Search strategy and selection criteria.

We followed two recent guidelines for the conduct of the systematic review and meta-analysis and Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines for reporting (S1 Table) [20–22]. The protocol of this study is registered in the PROSPERO database with the record code CRD42023399021.

MEDLINE/Ovid, EMBASE/Ovid, CINAHL/EBSCOhost, and Google Scholar were searched to identify relevant studies from January 1st, 2000, until March 18th, 2024, without language restrictions. Search terms related to AF transcriptomics were used, including atrial fibrillation, atrial cardiomyopathy, transcriptome, microarrays, and RNA-Seq, among many others. The complete search strategy is described in S1 Text. All observational studies comparing the transcriptomic profile (by microarrays or RNA-Seq) of patients with a diagnosis of AF with individuals suspectedly in SR were included. Only studies that assessed differential gene expression between AF and SR groups by analysis of myocardial tissue samples and provided any type of dataset suitable for reanalysis were included.

In parallel to the bibliometric database search, we queried four additional repositories that store biomedical studies datasets: the Gene Expression Omnibus database, the European Nucleotide Archive, ArrayExpress, and the European Genome Phenome Archive. The following search terms were used: “atrial fibrillation,” “atrial cardiomyopathy,” “atrial arrhythmia,” and “afib.” We included a dataset if it assessed myocardium samples, if data from at least four total patients were available, and if there was a protocol describing sample gathering and processing methodology available. S1 Fig summarizes the main steps in the analysis process of this study [18,19].

Database screening and data extraction

Four reviewers screened titles and abstracts in duplicate to systematically identify potentially includable studies according to the selection criteria. Subsequently, four reviewers independently extracted the following data from each included study: name of the first author, the country in which the study was performed, year of publication, study design, the total number of patients, number per group (AF vs. SR), demographic characteristics of the population evaluated, type of AF evaluated, left ventricular ejection fraction, sample sources, sequencing and analysis methods, and differentially expressed genes between groups.

Quality assessment

Four authors independently assessed each study in duplicate using an adapted version of the Checklist for Analytical Cross-Sectional Studies from the Joanna Briggs Institute (JBI). This adapted checklist was designed due to the lack of specific quality assessment tools for studies performing unbiased transcriptomic analysis in tissue samples and allowed an objective assessment of relevant methodological and reporting aspects of unique importance in this type of studies. It comprises ten items covering study design and participant characterization, sample processing, statistical analysis, and data reporting. Conflicts were resolved by consensus or the intervention of an additional reviewer.

Data processing

The raw data from studies found through the systematic search strategy and from queried repositories were downloaded and processed uniformly. Non-normalized count matrices were downloaded directly, whereas, for studies with only sequencing files (.fastq files), sequence alignment was performed using the HISAT2 alignment algorithm to acquire count matrices for data derived from RNA-Seq experiments. Standard processing and normalization were performed using the limma package for datasets derived from RNA-seq, while the oligo package was also used for those derived from microarrays.

Sample classification and contrasts

We classified the collected samples according to AF rhythm (AF vs. SR) and according to sample origin (left atrial appendage and right atrial appendage [LAA and RAA, respectively]). From this, the differential gene expression between AF and SR in each chamber was analyzed.

Statistical analysis

Overall, a consensus transcriptional signature meta-analysis can be defined as an approach to identify consistent gene expression patterns (transcriptional signatures) across multiple independent studies focusing on a particular condition [18,19]. This approach allows for the identification of more robust and reliable molecular signatures, potentially elucidating pivotal transcriptional programs consistently correlated with specific states or responses, such as disease onset, progression, or reaction to specific interventions. For this purpose, a Fisher’s combined probability test was used to pool the differential expression analysis results per individual genes if they were evaluated in at least 50% of the studies. For each contrast, the degrees of freedom for the statistical tests were derived from the number of studies included. Since none of the included studies mentioned having performed any probabilistic sampling procedure, additional study weighting was not performed. After the Benjamini-Hochberg (BH) correction of the pooled P values, a gene ranking representing the AF transcriptomic signature was generated [23]. Using an inferential approach, we estimated the accuracy of using individual studies to classify samples of other studies into either “AF” or “SR” categories. For this purpose, a disease score was calculated by linearly combining the t-values of the differentially expressed genes of the reference studies (disease pattern) with their expression values in the individual test study.

The number of disease classifiers built was the same as the total study number for each comparison based on the t-values of the differentially expressed genes in each contrast. Once each sample-level disease score was calculated, the area under the receiver operating characteristic curve (AUROC) was used to test the score’s accuracy to discriminate samples from AF patients, reflecting the conservation of gene regulation patterns across studies. Finally, the consistency of the direction of transcriptional regulation in the disease score results was also assessed by separating the top differentially expressed genes based on their regulation direction and performing gene set enrichment analysis (GSEA).

In addition, we assessed the concordance of the top 500 genes of each chamber’s consensus signature using the Jaccard index. Complementarily, we compared the consensus rankings of LAA and RAA using the Rank-Rank Hypergeometric Overlap (RRHO) test. This threshold-free algorithm detects common trends between two ordered lists of differentially expressed genes. In this method, the genes are ordered based on the product of the negative logarithm (base 10) of the differential expression p-value and the effect size direction [24]. According to this ranking, genes with unchanged, upstream, and downstream effects are positioned in the middle, at the top, and at the bottom of the list, respectively. The degree of similarity between the gene lists from two different datasets is then quantified through a one-sided p-value, which is determined using the principles of the hypergeometric distribution. Missing data were not imputed.

Sensitivity analysis

Using a leave-one-out procedure, we evaluated how much each study impacted the overall findings. Therefore, the meta-analysis for each contrast was repeated the same number of times of the amount of studies present, each time leaving out the data from one study. After each round of analysis, we performed a Spearman rank correlation test comparing the leave-one-out result with the full meta-analysis to test the influence of the left-out study. A correlation value close to 1 would indicate that the results were almost identical, and therefore, the excluded study did not significantly impact the overall findings. Conversely, a low correlation value would highlight the studies that strongly influenced the meta-analysis. This approach allowed us to assess the robustness of the meta-analysis ranking rigorously.

Added value of the meta-analysis

To analyze the benefit of combining the available transcriptomic data related to AF in our meta-analysis, we examined whether the top 500 genes identified from the consensus signature provided a more representative transcriptional ‘fingerprint’ of AF than the top 500 genes picked from each study. In particular, we checked whether the AUROCs based on the consensus genes were significantly better than those derived from individual studies using the Wilcoxon paired test. At first, we divided the top 500 genes into two groups: those up- and down-regulated in AF. We then checked for their presence in each individual study’s gene activity ranking using GSEA. Then, we compared these enrichment scores to the ones obtained using the top 500 genes from each separate study with the Wilcoxon paired test.

Functional analysis

After the meta-analysis, each gene’s resulting -log10 p-value was weighted based on the mean direction of change across the studies to develop the directed consensus signature. To comprehensively characterize the biological processes associated with AF, we performed multiple complementary functional analyses. First, we conducted gene set enrichment analysis using the Molecular Signatures Database (MSigDB) to identify enriched gene ontology terms, pathways, and biological processes. This was implemented using the msigdbr package with a ranked list of genes ordered by their weighted -log10 p-values. We tested for enrichment against multiple MSigDB collections, including GO biological processes, molecular functions, cellular components, Hallmark gene sets, KEGG, and Reactome pathways. In parallel, we estimated transcription factor (TF) activity using DoRothEA, which infers TF activity based on the expression changes of their known target genes [25–27]. Furthermore, we analyzed signaling pathway activity using PROGENy, which estimates pathway activities based on pathway-responsive genes [28]. For all analyses, we applied the Benjamini-Hochberg (BH) procedure to adjust p-values for multiple testing, considering results significant at an adjusted p-value < 0.05.

Tissue-cell-type enrichment of the AF-CS

Using the differentially expressed genes (BH p-value < 0.05) in each CS, we aimed to explore the potential cellular context in each chamber using WebCSEA. This tool provides a specialized query interface for analyzing gene sets against an extensive repository of cell expression profiles specific to various human tissue types. For each query run, the system produces a series of analytical metrics, including the raw statistical significance value (p-value) specific to cell type enrichment analysis (CSEA), a consolidated p-value derived using a permutation-based statistical approach, and a list of shared genes between the query set and the cell type-tissue-specific expression profiles.

Performance of the AF-CS as predictors of AF status

Finally, we aimed to explore how the two consensus transcriptional signatures derived from atrial myocardium analyses would perform in classifying other types of samples (peripheral blood, left ventricle and epicardial adipose tissue) derived from AF and SR patients in publicly available datasets.

Study registration

The study was registered in the International Prospective Register of Systematic Reviews (PROSPERO) under the record code CRD42023399021. Available from: https://www.crd.york.ac.uk/prospero/display_record.php?ID=CRD42023399021.

Descriptive analysis

This systematic review and meta-analysis was performed in compliance with Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines for reporting (S1 Table). Of the 1243 references screened, we identified 36 studies that applied transcriptomic approaches in atrial tissue for the study of AF (S2 Table) [13,29–61]. Of these, only 15 studies met the complete selection criteria by providing complete gene counts or raw sequences for each of the assessed samples, therefore being included in the meta-analysis (Fig 1) [29,31,34,35,37–39,48–52,56,61,62].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Flowchart of study selection.

https://doi.org/10.1371/journal.pone.0323534.g001

The selected studies were published between 2005 and 2023, with the majority (9 [60%] of 15) published within the past five years. The data sets derived from these studies consisted of a total of 534 samples (353 AF and 181 SR) from 470 patients (mean age 58.5 years; 58% male) (S2 Fig and S3 Table) [48,63–75]. Most samples were collected in the context of valve surgery (14/15; 93%), followed by coronary artery bypass grafting (CABG) surgery (7/15; 47%) and cardiac transplantation (2/15; 13%). The majority of studies included patients with permanent or persistent AF, with only three including patients with paroxysmal AF. Nevertheless, the lack of individual-patient information limited the assessment of different AF types independently (Table 1). Fig 2 summarizes the general characteristics of the studies included in the meta-analysis, explicitly highlighting the reporting of relevant information (Fig 2A) and the sample sizes evaluated (Fig 2B). Unfortunately, only a small minority of studies reported individual information on the included patients, which limited the performance of analyses by relevant covariates. From the total studies, 11 analyzed samples from the left atrial appendage (LAA), while eight analyzed samples from the right atrial appendage (RAA) (Table 1). Finally, both microarray (n = 7; 47%) and RNA-Seq (n = 8; 53%) technologies were utilized (Table 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. General characteristics of the included studies.

https://doi.org/10.1371/journal.pone.0323534.t001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Summary of overall study characteristics.

A, Available metadata on relevant characteristics in the included datasets. B, Sample size per group categories. Abbreviations. Ind: Individual-patient data; Gr: Group data (measures of central tendency and dispersion); NR: Not reported.

https://doi.org/10.1371/journal.pone.0323534.g002

To identify bias in the included studies, we used a modified version of the Checklist for Analytical Cross-Sectional Studies from the Joanna Briggs Institute (JBI) which included 10 points (S4 Table). We observed a mean score of 7/10 points for the included studies, with no studies scoring less than four points (S4 Table). The following sections describe the AF-CS results for each evaluated cardiac chamber (L-CS for the LAA and R-CS for the RAA).

Overall characteristics and consistency measures

At first, we sought to characterize the technical characteristics of the different datasets by atrial appendage. Gene coverage analysis, which assesses the similarities between the assessed genes across studies, highlighted higher comparability of the LAA samples (mean Jaccard index of ≈0.70) than RAA samples (mean Jaccard index of ≈0.51), meaning that overall the LAA sequencing results in the different studies shared more genes between them than the RAA ones (S3 Fig). Moreover, an ANOVA test revealed a significant association between the technology type (“RNA-Seq” vs. “Microarrays”) and the first three PCs of the z-transformed samples (P < 0.05 for the first three PCs in both atria) (S4 Fig). Though, after performing read count standardization, the heterogeneity derived from technical aspects was reduced. Furthermore, we evaluated the proportion of variance explained by age, sex, and left ventricular ejection fraction (LVEF) in standardized PCA to explore the role of clinical and sociodemographic parameters in gene expression patterns. Considering that LVEF represents the most consistently reported echocardiographic parameter and has a direct relationship with atrial function, it was decided to utilize this measure in the absence of more specific parameters for atrial status. We observed that the proportion of men in the different groups was not associated with any principal component, explaining zero percent of the variance. Mean age in each group (AF and SR) explained 3.5% of the variance in LAA and 4.9% in RAA, while mean LVEF explained 4.5% in LAA and 15.3% in RAA. In comparison, the variance explained by AF diagnosis was 7.9% in the LAA and 13% in the RAA (S5 Fig).

We then evaluated the consistency of the results of the studies for genes that ranked highest after differential expression analysis in each individual study (S6 Fig). We compared the top 500 differentially-expressed genes (DEG) in each individual study, observing low concordance in the results of both cardiac chambers (mean Jaccard Index LAA: 0.02 and RAA: 0.03) (S7A and S7E Figs). However, when evaluating the predictive capacity of these top 500 genes of each study to discriminate the diagnosis of AF in other datasets using a disease score, we observed a relatively high performance according to the median AUROC values (LAA: 0.74 and RAA: 0.78) (S7B and S7F Figs). This disease score was calculated by linearly combining the t-values of the differentially expressed genes of the reference studies (disease pattern) with their expression values in the individual test study. This was built assuming that if two or more studies share AF transcriptional signatures, the disease score derived from these could be used to discriminate between sample groups in other independent datasets accurately. Therefore, this disease score would estimate how similar the expression profile observed in each sample with the disease phenotype is, going beyond changes in the mean expression of specific genes and focusing on the joint regulation of multiple genes.

As the molecular responses showed some degree of uniformity across the evaluated studies, we proceeded further and examined whether the direction of differential regulation of the top DEG in each dataset was consistent with their observed trends in the remaining studies. For this purpose, an independent enrichment analysis was performed on the gene-level statistics derived from the top 500 up- and down-regulated genes from each study (S7 Fig). This enrichment score (ES) measures how much the top DEG in one study is over-represented among the top DEG in all the other studies, providing a consistency check for the direction of gene regulation. A positive score means that a gene is often found to be overexpressed in AF across multiple studies. Conversely, a negative score would mean that a gene is often found to be less transcribed in AF samples across these studies. In our results, the median ES for the differentially up-regulated genes in the L- and R-CS were 0.38 and 0.31, respectively (S7C and S7G Figs). On the other hand, the ES values for the down-regulated genes in the L- and R-CS were -0.37 and -0.33, respectively (S7D and S7H Figs).

Moreover, a significant correlation between the AUROC values of the disease score and enrichment scores for DEGs was observed for both the L-CS (Pearson correlations of 0.77 for up-regulated genes [p < 0.001] and -0.67 for down-regulated genes [p < 0.001]) and the R-CS (Pearson correlations of 0.80 for up-regulated genes [p < 0.001] and -0.73 for down-regulated genes [p < 0.001]). This finding suggests that, despite study-specific variations, the direction of gene expression changes was consistent across studies. Additional analyses showed similar results when different numbers of top genes were selected (50, 100, 200, 500, and 1000) (S8 Fig).

Meta-analysis of transcriptional signatures in AF

After confirming the feasibility of integrating the assessed data, we developed the AF-CS for each atrial chamber by pooling the individual-study results of 21,722 genes for the LAA and 11,436 for the RAA contrasts using the Fisher’s Method (S5 Table). This consensus ranking represents the overall differential gene regulation patterns across the AF spectrum, highlighting genes consistently deregulated in common directions over multiple studies (S9 Fig). In the LAA, 650 genes showed a BH p-value < 0.05; this number was 203 for the RAA analysis (S5 Table). Several genes related to structural cardiac remodeling processes were observed among the top positions of the AF-CS in both cardiac chambers, including COLQ, ANGPTL2, LBH, and COL21A1 [76–79]. Other genes near the top of the consensus included those involved in myocardial electrical remodeling processes, such as CACNA1G, HCN4, and KCNK3 [80–82]. Multiple genes related to the activity of the innate immune system were also found to be significant and highly ranked in the AF-CS, especially in the LAA, with ASAH1, CKAP4, STING1, and FCER1G standing out. Finally, new genes not previously related to AF in the literature were identified as occupying leading positions in the consensus, including DHRS9, CRTAC1, RASL11B, and GRIP2 (S10 Fig).

The leave-one-out sensitivity analysis confirmed the robustness of the overall ranking results, as the resulting gene ranks were consistently preserved after removing one study at a time (mean Spearman rank correlation: LAA = 0.96 and RAA = 0.94) (S11 Fig). There were no significant associations between the sample size of each analyzed study and the enrichment of its specific DEG in the higher positions of the CS (LAA: Spearman correlation, 0.58, P = 0.07; RAA: Spearman correlation, 0.63, P = 0.15). Despite the small sample sizes observed in some studies, this result indicated that most studies were robust enough to provide real biological insights from the contrasts performed.

When analyzing the top 500 genes of each individual chamber’s consensus signature, we observed a low concordance of the genes occupying these positions (Jaccard index value: 0.17). However, a threshold-free approach (RRHO test) of the whole gene rankings revealed a significant overlap in both directions of gene expression between the two chambers, highlighting a robust overall concordance between the two signatures (S12 Fig).

We then assessed the added value of performing a transcriptional consensus meta-analysis in this context by comparing the performance of the disease score constructed using the top 500 meta-analysis genes vs. individual studies. We observed superior predictive performance of the measures derived from the meta-analysis in both LAA (mean AUC AF-CS: 0.82 vs. mean AUC individual studies: 0.74. p < 0.001) and RAA (mean AUC AF-CS: 0.90 vs. mean AUC individual studies: 0.77. p < 0.001). Similarly, the mean enrichment scores obtained from the AF-CS results were significantly better compared to the ones from individual studies for both up-regulated genes (LAA: mean ES 0.52 vs. 0.38. p < 0.001; RAA: mean ES 0.47 vs. 0.30. p < 0.001) and down-regulated genes (LAA: mean ES -0.42 vs. -0.33. p = 0.002; RAA: mean ES -0.50 vs. -0.37. p < 0.001). When testing whether the top meta-analysis genes of each chamber’s AF-CS could predict the disease condition in samples from the other chamber, we observed significantly better performance in terms of the median AUC compared with the pairwise classifier from individual study data (LAA CS in RAA samples: 0.89 vs. individual RAA studies: 0.77. p < 0.001; RAA CS in LAA samples: 0.77 vs. individual LAA studies: 0.74. p < 0.001). These results suggested more consistent biological signals with the meta-analysis results over individual analyses and supported the presence of common gene expression patterns in both chambers.

Functional analysis of the AF-CS

The AF-CS results allowed us to characterize the cellular processes consistently deregulated in AF. To this end, we estimated the activity of transcription factors, signaling pathways, and miRNAs in each of the cardiac chambers using a gene set enrichment analysis approach (Fig 3). We found significant enrichment for the LAA consensus signature for 366 Gene Ontology terms, 28 Hallmark gene sets, 42 Reactome pathways, and 25 KEGG pathways (S6 Table). Positively enriched gene sets in this chamber were predominantly related to epithelial-mesenchymal transition, extracellular collagen matrix organization, and cytoskeletal regulation. In contrast, negatively enriched sets were associated with diverse processes related to electrical/muscle function, neuronal signaling, muscle contraction, and ion channel complexes. On the other hand, the number of results in the R-CS enrichment was lower, with only 15 significantly enriched Gene Ontology terms. These included upregulated processes related to extracellular collagen matrix organization; however, most of the enriched terms were observed to be downregulated in the AF group, highlighting cardiac conduction, autonomic function, action potential, and cell-cell signaling processes. When comparing the top results between the chambers, collagen extracellular matrix organization processes were significantly upregulated in both atrial appendages during AF, underlining the role of extracellular matrix remodeling in AF pathophysiology. On the other hand, processes related to normal cardiac conduction and signaling mechanisms were universally downregulated in AF. Notably, we did not find processes that were significantly upregulated in LAA and simultaneously downregulated in RAA, or vice versa, suggesting that differential regulation of processes between atrial appendages, if present, is likely to be nuanced.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Functional analysis results of the AF-CS in each cardiac chamber.

Bar length indicates the statistical significance of the specific item (gene set, miRNA, TF, pathway) enrichment (−log10 BH P‐values) with the vertical dashed line indicating the statistical significance threshold (Benjamini‐Hochberg P = 0.25) and bar color indicating the direction of enrichment (A, B, E, and F) or the direction of activation (C, D, G, and H). Top 25 (A) most enriched canonical, ontology, and hallmark gene sets, along with Reactome and KEGG pathways, (B) miRNAs’ targets, (C) transcription factor activities, and (D) signaling pathway activities in each direction from the L-CS. Top 25 (E) most enriched canonical, ontology, and hallmark gene sets, along with Reactome and KEGG pathways, (F) transcription factor activities, and (G) all signaling pathway activities in each direction from the RAA AF-CS.

https://doi.org/10.1371/journal.pone.0323534.g003

Furthermore, the inferred transcriptional activity of 754 transcription factors (TF) allowed us to identify 100 differentially active TF in the LAA and 74 in the RAA. At the same time, 100 miRNA targets were differentially enriched in AF patients in the LAA (S6 Table). Finally, pathway analysis reported the up-regulation of the TGFβ, JAK-STAT, Hypoxia, and WNT pathways in the LAA, with a simultaneous down-regulation of the PI3K, TRAIL, EGFR, NF-kB, and Estrogen pathways in this chamber. In contrast, the RAA results suggested an up-regulation of the TGFβ, MAPK, PI3K and VEGF pathways, while the Androgen, p53 and NF-kB pathways were observed to be downregulated. These results highlighted the divergent transcriptional landscapes and the potential chamber-specific molecular changes that underlie the complexity of atrial fibrillation pathophysiology.

Tissue-cell-type enrichment of the AF-CS

To get insights into how processes enriched in the AF-CS may relate to the function of individual cell types, we performed a tissue- and cell-type-specific enrichment analysis using WebCSEA (see Methods). At the selected Bonferroni-corrected significance threshold (p = 3.69 x 10^-5), the signatures of two sets of fetal cardiomyocytes were significantly enriched in the L-CS (Fetal_HCA_Heart_Cardiomyocytes: p-value = 8.43 x 10^-6 and FetalHeart_Ventricle cardiomyocyte: p-value = 2.63 x 10^-5), while the signatures of a fibroblast cell-type (AdultArtery_Fibroblast: p-value = 1.61 x 10^-5) and a fetal mesenchymal progenitor (FetalBrain_Fetal mesenchymal progenitor: p-value = 2.67 x 10^-5) were significantly enriched in the RAA CS (S13 Fig). These results highlight the reactivation of fetal cellular programs during the cardiac remodeling process in AF-affected atria.

Enrichment of GWAS-derived AF genes in the AF-CS

We explored whether the AF consensus signature included genes previously observed to be relevant for AF susceptibility in GWAS studies. To this end, we tested whether the number of AF susceptibility genes by GWAS in the NHGRI-EBI GWAS Catalog was significantly higher in the consensus AF transcriptomic signature of each chamber than would be expected by chance, observing a significant enrichment of AF susceptibility genes in both the LAA (n = 13; p < 0.001) and RAA (n = 8; p = 0.001) consensus signatures [83]. Among the shared genes we highlight the interleukin-6 (IL-6) receptor (IL6R), as single nucleotide variants (SNV) at this gene locus have been associated with AF development and recurrence after catheter ablation and its expression was consistently dysregulated in our AF consensus signatures for both chambers [84–88].

Performance of the AF-CS as predictors of AF status

We evaluated the predictive potential of the AF-CS derived from atrial myocardium in two distinct contexts: peripheral blood and left ventricular myocardium derived from AF and SR patients in publicly available datasets [89–91]. This approach aimed to assess the broader applicability and potential clinical utility of these signatures.

First, we explored the performance of AF-CS in peripheral blood samples, considering its potential as a non-invasive biomarker for AF. For this purpose, we tested different numbers of top genes from the consensus regarding their performance in sample classification to find the optimal count for each chamber’s signature. As a results, utilizing the top 300 genes from the L-CS and the top 200 genes from the RAA AF-CS, we observed promising discriminatory power between AF and SR patients (AUROC: 0.66–0.75 for LAA; 0.56–0.75 for RAA).

We then extended our analysis to left ventricular myocardium samples from patients with nonischemic (NICM; n = 8) and ischemic (ICM; n = 8) heart failure reported in Yang et al [89]. f. Interestingly, the AF-CS maintained discriminatory power in this context, suggesting a broader cardiac remodeling process in AF that extends beyond the atria.

These findings highlight the robustness of our AF-CS and its potential translational value, particularly in developing blood-based biomarkers for AF detection and monitoring. Detailed performance metrics for each tissue type and dataset are presented in S7 Table (L-CS) and S8 Table (R-CS), highlighting a significantly better performance of both CS for identifying AF cases from LV samples from the study of Yang et al compared to individual studies. On the other hand, only the L-CS had a superior performance than individual studies in classifying samples to AF or SR status from peripheral blood samples.

Strengths and limitations

As a meta-analysis with a standard bioinformatics analysis process, our study was strengthened by the increased sample size compared to individual studies, as well as the feasibility of pooling heterogeneous data, taking into account relevant factors such as the sampling/technical variability and variability derived from significant demographic and clinical covariates. Furthermore, our disease scoring metric enabled cross-study gene expression comparability. Significant directional consistency across studies supported the robustness of our differential gene regulation findings. Notably, including both atrial chambers in our meta-analytic framework provided a granular, chamber-specific perspective on AF pathophysiology, while comprehensive functional analyses yielded actionable insights into the implicated molecular processes. Finally, the validation of the consensus performance in classifying AF cases from other types of samples, specially blood samples, highlights not only the consistency of the signal across tissues and sources, but the potential of transcriptomics-derived markers for future applications in the clinical setting.

Notwithstanding its strengths, our study is subject to several limitations. First, the lack of individual patient data in most incorporated studies limited the possibility for subgroup analysis, and precluded optimal covariate adjustment in multivariate models. Additionally, the studies that met our inclusion criteria included predominantly patients with persistent or permanent AF. Moreover, the low proportion of dual-chamber samples and disparities in gene coverage and sample size between the LAA and RAA impede direct chamber-to-chamber comparability. Another limitation relates to the meta-analysis process itself, as our gene standardization approach could only partially eliminate technical heterogeneity arising from using varied platforms.

Implications for clinical practice and future directions

The consensus transcriptional signatures derived from this meta-analysis, which consolidates the available knowledge from genome-wide transcriptomics, offer an opportunity to refine the current clinical approach to AF. Identifying consistently dysregulated genes and pathways across heterogeneous patient samples may represent the first step towards identifying and validating novel targeted diagnostic/prognostic biomarkers and provides a molecular framework for designing mechanism-based treatment strategies.

Despite these advances, it is paramount to address the current study’s limitations to foster future research directions. The scarcity of individual patient-level data necessitates further studies to confirm these findings in diverse patient populations, focusing on subgroup analyses to determine the potential effect of concomitant comorbid conditions and to characterize the conserved gene expression patterns in AF across the different types (paroxysmal, persistent, and permanent) to better understand the mechanisms potentially associated with arrhythmia progression and perpetuation. Finally, while we believe our study marks a significant step forward in our understanding of AF, it also delineates the critical avenues for future research aiming to translate these molecular insights into actionable clinical interventions. The stable reference provided by our consensus signatures can serve as a foundation for developing and validating targeted therapies, as well as for identifying novel biomarkers for early detection and risk stratification in AF. This underscores the importance of our work not only in advancing our understanding of AF pathophysiology but also in paving the way for more personalized and effective management strategies for this complex arrhythmia.

Key messages

1. Clinical Relevance for Patient Care:
2. Our meta-analysis of 534 atrial tissue samples provides a comprehensive molecular characterization of AF, revealing consistent biological changes across diverse patient populations.
3. Key Pathological Mechanisms Disrupted in AF:
4. Structural Remodeling

• Extensive extracellular matrix changes occur in both atria
• Significant activation of fibroblasts and enhanced fibrosis
• Reactivation of fetal gene programs suggests maladaptive cardiac remodeling

Electrical Remodeling

• Widespread dysregulation of ion channels and calcium handling proteins
• Altered expression of genes controlling cardiac conduction
• Changes in both action potential and cellular signaling pathways

Metabolic Remodeling

• Evidence of altered energy metabolism in AF cardiomyocytes
• Simultaneous upregulation of both glycolysis and oxidative phosphorylation
• Suggests inefficient/insufficient energy production in AF tissue

Inflammatory Response

• Activation of innate immune responses, particularly neutrophil-related pathways
• Supports the role of inflammation in AF pathogenesis
• Provides molecular basis for observed elevated inflammatory markers in AF patients

Diagnostic Applications:

• The AF molecular signature identified AF patients using blood samples with modest success (AUROC: 0.66–0.75)
• Proof of principle of the potential of blood transcriptomics in AF patient assessment.