Atrial fibrillation (AF), the most common sustained cardiac arrhythmia, is characterized by rapid, irregular atrial depolarizations that lead to an increased risk of clot formation and subsequent stroke (1). Long-term risks of myocardial ischemia, heart failure and dementia are elevated in patients with AF (2–6). Clinical management of AF involves a combination of pharmacological interventions (such as calcium channel blockers) and interventional techniques (such as radiofrequency ablation) to dampen or eliminate the substrates for arrhythmogenic activity (7).

AF is a frequent arrhythmia after cardiac surgery on cardiopulmonary bypass. Approximately 30–40% of cardiac surgeries are complicated by post-operative atrial fibrillation (POAF), which is associated with greater hospital resource utilization, longer ICU and hospital stays and hence greater healthcare costs (8–10). POAF also significantly elevates the risk for more serious long-term cardiovascular complications including myocardial infarction, cardiac arrest, bleeding, renal failure, stroke and death. For coronary artery bypass grafting (CABG) surgery alone, the Society of Thoracic Surgeons database recorded 161,816 procedures in the United States in 2019, plus thousands more procedures in which mitral or aortic valve replacements were performed alone or concomitant with CABG (11).

DNA methylation frequently occurs on cytosines followed by guanine (CpGs) and has been shown to correlate with gene expression in development and disease (12). DNA methylation has been implicated in diseases such as cancer (13), and recent studies have implicated altered global DNA methylation in cardiovascular disease (14–18). Because it has been shown in the Framingham cohort that DNA methylation in peripheral blood samples is associated with AF in the outpatient setting (19), we sought to find CpGs associated with POAF, a clinically distinct manifestation of AF, after cardiac surgery. It would be useful for mechanistic studies to determine the association of these DNA methylation loci with altered gene expression in left atrial tissue; however, it is not possible to obtain left atrial tissue samples from all patients undergoing cardiac surgical procedures without a separate biopsy of the left atrium, an undue risk to patients who otherwise would not have an incision on the left atrium. In this study, we enroll patients undergoing a variety of cardiac surgical procedures (Supplementary Table 1) because these are the patients we care for in the operating room and for which we would like to develop perioperative epigenetic assessment of POAF risk. Thus, we sought to assess the association of DNA methylation in peripheral blood samples with POAF.

While the genetic contribution to AF has been well studied (20), including identification of sequence variation associated with POAF (21), the epigenetic contribution through DNA methylation remains under-explored and represents an appealing target for biomarker discovery based on the following rationale: DNA methylation is relatively stable but not immutable (it can be modified throughout life and thus may serve as a molecular beacon of modifiable risk), and some features of DNA methylation are conserved across tissue types while still exerting cell type specific effects (14, 22, 23). For example, in one study (14) a subset of DNA methylation loci associated with dilated cardiomyopathy in left ventricular myocardial tissue also displayed conserved methylation associated with dilated cardiomyopathy across other tissues, including peripheral blood samples. In the present study, we sought to identify pre-operative epigenomic biomarkers that could be deployed in a precision medicine context to predict a specific and prevalent post-operative complication. We report herein two prediction models combining pre-operative clinical variables and DNA methylation marks that predict POAF in cardiac surgery patients.

Of the 110 patients analyzed in the discovery cohort, 47 patients developed POAF for an incidence of 42.7%. There were 101 patients analyzed in the validation cohort, 43 of which experienced POAF for a similar incidence of 42.6%. The incidence in both cohorts is similar to other cardiac surgical populations. The characteristics of the 211 patients analyzed in this study are displayed in Table 1 by cohort. Patients who developed POAF had a longer length of stay with a median difference of an additional 3 days (median of 9 days total vs. 6 days). The difference in length of stay in patients experiencing POAF in all 211 patients was statistically significant (p-value < 0.001) as calculated by the Mann-Whitney U test. The surgical procedures for this patient cohort are displayed in Supplementary Table 1. This table shows that patients who had 2 or more surgical procedures (any combination of the following: aortic valve surgery, mitral valve surgery, pulmonic valve surgery, tricuspid valve surgery, CABG surgery, open aortic surgery and myomectomy) on cardiopulmonary bypass had a higher risk of POAF and so this was included as a covariate in our analysis. The racial and ethnic composition of this cohort, another covariate in our analysis, is shown in Supplementary Table 2.

Because DNA methylation at some CpGs is cell-type specific, differences in cell type composition is a known source of confounding which can mistakenly attribute differences in cellular composition as biologic differences related to the disease being studied (28). To account for the heterogeneity of cells in human blood, we compared three methods to control for cell type composition, a reference-based method (22) and two reference-free methods [ReFACTor (30) and RefFreeEWAS (29)]. The reference-based method estimates were the most highly correlated with FACS data so therefore we used reference-based estimates of percent cell type composition as covariates in our analysis of the discovery cohort. Supplementary Table 3 shows the Pearson correlation coefficients between the measured cell type composition and those estimated in silico using the reference-based method. The reference-based cell type estimates are statistically significant for neutrophils, monocytes, CD4+ T-cells and NK cells but not B-cells or CD8+ T-cells.

We then analyzed autosomal CpGs in pre-operative blood samples from the discovery cohort for association with POAF while controlling for known clinical and methylation covariates (Figure 2). There were 2,557,388 CpGs with 10 × coverage in at least 50% of patients in our discovery cohort and an average of 2,151,705 of these CpGs per sample. We identified 12 CpGs with strong association with POAF (Table 2). Of these 12 significant CpGs, 7 were within regions of the genome amenable to targeted bisulfite sequencing. CpGs that were not amenable to targeted bisulfite sequencing were within sequences that shared homology with other genomic regions. Six of these CpGs had sequencing data that met sufficient coverage thresholds as defined in the discovery cohort. Of these 6 CpGs, 3 (chr17:20376849, chr17:20376904, and chr17:20376906) are within a differentially methylated region on chromosome 17 and have highly correlated methylation. This DMR, chr17: 20376813-20376906, was identified by the callDMR function in DSS using default parameters and contained a total of 12 CpGs in our discovery cohort dataset. In a separate exercise, we used ReFACTor to analyze our data in an unsupervised manner (Supplementary Table 4). We performed a beta binomial regression controlling for 4 principal components (as determined by the package's parameter optimization tool) and the clinical risk factors as previously described in our initial analysis. Notably, the most important loci (chr16:24640902 and chr17:20376849-20376906) used in our models are still significant after accounting for these principal components.

Out of the 6 CpGs significant in the discovery cohort that were able to be sequenced in the validation cohort, one, chr16:24640902, was statistically significant in the validation cohort (p-value = 0.026). Chr16:24640902 is within the gene body of TNRC6A, a protein expressed in multiple tissues, including in the left atrium and previously associated with permanent AF (36). The methylation status of chr16:24640902 in patients experiencing POAF and in those that do not is shown in Supplementary Figure 1. Supplementary Table 5 shows the Pearson correlation of CpGs used in the POAF prediction models with cell type estimates calculated in silico with the reference-based method for cell type deconvolution. In addition, there is minimal statistical inflation in our analysis with λ = 1.026 (See QQ plot in Supplementary Figure 2).

We created two POAF prediction models using clinical risk factors and methylation data. The methylation POAF model (Table 3) was developed using three of the most significant methylation loci and 3 of the most significant clinical risk factors in the discovery cohort. After feature selection in the discovery cohort, the methylation POAF model was trained on discovery cohort data only. The methylation POAF model had an AUC of 0.83 in the discovery cohort and 0.79 in the validation cohort. ROC curves comparing the POAF methylation model performance in the discovery and validation cohorts to published clinical models are shown in Figure 3. The AUC of the methylation loci without clinical risk factors was 0.67 in the validation cohort (data not shown). An additional POAF prediction model (Supplementary Table 6) uses chr17:20376906, the CpG within the chr17: 20376813-20376906 DMR that was most statistically significant in the discovery cohort, and 4 additional clinical risk factors. This model had an AUC of 0.81 in the discovery cohort and 0.67 in the validation cohort.

We compared the methylation POAF model (Table 3) to two clinical risk factor POAF models: the CHA₂DS₂-VASc POAF model (37) and the Vanderbilt Cardiac Surgery Registry clinical-only model (21). CHA₂DS₂-VASc, a tool used to predict the risk of embolic stroke in patients with AF, has been shown to predict POAF (37) and has a scoring system that is familiar to most clinicians. The Vanderbilt Cardiac Surgery Registry clinical-only model is a validated prediction model that utilizes seven risk factors for POAF prediction. The CHA₂DS₂-VASc model (37) and the Vanderbilt Cardiac Surgery Registry clinical-only model (21) had AUCs of 0.62 and 0.74, respectively, in the validation cohort in our dataset. The methylation POAF model demonstrated a statistically significant improvement in prediction in the validation cohort over the CHA₂DS₂-VASc model with p-value = 0.00535 by DeLong's test. Compared to the Vanderbilt Cardiac Surgery Registry clinical-only model (21), the methylation POAF model showed a statistically significant improvement in the discovery cohort (p-value = 0.00117) but not the validation cohort (p-value = 0.264).

In a separate analysis, we assessed the statistical significance of 21 CpGs previously shown to be associated with AF in the Framingham cohort (Supplementary Tables 7, 8) in our validation cohort. In the validation cohort, the most significant Framingham CpG was cg12739419 (p-value = 0.033) but it was not statistically significant after accounting for multiple hypothesis testing with the Benjamini-Hochberg procedure (38). The CpG cg12739419, which is known to be associated with the AF SNP rs3807989 and the gene CAV1 in the Framingham cohort, demonstrated hypomethylation associated with POAF, consistent with hypomethylation found at this CpG in the Framingham cohort.

AF is highly prevalent, ranging from 2.7 to 6.1 million affected individuals in the United States and inflicting a cost of ~$6 billion per year on the healthcare system (39). As the most common sustained cardiac arrhythmia, AF is associated with increased risk of stroke, thromboembolism, renal failure, and myocardial infarction. Approximately 40% of people diagnosed with AF or heart failure will develop the other condition (39).

POAF in particular exacerbates morbidity and healthcare costs for patients undergoing what are now routine cardiac surgeries in medical centers around the world (8–10). Previous studies have identified genetic (20) and DNA methylation differences (19, 40) that correlate with AF, but no previous work has provided a molecular measurement that can be made in the blood of presurgical patients that will aid prediction of AF prior to its occurrence. The present study provides such information for the setting of POAF, enabling the development of a simple pre-operative blood test that, when combined with common clinical measurements, can guide the physician in managing cardiac surgery patients according to their relative risk. The prospective design of this study is crucial in that it allows for assessment of epigenetic marks as risk factors for future disease—predominantly in patients without a history of AF—rather than sequelae of disease given the dynamic nature of DNA methylation.

We present a methylation POAF model with 3 clinical factors and 3 CpGs that has an AUC of 0.79 in our validation cohort. We compared this novel prediction model to the CHA₂DS₂-VASc POAF model and the Vanderbilt Cardiac Surgery Registry clinical-only model. Though CHA₂DS₂-VASc was developed to determine the risk of embolic stroke in patients with AF, it has been shown to predict POAF (37), is simple to implement and has a scoring system that is familiar to most clinicians. The Vanderbilt Cardiac Surgery Registry clinical-only model is a validated prediction model that utilizes seven risk factors for POAF prediction (21). The combined clinical and DNA methylation model had higher AUCs than the CHA₂DS₂-VASc model (37) and a leading clinical POAF prediction model (21) which had AUCs of 0.62 and 0.74, respectively, in our validation cohort. The multicenter risk index for atrial fibrillation after cardiac surgery (32) is another important POAF model but was not compared to the methylation model because post-operative risk factors were not a component of the current study.

The methylation POAF model had a statistically significant improvement in prediction over the CHA₂DS₂-VASc model in both the discovery and validation cohorts. The methylation POAF model showed statistically improved prediction compared to the Vanderbilt Cardiac Surgery Registry clinical-only model in the discovery cohort only. This demonstrates that DNA methylation can add predictive value to clinical risk factors, though the extent to which this is possible depends on how well the particular clinical risk factors are represented in the patient population in question. The methylation POAF model's performance in the validation cohort shows clearly improved prediction over the CHA₂DS₂-VASc model, demonstrating an improvement incorporating DNA methylation biomarkers with clinical predictors. Although the goal of this research was to identify DNA methylation biomarkers that would be applicable to patients who present for a variety of elective cardiac surgical procedures, the diversity of cardiac disease within this cohort such as those with different forms of valvular heart disease may limit the application of these models to cardiac surgical populations which may have a relatively higher proportion of isolated coronary artery bypass grafting procedures compared to combined surgical procedures more typical in the University setting.

In this study, we performed targeted bisulfite sequencing of CpGs associated with AF in the Framingham cohort (19). The finding that the Framingham CpGs were not statistically significant in our validation cohort is consistent with the notion that POAF (not measured in the Framingham study) is a clinically distinct manifestation of AF and has risk factors that are unique compared to AF in the outpatient setting (41). In addition, because we utilized RRBS rather than the Infinium HumanMethylation450 BeadChip Array technology used in the Framingham cohort, the sites associated with POAF in this study were not measured in the Framingham cohort dataset and cannot be analyzed to assess their association with AF in the outpatient setting.

The methodology presented in this study could be implemented in a precision medicine environment where an epigenomic risk metric of POAF and other adverse peri-operative outcomes could be determined from a blood draw with other routine pre-operative labs. In particular, epigenomic risk profiles could be of significant value in perioperative risk stratification and planning perioperative care for each patient. Further research is needed to refine prediction models and associations with additional outcomes of interest. The epigenomic profile of each patient could also help assess perioperative risk and guide clinical decision making for patients at risk for POAF and other adverse outcomes. In addition, it could help guide which drugs or treatments patients may respond better to given their own perioperative risks and epigenomic profile. DNA methylation risk loci could also be targeted directly. In a rat model of spontaneous AF, administration of the DNA methylation inhibitor decitabine reduced atrial tachycardia, demonstrating the possibility of targeting DNA methylation for therapeutic effect (42). It will also be informative to evaluate whether these or other epigenomic marks may serve as biomarkers for AF outside of the context of cardiac surgery. With improvements in sequencing technology, patient samples can be pooled to simultaneously undergo targeted bisulfite sequencing for each individual sample at many loci of interest for reduced cost compared to RRBS and whole genome bisulfite sequencing. Targeted bisulfite sequencing costs $25–50 per sample with results available in ~2 weeks. Targeted bisulfite sequencing would ideally be utilized for patients undergoing surgeries that have a relatively higher risk of adverse outcomes—such as cardiac surgery—and include epigenetic risk factors for multiple adverse outcomes that could be used together for optimal perioperative planning.

Though this is an association study, the prospective study design and utilization of epigenetic risk factors have the potential to identify mechanisms impacting POAF susceptibility with greater granularity than clinical risk factors alone. What are the molecular mechanisms by which the association between DNA methylation and risk of AF is established? DNA methylation in so-called CpG islands is thought to control gene expression by silencing promoters and regulatory regions (43). Gene body methylation, however, is associated with increased transcription (12). The genes nearest to the regions of altered methylation observed in this study (Table 2) provide some hints about a potential molecular mechanism acting in cis to regulate the nearest gene. Chr16:24640902, a CpG statistically significant in both the discovery and validation cohorts, is within the gene body of TNRC6A, a protein expressed in many tissues including the left atrium, associated with miRNA-mediated gene silencing (44) and previously associated with permanent AF (36). The significance of micro-RNAs has emerged in the pathophysiology of AF through electrical and structural remodeling (45). The other CpGs in the POAF prediction model have unknown relation to AF. Overall, these sites may reflect molecular mechanisms in the left atrium that create substrates for POAF, which then become activated by the physiologic stress of cardiopulmonary bypass and surgery. Another mechanism by which altered CpG methylation may control gene expression and organ phenotype is through chromatin structure, recently appreciated to be globally remodeled in the setting of heart failure in human cells and animal models (46, 47). Future studies will be required to resolve the molecular mechanisms definitively—however, this study provides a novel set of biomarkers to be explored clinically in larger cohorts for their ability to aid in the prediction of AF, toward the goal of preventing many of the downstream disease sequelae precipitated by this arrhythmia.

The datasets presented in this study can be found in online repositories. The names of the repository/repositories and accession number(s) can be found below: GEO; GSE194156.

The studies involving human participants were reviewed and approved by University of California, Los Angeles. The patients/participants provided their written informed consent to participate in this study.

MF performed bioinformatic, read mapping/alignment and statistical analyses with contributions from DM, DC, and DE. MC, TK, ES, and EM performed experiments including genomic DNA library preparation. MM, CF, and MP performed bisulfite sequencing. JS, NM, and CB obtained patient data. MF and RS recruited patients and performed clinical assessment. MF, AM, and TV conceived of the study which was directed by MF and TV. MF and TV wrote the manuscript. All authors contributed to the article and approved the submitted version.

This project was supported by the UCLA Department of Anesthesiology, the UCLA Cardiovascular Theme and NIH grants UL1TR001881 (UCLA CTSI, PI: Steven Dubinett), HL 105699 (TMV), HL150667 (TMV), HL 115238 (TMV). MF was supported by a Starter Grant from the Society of Cardiovascular Anesthesiologists and a mentored research training grant by the Foundation for Anesthesia Education and Research.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.