of ISCHEMIA molecular subtypes in the PACE and YFS cohorts.

A Pictorial depiction of the creation and implementation of an RNA subtype (RS) prediction model using ISCHEMIA RS scores for training and applying on PACE RS score to define PACE RNA subtypes. B Heatmap depicting the KM results for each PACE event comparing each PACE RS vs the rest of the population. All non-gray values have a KM log-rank p < 0.05 with the color representative of the p-value and direction of the survival test result. C Kaplan–Meier (KM) event-free survival analysis for all PACE RS comparing the major adverse cardiovascular or limb event (MACLE) composite and (D) the death free probability. Panel A was created in BioRender. Muller, M. (2026) https://BioRender.com/6k2vgkx, licensed under CC BY 4.0. E Picturial depiction of ISCHEMIA methylation subtypes (MS1–MS3), with an ISCHEMIA-trained classifier applied to the YFS cohort to assign corresponding methylation subtypes. F KM MI-free survival analysis by age for YFS methylation subtypes (log-rank p = 0.05).

Applying this to the independent cohort of patients with PAD, we used the RS SingScore calculation to create 4 RS scores for each PACE participant. The ISCHEMIA-trained model was then used to classify each PACE participant with PAD into one of the 4 ISCHEMIA RS subtypes (Fig. 7A). Next, we investigated the association of these ISCHEMIA-derived RNA subtypes with major adverse cardiac and limb events (MACLE) in PACE (Supplementary Data 1). In this external dataset, PACE participants with predicted RNA subtype 1 (RS1) had fewer CV events, whereas those with predicted RS2 and RS4 were at higher risk for CV events, consistent with data from ISCHEMIA (Fig. 7B). A Kaplan–Meier plot for MACLE in PACE demonstrated a gradient of risk from low to high across ISCHEMIA derived RS subtypes in the PACE cohort (Fig. 7c). Subtype RS4, although small, showed a significantly high event rate with 4 of 5 patients dying within 6 months of study enrollment. (Fig. 7C–D). Further, we found that participants with predicted RS3 subtype were protected from death (Fig. 7D). Projection of our co-expressed gene modules from the ISCHEMIA cohort onto the PACE cohort also found consistency in the subtype-specific expression signatures (Supplementary Fig. 9). Collectively, our external PAD dataset was able to recapitulate the transcriptomic risk profiles that we identified in ISCHEMIA. This externally validated transcriptomic subtype-prediction is publicly available at https://github.com/ruggleslab/ischemia_multiomics and can be used to calculate the probability of external samples being a member of each RNA subgroup in other patient cohorts.

Validation of methylation subtypes in a population-based cohort

To determine if our methylation-based molecular subtypes have prognostic relevance in other cohorts, we trained a methylation subtype classification model using the ISCHEMIA cohort. The top 1000 inverse normal transformed probes were used to construct a k-nearest neighbors (KNN) classification model which was trained on 70% of ISCHEMIA biorepository participants with methylation data and tested on the remaining 30% hold-out validation cohort to determine if the probe-based signatures could accurately recapture the methylation subtypes. This multi-classifier model performed well in the internal validation cohort, demonstrating robust classification accuracy across methylation subtypes (Accuracy: 0.891).

To validate the methylation-based molecular subtypes in an independent cohort, we applied the ISCHEMIA-derived methylation subtype (MS) classification to participants in the Young Finns Study (YFS), a Finnish longitudinal study designed to investigate cardiovascular health and related complications. Using 1251 YFS participants with DNA methylation data available³⁷, we applied our ISCHEMIA methylation subtype classification to classify participants into MS1 (N = 392), MS2 (N = 319), and MS3 (N = 540). Heatmaps of CpG site methylation across clusters using inverse normal rank-normalized values demonstrated distinct methylation patterns consistent with those observed in ISCHEMIA (Fig. 7E). We then tested the association between ISCHEMIA-derived methylation subtypes and incident myocardial infarction. In unadjusted Cox proportional hazards models, MS3 demonstrated significantly higher risk compared to MS1 and MS2 subtypes (HR = 4.71 [0.98–22.66], p = 0.05; Fig. 7F–G). After adjusting for sex, this association remained nominally significant (HR = 4.48 [0.90–22.23], p = 0.07), consistent with our findings showing MS3 as at higher risk of MI (Fig. 4D).

Discussion

Clinical heterogeneity for patients with established CCD poses a significant challenge for estimating future risk of cardiovascular events. Our study addresses this challenge by leveraging the rich clinical phenotype and high-dimensional molecular omics data from the ISCHEMIA biorepository to characterize molecular features associated with CV event risk and disease severity for patients with CCD. These findings demonstrate how transcriptomic and methylation-based profiling reveals distinct molecular signatures of risk that were not accurately captured by traditional risk scores (e.g., Framingham Risk Score or ASCVD), emphasizing the utility of this approach in identifying the hidden molecular drivers of CCD.

We first characterized transcripts and methylation signatures of ischemia and CAD severity to better understand the molecular underpinnings associated with contemporary methods of risk assessment for CCD. Both comparisons demonstrated changes in the WB transcriptome and methylome, though only comparisons of ischemic severity had molecular features statistically associated with CV phenotype. Patients with severe ischemia exhibited alterations in pathways of B cell mediated immunity, immunoglobulin receptor binding and platelet activation, which were all found to be hypomethylated and transcriptionally upregulated compared to those with mild to no ischemia. When comparing patients with severe multivessel CAD to those with single vessel CAD, we observed similarly enriched genes and pathways at the RNA and methylation level, including pathways of mediated immunity, immunoglobulin receptor binding and humoral immune response. Together, these findings corroborate prior studies demonstrating a key role for the inflammasome and altered immunity in atherosclerosis^38–41 and may highlight potential benefits of inflammation targeted therapies to reduce cardiovascular risk in selected patients based in multiomic findings⁴².

We then applied an unsupervised approach to identify distinct WB molecular subtypes of CCD orthogonal to state-of-the-art clinical testing. Here, 4 transcriptomic (RS) and 3 methylomic (MS) subtypes were described, several of which associated with clinical variables and differential CV risk. Combining the RNA and methylation subtyping, we were able to further resolve this risk into two smaller combined subtype subgroups, with CS1 and CS2 found to be at particularly high risk. CS1 was found to be immune driven, with enrichment in neutrophil and macrophage populations, immune pathway signaling and elevations in MPO, while CS2 identified participants with elevations in several established CV risk biomarkers and increased expression of coagulation and wound healing pathways. The molecular signatures highlighted in these divergent high risk patient subsets provides an opportunity for a more personalized and directed medical therapy approach based on a patients WB molecular profile⁴³. Starting to address this potential, druggability analysis of CS1 and CS2 highlighted distinct subtype-specific drug target profiles. In one example, CS1 demonstrated hypermethylation of PDCD1, encoding for PD-1 protein, a T- and B-cell surface receptor integral in immune response and experimental models. Studies have shown that PD-1 deficiency accelerates atherosclerotic plaque formation^44–46 and causes T cell inactivation⁴⁷. Methylation and transcriptomic profiles of this FDA-approved target in the high risk CS1 group indicate that targeted upregulation of this gene may be a potential therapeutic avenue for CCD patients falling within this molecular subtype⁴⁸.

Importantly, risk models built using our multiomic subtypes demonstrated improved risk prediction for MACE, CV death and MI compared to models built using demographics and clinical risk factors alone. Further, we were able to validate our findings in two external cohorts, the high-risk PAD transcriptomics cohort and the Young Finns Study epigenomics cohort demonstrating the potential applicability of molecular characterization for secondary prevention in extracardiac CV disease. Together, these validations highlight the potential utility and applicability of these molecular signatures in risk assessment and therapeutic target nomination.

Several limitations should be acknowledged in interpreting these findings. The integrative analysis combining transcriptomic and methylomic subtypes resulted in relatively modest sample sizes for combination subtypes CS1 (N = 61) and CS2 (N = 15), which may limit the robustness of risk associations and statistical power for detecting effects in these groups. While we successfully validated individual transcriptomic subtypes in the PACE cohort and methylomic subtypes in the Young Finns Study, the lack of an external cohort with both transcriptomic and methylomic data precluded direct validation of our integrative combination subtypes, though the consistent performance of individual omics layers supports their biological and clinical relevance. The limited demographic diversity of our study cohort also highlights the need for validation in more racially, ethnically, and geographically heterogeneous populations. Additionally, our cross-sectional study design captures molecular signatures at a single time point, preventing assessment of temporal changes in molecular subtypes or their dynamic relationship to disease progression and treatment response. Further, we only had access to blood-based biomarkers and were unable to investigate omics from cardiac tissue, spatially resolved data or imaging data which would be complementary in mapping systemic blood markers to localized cardiac pathology. Future studies incorporating serial sampling in larger, more diverse cohorts with spatial resolution will be essential for translating these molecular subtypes into clinical practice.

In conclusion, through integrating multiomic data with rigorous clinical phenotyping and adjudicated cardiovascular events, our findings represent an important step towards a more comprehensive approach to a personalized care and secondary prevention for patients with established CCD. This study provides a molecular characterization of the ISCHEMIA cohort, revealing distinct subtypes of CCD with clinically relevant phenotypes and differential risk and unique opportunities for secondary prevention. The gene and methylation probe signatures identified within these combined subtypes not only provide important insights into the underlying mechanisms of risk but also help to advance and refine molecular diagnostics of risk prediction in patients with established CCD. These findings underscore the potential of blood-based omics to advance our understanding of secondary prevention strategies for CV events for patients with CCD, and to inform personalized therapeutic approaches.

Methods

The study was approved by the institutional review boards at NYU Grossman School of Medicine and at each participating site. All patients provided informed consent to participate in the ISCHEMIA trial.

Experimental model and study participant details

The ISCHEMIA cohort

The design and results of the ISCHEMIA trials have been previously reported^9,11,49. Briefly, ISCHEMIA enrolled patients with CAD based on the ischemia assessed by stress imaging or on exercise electrocardiography⁹. Blinded coronary computed tomography angiography (CCTA) was performed in patients to exclude patients with left main occlusion >50%, and to assess the number of diseased vessels^9,50. Blood samples were obtained at baseline within 6 weeks of the initial enrollments and prior to the assigned treatment strategy. The primary outcome investigated in our study was MACE, which refers to CV (cardiovascular) death, MI (myocardial Infarction), hospitalization for unstable angina, resuscitated cardiac arrest, or heart failure (Supplementary Table 1)⁴⁹. Prior to randomization, complete clinical data comprising history, labs, and CV events (MI, stroke, CV death) were collected. The risk profile of this dataset presents a unique opportunity to molecularly profile chronic coronary disease.

Method details

The design and primary outcomes of the ISCHEMIA trials have been previously reported^9,11,49. The ISCHEMIA trial enrolled patients with known or suspected coronary artery disease (CAD) who exhibited moderate or severe ischemia on stress imaging modalities such as echocardiography, nuclear perfusion, or cardiac magnetic resonance imaging, or severe ischemia on exercise electrocardiography⁹. Ischemia severity was categorized as mild, moderate, or severe based on the extent and degree of reversible perfusion defects or wall motion abnormalities on stress imaging, or the magnitude of ST-segment depression and time to onset during exercise testing⁹. Atherosclerosis severity was assessed by CCTA and graded by stenosis severity (<50%, 50–69%, ≥70%) and number of vessels involved⁹. Patients were scored using the Framingham Risk Score for General Cardiovascular Disease (FRS General) and Atherosclerotic Cardiovascular Disease 10-year risk Pooled Cohort Equation (ASCVD 10 yr)^24,25. The administration and function of all ISCHEMIA core labs involved in the clinical phenotyping and adjudication of clinical events are outlined below^9,11,15.

The design of the ISCHEMIA trial incorporated multiple independent core labs for clinical testing to ensure rigorous, unbiased, and standardized data interpretation across a diverse international participant cohort. These core labs were essential for maintaining internal validity and minimizing site-level variability. Core laboratories for ISCHEMIA include the stress test core, the coronary computed tomography angiography (CCTA) core, the angiographic core, the electrocardiographic (ECG) core, and the clinical event adjudication committee (CEC).

The stress core was tasked with independent interpretation of pre-enrollment stress tests used to determine ischemia severity and included the following modalities: nuclear myocardial perfusion imaging (SPECT/PET), stress echocardiography, cardiac magnetic resonance imaging (CMR), and non-imaging, ECG-based, exercise tolerance testing (ETT). The stress core lab independently validated site assessments of ischemia severity, ensured standardized interpretation criteria were applied, provided quality control and site retraining for any discrepancies, steps crucial for confirmation of eligibility and analyses based on ischemia severity⁹.

The CCTA core determined anatomic eligibility based on obstructive CAD severity¹⁵. Sites were blinded to the results of the enrollment CCTA and only informed as to the presence of obstructive CAD meeting enrollment criteria or other exclusionary criteria, as detailed elsewhere^9,15. Importantly, the CCTA and all other cores operated independently, without knowledge of clinical data or other trial-based clinical testing.

The angiographic core reviewed protocol-assigned invasive coronary angiograms, including those performed post-randomization such as a clinical event⁹. The responsibilities of the angiography core included verifying exclusion of patients with left main disease, assessing CAD burden and distribution, quality and completeness of revascularization and appropriateness of percutaneous or surgical revascularization. The ECG core lab interpreted resting and stress ECGs, particularly for patients enrolled by ETT. The ECG core confirmed ST-segment changes, exercise capacity thresholds, and trial eligibility based on predefined ischemic criteria⁹. Finally, the CEC functioned similar to a core lab by performing independent adjudication of clinical endpoints (e.g., MI, cardiovascular death, hospitalization for heart failure or unstable angina) blinded to treatment assignment in the either the invasive or conservative treatment arms⁹.

Participants with an estimated glomerular filtration rate (eGFR) of ≥30 mL/min/1.73 m² were included in the main ISCHEMIA trial, while those with an eGFR <30 mL/min/1.73 m² were included in the ISCHEMIA-CKD trial^9,51. Both trials compared an initial invasive strategy to a conservative approach and found that the invasive strategy did not reduce the risk of primary or secondary endpoints^10,11.

ISCHEMIA trials biorepository sample collection and library preparation

Venous blood samples were collected from consenting participants at baseline, within six weeks of enrollment, and prior to the initiation of the assigned treatment strategy. Plasma samples were aliquoted and stored at −70 °C or colder until data generation. Total RNA was extracted and checked for concentration and RNA integrity on the Agilent 2100 Bioanalyzer system using Nano chip cartridges. Samples were normalized to 200 ng total per sample and made into libraries on the Beckman Coulter FXP system with the Illumina automated Truseq v2 stranded mRNA kit (96rxn). Samples underwent 12 cycles of PCR amplification. Following library prep, samples were checked on the Agilent 4200 tapestation with high sensitivity DNA screentape to visualize the libraries and the concentration was determined by Quant-it quantitation kit (Invitrogen). Samples were pooled equimolar into 1 tube and run on the Illumina Novaseq 6000 sequencing system with an S2 100 cartridge kit and flow cell. The run length was paired end 50. Genome-wide DNA methylation profiling was conducted on the samples using the Illumina Infinium MethylationEPIC (850K)  BeadChip array, covering over 485,000 methylation sites. Assay performance and quality were evaluated as described below. Cardiovascular biomarkers were measured at the Uppsala Clinical Research Center Laboratory at Uppsala University (Uppsala, Sweden), accredited to SS-EN ISO 15189^52,53. Plasma biomarkers including interleukin-6 (IL-6), high-sensitivity troponin T (hsTnT), growth differentiation factor 15 (GDF-15), N-terminalpro-B-type natriuretic peptide (NT-proBNP), lipoprotein(a) (Lp(a)), high-sensitivity C-reactive protein (hsCRP), cystatin C, soluble CD40 ligand (sCD40L), myeloperoxidase (MPO), and matrix metalloproteinase 3 (MMP3) were measured from baseline samples. All biomarkers were natural log transformed, and values below the limit of detection were substituted for half the limit of detection¹³.

Transcriptomic analyses

FASTQ files for RNA-sequencing were first analyzed using the Seq-N-Slide pipeline⁵⁴. Summarized steps of the pipeline are as follows: Initially, read quality was assessed using FASTQC (v0.12.1)⁵⁵, fastqscreen (v0.15)⁵⁶, and Picard (v3.1.1)⁵⁷. FASTQ files were trimmed using trimmomatic (0.39)⁵⁸ and were then aligned to the hg38 genome using STAR(v2.7.11b)⁵⁹, and reads were quantified using featurecounts (v2.0.6)⁶⁰. All downstream analyses were performed in R (v4.4.0)⁶¹. Due to the compositional nature of RNA-seq data, hemoglobin reads made up a large proportion of reads, and were removed before all other analysis were performed. Samples had to contain a minimum number of 2,000,000 reads. For a gene to be included in downstream analysis, it had to have at least 16 reads in at least half of all samples. Counts were normalized using DESeq2 (v.1.44.0)⁶².

All figures including heatmaps, scatterplots, boxplots, alluvial plots were created using ggplot2 (v3.5.1)⁶³ and ComplexHeatmap (v2.20.0)⁶⁴. Differential expression analyses were performed using DESeq2⁶². Gene Set Enrichment Analysis (GSEA) was performed using clusterProfiler (4.12.0)⁶⁵. Outlier analysis was performed using the BlackSheep (v1.0.0)²⁶ R package. Survival analyses were performed using Survminer (v0.4.9)⁶⁶. Cell type abundances were estimated from RNA-seq data using CIBERSORT deconvolution with the LM22 leukocyte signature matrix⁶⁷.

Patient-level unsupervised analysis

Unsupervised analyses were performed using NMF (v0.27)²² clustering. Data were prepared for NMF by first z-score normalization. Since NMF requires non-negative input matrices, the z-scored data were converted to non-negative form by decomposing each matrix into two components: one with all negative values set to zero (retaining only positive values), and a second matrix with all positive values set to zero and signed removed from negative values (converting them to positive). These two matrices were then combined row-wise, resulting in a doubled feature space containing non-negative values suitable for NMF factorization⁶⁸. NMF was performed with a factorization rank of 4 for RNA-seq and 3 for methylation. These numbers were determined by testing a range of clusters between k = 3 and k = 8, where each rank was tested and evaluated using the cophenetic correlation coefficient. Final rank was chosen by the maximal cophenetic coefficient, indicating the normalized data is well represented by the clusters⁶⁸.

Weighted gene correlation network analysis

Briefly, WGCNA (v1.73)²³ was used to create gene modules using default settings. Given the large number of features in the methylation probe set (670,020) compared with RNA-Seq (13,332). WGCNA of methylation probes yielded 119 modules and were subsequently filtered for modules that had a minimum of 100 probes, and with at least one GO term with significant enrichment. Importantly, a positive eigengene or normalized expression is indicative of increases in methylation. This is notable because MSs have associated modules based on the highest degree of methylation, which by convention is associated with decreased gene expression. Probe modules were annotated using Probe based GO term analysis. (Supplementary Data 2, Fig. 5B).

Methylomic bioinformatic analyses

The IDAT files, which are in the raw format and contain intensity values of methylated site probes, were imported using the R package Champ (v2.34.0)⁶⁹. The degree of methylation was represented by the beta value of the probes⁷⁰. To ensure data quality, probes with a detection p-value greater than 0.01 were removed, as well as probes with a bead count less than 3. Other default filter options in Champ were applied. The normalization of type 1 and type 2 probes was performed using the BMIQ method⁷¹. Batch effects were removed using the R package limma (v3.60.0)⁷² with covariate control variables including sex, race, and age. Differentially methylated probe analysis was conducted using the limma package. Differentially methylated probes were utilized for pathway analysis using the Champ package.

For unsupervised analyses, WGCNA and k-means + + clustering were employed⁷³. WGCNA was used to create probe modules using default settings. Probes were annotated using GO terms from MSigDB⁷⁴ that were converted to probe sets by aligning all probes near the associated gene of the gene set. And we converted all gene in reference database to methylation probe. For k-means + +, The number of cluster centroids was determined by testing a range of clusters from 1 to 20. The best number of centroids was determined by evaluating the sum of variance within clusters. Methylation probes were annotated by their closest gene identifying 488,985 of the 670,020 probes with an annotated gene.

Probe Set Enrichment Analysis (PSEA) was performed by aligning probes with their nearest gene annotation with gene sets provided by msigdbr (v7.5.1)⁷⁵. PSEA was performed using the GSEA algorithm with probes instead of genes, and change in methylation was used as the ranking metric⁷⁶.

Machine learning analyses

All machine learning was performed using the glmnet (v4.1)⁷⁷ method in conjunction with the caret (v.6.0)⁷⁸ package. For differential signature prediction for imaging degree of ischemia and number of diseased vessel, the data was split to use 60% of data for training and 40% of data was set aside testing. For the creation of the SingScore RNA subtype score classifier, 70% of data was used for training and 30% of data was set aside for testing. Training was performed on the train-set with 10-fold cross validation repeated 5 times. To create a methylation subtype classifier, a k-nearest neighbors (KNN) model was cross validated on inverse normal transformed beta values to determine the top 1000 features for maximizing accuracy and subsequently trained on these features. The trained models were assessed on the set aside test-set, and all results reported are test-set statistics.

SingScore (v1.24.0)³⁶ was used to create the RNA subtype scores, using the UpScore from the top 50 differentially expressed gene for each RNA subtype differential expression analysis. The genes were ranked using a combination of differential expression metrics (p. value and log2foldchange) and the base mean expression of the gene. SingScore was used to collapse groups of genes into single expression values representative of the global expression of each subtype³⁶. To do this, differential analysis was performed within ISCHEMIA for each RS vs. the rest of the cohort, yielding 1000 s of differentially expressed genes per subtype. The 50 most significantly increased genes (based on p-value, log2 fold change and base mean) were chosen to represent each subtype. These 50 gene signatures were then used to derive RS Scores for each subtype, resulting in a total of 4 scores for each ISCHEMIA patient.

For the PACE validation study, the PACE dataset was treated as a completely independent validation dataset. At no point was the PACE data used for training or the initial testing in the creation of the RNA subtype prediction model. Briefly, patients with symptomatic peripheral artery disease (PAD) were recruited at NYU centers and scheduled for lower extremity revascularization. WB was collected immediately before the procedure and had WB RNA-sequencing performed on the sample. Events were followed for years afterwards and logged accordingly. MACE events were defined as death, MI, or stroke. MALE events were a combination of major amputation or surgical reintervention. MACLE included all of the listed events in MALE and MACE in addition to minor amputations (Supplementary Table 1).

For the Young Finns Study (YFS) validation study, the YFS dataset was treated as a completely independent validation dataset. The YFS is a longitudinal Finnish cohort designed to investigate cardiovascular health trajectories from childhood into adulthood. Participants (N = 1251) had Illumina EPIC array DNA methylation (DNAm) data and were assigned to methylation subtypes (MS1-MS3) using a k-nearest neighbors (KNN) classification model trained on ISCHEMIA MS signatures and applied to YFS inverse normal transformed beta-values. MI events were ascertained from Finnish national health registry data through June 2019^37,79. Cox proportional hazards regression models with pairwise subtype comparisons were performed to assess associations between molecular subtypes and incident MI in both unadjusted models and adjusted for sex.

Drug targets of multiomic subtypes

To identify therapeutically actionable genes within molecular subtypes, we probed the Drug Gene Interaction Database (DGiDb v5.0)⁸⁰. For each subtype, genes meeting significance thresholds (adjusted p < 0.05) were queried for FDA approved targets. Interaction scores represent DGiDb’s confidence metric integrating evidence strength across multiple databases, literature co-mentions, and interaction type specificity (e.g., inhibitor, antagonist, antibody, modulator). The top 10 genes per subtype were selected based on maximum significance score (-log10(adjusted p-value). For composite subtype analyses (CS1, CS2), druggability assessment was restricted to genes identified through outlier detection and subsequently filtered for targets of FDA-approved drugs to prioritize clinically translatable therapeutic opportunities.

Additive modeling of multiomic subtypes

To explore the additive value of -omic subtypes beyond traditional risk factors, we performed sensitivity analysis to compare 3 different sets of logistic regression models of increasing complexity: (1) a base demographic model (age, sex, race, and ethnicity), (2) a clinical risk factor model (age, sex, race, ethnicity, eGFR, dialysis, and LVEF), (3) and an -omic subtype model (age, sex, race, ethnicity, eGFR, dialysis, LVEF, and -omic subtype). In all models with eGFR an interaction term to eGFR and dialysis status was included to account for the influence of dialysis on eGFR measurements¹³. Models were created using a binomial logit within the glm function for outcomes of MACE, CVD or MI, ACD or MI, CVD, and ACD. Multiomic models were created using the RNA subtypes, Methylation subtypes, and composite subtypes. Receiver operating characteristic (ROC) curves were computed using the roc function with smoothing, and differences between area under the curve (AUC) were compared with the roc.test function to show significant differences between model performance. Analyses were conducted in R using the pROC (v1.18.5) package.

Additional resources

The ISCHEMIA study homepage: https://www.ischemiatrial.org/.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Author contributions

Conceptualization: M.G.C.; K.V.R.; M.A.M.; J.D.N.; J.S.B.; D.J.M.; Data curation: M.G.C.; K.V.R.; M.A.M.; J.D.N.; S.R.; H.Y.; formal analysis, M.G.C.; M.A.M.; Z.C.; D.C.; S.R.; P.S.; G.D.; H.Y.; funding acquisition, M.G.C.; K.V.R.; J.D.N.; J.S.B; J.S.H.; D.J.M.; investigation, M.G.C.; K.V.R.; M.A.M.; J.D.N.; J.S.B.; S.R.; methodology, M.G.C.; K.V.R.; M.A.M.; J.D.N.; J.S.B.; project administration, M.G.C.; M.A.M.; K.V.R.; J.D.N.; J.S.B.; resources, M.G.C.; K.V.R.; J.D.N.; J.S.B.; O.R.; T.L.; software, M.G.C.; K.V.R.; M.A.M.; Z.C.; D.C.; S.R.; supervision, M.G.C.; K.V.R.; J.D.N.; J.S.B.; D.J.M.; validation, M.G.C.; M.A.M.; J.D.N.; J.S.B.; P.S.; visualization, M.G.C.; K.V.R.; M.A.M.; Z.C.; D.C.; S.R; P.S.; writing—original draft, M.G.C.; K.V.R.; M.A.M.; J.D.N.; J.S.B.; S.R.; writing—review & editing: All Authors.

Peer review

Peer review information

Nature Communications thanks Vivek Das, Magdalena Harakalova, who co-reviewed with Rogier Veltrop, and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Funding

NIH grants U01HL105907, U01HL105462, U01HL105561, U01HL105565, R56HL1555528, R01HL165208, R35HL144993 and the Medical Scientist Research Service Award T32: GM136573. We would like to thank the NYU Genome Technology Center (GTC) and the NYU High Performance Compute (HPC) Center for providing invaluable help in the processing and preparation of samples and data analysis, respectively. This project was supported in part by Clinical Translational Science Award No. 11UL1 TR001445 from the National Center for Advancing Translational Sciences and by grants from Arbor Pharmaceuticals LLC and Astra Zeneca Pharmaceuticals LP. Devices or medications were provided by Abbott Vascular (previously St. Jude Medical, Inc); Medtronic, Inc.; Royal Philips NV (previously Volcano Corporation); and Omron Healthcare, Inc.; medications provided by Arbor Pharmaceuticals, LLC; AstraZeneca Pharmaceuticals, LP; and Merck Sharp & Dohme Corp. Its contents are solely the responsibility of the authors and do not necessarily represent official views of the National Center for Advancing Translational Sciences, the National Heart, Lung, and Blood Institute, the National Institutes of Health, or the Department of Health and Human Services.

Data availability

Due to participant privacy and consent restrictions, ISCHEMIA Biorepository molecular data are available to qualified researchers under a controlled-access data use agreement from the corresponding author(s) upon reasonable request. The YFS dataset comprises health related participant data and their use is therefore restricted under the regulations on professional secrecy (Act on the Openness of Government Activities, 612/1999) and on sensitive personal data (Personal Data Act, 523/1999, implementing the EU data protection directive 95/46/EC). Due to these legal restrictions, the data from this study cannot be stored in public repositories or otherwise made publicly available. However, data access may be permitted on a case-by-case basis upon request only. Data sharing outside the group is done in collaboration with YFS group and requires a data-sharing agreement. Investigators can submit an expression of interest to the YFS coordinator (Olli Raitakari, University of Turku, Turku, Finland, olli.raitakari@utu.fi). PACE whole blood RNA sequencing data are available in The Gene Expression Omnibus under accession number GSE310095. Source data are provided with this paper.

Code availability

All original code has been deposited at https://github.com/ruggleslab/ischemia_multiomics⁸¹ and is publicly available as of the date of publication.

Competing interests

Judith S. Hochman was PI for the ISCHEMIA trial for which, in addition to support by National Heart, Lung, and Blood Institute grant, devices and medications were provided by Abbott Vascular; Medtronic, Inc.; Abbott Laboratories (formerly St. Jude Medical, Inc); Royal Philips NV (formerly Volcano Corporation); Arbor Pharmaceuticals, LLC; AstraZeneca Pharmaceuticals, LP; Merck Sharp & Dohme Corp.; Omron Healthcare, Inc, and financial donations from Arbor Pharmaceuticals LLC and AstraZeneca Pharmaceuticals LP. She is PI for the ISCHEMIA-EXTEND trial. David J. Maron reports grants from National Heart, Lung, and Blood Institute, during the conduct of the study, independent contractor fees from Abiomed, stock in Ablative Solutions, research funding from Cleerly, and consulting fees from Regeneron. Jeffrey S. Berger reports grants from National Heart, Lung, and Blood Institute during the conduct of the study. Jonathan D. Newman reports grants from National Heart, Lung, and Blood Institute during the conduct of the study. All other authors have no conflicts to disclose.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-026-73815-5.