Ethics approval

The collection and genomic analysis of the DNA were approved by the institutional review board at the University of Utah, and written informed consent was obtained from all study participants. Research was carried out in compliance with the Helsinki Declaration.

Sample collection

The samples used in our study all came from families of human research participants ascertained in clinics at the University of Utah (see Additional file 1, Figure S1 for pedigrees). Blood samples were collected and genomic DNA extracted using alkaline lysis and ethanol precipitation (Gentra Puregene; Qiagen Corp., Valencia, CA USA). DNA was quality-checked on agarose gels and quantified using a microvolume spectrophotometer (NanoDrop 2000; Thermo Fisher Scientific Inc., West Palm Beach, FL, USA).

Whole-genome sequencing and analysis with Complete Genomics

After quality control to ensure lack of genomic degradation, we sent DNA samples (10 ug) to Complete Genomics (CG) (Mountain View, CA, USA) for sequencing. The whole-genome DNA was sequenced using nanoarray-based short-read sequencing by ligation technology [14], including an adaptation of the pairwise end-sequencing strategy [15]. Reads were mapped to the Genome Reference Consortium assembly GRCh37. Owing to the proprietary data formats, all the sequencing data quality control, alignment, and variant calling were performed by CG as part of their sequencing service, using their version 2.0 pipeline [16].

Exome capture and sequencing with Illumina HiSeq2000

Exome capture for all 15 samples was carried out using a commercially available in-solution method (SureSelect Human All Exon v2; Agilent Technologies Inc., Wilmington, DE, USA), following the manufacturer's guidelines. This method is designed to target all human exons, regions totaling approximately 44 Mb, covering 98.2% of the Consensus Coding Sequence (CCDS) database. For the capture, a DNA-shearing instrument (focused-ultrasonicator; Covaris Inc., Woburn, MA, USA) was used to randomly fragment the pure and high molecular weight genomic DNA samples (experiments carried out by BGI-Shenzhen, Shenzhen, China), resulting in DNA fragments with a base-pair peak of 150 to 200 bp. Adaptors were then ligated to both ends of the resulting fragments. The adaptor-ligated templates were purified by magnetic beads (Agencourt AMPure SPRI; Beckman Coulter Inc., Brea, CA, USA), and fragments with an insert size of approximately 250 bp were excised. Extracted DNA was amplified by ligation-mediated (LM)-PCR, purified, and hybridized (SureSelect Library; Agilent Technologies) for enrichment. Hybridized fragments bound to the strepavidin beads, whereas the unbound non-hybridized fragments were washed out after 24 hours of hybridization. Captured LM-PCR products were analyzed using a microfluidics-based platform (2100 Bioanalyzer; Agilent Technologies) to estimate the magnitude of the enrichment. Paired-end sequencing was performed using a sequencing platform (HiSeq2000; Illumina Inc., San Diego, CA, USA) with average read lengths of 90 bp. Raw image files were processed (Pipeline version 1.6; Illumina Inc.) for base-calling, using the default parameters. FASTQ files were produced from the pipeline for downstream sequence data analysis. A gender check was compatible with the known genders of the collected human participants.

SNP arrays

DNA samples were genotyped on the SNP arrays (Human610-Quad, version 1; Illumina Inc.) with approximately 610,000 markers (including approximately 20,000 non-polymorphic markers) at the Center for Applied Genomics (Children's Hospital of Philadelphia, Philadelphia, PA USA). Total genomic DNA extracted from whole blood was used in the experiments. Standard data-normalization procedures and canonical genotype-clustering files provided by Illumina were used to process the genotyping signals. Concordance between SNPs from the arrays and SNPs from exome sequencing was determined by calculating the percentage of variants from exome sequencing and comparing this with the same genotype derived from the SNP arrays.

Alignment and variant calling

Post-variant calling analyses

Post-variant-calling analyses were performed using Golden Helix SVS (version 7.6.10 [25], ANNOVAR [26], the R suite of statistical programming tools http://www.r-project.org, and custom Perl scripts.

MiSeq sequencing for validation

Validation variants were randomly selected from sets of particularly controversial variants, indels and SNVs unique to GATK, indels and SNVs unique to SOAP, and variants (both SNVs and indels) shared by these two pipelines. PCR primers were designed using the software program Primer 3 http://sourceforge.net, to produce amplicons (ranging in size from 100 to 200 bp) containing variants of interest in approximately the center of the amplicon. Primers were obtained in 96-well plate format, 10 μmol/L dilution each (Sigma-Aldrich, St Louis, MO, USA). All primers were first tested for PCR efficiency using a HAPMAP DNA sample (Catalog ID NA12864l Coriell Institute for Medical Research, Camden, NJ, USA) and DNA polymerase (LongAmp^® Taq DNA Polymerase; New England Biolabs, Beverly, MA, USA). k8101-49685 genomic DNA was used as template for the validation experiment. After quality-control steps using agarose gel, the product was purified (ExoSAP-IT^® reagentsl Affymetrix Inc., Santa Clara, CA, USA) and pooled. Final PCR products were quantified (Qubit^® dsDNA BR Assay Kitl Invitrogen Corp., Carlsbad, CA, USA), then library construction for the sequencer platform (MiSeq Personal Sequencer; Illumina Inc.) was performed. Finally, before being loaded onto the MiSeq machine, the quality and quantity of the sample was verified using the Bioanalyzer (Agilent Technologies) and quantitative PCR (Kapa Biosystems Inc., Woburn, MA, USA).

Accessing data

All of the data have been submitted to the Sequence Read Archive under project accession SRP019719, corresponding to the 15 exomes and the single whole genome analyzed during the course of our study (see Additional file 2).

Data production summary

SNV analysis

Variant-calling pipeline concordance

SNV concordance between all 5 Illumina pipelines across all 15 exomes was 57.4% on average, and Ti/Tv ratios showed a generally increasing trend for sets of variants intersected by an increasing number of variant-calling pipelines (Figure 1). We found that for novel SNVs (those not found in dbSNP135) the overall concordance (11.4%) was much lower than the overall concordance between known SNVs (59.6%) (Figure 1). In a previous paper, we validated with Sanger sequencing or Sequenom genotyping 17 SNVs found in 3 of the current pilot samples [32]. Of these 17 validated SNVs, 16 were detected by all 5 pipelines, and the remaining variant was called by 4 of the 5 pipelines. Additional validation analyses are presented later in this paper.

A more detailed analysis of SNVs of one sample (k8101-49685) revealed that the exome variant calls had moderate to high depth of coverage (see Additional file 1, Figure S3). The range of read depths along with read-depth uniformity of variant calls varied between pipelines (see Additional file 1, Figure S3). Overall concordance between all five pipelines for sample k8101-49685 was 57.5%; however, sub-setting variants called by Illumina pipelines using increasingly stringent read-depth thresholds did not increase SNV concordance (see Additional file 1, Figure S4), and overall concordance was lowest when read-depth threshold values were at their highest (32.7% concordance when depth was required to be greater than 30 supporting reads).

Sequencing platform concordance

For sample k8101-49685, we selected variants generated by the CG pipeline that fell within the exon-capture regions of the Agilent SureSelect version 2 capture kit. We found that of the 21,050 SNVs identified by CG and located within the UCSC refGene regions, 19,407 (92%) were also within regions targeted for capture by the Agilent SureSelect version 2 kit. Of these, 2,085 (11%) were not called by any of the Illumina-based exome-analysis pipelines, despite computed high mappability scores for these variants (Figure 2) [33]. Of these 2,085 SNVs uniquely called by CG, an average of 558 had no sequence coverage as mapped by any of the Illumina-based exome-analysis pipelines. The Illumina exome read-depth for the remaining 1,527 CG-unique SNVs was calculated, and the majority of these SNVs were found to be in regions of very low Illumina sequence coverage (< 20 reads) in the exome datasets (see Additional file 1, Figure S5).

We found that 89.3% of CG SNVs (17,322 of 19,407) were contained within the union of all five Illumina pipelines (35,653 putative SNVs), whereas 18,331 of these 35,653 putative Illumina SNVs were not called by CG, suggesting a high false positive rate in the union of the Illumina calls and/or conversely a high false-negative rate in the CG calls (Figure 2). Overall concordance displayed marginal increases when VQLOW SNVs (low-quality CG variants) were removed from the pool of CG SNVs (Figure 2). Overall concordance remained stable as the depth of coverage threshold value associated with Illumina data calls increased (see Additional file 1, Figure S6A).

When only highly concordant Illumina SNVs (SNVs called by all five Illumina pipelines) were compared with the CG SNVs, only 64.4% (12507) of CG SNVs were contained within the concordant Illumina set, suggesting a high false-negative rate in this highly concordant Illumina set. Overall agreement decreased as the depth of coverage threshold value for Illumina calls increased, consistent with an increasing false-negative rate (see Additional file 1, Figure S6B).

Indel analysis

Variant-calling pipeline concordance

For indel calls, initial agreement between SOAPindel, SAMtools and GATK was very low at 3.0% (see Additional file 1, Figure S8). Indel coordinates were subsequently left-normalized and intervalized using a total range of 20 genomic coordinates (10 bp in each direction of their genomic coordinates). We found that increasing the intervalized indel range to as much as 60 genomic coordinates only marginally differed from having 20, so we chose to use 20 as a reasonable and conservative range for intervalizing indels. This method increased the overall concordance to 26.8% between the three indel-calling pipelines (Figure 3). For novel indels, the concordance (4.7%) was much lower than the overall concordance among known indels (43.3%). In an earlier paper, we previously validated with Sanger sequencing three indels found in three of the current pilot samples [32]. These three validated indels were detected by all three indel-calling pipelines.

Sequencing platform concordance

Indels falling within the range of the Agilent SureSelect version 2 exons were excised from the whole genome of sample k8101-49685 sequenced and analyzed by CG. Indels were again left-normalized and intervalized using a total range of 20 genomic coordinates, 10 in either direction. The normalized and intervalized CG indel calls were compared with all normalized and intervalized indels detected by the three Illumina data pipelines, and 32% were in agreement (Figure 4).

Cross-platform comparison of unique-to-pipeline indels

For the three Illumina data indel-calling pipelines, unique-to-pipeline indels were compared with indels discovered by CG (see Additional file 1, Figure S9). Concordance with CG indels was relatively low for all three pipelines with unique-to-GATK indels, showing a concordance of 24% with CG indels (324 of 1366), unique-to-SAMtools showing 29% concordance with CG indels (142 of 498) and unique-to-SOAPindel having 12% of called indels being concordant with CG indels (147 of 1246). Concordance rates improved for variants that were called by two Illumina data pipelines and further improved for variants called by all three of the indel-calling pipelines, 63% (1241 of 1986) and 90% (963 of 1069), respectively.

Novel (to dbSNP135) and unique-to-pipeline indels were also compared to CG indel calls, and the concordance rates for each pipeline were similar to those of the unique-to-pipeline only variants (see Additional file 1, Figure S9). Novel and unique-to-GATK indels had 24% of its indels concordant with CG indels (299 of 1236), novel and unique-to-SAMtools had 28% of its indels concordant with CG indels (96 of 343) and novel and unique-to-SOAPindel had 5% of its indels concordant with CG indels (53 of 1056). Novel indels that were called by two Illumina data pipelines displayed an increased concordance rate, 54% (229 of 423). Variants called by all three of the indel-calling pipelines showed the highest concordance rate for novel, unique-to-pipeline indels, 84% (103 of 122).

MiSeq validation of pooled PCR amplicons

To validate variants called by the two more widely used pipelines (SOAP and GATK), we used the orthogonal approach of PCR amplification of genomic DNA regions containing selected SNVs and indels, followed by pooled MiSeq sequencing. The PCR amplification (instead of exon capture), longer read lengths on the MiSeq platform, and the much higher depth of coverage provided a strong method of validation for SNVs and indels. A total of 1,140 SNVs found in sample k8101-49685 were selected for MiSeq validation; 760 of these SNVs were randomly selected from the set of SNVs that were unique to the GATK version 1.5 and SOAPsnp version 1.03 pipelines, 380 SNVs from each pipeline respectively. An additional 380 SNVs were randomly selected from the set of variants that were in agreement between GATK and SOAPsnp. After some analysis of the quality of the MiSeq data using FASTX, the MiSeq paired-end read data (version 2 sequencing kit, 250 × 250 bp reads) was trimmed to 150 bp and then aligned with BWA version 0.6.2 to the human reference genome sequence GRCh37, and variants were called with GATK UnifiedGenotyper version 2.3-9.

Of the 1,140 SNVs targeted for MiSeq validation, 919 (81.0%) were successfully amplified and sequenced, with an average read depth of 5,392. Validation rates for unique-to-GATK SNVs were high, with 306 of 315 (97.1%,) being successfully validated. For unique-to-SOAPsnp, 174 of 289 SNVs (60.2%) were validated. SNVs that were called by both GATK and SOAP were validated in 312 of 315 instances (99.1%) (Figure 5).

For indels found in sample k8101-49685, 960 were randomly selected for validation. Of these, 386 were randomly selected from the unique-to-GATK indel set, 387 were randomly selected from the unique-to-SOAPindel set, and 187 were randomly selected from set of indels overlapping between the two (SOAPindel and GATK). Of the 960 indels that were targeted for sequencing, 841 (83.5%) were successfully amplified and sequenced, with an average coverage of 4,866.

Unique-to-GATK indels had a validation rate of 180 of 336 (54.0%), being validated. The validation rate for unique-to-SOAPindel was found to be 44.6%, with 148 of 332 validating. For indels that were called by both SOAPindel and GATK, 132 of 169 (78.1%) were successfully validated (Figure 5).

GATK v2.3-9 and the new HaplotyperCaller

Newer implementations of SNV and indel-calling pipelines continually advance the field of variant discovery and analysis by increasing the accuracy by which variants can be reliably called. Here, we show an example of the differences between previous versions of GATK with respects to SNV calls and indel calls on the same sample, k8101-49685. The vast majority of SNV calls made by both the GATK UnifiedGenotyper version 2.3-9 and the GATK HaplotypeCaller version 2.3-9 modules overlapped with the SNV calls made by the GATK UnifiedGenotyper version 1.5, showing an overall concordance of 91.0% (27,150 of 29,912) and 87.0% (26,751 of 30779) respectively. However, for indel calls, the picture was quite different, with the GATK UnifiedGenotyper version 2.3-9 and GATK HaplotypeCaller version 2.3-9 modules showing an overall concordance with the GATK UnifiedGenotyper version 1.5 calls of 54.7% (1,688 of 3,085) and 54.6% (1,858 of 3,404) respectively (see Additional file 1, Figure S10).

A discussion about variant quality and the case for multiple methods for sequence analysis in personal genomes

For the five variant-calling pipelines included in our study, a large number of calls (both SNV and indel) are shared among them in each exome, 21,146 on average for SNVs. Although all five pipelines converge on a relatively large number of SNVs, this still represents less than 60% of the total SNV call set determined by all five pipelines, and hence there still exists a considerable degree of variation between pipelines used, with near-default parameterization on the same raw sequencing data. This disagreement is likely to be the result of many factors including alignment methods, post-alignment data processing, parameterization efficacy of alignment and variant-calling algorithms, and the underlying models utilized by the variant-calling algorithm(s).

SAMtools, SOAPsnp, and GATK use similar Bayesian methods to compute the posterior probability of the true genotype [19, 34–37], but they differ in the prior probabilities used. For example, SAMtools (MAQ) sets the prior probability for a heterozygous SNV at 0.001 for novel variants, and 0.2 for known SNVs. SOAPsnp uses a more complex method of assigning prior probabilities by distinguishing the homozygous genotype for the reference allele from the homozygous genotype for the alternative, and distinguishing transition (A↔G, C↔T) mutations from transversion (A/G↔C/T) mutations. GATK is similar to SAMtools but utilizes more advanced pre-processing and post-processing steps, such as local re-alignment around possible indel loci, and quality recalibration of both base quality and variant quality to improve overall variant-call performance. By modeling allele frequency, SNVer uses a frequentist framework for calling variants [24]. SNVer formulates variant calling as a hypothesis-testing problem so that a prior probability is not required, and SNVer could act as a complementary method to Bayesian methods. GNUMAP, which employs a probabilistic Needleman-Wunsch algorithm, might also be considered an orthogonal method, as even its computational framework for sequence alignment is novel [23].

Each variant-calling pipeline detects variants that others do not, and the accuracy of these dis cordant variants is expected to be low, but not zero. Indeed, for our comparison of SNVs called by SOAP, GATK, or both, MiSeq validation of unique-to-GATK and unique-to-SOAP variants demonstrated relatively high rates of validation, with 306 of 315 of the randomly selected SNVs from unique-to-GATK and 174 of 289 SNVs from unique-to-SOAPsnp being validated. Indels had lower, but still non-zero, validation rates, with 180 of 336 unique-to-GATK indels and 148 of 332 unique-to-SOAPindel indels being validated. Given that 'unique-to-pipeline' variants exist even in regions of relatively high sequence coverage, it is necessary to develop other approaches for including or excluding these variants from downstream analyses.

In the realm of biomedical research, every variant call is a hypothesis to be tested in light of the overall research design. Missing even a single variant can mean the difference between discovering a disease-contributing mutation or not [38]. For this reason, our data suggest that using a single bioinformatics pipeline for discovering disease-related variation is not always sufficient. A more comprehensive approach can be taken; all variants discovered by multiple variant-calling pipelines, when coupled with appropriate no-calling and quality filtering, could be included in downstream analyses, so as to not miss potentially disease-contributory variants. This is something that we intuitively implemented in a prior project [39], but for which these data now provide empirical support.

One can minimize false positives by increasing stringency filters, but this automatically and correspondingly increases the false-negative rate. The intersection between multiple variant-calling pipelines will minimize the false-positive rate, but we have also shown that each pipeline uniquely identifies some true variants. Hence, clinicians and policy-makers should be aware that propagating forward the intersection of variant calls will result in a high false-negative rate with exome sequencing and WGS, and we discuss below the advantage brought to bear on this issue by studying families. Therefore, processing single sample datasets from one sequencing platform with multiple variant-calling pipelines should not be the long-term solution for generating variant calls with high sensitivity and specificity, and we discuss alternative approaches below.

The case for indel standardization

Although the focus of most variant-calling software has been on detecting SNVs, it is the case that large-scale structural copy number variants and small indels are known to also be a biologically relevant and prevalent form of genetic variation [40, 41]. Indeed, initial indel mapping efforts revealed upwards of 800,000 indels in a diverse population that map to known human genes, some of which can be associated with genetic disease [42, 43], while recent estimates from the 1000 Genomes Project [44] suggest a 10:1 ratio of SNVs to indels in individual human genomes. Reliably detecting indels is therefore a crucial component of constructing a comprehensive set of clinically relevant genetic variants.

In contrast to SNVs, few indel-calling tools have been developed, so current knowledge of the existing variation due to indels, as well as the clinical implications of indels, has lagged. In spite of the fact that indel detection is becoming an important aspect of structural-variant analysis [40], indel calling is relatively imprecise and inaccurate. For example, the position of an indel with respect to its reference is, in many cases, ambiguous. An indel can often be represented at any of multiple locations. Krawitz et al. [45] designed an indel coordinate comparison metric, the equivalent indel region (eir), for comparing indel calls between pipelines, and GATK provides a tool which attempts to normalize indel position by left-justifying the indel within its multiple possible coordinate representations. Indeed, commonly used databases such as dbSNP have not yet entirely addressed the imprecision of indel-calling pipelines [45] and report only a single position for an indel, which could lead to disparate clinical diagnoses/outcomes between similarly affected individuals. We suggest that a more comprehensive approach should be taken, with all potential positions for each indel expressed and accounted for, so that downstream analysis can take advantage of the known existing ambiguity.

Our data demonstrates large discrepancies between indel-calling pipelines and suggests potentially high numbers of false positives and/or false negatives. Although putative false positives can be tested via modest resequencing efforts, false negatives and 'no-calls' require large-scale, often impractical, resequencing projects to discover them. Because of this limitation, few data exist on false-negative rates across pipelines, and inferring these rates from unique calls between pipelines is likely to be inadequate. Indel frequency, indel size, read length, and read depth are all known to affect the accuracy of indel-calling pipelines, and the performance of different pipelines also depends, in part, on experimental conditions [46]. We show that a relatively simple method can increase comparison accuracy for indels between pipelines and between individuals; left-normalized and intervalized indel calls allow rapid and reasonable comparison of called indels between different indel-calling pipelines, as well as between individuals who have had their genome or exome sequenced. The issues highlighted in our indel comparisons demonstrate the difficulties associated with attaining accurate and standardized indel calls, and our data illustrate the need for robust and ubiquitous indel standardization metrics/methods to allow for objective comparisons across pipelines and across sequencing projects.

The case for studying large families for discovering disease-related genetic variation

Our analysis of two families, one containing only two generations (parents and children) and the other containing three generations (one grandparent, both parents, and children) demonstrates that the ability to accurately distinguish de novo variants from familial inherited variants may be more strongly limited by high false-negative rates in the parents than by high false-positive rates in the children. This can be significantly improved by having sequence data from one or more grandparents or other relatives. This finding is particularly salient for single-generation de novo studies, which attempt to characterize novel variants that are associated with genetic disease observed in the children of healthy parents [47–51]. While such 'no-call' or false-negative errors in parents can be ameliorated somewhat with higher sequencing depth and/or more comprehensive variant-calling strategies, larger and more comprehensive pedigrees provide a powerful, complementary source for discovering and studying human genetic variation. Although most studies utilizing NGS data to date have focused on 'quads' or 'trios' [47–51], studying large and/or consanguineous families can maximize the utility of filtering strategies and statistical approaches for identifying disease-contributory variants in genetic disease [52].

The case for different platforms for a more comprehensive "exome"

The relative merits of WGS are expanding as both the cost of the technique decreases and as more scientists and clinicians use the technique in an increasing number of studies/analyses [5, 30, 53, 54]. We have shown that WGS with the version 2.0 CG pipeline delivers SNVs and indels not discovered by the Agilent exon capture and Illumina sequencing, and that these variants have a very low number of reads (< 20) at those positions in the Agilent exomes, arguing that the capture of those regions was not efficient. Conversely, there were a significant number of variants in our data that were unanimously called by the five Illumina variant-calling pipelines but not by the CG WGS. Exon capture and sequencing at high depth with one platform can yield a much higher depth of coverage in most exonic regions, whereas WGS offers a more uniform and comprehensive coverage that appears to cover regions missed by exon capture and sequencing. To attain a truly comprehensive set of exonic variants, WGS on one platform could be combined with exon capture and sequencing on a different platform. This combination of the depth of exonic sequencing provided by exon capture with the breadth of coverage of WGS on a different sequencing platform, alongside the use of multiple variant-calling pipelines, provides a powerful means to maximize sensitivity and specificity for any one personal genome. However, as costs for sequencing are reduced, we anticipate that sequencing whole genomes on two or more platforms may become a feasible option for similarly maximizing accuracy.

The current state of variant discovery

Many of the most recent advancements that have been made in variant discovery and sequencing analysis are those related to indel discovery and analysis. Indeed, newer versions of GATK have improved upon the false-negative and false-positive rates of their indel calls in both UnifiedGenotyper and with the newer HaplotypeCaller. By leveraging local de novo assembly, similarly to the SOAPindel pipeline used in this study, the new HaplotypeCaller from GATK potentially greatly improves upon its indel-calling accuracy. The more distinct differences observed between GATK indel calls by different versions reflects the fact that indel discovery is in the earlier stages of development, when large differences are often observed between and within pipelines, with accuracy potentially remaining relatively low. For example, in each pipeline, SNV calling relies on set algorithms, which are not dramatically changed in updated versions of the software. Therefore, we do not see great leaps in accuracy for SNVs with newer versions of GATK, despite the fact that we found in the current study that at least one other pipeline (SOAP) did uniquely discover some validated SNVs that were not discovered by any other pipeline or any version of BWA-GATK tested here, and vice versa. We also note that GATK discovered comparatively more unique SNVs not discovered by any other pipeline. One caveat is that we processed the MiSeq data with the newest BWA-GATK pipeline, so this might favor the exome variants previously called by GATK, but it is nonetheless the case that SOAP identified variants that GATK missed in the same exome data using near-default parameters.

Some limitations of our study

It is important to note that our study did not examine somatic mutations in tumor samples, so our conclusions and our testing pipelines focus only on germline mutations from diploid genomes. We recognize that variant calling in cancer genomes represents similar but somewhat distinct challenges, and that software tools (such as SNVMix [55]) are typically developed specifically for somatic mutations in cancer genomes. Our efforts were designed to evaluate whether or not rare variants within personal genomes can be reliably and comprehensively generated from sequencing data, with or without sequencing data from other family members. Hence, we did not evaluate pipelines that specifically employ imputation or multi-unrelated-sample variant-calling algorithms, such as Thunder [56], IMPUTE2 [57–59], or BEAGLE [60, 61], or specific procedures within the software tools used that allow for calling of multiple unrelated sample (such as those available by GATK) [13]. Additionally, software tools that are only commercially available (such as CLCBio) are not evaluated in our comparative study.