Consecutive consenting adult patients ≥ 18 years old undergoing cardiac surgery with left atrial appendage excision at Sutter Hospital, Sacramento Medical Center were recruited between October 1, 2010 and November 1, 2012. Patients were excluded if they had congenital heart disease, any history of rheumatic valve disease or mitral stenosis, if a right thoracotomy approach was employed, if they were unable to provide informed and witnessed signed consent, or if they were pregnant or incarcerated. Participant demographics and medical details were obtained using a study questionnaire and were verified with a subsequent chart review. All study participants provided informed written consent under protocols that were approved by the University of California, San Francisco (UCSF) and Sutter Hospital, Sacramento, CA.

The genetic panel was designed in order to include all ion channels within the genome and genes previously implicated in Mendelian forms of cardiac disease as of November, 2013. The list of ion channels was constructed through a search of the Uniprot Knowledgebase using the terms ion channel and human. The 502 candidates were further manually curated to verify that the listed gene encoded an ion channel. This strategy led to the identification of 398 separate genes that were incorporated into the genetic panel (Data Supplement, Table 1).

An additional 162 genes were selected based on their documented or potential involvement in primary cardiac disease (Data Supplement, Table 1). We constructed this aspect of the genetic panel through a review of the genetic culprits associated with the following conditions: Long QT syndrome, Short QT syndrome, Brugada syndrome, Catecholaminergic Polymorphic Ventricular Tachycardia, Early Repolarization Syndrome, Idiopathic Ventricular Fibrillation, Arrhythmogenic Right Ventricular Cardiomyopathy, Hypertrophic Cardiomyopathy, Dilated Cardiomyopathy, Restrictive Cardiomyopathy, Left Ventricular Non-Compaction, and Mitochondrial Cardiomyopathy. In addition, we included all genes implicated by proxy in the pathogenesis of AF from genome wide association studies and other genes whose protein products have been implicated in the pathophysiology of the arrhythmia.

We extracted all known exons of the 560 genes using the Ensembl General Transfer Format (gtf) file annotating all transcripts in the human genome (release 68). In order to obtain exhaustive coverage of protein coding regions of genes of interest, a customized set of hybridization probes was designed and constructed using the Nimblegen SeqCap EZ Library kit (Roche NimbleGen, Madison, WI). In total, 3,218,095 of the 3,330,918 bases of interest were covered by one or more probes.

Intraoperatively, left atrial appendage samples were immediately flash frozen in liquid nitrogen in a sterile fashion. Genomic DNA was isolated from atrial tissue using the AllPrep DNA/RNA Mini Kit (Qiagen, Valencia, CA). Matching lymphocyte DNA was purified from the buffy coat using the GentraPuregene Blood Kit (Qiagen) obtained from phlebotomy performed prior to surgery.

In order to generate sequencing libraries, 1 microgram of DNA from each sample was randomly sheared to approximately 200 base pairs using a Covaris S2 Ultrasonicator (Covaris, Woburn, MA). Subsequent library preparation was performed using the KAPA Library Preparation Kit (Kapa Biosystems, Wilmington, MA). Briefly, genomic DNA fragments were end-repaired and underwent A-tailing prior to adaptor ligation with 24 unique NEXTflex DNA Barcodes (Bioo Scientific, Austin, TX). Library enrichment was then performed through polymerase chain reaction (PCR) amplification, followed by analysis of the size and quantity of ligated fragments using the Agilent 2100 Bioanalyzer (Santa Clara, CA). Recommended clean-up was performed at each step using Agencourt AMPure XP Beads (Beckman Coulter, Indianapolis, IN).

The barcoded DNA library for each sample was then pooled with equal quantities of 23 other unique barcoded libraries to a total of 1 microgram. The corresponding 24 NEXTflex DNA Barcode Blockers and COT human DNA were added to each pooled sample and then heat dried using a DNA vacuum concentrator. Hybridization of the genomic libraries with the custom designed genetic panel was then performed consistent with manufacturer specifications.

Following hybridization, the targeted fragments were pulled down and recovered using Streptavidin-coupled Dynabeads (Life Technologies, Grand Island, NY). Library enrichment and product analysis was repeated as detailed above. 101 base pair paired-end sequencing was performed on an Illumina HiSeq 2500 sequencer with 24 samples to each lane of a flow cell. Samples were demultiplexed prior to analysis.

The Burrows-Wheeler Aligner (bwa) was used to align paired-end short reads to the human genome, while the Genome Analysis Toolkit (GATK) was used for local realignment and recalibration.^22,23 Sequencing quality metrics, including number of mapped reads, number of duplicate reads, and number of mappable bases were performed with samtools, Picard tools, and GATK (Data Supplement, Figures 1 and 2).^23,24 A total of 90 genes contained isolated base pair regions with inadequate coverage precluding reliable variant calling, collectively accounting for no more than 0.7% of the region of interest (Data Supplement, Figure 3).

As a quality assurance measure, we compared single nucleotide polymorphism (SNP) minor allele frequencies observed from our data with those from the 1000 Genomes CEU population.²⁵ We identified 2369 SNPs within our coverage area with minor allele frequency estimates of at least 5%. As inclusion of non-European individuals would generate inconsistent frequency estimates, we performed principal components analysis and removed 3 individuals based on the first principal component. Minor allele frequency estimates were generated using the remaining samples and plotted against 1000 Genomes CEU frequencies.

To improve variant calling, an estimate of cross-contamination of each sample was first obtained using the ContEst program (Data Supplement, Figure 4).²⁶ Per sample estimates were then input into MuTect, allowing more accurate models for variant detection. MuTect is a state-of-the-art somatic mutation caller designed to maximize sensitivity and minimize the impact of technical sources of false positives.²⁷ The software program uses a probabilistic model to call differences based on number of reads, mapping quality, strand bias, and estimates of cross-contamination. Importantly, “tumor” and “normal” sources are analyzed in parallel, preventing the false positives and negatives that typically arise from performing a post hoc comparison of variant files after variant calling has already been performed on the samples independently.

We applied the default MuTect parameters to identify potentially discordant variants between atrial and lymphocyte DNA. We observed that the overwhelming majority of variants that emerged from this analysis corresponded to G>T transversions (Data Supplement, Figure 5). Low level G>T transversions have become recognized as an important artifactual change that occurs with high coverage next-generation sequencing and are felt to arise secondary to oxidative damage that occurs during acoustic shearing of genomic DNA during sample preparation.²⁸ This oxidative damage occurs regardless of DNA source (ie. tissue versus lymphocyte). Although not an issue for germ-line mutation calling due to their trace quantities not being consistent with a heterozygous state, their low levels may be confused with somatic mutations. As a result, the recommended bioinformatic approach to G>T transversions within the cancer literature has been to filter them from the analysis given the overwhelming probability that they represent sequence artifacts.²⁸ Consistent with this methodologic approach, we also elected to filter G>T transversions from our analysis in an effort to minimize false positive findings.

Among the final discordant atrial variants, SnpEff (v. 3.3) was used to assess their impact on protein coding sequence.²⁹ Atrial-lymphocyte discordant variants expected to change protein sequence were further scrutinized with manual annotation, including examining their frequency in control populations, visualizing mapped reads, and analyzing whether the regions of interest map ambiguously in the genome. Discordant variants were also analyzed for their presence in multiple participants within the cohort and for evidence of occurrence in the “reverse direction” (absent in atrial cells and present within lymphocytes) among other members of the cohort; findings suggestive of systematic sequencing errors. A summary of our overall analytical approach to discordant variant calling is outlined in Figure 1. Somatic fractions, defined as the percentage of total reads within an atrial sample, were determined for each remaining discordant atrial variant.

Attempted verification of potential discordant variants was pursued with Sanger sequencing of atrial DNA from relevant study participants. Amplification of targeted genomic regions was performed using polymerase chain reaction (primer sequences provided in the Data Supplement) followed by DNA sequencing using the ABI PRISM dye terminator method (Applied Biosystems, Foster City, CA, USA).

Because of the sensitivity of the sequencing employed, it is not possible to definitively distinguish potential false positive discordant variants that pass through our filtering protocol from actual somatic mutations with low somatic fractions. In order to obtain a conservative estimate, we based our calculations of somatic mutation rates on the assumption that all possible somatic mutations that passed our a priori filtering processes were real. These rates were calculated by dividing the total number of somatic mutations by the total number of nucleotides examined in both AF cases and controls. The mean somatic mutation rate was reported as the number of somatic mutations per 100 million nucleotides.

Normally distributed continuous variables are presented as means ± standard deviation and the Student’s t-test. Comparison of categorical values was performed using the Chi-squared and Fisher’s Exact tests. SNP minor allele frequencies within our cohort and the 1000 Genomes Project CEU subpopulation were compared using the Pearson pairwise correlation coefficient. In order to evaluate the possibility that filtering of G>T transversions may have resulted in a reduced sensitivity for detecting bona fide somatic mutations, we conducted a Monte Carlo simulation analysis to estimate the anticipated number of G>T transversions that would result in functional coding changes. Given that there are 12 nucleotide changes that could be observed for somatic mutations, with each assumed to be equally likely, we were able to generate a robust bootstrap estimate of the number of G > T changes expected under the assumption of no oxidative artifact. We used the total number of pre-filtered observed changes (synonymous and non-synonymous) as an empirical distribution and performed 10,000 random draws. For each draw and for each potential variant, we estimated a probability of missense mutation (using the observed missense rate and allowing some uncertainty) and totaled the result. A posterior mode and 95% credible interval were computed from the posterior distribution.

Two-tailed p-values < 0.05 were considered statistically significant. Statistical analyses were performed using Stata version 12 (College Station, Tx, USA) and R.

A total of 34 patients undergoing cardiac surgery with left atrial appendage excision at Sutter Hospital, Sacramento provided both atrial tissue and peripheral blood for sequencing analysis. Twenty-five had a history of AF, and 20 were undergoing cardiac surgery exclusively for a minimally invasive AF ablation. Among the participants with AF, the mean age at diagnosis was 63.0 ± 11.9 years and 16 (64%) were male. Seventeen participants had AF in the absence of structural heart disease, 13 of whom had no family history of the arrhythmia. There were 11 individuals with AF in the absence of all known AF risk factors (including hypertension), 9 in the absence of a known family history. The remaining baseline characteristics of the participants are summarized in Table 1.

Targeted gene capture and high-throughput sequencing permitted alignment of 16.5 million reads per sample (IQR 12.2–19.5 million) at a median 265-fold coverage depth (IQR 164–369) (Figure 2). With respect to GJA1 and GJA5, the median fold-coverage depths were 267 (IQR 171–422) and 234 (IQR 149–369), respectively. The median number of mapped bases per sample was 3.25 million (IQR 3.246–3.259 million), or 99.3% of the bases of interest. Comparison of the minor allele frequencies of 2369 SNPs from our cohort and the 1000 Genomes Project CEU subpopulation revealed a strong correlation (rho = 0.96; Data Supplement, Figure 6).

Bioinformatic analysis using the somatic mutation caller MuTect initially identified 8710 discordant base calls when treating lymphocytes as the reference (“germline”) DNA source and atrial tissue as the somatic DNA source. Notably, 8604 (98.8%) of these discordant base calls represented G>T transversions, an aforementioned common source of artifactual DNA mutations arising secondary to oxidative damage during sample preparation.²⁸ No G>T transversions resulting in non-synonymous missense mutations were observed within GJA1 or GJA5. Selective filtering of false positive G>T transversions was precluded by the absence of previously reported contextual and strand bias. Monte Carlo simulation analysis revealed a 63% probability that none of the previously filtered G>T transversions reflected bona fide functional somatic mutations (95% CI: 0–2) (Data Supplement, Figure 7). Classification of G >T transversions as false positives reduced the list of discordant base calls to 106. From this list, an additional 5 were flagged by MuTect as having poor coverage, and thus reduced reliability for variant calling.

The above filters resulted in a total of 101 potential somatic atrial variants. A “reverse” analysis, treating atrial samples as reference and lymphocytes as the somatic tissue revealed a comparable number (93) of variants. Within the overall list, 12 represented non-synonymous SNPs predicted to impact the protein coding regions of a total of 11 genes. The remaining discordant base calls represented synonymous SNPs or were located within intronic regions.

Of the 12 potential non-synonymous SNPs observed within atria and not in lymphocytes, an additional 7 were found to be consistent with sequencing artifact on the basis of their 1) being observed in multiple participants (a systematic error associated with the sequencing protocol was felt to be the likely explanation, particularly given that certain of these variants were present within highly repetitive regions of DNA prone to alignment errors); 2) being observed in the “reverse direction” (present within lymphocytes and absent from atrial cells) among other participants; 3) having a greater than 50% carrier frequency within the general population. The 5 remaining discordant genetic variants did not have population data frequency available indicating that they were either rare or novel (Table 2). The discordant variants were carried by 3 of the 25 participants with AF and 2 were present within a single control participant with no prior history of the arrhythmia (12% vs. 11%, p=1).

The somatic fractions, defined as the percentage of total reads within the relevant atrial sample, for the 5 remaining potential non-synonymous cardiac somatic mutations ranged from 2.3 to 7.3%. Sanger sequencing of atrial samples from each patient carrying a potential discordant variant yielded electropherograms with no evidence of a somatic mutation (Figure 3).

Among the 25 AF cases, there were 3 potential somatic mutations and an average of 3.25 million mapped base pairs per sample corresponding to an average somatic mutation rate of 4 per 100 million nucleotides (range: 0–31, standard deviation: 10). A total of 2 potential somatic mutations were observed among the 9 control participants, and the average mapped base pairs per sample was also 3.25 million. This yields an average somatic mutation rate among controls of 7 per 100 million (range: 0–62, standard deviation:21).

Although our study examining for atrial somatic mutations involves both the largest number of patients and genes tested to date, our cohort size of 25 AF patients and 9 controls is still modest. In addition, our bioinformatic methods for identifying somatic mutations with next-generation sequencing, although highly sensitive and state-of-the-art, could potentially have failed to identify bona fide atrial somatic mutations (particularly those with an atrial somatic fraction ≤2% when our sensitivity is estimated to drop below 95% given our median 265-fold coverage depth). Although the frequency of potential somatic mutations was similar in both cases and controls, we cannot exclude the possibility that the discordant variants among the AF cases were pathogenic while those in controls were benign. Finally, although our genetic panel covered more than 3 million base pairs and an exhaustive number of genes related to cardiac pathophysiology (including the genes previously implicated in somatic mutations), it remains possible that genetic mosaicism involving undiscovered variants related to AF could yet be important. We chose to restrict our analysis to 560 genes in order to assure high depth coverage, thereby maximizing our sensitivity and specificity for accurately identifying somatic mutations.