Epstein-Barr virus (EBV) is a herpesvirus that infects epithelial and B cells and has been associated with the development of various tumors, including Hodgkin’s and Burkitt’s lymphoma[1–4]. Since EBV is able to transform B cells into continuously proliferating lymphoblastoid cell lines (LCLs), it is commonly used as a tool in clinical research for creating an unlimited source of patients’ material[5–9]. Although LCLs have been used frequently as a source of DNA in genetic studies, controversy still exists about their reliability in faithfully replicating the variation present in the donor germ-line (e.g.[6, 9–13]).

In all latently infected LCLs, the EBV genome is present in the form of extra-chromosomal copies (episomes) from which the viral genome is replicated[4]. Most research on the transforming abilities of EBV has been focused on the expression of viral gene products and the host’s transcriptional response. From this body of research it is now well understood that the viral transcription factor EBV nuclear antigen (EBNA)-2 activates the expression of several EBV proteins and non-coding RNAs - the growth transcription program - that interfere with the host’s signaling pathways[2, 14]. Specifically, the growth transcription program drives cell transformation by activating cellular proliferation, while suppressing growth inhibitors[14]. Even though the viral gene products exert their transforming functions primarily by interacting with host proteins, recent evidence suggests that EBV also promotes genomic instability in the host[15]. Furthermore, EBV has the potential to cause mutations through integration and disintegration into the host’s genome (e.g.[16, 17]). These putative early (pre-immortal) genetic consequences of EBV infection are less well studied. If ascertained, those structural genomic changes would have important implications for the interpretation of a large number of genetic studies that assume LCLs are a bona-fide source of genomic DNA.

In recent years, massively parallel DNA (i.e. next generation) sequencing has become increasingly affordable[18], enabling the discovery of causative DNA variants in rare genetic diseases[19–21] and providing new insights into carcinogenesis and autoimmunity[22–32]. A recent study reported the comparison of DNA isolated from peripheral lymphocytes and from an LCL from the same individuals by means of exome sequencing[33]. This study concluded that the exome fractions of genomic and lymphoblastoid cell line DNA (roughly 1-2% of the genomes) are more than 99% identical.

Whole genome sequencing can, in addition to identifying variation in coding regions, reveal copy number variation (CNV) and chromosomal rearrangements at high resolution. A whole genome sequence data set is therefore multilayered and can be queried for different aspects of genomic organization, making it a valuable tool for our study. Here we report the analysis of complete genome sequencing (40X) of a single individual to gain a better understanding of putative genetic alterations introduced by EBV transformation.

Here we report the complete sequencing and analysis of the normal genomic and lymphoblastoid cell line DNA from the same individual. We organized our analysis so as to proceed from investigation of major chromosomal rearrangements to small sequence variation, to single base changes. As a preliminary step, we performed high resolution karyotyping of the cell line to identify major chromosomal abnormalities. Fifteen out of 20 analyzed metaphases showed a normal karyotype (Additional file1). Only 5 cells showed random changes in chromosome number, consistent with what would be expected for an early-passage LCL. Once sequencing data was obtained, we evaluated quality metrics of the two sequencing data sets, such as overall coverage, ratio of hetero-to-homozygous single nucleotide polymorphisms (SNPs) and the ratio of SNP transitions to transversions. These parameters are similar to those reported for previously sequenced genomes using the same technology (e.g.[20, 21, 34]) and other human datasets using next generation sequencing (Additional file2). We also assessed the presence of viral DNA. To this end, we extracted raw reads of sequences that did not map to the human genome and aligned these to the EBV genome. We found an average of 2 copies of EBV DNA reads in the cell line as compared to 0 copies in the genomic DNA (data not shown), confirming that viral DNA was present only in the cell line sample. Since EBV has been described to integrate into host DNA (e.g.[35–40]), we also checked for viral DNA in all longer (>8 nt) insertions and substitutions that passed a pre-determined quality filter (SomaticScore of at least 0.1), as well as in all “non-reference” DNA stretches joining the two arms of a chromosomal rearrangement in the cell line (transition sequences). None of these calls supported the presence of inserted EBV genomic information.

A total of 217.33 Gb of sequence were generated from the cell line and 238.95 Gb from the genomic DNA. This difference in coverage did not alter most of the parameters analyzed, as these were highly similar between the two genomes (Table1). While the slightly deeper coverage of the genomic DNA did result in more variants being called in this sample, the overall depth of coverage was highly similar for the two genomes (~ 80% of both genomes were covered at least 40 times, Figure1A and C). Also, the regional distribution of reads was highly concordant (Figure1B). Once we established that quality measures were not significantly different, we set out to make a detailed comparison between the two genomes.

	genomic	cell line
Gender	male	male
Gross mapping yield (Gb)	238.95	217.33
SNP Transitions/transversions	2.14076	2.1468
SNP het/hom ratio	1.582496107	1.571157051
INS het/hom ratio	1.344223031	1.289809921
DEL het/hom ratio	1.663654705	1.654600015
SUB het/hom ratio	1.715782613	1.713757678
SNP total count	3346813	3271797
INS total count	187392	170265
DEL total count	204807	186626
SUB total count	71344	66370
SNP novel rate	0.0487541	0.0472266
INS novel rate	0.189213	0.17794
DEL novel rate	0.239469	0.235096
SUB novel rate	0.309865	0.295857
Fully called genome fraction	0.967831529	0.961877785
Partially called genome fraction	0.004574129	0.006604011
No-called genome fraction	0.027594342	0.031518204
Synonymous SNP loci	9778	9387
Missense SNP loci	9329	8935
Nonsense SNP loci	90	89
Nonstop SNP loci	13	13
Frame-shifting INS loci	121	110
Frame-shifting DEL loci	108	100
Frame-shifting SUB loci	17	13
Frame-preserving INS loci	113	101
Frame-preserving DEL loci	107	91
Frame-preserving SUB loci	258	227
Frame-shifting/preserving ratio	0.514644351	0.53221957
Nonsyn/syn SNP ratio	0.954080589	0.951848301
Insertion/deletions ratio	0.914968727	0.912332687
Ins + del/SNP ratio	0.117185812	0.109081034
Coding insertion/deletions ratio	1.060465116	1.078534031
Coding SNP/all SNP ratio	0.00626387	0.006151054
Coding (ins + del)/all (ins + del) ratio	0.001129529	0.001112384

This table also lists information on the number and kind of indels identified per genome, but does not contain data on structural variances.

Gb: gigabase, SNP: single nucleotide polymorphism, INS: insertion, DEL: deletion, SUB: substitution, (Non)syn: (non-)synonymous.

First, we assessed whether the two genomes differed in structural variants, i.e. chromosomal rearrangements and copy number variations (CNVs). Chromosomal rearrangements can be inferred from reads spanning over a stretch of DNA that is not contiguous in the reference genome (discordant reads). For example, if one half of a sequence read maps to chromosome 1 and the other half to chromosome 3, this might be indicative of a translocation event. By visual inspection of reported DNA junctions, we found that both genomes exhibited a largely similar number of chromosomal rearrangements compared to the reference and that they shared most of the inter-chromosomal junctions as well (Figure2A). Indeed, a similar but small number of unique chromosomal rearrangements were identified in each genome (131 in genomic, 78 in cell line; see Additional files3,4.5); 34% and 45% of these unique junctions, respectively, had been detected in more than 75% of all other genomes sequenced by CGI at that time and hence might represent false positives (Additional files4 and5). To examine whether observed junctions might be linked to the transformation process, we reasoned that if the observed dif-ferences were of biological origin (e.g. driven by the transformation), a set of genes involved in cell cycle regulation would be among the affected loci. GO and KEGG analyses revealed very similar gene categories affected by chromosomal rearrangements in both genomes (data not shown), thus suggesting these differences were likely random or false discoveries.

We next inspected ploidy (as a surrogate for CNVs) from genome coverage files in windows of 2 kb (Figure2B). Both genomes shared almost all CNVs throughout the genome; of note, most DNA stretches with ploidy > 2 were observed near telomers and centromers, likely reflecting the difficulty to properly align reads in these highly repetitive DNA regions. The cell line showed a decreased copy number in only 4 regions (3 haploid regions, one deletion) as compared to the genomic DNA (Additional file6). Four genes (KIAA0125, PRAME, ZNF280A, ZNF280B) and one pseudogene (ADAM6) were affected by the CNVs (Additional file6). None of these is reported to have a negative impact on cell proliferation. Hence, it is unclear whether their reduced ploidy plays any role in the transformation process. Notably, a 9-fold increase in copy number of mitochondrial DNA was observed for the cell line, likely reflecting the increased energy demand of the actively dividing transformed cells (Figure2B). This finding is consistent with a previous study[41].

We finally turned to an in-depth analysis of single nucleotide (SNP) and insertion/deletion (indel) polymorphisms. Using an automated whole-genome comparison algorithm (calldiff from cgatools) we found that 99.2% of the variant calls were identical between the two genomes (3,782,487 shared variants). Only 0.4% (15,364) and 0.3% (11,435) of variants were unique to the genomic and the cell line-derived DNA, respectively (Additional file7, panel B). Of note, this level of discrepancy is within range of the error rate between technical replicates using CGI technology (SY, unpublished observations). Although the number of expected differences between the 2 genomes from the same individual was low a-priori, we continued searching for potentially functional differences, namely non-synonymous variants such as “missense” (amino acid changing mutation), “nonsense” (creating a stop codon where there was none before), “nonstop” (removing an existing stop codon) or “frameshift” (changing the reading frame of a gene). For each class of variant we identified the genes that were affected in each sample and then determined the overlap between the two genomes. While we found that 92% of the affected genes overlapped (5,995 genes total), these genes were not always affected by exactly the same mutations in both genomes. To test the reliability of these called differences, we inspected the sequence reads of 307 selected genes (exhibiting in total 647 non-synonymous variants) using the Integrative Genomics Viewer (IGV)[42]. We observed that in most instances local coverage in one of the two genomes was very low, thus no variation call could be made with high confidence at that position. In other cases, one of the calls in the two genomes was wrongly reported due to ambiguity in the alignments (i.e. the reads could have been aligned differently, so that the position of a variant differed between the genomes, even though the resulting non-reference sequence was the same). Another common discrepancy arose from the fact that a variant was determined to be homozygous in one of the genomes and heterozygous in the other, even though both genomes were homozygous (Figure3A). Only 11% of the called differences between the two genomes were supported by visual inspection of actual reads (10 variants; Figure3B), implying that a considerable fraction of the reported differences between genomic and cell line DNA represents false positives.

In order to better control the false positive rate, we used the option -SomaticOutput within calldiff, in which a SomaticScore is computed for every variant that permits adjusting for sensitivity and specificity (sensitivity = 1 - SomaticScore). The SomaticOutput analysis requires specification of one sample as “normal” and another as “tumor” and generates an output containing all loci that are non-reference in the “tumor” sample. Since the transformed cell line can be regarded as a tumor sample derived from the normal genomic sample, definitions were set accordingly (“Cell line - > Tumor (CT)” analysis). The reciprocal definitions were also analyzed as a control (“Genomic - > Tumor (GT)” analysis). Since most of the variants only found in the genomic DNA sample are expected to be the result of sequencing or calling errors, the GT analysis provides a reasonable estimate of the experimental noise. For both comparisons, the number of variant calls was assessed using different SomaticScore cut-off values. As shown in Figure4A, increasing the SomaticScore cut-off increased the proportion of CT to GT variants, thus potentially maximizing true positive findings. Even though the total number of variant calls unique to the genomic DNA (retrieved by the GT analysis) is larger (Figure4B), a larger number of variants was detected in the CT analysis at all tested cut-off values (Figure4A). To minimize false positives, we chose a stringent cut-off of 0.5 [at this level, the number of differences in the CT analysis (417) almost doubles those found in the GT analysis (269)]. Assessing the regional distribution of these mutations revealed that variants unique to the cell line were randomly distributed throughout the genome; in contrast, a high proportion of variants unique to the genomic DNA seemed within or near telomeric or centromeric regions (Additional file8). Interestingly, 52% of variants in the CT analysis represented SNPs, compared to only 6% identified in the GT analysis (Figure4B). Since SNPs are more reliably called than other classes of variants, they are less likely to constitute noise. This could explain the low fraction of (confidently called) SNPs in the GT analysis, which is expected to mainly represent technical noise. We next compared the proportion of SNPs that were novel (not present in the dbSNP database[43]) in both analyses. Strikingly, whereas all SNPs that were only present in the genomic DNA have been reported before, none of the SNPs unique to the cell line were annotated in dbSNP, with the exception of one variant (Figure4B). Although the low number of identified variants between LCL and genomic DNA is within technical noise their novelty suggests that, if real, most of these differences would be random mutational events, driven by the accelerated proliferation of transformed cells. When assessing whether these SNPs altered coding sequences, we found that while 40% of them overlapped with genes, none had an impact on mRNA sequences. Specifically, except for one SNP, all variants are either located in introns or untranslated regions, thus their consequences are not straightforward (Additional file9). None of the 15 SNPs unique to the genomic DNA fell within a gene. In order to estimate the exact error rate of this technology, we randomly selected 60 SNPs unique to the cell line and re-analyzed these by Sanger sequencing. We could not confirm any of the identified variants (data not shown), suggesting that the real number of differences between the two genomes is even smaller than that implied by the SomaticOutput analysis.

Taken together, these results suggest that by using this technology at 40X resolution DNA from the cell line is mostly undistinguishable from genomic DNA from the same individual. The few putatively true differences are randomly distributed across the genome (Additional file8) and do not seem to drive the transformation process.

Here we report on the first genome-wide sequence-based analysis of the immediate genetic consequences of EBV transformation on a low-passage lymphoblastoid cell line from a subject with MS. While genomes from MS and healthy individuals may differ slightly, we deemed that this would not affect the conclusions of our study, which focused on characterizing the genomic consequences of EBV transformation. These effects should be clear-cut and insensitive to the source of the sample (with the exception of certain tumors, where DNA may contain abundant somatic mutations). For decades, the cell-transforming abilities of EBV have been used in genetic research to create repositories of subjects’ DNA. While the roles of viral gene products in the transformation process have been described in detail, whether genetic alterations are introduced as a consequence of EBV transformation is less well understood.

Several studies have systematically compared LCLs directly to their parental cells using traditional molecular techniques such as SNP typing, gene expression, and whole chromosome analysis[12, 44, 45]. The overall consensus is that no reproducible differences were identified[7, 45–47]. For instance, Redon et al. assessed differences between DNA from HapMap LCLs and their genomic counterparts and found that only 0.5% of observed CNVs were caused by transformation[46]. Another study confirmed that most CNVs in LCLs can also be seen in normal B cells[47]. Two of the CNVs reported in this latter study overlap with those detected in our LCL. This is not surprising since the genomic DNA was isolated from PBMCs, whereas the LCL is a B cell line – hence B cell specific CNVs were identified as differences to the genomic DNA in our study. The mitochondrial DNA CNV reported here was also seen as a “cell line-specific” CNV in a previous study[41].

Next generation sequencing was recently used to compare genomic and LCL DNA, although only the coding sequence (~1-2% of the genome) was assessed in that study[33]. Specifically, authors used exome sequencing to compare DNA from four LCLs with their corresponding genomic DNA (extracted from peripheral blood mononuclear cells). Focusing their analysis on SNPs and small indels, authors reported a 99.82% concordance between the parental DNA and the cell lines, with all discordant calls stemming from a single LCL-donor pair[33]. Given the relevance of non-coding regions in the modulation of gene expression and thus cell stability and function, whether the high level of concordance between genomic and LCL DNA extended to the whole genome remains an important question. By analyzing the whole genomes (at 40X) of an LCL and its genomic DNA counterpart, we identified 9,468 and 12,719 unique variants with respect to the reference genome, respectively. In-silico analysis reduced the number of differences to 417 (216 SNPs, 201 indels) in the LCL and 269 (15 SNP, 254 indels) in the genomic DNA. However, none of the 60 variants chosen for validation by Sanger sequencing were confirmed, thus suggesting that the number of real differences ought to be significantly smaller.

The known error rate of the ligation-based sequencing technique we employed was empirically determined to be approximately 1 in 100–200 kb (SY, unpublished observations). Hence, we can expect 20,000 to 30,000 errors in each genome. In a previous comparison of two technical replicate samples (sequencing the same sample twice) using the same technology, 27,893 differences were detected (SY, unpublished observations); we observed a similar number of differences (22,187) between the genomic DNA and the cell line genome in our analysis. These numbers highlight how close the difference between the two genomes is to technical noise. We therefore chose to minimize false positives in our analysis by applying stringent filters to the lists of called differences. By these means we identified a number of SNPs that seemed enriched in the cell line (216 SNPs in the CT analysis versus 15 SNPs in the GT analysis). These variants represented silent mutations and appeared to be random mutational events, possibly resulting from the accelerated division rate of the transformed cells. However, we were not able to confirm a selected subset of those SNPs and further validation is needed to establish any true discrepancies between the genomic DNA and the cell line.

In conclusion, our results indicate that EBV transformed cell lines at low passage number/short time in culture are genetically indistinguishable to the parental cells, suggesting that discoveries in genetic studies conducted using low-passage LCLs with a normal karyotype can be extrapolated to the parental patient samples. We determined with high confidence that 99.2% of the genomes were identical, with no reproducible changes in structural variation (chromosomal rearrangements and CNV) or indels. While we identified 231 differences (216 from the LCL, 15 from the genomic DNA) in single nucleotide variants, none of the 60 randomly selected variants validated by Sanger sequencing. These findings suggest that the true differences between the two genomes ought to be less than 5.8% of the candidates identified by bioinformatics filters (proportions test; 95% confidence interval p-value: 1.3 x 10^-14). We acknowledge that the sample size is a limiting factor of this study, and while similar results were reported by independent groups using different technologies[7, 9, 33, 45–47], larger studies will be needed to precisely determine the genomic effects of EBV transformation.