Myocardial Proteome in Human Heart Failure With Preserved Ejection Fraction

Abstract

Background

Heart failure with preserved ejection fraction (HFpEF) constitutes more than half of all HF but has few effective therapies. Recent human myocardial transcriptomics and metabolomics have identified major differences between HFpEF and controls. How this translates at the protein level is unknown.

Methods and Results

Myocardial tissue from patients with HFpEF and nonfailing donor controls was analyzed by data‐dependent acquisition (n=10 HFpEF, n=10 controls) and data‐independent acquisition (n=44 HFpEF, n=5 controls) mass spectrometry‐based proteomics. Differential protein expression analysis, pathway overrepresentation, weighted coexpression network analysis, and machine learning were integrated with clinical characteristics and previously reported transcriptomics. Principal component analysis (data‐dependent acquisition‐mass spectrometry) found HFpEF separated into 2 subgroups: one similar to controls and the other disparate. Downregulated proteins in HFpEF versus controls were enriched in mitochondrial transport/organization, translation, and metabolism including oxidative phosphorylation. Proteins upregulated in HFpEF were related to immune activation, reactive oxygen species, and inflammatory response. Ingenuity pathway analysis predicted downregulation of protein translation, mitochondrial function, and glucose and fat metabolism in HFpEF. Expression of oxidative phosphorylation and metabolism genes (higher) versus proteins (lower) was discordant in HFpEF versus controls. Data‐independent acquisition‐mass spectrometry proteomics also yielded 2 HFpEF subgroups; the one most different from controls had a higher proportion of patients with severe obesity and exhibited lower proteins related to fuel metabolism, oxidative phosphorylation, and protein translation. Three modules of correlated proteins in HFpEF that correlated with left ventricular hypertrophy and right ventricular load related to (1) proteasome; (2) fuel metabolism; and (3) protein translation, oxidative phosphorylation, and sarcomere organization.

Conclusions

Integrative proteomics, transcriptomics, and pathway analysis supports a defect in both metabolism and translation in HFpEF. Patients with HFpEF with more distinct proteomic signatures from control more often had severe obesity, supporting therapeutic efforts targeting metabolism and translation, particularly in this subgroup.

Nonstandard Abbreviations and Acronyms

DDA‐MS

data‐dependent acquisition

DIA‐MS

data‐independent acquisition

GO:BP

Gene Ontology: Biological Processes

HFpEF

heart failure with preserved ejection fraction

HFrEF

heart failure with reduced ejection fraction

MS

mass spectrometry

OxPhos

oxidative phosphorylation

Clinical Perspective.

What Is New?

• Proteomics of endomyocardial biopsies from patients with heart failure with preserved ejection fraction (HFpEF) have 2 distinct subgroups, one more disparate from nonfailing controls.

• Overall, HFpEF myocardium exhibits lower protein expression related to protein translation, oxidative phosphorylation and fuel metabolism, and higher expression of proteins in inflammation/immune activation.

• The subgroup with HFpEF most disparate from controls more often have severe obesity and exhibit more downregulation of protein translation, mitochondrial oxidative phosphorylation, and fuel metabolism pathways.

What Are the Clinical Implications?

• There are substantial differences in myocardial protein expression among patients with HFpEF, with those with extreme obesity more likely displaying greater differences from nonfailing controls.

• Overall, HFpEF proteomics highlights defects in mitochondrial function, protein translation, and metabolism, which suggest multiple novel therapeutic approaches.

Heart failure with preserved ejection fraction (HFpEF) will soon surpass HF with reduced EF (HFrEF) as the most prevalent form of HF. ¹ Importantly, HFpEF is associated with poor clinical outcomes, with a 5‐year survival of 24% after HF hospitalization. ² , ³ Development of therapeutics to improve morbidity and mortality has been challenging partly due to our limited understanding of underlying molecular mechanisms. This has been complicated by a major shift in the HFpEF phenotype from one characterized by systolic hypertension, left ventricular (LV) hypertrophy, and diastolic dysfunction ⁴ , ⁵ , ⁶ to one dominated by severe obesity and metabolic syndrome. ⁷ In recent studies, we identified unique myocardial abnormalities associated with obesity in HFpEF including higher oxidative phosphorylation (OxPhos) gene expression ⁸ and depressed sarcomere contractile function (to levels observed in end‐stage HFrEF). ⁹ Myocardial metabolomics suggested reduced metabolism of branched‐chain amino acids and fatty acids in HFpEF, the latter similar to HFrEF. ¹⁰ These findings both improved appreciation of the impact of severe obesity on the HFpEF heart and may provide insight into the myocardial pathobiology shared by HFpEF and HFrEF versus those that are more unique.

Although mRNA signatures can suggest disrupted cellular pathways, they do not necessarily predict protein abundance or posttranslational modifications that confer physiological effects. Furthermore, metabolomics captures a snapshot rather than activity or flux in a metabolic pathway. Myocardial proteomic data in HFpEF are currently limited to 1 study that did not find any differentially expressed proteins (after adjustment for multiple comparisons), likely due to a mixed phenotype in both the HFpEF (severe coronary artery disease and suspected inflammatory/infiltrative cardiomyopathy) and control (severe aortic stenosis and donor heart) groups. ¹¹ Accordingly, the present study made use of the Johns Hopkins University's human HFpEF myocardial tissue biobank to perform 2 independent proteomic analyses each compared with nonfailing controls. Our results show 2 distinct HFpEF subgroups, 1 being more similar to controls. Overall, both analyses of the HFpEF proteome independently highlight defects in protein translation, mitochondrial electron transport (OxPhos) and metabolic pathways, the latter most suppressed in the proteomic‐defined subgroup of HFpEF associated with severe obesity.

METHODS

Data Availability

The raw proteomics DDA‐MS and DIA‐MS data are available in Proteomics Identifications Database (PXD045677 and PXD060431, respectively) and the deidentified metadata are available in the supplemental material. Python and R scripts will be provided upon reasonable request.

Study Populations

This was a single‐center, cross‐sectional study. Patients with HFpEF referred to the Johns Hopkins University HFpEF Clinic from July 2016 to January 2020 were screened for inclusion. All patients provided informed consent and underwent research endomyocardial biopsy under a Johns Hopkins University institutional review board‐approved protocol. HFpEF diagnosis was based on consensus criteria, ² , ¹² , ¹³ , ¹⁴ as described previously. ¹⁵ Detailed inclusion and exclusion criteria are provided in Data S1. Detailed history and physical exam (including height, weight, and body mass index) were obtained at the initial clinic visit to the Johns Hopkins University HFpEF clinic. For controls in both data sets, we used unused donor hearts obtained under an institutional review board‐approved protocol from the University of Pennsylvania. In analysis where these proteomic data were integrated with transcriptomics, patients were not identical but from the same cohort.

Myocardial Tissue Procurement and Processing

Endomyocardial tissue from the right ventricular septum was obtained by a standard clinical bioptome (Jawz Bioptome, Argon Medical, Frisco, TX) in patients with HFpEF as described. ⁸ Biopsies were rapidly immersed in liquid N₂ and maintained at C−160 °C.

Proteomics Experiment, Quality Control, and Data Cleaning

Two independent mass spectrometry (MS) proteomic analyses were performed: Data‐dependent acquisition MS (DDA‐MS, n=10 controls, n=10 HFpEF) and data‐independent acquisition MS (DIA‐MS, n=5 controls, n=46 HFpEF). Details of both assays are provided in Data S1. ¹⁶ Briefly, DDA‐MS is an untargeted acquisition method that allows for discovery of differentially expressed proteins, but peptides need to be of sufficient abundance to be selected for further analysis and identification by the mass‐spectrometer. DIA‐MS is higher throughput, more consistent than DDA‐MS, and not as limited by peptide abundance since it relies upon prespecified mass‐to‐charge windows for peptide selection and further analysis. However, it generally relies upon peptide libraries unless library‐free approaches are used. The DDA‐MS analysis was used for sensitive acquisition of proteins abundant in myocardium, including myofilament, ribosomal, and mitochondrial proteins, and is presented as our primary differential expression analysis. The DIA‐MS acquisition method allowed for identification of proteins less abundant in the myocardium and for subgroup analysis of patients with HFpEF.

DDA‐MS Processing

DDA‐MS outputs were median normalized and processed with LFQAnalyst. ¹⁷ Details for this processing pipeline are provided in Data S1. Proteins with >50% missing values in both groups were excluded. Data were median normalized and imputed using the default “Missing not At Random” method, which randomly samples from a Gaussian distribution left shifted by 1.8 SDs from the mean intensity value for that protein. Normalized and imputed data were then log transformed. For statistical analysis, protein wise linear models using the Bioconductor “limma” package were used for pairwise comparisons.

DIA‐MS Processing

DIA‐MS outputs were median normalized, batch‐corrected with comBat, and processed as described previously ¹⁸ and in Data S1. Briefly, proteins with ≥50% missing values in any group (controls, HFpEF) were removed. Proteins related to blood contamination were removed as described in Data S1. Missing data were imputed using MissForest (missingpy 1.1.3) in Python (ver 3.5), which involves fitting a series of random forests to impute the missing data. ¹⁹ Imputed data were log‐transformed for normalization before subsequent analysis.

Statistical Analysis

Demographic, clinical, laboratory, and echocardiographic data were compared using a Wilcoxon rank‐sum test or Fisher's exact test for continuous and categorical variables. Statistical analyses were performed using Python or STATA version 15.1 (STATA Corp LP, College Station, TX) with a 2‐sided P value <0.05 to define significance. Principal component analysis on the imputed, log‐transformed data was performed using the sklearn package in Python. In the DDA‐MS data, 1 outlier (control) was excluded from the principal component analysis. Parallel components plots for principal components analyses (for both DIA‐MS and DDA‐MS) are provided in Figure S1. Differential protein expression analysis for the DDA‐MS data was performed using LFQ analyst, and P values were adjusted for multiple comparisons using Benjamini–Hochberg. Differential protein expression analysis for the DIA‐MS data between the 2 subgroups with HFpEF and controls was performed using a general linear model with no ties, with batch as a covariate (using nonbatch corrected data) and Benjamini–Hochberg P value adjustment. Additional analysis was performed using ordinary least squares regression with heteroskedasticity‐consistent SEs to confirm heteroskedasticity did not influence the results after log transformation. The threshold for significance was adjusted P value <0.05. Gene Ontology: Biological Processes (GO:BP) enrichment was performed in R (4.0.2) using the enrichGO function in clusterProfiler (3.18.1), restricting the background universe to proteins measured. GO:BP terms were manually curated based on the Jaccard index to avoid redundancy and function of individual proteins were interpreted. The full GO:BP enrichment is provided in Data S1 For all analyses, human genome annotation org.Hs.eg.db (2.12.0) and biomaRt (2.46.3) were used. Ingenuity pathway analysis (QIAGEN) was used to identify relevant pathways. Proteins and their corresponding genes were compared using their log₂‐fold change and P value, divided 4 four groups: (1) both higher versus controls, (2) both lower versus controls, (3) proteins higher/genes lower versus controls, and (4) proteins lower/genes higher versus controls.

Subgroups of HFpEF based on proteomic signatures from the DIA‐MS data were identified using unsupervised machine learning. Imputed and log transformed DIA‐MS data from only HFpEF patients were used for all cluster analysis. Six unsupervised machine learning algorithms were tested—K‐means clustering, ²⁰ hierarchical clustering, ²¹ Gaussian mixture model, ²² Dirichlet process model, ²³ spectral clustering, ²⁴ and nonnegative matrix factorization. ²⁵ Each model was run using between 2 and 15 clusters resulting in total 84 models, from which only 1 is chosen for the final cluster analysis. Detailed methods for implementation are shown in Data S1. The optimal model (K‐means) and cluster number (2) was determined by maximizing the mean intercluster difference while minimizing the mean intracluster difference (silhouette score). Weighted gene coexpression network analysis was used to construct protein modules based on protein–protein expression correlation in HFpEF samples using the imputed, log‐transformed DIA‐MS data (power: 9, selected from soft threshold 0.85, minimum module size: 100, dendrogram cut height: 0.20, maximum block size: 700, block size penalty: ∞). ²⁶ Clinical traits were correlated with the first principal component (ie, module eigengene) using Pearson correlation.

RESULTS

Proteomic Differences in Human HFpEF (DDA‐MS Analysis)

Clinical characteristics for the DDA‐MS proteomics data are provided in the Table and Data S1. Age (P=0.54) and sex (P=0.35) did not significantly differ between groups, and more patients with HFpEF self‐identified as Black race (60% versus 0%, P=0.01) and had hypertension (90% versus 50%, P=0.03) versus controls. Patients with HFpEF had a higher body mass index (BMI, 41 [36–45] versus 25 [21–31] kg/m², median [25%–75%], P=0.001). Invasive hemodynamics were obtained only in patients with HFpEF and values are similar to those previously reported. ¹⁰ , ¹⁵ Normalized DDA‐MS data are provided in Data S1 through S32. Figure 1A shows principal component analysis based on DDA‐MS proteomics in HFpEF (n=10) and controls (n=9). A total of 336/1960 (17.1%) proteins (volcano plot, Figure 1B) were differentially expressed; 88 (4.5%) upregulated and 248 (12.7%) downregulated versus controls. Most of the variance between groups (52%) was encoded by principal component 1 (PC₁) which separated HFpEF from controls and separated HFpEF into two sub‐groups, one closer to controls. The top 10 positive and negative contributors to PC₁ are in Figure 1C and full results for PC1 in Data S3. GO:BP enriched in proteins positively correlated with PC₁ (n=89, ie, higher in the subgroup with HFpEF more distant from controls) immune response, actin cytoskeleton organization, cell adhesion/migration, glycolysis, and response to lipids (Figure 1D, Data S4 and S5). GO:BP enriched in proteins negatively correlating with PC₁ (n=373) included DNA conformation change and ribosome biogenesis (Figure 1D). Intriguingly, the 2 subgroups were not distinguishable by clinical parameters (Table S1).

Figure 1. Proteomic signatures in HFpEF and nonfailing controls.

A, Principal component analysis of DDA‐MS proteomics from 9 nonfailing controls (black) and 10 patients with HFpEF (purple). B, Volcano plot (negative logarithm base 2 of the adjusted P value [−log₂ P _adj] vs the logarithm base 2 of the fold change [log₂FC]) showing proteins significantly lower or higher in HFpEF vs controls (black). C, Dot plot of all nonzero loadings from principal component 1 from all proteins in the data set. The x axis corresponds to the loading. The 20 most positive (blue) and most negative (red) proteins are shown. D, Corresponding enrichment of GO:BP of all proteins correlated to PC1. The x axis is the log transformation of the adjusted P value. Circle size reflects protein ratio, which is the proportion of differentially expressed proteins in the pathway vs all differentially expressed proteins. GO:BP terms were curated using the Jaccard index to avoid redundancy; the full list of GO:BP terms is found in Data S1. Color coding reflects P value after Benjamini–Hochberg adjustment for multiple comparisons. GO:BP indicates Gene Ontology: Biological Processes; HFpEF, heart failure with preserved ejection fraction; HFrEF, heart failure with reduced ejection fraction; PC1, principal component 1; and PC2, principal component 2.

Proteomic Signatures in Myocardium of HFpEF

Figure 2A compares GO:BP enrichment of up‐ and down‐regulated proteins in HFpEF versus respective controls. The full differential protein analysis is provided in Data S6. The full GO:BP enrichment is provided in Data S7 and S8. Downregulated proteins in HFpEF versus controls were enriched in processes related to ribosomal protein translation, oxidative phosphorylation, fatty acid oxidation, mitochondrial transport and organization, and fuel metabolism (encompassed in “organic acid catabolic process” and “purine nucleotide biosynthetic process”). Upregulated proteins in HFpEF versus controls were enriched in processes related to actin cytoskeletal organization, regulation of protein kinase activity and phosphorylation, G protein‐coupled receptor signaling, fatty acid biosynthesis, and inflammatory/immune response.

Figure 2. Pathways enriched by protein differences between HFpEF and respective controls (DDA‐MS).

Significantly enriched pathways were curated to diminish redundancy; the full pathway lists are provided in Data S1. Overrepresentation analysis of differentially expressed proteins is displayed. A, GO:BP enrichment of differentially expressed proteins in HFpEF vs Control. Circle size reflects the proportion of differentially expressed proteins in the pathway vs the total number of differentially expressed proteins. Color coding shows the P value after Benjamini–Hochberg adjustment for multiple comparisons. B, GO:BP pathway enrichment of differentially expressed genes/proteins with concordant or discordant directionality vs controls. Only genes with corresponding protein data were used in this analysis. Proteins and their corresponding genes were divided into 4 groups: (1) both higher expression vs controls, (2) both lower expression vs controls, (3) proteins higher/genes lower vs controls, and (4) proteins lower/genes higher vs controls. Color coding reflects P value after Benjamini–Hochberg adjustment for multiple comparisons. DDA‐MS indicates data‐dependent acquisition; GO:BP, Gene Ontology: Biological Processes; and HFpEF, heart failure with preserved ejection fraction.

Discrepancies in the HFpEF Proteome and Transcriptome

Our prior transcriptomic analysis ⁸ of HFpEF identified genes associated with electron transport and ATP synthesis (eg, OxPhos) as being upregulated, although after adjusting for BMI, these pathways were no longer enriched. Using DDA‐MS proteomic data and prior transcriptomic data ⁸ from the same 2 groups, we identified differentially expressed genes and corresponding proteins that were directionally discordant (Data S9). Pathway enrichment analysis (Figure 2B, Data S10 and S11) identified the upregulated genes with downregulated protein levels in HFpEF were enriched in oxidative phosphorylation and several metabolic pathways, including the tricarboxylic acid cycle, and fatty acid and amino acid metabolism. Downregulated genes with upregulated protein levels in HFpEF were enriched in actin and cytoskeletal structure (actin filament polymerization, cell adhesion), regulation of ion transport, and activation of the immune response.

Defective Mitochondrial Function, Fuel Metabolism, Peroxisome Proliferator‐Activated Receptor Signaling, and Protein Synthesis in HFpEF

To further identify critical pathways and their associated upstream modulators altered in subgroups with HFpEF versus controls, we employed Ingenuity pathway analysis. ²⁷ This identified inhibition of PPAR $\gamma$ (peroxisome proliferator‐activated receptor gamma) as a node for multiple defects, including proteins involved in actin cytoskeletal signaling, mitochondrial dysfunction, disordered glucose metabolism, defective protein synthesis, and nucleotide metabolism (Figure S2). This is consistent with the GO:BP enrichment presented in Figure 2.

Proteomic Signatures in Subgroups With HFpEF (DIA‐MS Analysis)

In the second analysis, we performed DIA‐MS of myocardium using a larger cohort of patients with HFpEF (n=44) versus 5 nonfailing controls (Table, Data S12). No differences were observed in relevant characteristics between controls used in DIA‐MS versus DDA‐MS or HFpEF used in DIA‐MS versus DDA‐MS. (Table S2). A total of 2175 proteins were quantified and passed quality control benchmarks. Normalized DIA‐MS data are provided in Data S13. Unsupervised clustering of subgroups with HFpEF using the proteome identified 2 groups. Figure 3A shows principal component analysis identifying a Group 1 (blue, n=31) with a proteomic signature closer to controls versus Group 2 (red, n=13) which is shifted principally along PC₁. In total, 545 (25%) proteins were differentially expressed between Group 1 and controls, 1294 (59%) between Group 2 and controls, and 1643 (76%) between subgroups with HFpEF (Data S14 and S15). Volcano plots are shown in Figure S3. GO:BP enrichment of differentially expressed proteins between subgroups with HFpEF and controls is shown in Figure 3B and Data S16 through S21. HFpEF Group 1, the larger subgroup more similar to controls, still had several proteins that were differentially expressed (n=30 up, 515 down). Downregulated proteins were enriched in anaplerosis and metabolism of branched‐chain amino acids, glycolysis, fatty acids, and ketones. Although no enriched pathways were identified, some of the upregulated proteins were related to glucose and fatty acid metabolism (eg, SLC2A1 [solute carrier family 2, facilitated glucose transporter member 1], MCCC1 [methylcrotonyl‐CoA carboxylase subunit 1], ACSF3 [acyl‐CoA synthetase family member 3], LDHD [lactate dehydrogenase D]), intracellular trafficking (eg, EHD4 [EH domain containing 4], PDIA6 [protein disulfide isomerase family A member 6], EPN1 [epsin 1]), and fibrosis (eg, COL[collagen type]14A1, COL6A2, COL6A6, MFAP4 [microfibrillar‐associated protein 4]).

Table 1.

Clinical Characteristics of Nonfailing Controls and Patients With HFpEF Included for DDA‐MS and DIA‐MS Proteomics Analyses

Variable	DDA‐MS cohort			DIA‐MS cohort
	Nonfailing (n=10)	HFpEF (n=10)	P value	Nonfailing (n=5)	HFpEF (n=44)	P value
Age, y	59 (55, 63)	61 (52, 70)	0.54	63 (60, 64)	63 (53, 69)	0.94
Sex, % female	5 (50%)	8 (80%)	0.35	3 (60%)	29 (66%)	1.00
Race
Non‐hispanic Black, n (%)	0 (0%)	6 (60%)	0.01*	0 (0%)	28 (63%)	0.006*
Non‐hispanic White, n (%)	9 (90%)	4 (40%)		5 (100%)	13 (30%)
Hispanic White, n (%)	1 (10%)	0 (0%)		0 (0%)	3 (7%)
Medications
Angiotensin‐converting enzyme inhibitor/angiotensin receptor blocker, n (%)	2 (20%)	5 (50%)	0.35	3 (60%)	30 (68%)	0.33
Beta blocker, n (%)	3 (30%)	2 (20%)	1.00	1 (20%)	24 (55%)	0.19
Loop diuretic, n (%)	0 (0%)	9 (90%)	<0.001*	0 (0%)	39 (0.89)	<0.0001*
Past medical history
Hypertension, n (%)	5 (50%)	9 (90%)	0.03*	2 (40%)	41 (93%)	0.002*
Diabetes, n (%)	0 (0%)	4 (40%)	0.09	0 (0%)	31 (70%)	0.004*
Atrial fibrillation/flutter, n (%)	1 (10%)	1 (10%)	1.00	0 (0%)	5 (11%)	1.00
Clinical features
Body mass index, kg/m2	25 (21–31)	41 (36–45)	0.001*	29 (25–31)	39 (33–43)	0.02*
Systolic blood pressure, mm Hg	…	132 (123–158)	…	…	153 (132–165)	…
Diastolic blood pressure, mm Hg	…	69 (64–74)	…	…	77 (71–92)	…
Echocardiography
LV ejection fraction, %	65 (60–68)	65 (60–75)	0.28	63 (59–68)	65 (63–70)	0.40
LV end diastolic dimension, cm	3.9 (3.9–4.5)	4.3 (4.2–4.5)	0.61	3.9 (3.6–3.9)	4.5 (4.1–4.9)	0.03*
LV mass index, g/m2	120 (95–133)	89 (77–103)	0.10	90 (83–97)	98 (76–119)	0.63
Laboratory values
N‐terminal pro‐B type natriuretic peptide, pg/mL	…	179 (58–741)	…	…	220 (58–691)	…
Hemoglobin A1c, %	…	6.2 (5.9–6.8)	…	…	6.6 (5.9–7.8)	…
Creatinine, mg/dL	0.9 (0.8–1.0)	1.3 (1.0–2.0)	0.03*	1.0 (1.0–1.3)	1.3 (1.0–1.8)	0.29
Estimated glomerular filtration rate, mL/min per 1.73 m2	87 (71–101)	51 (45–71)	0.04*	64 (63–68)	51 (35. 74)	0.23
Invasive hemodynamics
Right atrial pressure, mm Hg	…	8 (5–10)	…	…	11 (8–14)	…
Mean pulmonary artery pressure, mm Hg	…	24 (19–36)	…	…	29 (21–35)	…
PA systolic pressure, mm Hg	…	34 (30–52)	…	…	44 (32–54)	…
PA diastolic pressure, mm Hg	…	17 (12–28)	…	…	22 (16–26)	…
PA wedge pressure, mm Hg	…	15 (9–28)	…	…	20 (15–23)	…
Cardiac output, L/min	…	5.2 (4.6–7.9)	…	…	5.5 (4.5–6.4)	…
Cardiac index, L/min per m2	…	2.4 (2.0–3.3)	…	…	2.5 (2.1–3.0)	…
Pulmonary vascular resistance, Wood units	…	1.1 (0.9–1.7)	…	…	1.6 (1.0–2.4)	…
Right atrial to pulmonary capillary wedge pressure ratio	…	0.5 (0.4–0.6)	…	…	0.6 (0.5–0.7)	…
Pulmonary artery pulsatility index	…	2.2 (1.3–3.8)	…	…	2.2 (1.5–2.7)	…

Groups comparisons were performed with a Wilcoxon rank‐sum test or Fisher's exact test for continuous and categorical variables, respectively. Data are reported as median (interquartile range) or number (%). estimated glomerular filtration rate is based on the CKD‐EPI 2021 race‐free equation. DDA‐MS indicates data‐dependent acquisition; DIA‐MS, data‐independent acquisition; HFpEF, heart failure with preserved ejection fraction; LV, left ventricle; and PA pulmonary artery.

^* p <0.05

Figure 3. Identification of subgroups with HFpEF by agnostic clustering of protein expression measured by DIA‐MS proteomics data set.

Subgroups of HFpEF were identified by K‐means clustering using proteomic data. A, Principal component analysis using all proteins measured identifies 2 distinct HFpEF proteomic signatures, with HFpEF Group 1 (blue) much closer to controls compared with HFpEF Group 2 (red). B, Gene Ontology: Biological ProcessesGO:BP enrichment for each subgroup with HFpEF compared with controls and to each other. Significantly enriched pathways were curated to diminish redundancy; the full pathway lists are provided in Data S1 Circle size reflects protein ratio, which is the proportion of differentially expressed proteins in the pathway vs all differentially expressed proteins. Color coding reflects P value after Benjamini–Hochberg adjustment for multiple comparisons. DIA‐MS indicates data‐independent acquisition; ER, endoplasmic reticulum; HFpEF, heart failure with preserved ejection fraction; PC1, principal component 1; and PC2, principal component 2.

The smaller HFpEF Group 2 was more disparate from the controls and had downregulated proteins enriched in fuel metabolism but also proteins involved with the proteasome, ribosome/translation, tricarboxylic acid cycle, and OxPhos. Upregulated proteins in both groups were associated with glucose and fatty acid metabolism (eg, SLC2A1, MCCC1, ACSF3, LDHD) and intracellular trafficking (eg, EHD4, PDIA6, EPN1). Proteins uniquely upregulated in Group 2 HFpEF included MYPT1 (myosin phosphatase target subunit 1; l₂fc=1.1, P=0.04) and mitochondrial proteins (eg, NLN [neurolysin], FARSB (phenylalanyl‐TRNA synthetase subunit beta], UQCR10 [ubiquinol‐cytochrome C reductase, complex III subunit X]). Pathways distinguishing Group 1 and 2 involved pathways associated with metabolism, ribosome/translation, OxPhos, and proteasomal proteins.

Given the disparity in the proteome between controls and Group 2, we further examined whether if this was coupled to a particular clinical phenotype (Table S3). A larger proportion of Group 2 patients had a BMI over the median of 40 kg/m² (69% versus 35% for Group 1, 1‐sided Fisher's exact P=0.043). Nearly all other demographic, clinical, echocardiographic, and hemodynamic measures did not differ between subgroups, other than LV end diastolic dimension that was higher in Group 1 (4.7 versus 4.3 cm, P=0.04, Table S3). No proteins were differentially expressed when comparing patients with HFpEF by sex or race (Data S22).

Given the association between proteomic phenogroup and BMI ≥40 kg/m², we examined which of the differentially expressed proteins correlated with BMI (Table S4, Data S23). Several that were inversely correlated (lower in patients with severe obesity and HFpEF) related to transcription and protein translation including glycine‐tRNA ligase, EIF3B (eukaryotic translation initiation factor 3B), EIF3M (eukaryotic translation initiation factor 3M), DDX23 (DEAD box protein 23), and NUD16 (U8‐snoRNA decapping enzyme). Others related to the citric acid cycle, such as MAOM (NAD‐dependent malic enzyme) and 2‐oxoglutarate dehydrogenase complex component.

To further test the relationship between clinical features and proteomic signatures, we performed weighted gene coexpression network analysis (Figure 4, Data S24 through S26). Five protein modules were identified. GO:BP enrichment of the modules overlapped with Figures 2 and 3 (Data S27 through S31). The “turquoise” module was the largest and correlated with higher LV end‐diastolic dimension and LV mass index. It was enriched in predominantly metabolic pathways, including pyruvate metabolism, glycolysis, the tricarboxylic acid cycle, and oxidation of branched‐chain amino acids and fatty acids. The “yellow” module correlated with higher LV end‐diastolic dimension, higher LV mass index, and male sex and was enriched in proteolysis (proteasome proteins), pentose‐phosphate shunt, and nucleoside metabolism. The “brown” module correlated with higher LV end‐diastolic dimension and mean pulmonary artery pressure; proteins were enriched for translation initiation and ribosome biogenesis, OxPhos (“mitochondrial respiratory chain complex assembly”), and sarcomere proteins (“muscle filament sliding”). The “blue” module correlated with lower right atrial to pulmonary capillary wedge pressure ratio and was enriched in OxPhos, mitochondrial transport, and sarcomere proteins (“regulation of heart contraction”). The “green” module did not correlate with any clinical features yet was enriched in actin cytoskeletal proteins related to the immune response and cell motility.

Figure 4. Clinical correlates of HFpEF proteomic phenotypes from DIA‐MS proteomics.

A, Gene ontology enrichment of protein modules identified by weighted gene coexpression network analysis and corresponding cluster dendrogram. Significantly enriched pathways were curated to diminish redundancy; the full pathway lists are provided in Data S1. B, Heatmap of clinical variables that correlated with protein modules in the DIA‐MS data set for HFpEF only. Pearson correlation coefficient between clinical variables and protein eigenvalues (the first principal component of each group of correlated proteins) are displayed. Correlation coefficients for P values <0.05 are shown. Full correlation results are in Data S1. Sex is coded as 1—female, 0—male (ie, negative correlation implies protein module is associated with male sex). DIA‐MS indicates data‐independent acquisition; HFpEF, heart failure with preserved ejection fraction; LVEDD indicates left ventricular end diastolic diameter; LVMi, left ventricular mass index; MAPK, mitogen‐activated protein kinase; mPAP, mean pulmonary artery pressure; NIF/NK‐κB, nuclear factor kappa B‐inducing kinase; and RAP/PCWP, right atrial to pulmonary capillary wedge pressure ratio.

DISCUSSION

The present analysis of myocardial proteomics from patients with clinically well‐documented HFpEF has several important findings. First, we find consistent downregulation of proteins related to ribosomal structure, protein translation, fuel metabolism, and OxPhos in subgroups with HFpEF versus Controls, particularly in a subgroup with HFpEF that was most disparate from controls. These distinctions were independently confirmed in 2 different proteomics analyses. Second, we find a novel discrepancy between differential gene and protein expression in HFpEF related to OxPhos and metabolism. The specificity of the gene‐protein expression discordance for metabolism proteins in human HFpEF suggests defects in mRNA stability or protein translation/clearance. Third, machine learning identified a subgroup of HFpEF with a proteomic signature most disparate from controls, highlighting abnormal processes related to metabolism and protein translation. Patients with HFpEF in this subgroup were more likely to have severe obesity, and BMI itself correlated with proteins related to metabolism and translation. Although both subgroups had downregulated metabolism and translation proteins versus controls, this expression was even lower in the more obese subgroup. Finally, protein network analysis identified modules related to metabolism and translation that correlated with measures of LV size, pulmonary hypertension, and right ventricular hemodynamics, highlighting potentially relevant clinical features related to the HFpEF proteome. Together our data highlight several novel, distinct therapeutic targets—mainly metabolism and protein translation—for HFpEF with severe obesity, an increasingly prevalent cohort.

Prior proteomic analyses in HFpEF have nearly all been based on serum/plasma samples, often using targeted proteomics (Olink, SomaLogic, SOMAscan) where HFpEF had higher levels of proteins associated with extracellular matrix remodeling and inflammation. ²⁸ , ²⁹ , ³⁰ Serum/plasma proteins do not necessarily reflect what is secreted by the heart, with established exceptions being natriuretic peptides and sarcomere proteins released in response to myocardial injury (eg, troponin), but rather the systemic response to comorbidities/multiorgan disease. In fact, in a prior metabolomics study of human HFpEF, we found important discrepancies between plasma and myocardium. ¹⁰ We did find enrichment of immune/inflammation pathways in the upregulated proteins in subgroups with HFpEF versus controls (DDA‐MS). We did not find enrichment of extracellular matrix remodeling in our analysis; however, there were individual proteins related to fibrosis that were upregulated in HFpEF.

To our knowledge, there are only 2 prior published proteomic studies of human striated muscle in HFpEF. One examined skeletal muscle biopsies from a cohort with less obesity (median BMI ~32 kg/m²) and reported reduced expression of OxPhos and other metabolic proteins (tricarboxylic acid cycle, fatty acid, and ketone metabolism) in HFpEF that correlated with reduced exercise capacity. ³¹ Similar abnormalities were observed in our study in the heart, potentially highlighting defects shared between skeletal and cardiac muscle. The only myocardial proteomics study ¹¹ was small and identified only 10 differentially expressed proteins out of 1015, without multiple comparison adjustment, and of these, most were unchanged or discordant in directionality in our data set. Eukaryotic initiation factor 4A2 was upregulated in their HFpEF samples whereas we found the eukaryotic initiation factors generally downregulated in HFpEF versus controls. Proteins found upregulated in their HFpEF samples were related to cytoskeleton (tubulin, cytoplasmic actin), neither of which were upregulated in our data. However, we did find enrichment of actin cytoskeletal proteins in the upregulated proteins in HFpEF. The lack of overlap in our findings could reflect the difference in phenotypes studied, as most participants with HFpEF in the prior study had severe ischemic heart disease and less obesity (60% had BMI <30 kg/m²), and controls had severe aortic stenosis. By contrast, our data used prospectively diagnosed patients with HFpEF and compared with controls without valvular or myocardial disease.

Our findings that HFpEF had lower protein expression related to OxPhos and other metabolic pathways is consistent with prior myocardial proteomics reported from patients with HFrEF with LV assist devices, ³² ischemic and nonischemic HFrEF, ³³ hereditary dilated cardiomyopathy, ³⁴ , ³⁵ and hypertrophic cardiomyopathy. ³⁶ Although resting systolic function differs among these phenotypes, contractile reserve is depressed in each, including HFpEF, ⁹ as is diastolic dysfunction. ³⁷ These features could be linked to depressed OxPhos and thus myocardial reserve and active relaxation. Notably, these prior studies were of end‐stage HFrEF or refractory obstructive hypertrophic cardiomyopathy, whereas our study data were obtained from endomyocardial tissue at an earlier stage of disease.

Unsupervised clustering of the proteome identified 2 unique subgroups with HFpEF—one with larger hearts, the other with more severe obesity and lower expression of metabolic and protein translation proteins. Our transcriptomic study in HFpEF ⁸ found a subgroup of HFpEF with higher OxPhos gene expression versus control that also had the highest BMI. This transcriptomic subgroup shares similarities with our proteomic HFpEF Group 2. This suggests that severe obesity with HFpEF may constrain translation of these proteins and augmented mRNA transcripts are a counter response. There are several potential mechanisms including depressed mitochondrial fission and fusion due to downregulation of their regulators (eg, OPA1 [optic atrophy 1 protein], ³⁸ P=0.04), insulin desensitization that can downregulate mitochondrial proteins, ³⁹ or reduced mitochondrial protein translation. This deficit may not be isolated to mitochondrial proteins, as more proteins were downregulated overall than those that were upregulated, and other metabolic pathways displayed discordant expression of gene versus protein. Weighted gene coexpression network analysis identified a few relationships between clinical indices of ventricular hypertrophy and pulmonary hypertension/right ventricular dysfunction and protein modules, highlighting potentially relevant clinical features.

The heart primarily favors fatty acids over other fuels ⁴⁰ ; however, the HFrEF heart compensates for lower fatty acid oxidation with increased consumption of ketones and lactate. ⁴¹ The extent to which this limits reserve has been questioned ⁴² but the shift in substrate is not debated. In HFpEF, prior metabolomic ¹⁰ and the current proteomic analysis suggest fatty acid metabolism is reduced, particularly in the subgroup associated with more severe obesity. This raises an intriguing question of whether such defects are reversible with weight loss. The recent GLP1‐agonist semaglutide (STEP‐HFpEF ⁴³ ) improved symptoms and exercise capacity in HFpEF, and semaglutide has been found to increase mitochondrial protein expression in a mouse model of HFpEF. ⁴⁴ Further clinical HFpEF studies after weight loss assessing myocardial proteomic and metabolic signatures are needed to determine these associations.

Few animal models of HFpEF have associated proteomics data, and only one has included obesity among its features (ZSF1 rat). ⁴⁵ The other models are salt/volume hypertensive rodents. ⁴⁶ , ⁴⁷ , ⁴⁸ Consistent with greater hemodynamic loading, these generally have increased sarcomere protein expression consistent with cardiac hypertrophy and depressed OxPhos proteins. ⁴⁵ , ⁴⁶ , ⁴⁷ , ⁴⁸ These features are similar to those reported here in human HFpEF. However, we did not observe upregulated nitrosative stress as found in the ZSF1 rat ⁴⁵ or increased fatty acid oxidation ⁴⁹ found in the L‐NAME/highfat diet HFpEF model. ⁵⁰ Increased fatty acid oxidation appears is a prominent feature in the L‐NAME/high‐fat diet model suggesting fatty acids are the primary fuel source, ⁵¹ whereas in human HFpEF, the metabolome ¹⁰ and current proteome data suggest reduced fatty acid oxidation.

Our study has some limitations. We measured myocardial proteins at 1 time point so stability/progression over time remains unknown. We also studied right ventricular biopsies; however, prior gene expression analysis comparing right ventricle to left ventricle found very strong correlations in HFrEF and controls. ⁸ Our transcriptomic and proteomic data included some of the same patients, but importantly all were from the same general cohorts (HFpEF samples from the Johns Hopkins University HFpEF program and control samples from the University of Pennsylvania biobank). No differences were observed in relevant characteristics between controls used in DIA‐MS versus DDA‐MS, or HFpEF used in DIA‐MS versus DDA‐MS. The list of proteins in the proteomic data set was shorter than the transcriptomic data set, limiting analysis of gene/protein expression discordance. Patient groups were not matched to age, sex, race, or clinical comorbidities, although we did not find differences in the proteome related to sex or race. Protein activity is influenced by posttranslational modifications, and clarifying how this is altered in HFpEF is an ongoing focus of investigation. Lastly, the 2 MS data sets used different data acquisition approaches with their own strengths and limitations. With DDA‐MS, only certain peptides are selected and then fragmented for sequence information, allowing for new protein identification with each sample. With DIA‐MS, all peptides are fragmented but it is limited, in this case, to a preexisting peptide library. ⁴⁴ , ⁵² These differences prevent the 2 analyses from being combined, but they can be integrated at the pathway level.

CONCLUSIONS

In summary, we report reduction of protein expression involved with protein translation and oxidative metabolism in the HFpEF myocardium that is more pronounced with severe obesity. At the protein level, patients with HFpEF separated into 2 groups, one with protein expression profiles closer to nonfailing controls and the other quite different. Future experiments are needed to validate our exploratory findings.

Sources of Funding

The study was supported by National Institutes of Health National Heart, Lung, and Blood Institute: R35HL135827, National Institute of Allergy and Infectious Diseases RAI156274A, American Heart Association: 20SRG35490443 (David A. Kass); 16SFRN28620000 (David A. Kass); 16SFRN28620000 (Kavita Sharma); National Institutes of Health 2T32HL007227 (Edwin J. Yoo, Virginia S. Hahn); National Institutes of Health 1L30HL165593‐01 (Edwin J. Yoo); Predoctoral Fellowship Grants: American Heart Association 23PRE1026275 and F31 HL168850 (Vivek P. Jani); Amgen Research Support (Kavita Sharma, David A. Kass); National Heart, Lung, and Blood Institute 1K23HL166770‐01, 1L30HL138884, Sarnoff Cardiovascular Research Foundation 138828 (Virginia S. Hahn); Einstein‐Mount Sinai Diabetes Center, Merck, Relay Therapeutics, Deerfield Management Company (Xseed award), National Institutes of Health S10OD030286 (Simone Sidoli), American Heart Association Postdoctoral Fellowship 829444 (Aleksandra Binek); 1R01 HL144509‐01, 1R01 HL155346‐01A1, Barbara Streisand Women's Heart Center and the Smidt Heart Institute (Jennifer E Van Eyk).

Disclosures

David A. Kass is on the advisory board for Amgen, Cardurion, Astra Zeneca, Cytokinetics; Consultant: Gordian, Lilly, Moderna; Kavita Sharma is an advisory board member and consultant to Alleviant, AstraZeneca, Bayer, Boehringer‐Ingelheim, Edwards Lifesciences, Janssen, Medscape, Novartis, NovoNordisk, RIVUS, and Regeneron. Kavita Sharma and David A. Kass receive funding from Amgen. Simone Sidoli receives grant funding from Merck and Relay Therapeutics. The remaining authors have no disclosures to report.

Supporting information

Data S1

Tables S1–S4

Figures S1–S3

Datasets S1–S31

References 53–56