Robustness of Massively Parallel Sequencing Platforms

The improvements in high throughput sequencing technologies (HTS) made clinical sequencing projects such as ClinSeq and Genomics England feasible. Although there are significant improvements in accuracy and reproducibility of HTS based analyses, the usability of these types of data for diagnostic and prognostic applications necessitates a near perfect data generation. To assess the usability of a widely used HTS platform for accurate and reproducible clinical applications in terms of robustness, we generated whole genome shotgun (WGS) sequence data from the genomes of two human individuals in two different genome sequencing centers. After analyzing the data to characterize SNPs and indels using the same tools (BWA, SAMtools, and GATK), we observed significant number of discrepancies in the call sets. As expected, the most of the disagreements between the call sets were found within genomic regions containing common repeats and segmental duplications, albeit only a small fraction of the discordant variants were within the exons and other functionally relevant regions such as promoters. We conclude that although HTS platforms are sufficiently powerful for providing data for first-pass clinical tests, the variant predictions still need to be confirmed using orthogonal methods before using in clinical applications.


Introduction
The robustness and the reproducibility are the sine qua non of every data intended to be used for clinical applications. These factors have been the main issue hindering large scale applicability of array-based technologies for clinics. High throughput sequencing (HTS) offers alternative solutions to array based technologies with respect to genotyping, and HTS data are Data Availability Statement: All SRA files are available from the NCBI database (accession number (s) SRP021510).: http://www.ncbi.nlm.nih.gov/sra/? term=SRP021510. considered to be more robust and comprehensive. The performance of HTS platforms has been tested in various studies [1][2][3], but the robustness of HTS platforms still need to be systematically assessed. More specifically, it is of crucial importance to obtain accurate single nucleotide polymorphism (SNP), indel, and structural variation (SV) call sets in the sense that the calls made for specific SNPs or SVs should be solely dependent on the actual genotypes of sequenced individuals but not the location, time, or the platform of choice of the study.
Here we investigate the robustness of the Illumina HiSeq platform, currently the most widely used HTS technology in genome sequencing. In order to achieve this, we resequenced the genomes of two individuals from the Turkish Genome Project [4] twice. The two genomes were previously sequenced once [4], using the Illumina HiSeq 2000 platform in BGI Shenzhen, and a second time through the same platform set up at the Turkish Advanced Genomics and Bioinformatics Research Group (TÜBİTAK İGBAM). Although the same model sequencing machines were used, roughly the same level of coverage was achieved, and identical tools were used with identical parameters, independent analysis of the SNP and indel calls revealed significant number of differences between the two trials. In particular, we noticed that roughly 280 thousand of the 3 million SNPs genotyped by the GATK [5] tool in one trial (e.g. BGI) or the other (e.g. TÜBİTAK) are unique to only one callset-implying that the reproducibility rate of SNP calls is * 92%. Interestingly, the multisample calling option of GATK that jointly analyzes two WGS datasets simultaneously does not seem to substantially improve the reproducibility and thus accuracy of the results. In this study, we explore the "sources" of this loss of accuracy as a function of both quality scores and coverage levels in each of the samples. Although increase in coverage levels in each sample typically decreases the differences between the GATK calls for specific loci on the two samples, there are still some cases in which differences can not be attributed to low coverage or quality score differences.
Our main contribution in this paper is a detailed investigation of the types and causes of exclusive variants within the call sets that are expected to be substantially the same. In addition, we try to identify strategies to handle such discrepancies when there is a second WGS dataset generated from the genome of the same donor. With further technological advancements and the cost improvements, sequencing a sample many times can be expected to be prevalent, as storing the data may become more expensive than resequencing the same sample. Here the same donor sample is sequenced twice, to evaluate the outcome of this highly possible situation in the future. For such cases, when there are more than one WGS sequence of the same donor, we state our remarks on how to exploit all the data fruitfully. In Section 1, we describe the methods used in the study. In Section 2, we present the results of the study and show the shared and exclusive sets of different SNP groups. And finally, in Section 3, we provide our remarks on the results and conclude.

DNA Samples and Ethics Statement
Genomic DNA from two individuals were collected and purified in 2011, only once from the blood of two volunteers for a previously published study [4]. The source (i.e. blood), DNA extraction time and location are the identical. As indicated in [4], institutional review board permission was obtained from INAREK (Committee on Ethical Conduct in Studies Involving Human Subjects at the Boğaziçi University) before data collection, and all participants including those that are included in this study provided informed consent. preparation kits for data production and the funder BİLGEM-TÜBİTAK (The Scientific and Technological Research Council of Turkey) (Award Number: T439000) provided support in the form of salaries for authors (PK, MOK, MŞS), and the funder Marie Curie Career Integration Grant (Award Number: 303772) provided support in the form of salaries for author (CA) but did not have any additional role in the study design, data collection and analysis, decision to publish, or preparation of the manuscript. The specific roles of these authors are articulated in the "author contributions" section.
Competing Interests: The authors have declared that no competing interests exist. One of the author's current affiliation is a commercial company. This does not alter the authors' adherence to PLOS ONE policies on sharing data and materials.

Sequencing
The genomes of the two individuals were already sequenced using Illumina HiSeq2000 in 2011 at BGI Shenzhen [4]. The same samples were resequenced for a second time using another Illumina HiSeq2000 in 2012 at the TÜBİTAK İGBAM located in Kocaeli, Turkey. For the first sequencing data set, DNA samples were fragmented to 500bp, and paired-end sequencing data were generated with a read length of 90bp. For the second sequencing experiment at TÜBİ-TAK, we used the same protocols and sheared the DNA to 500bp fragments, and sequenced 104bp paired-end reads. In the remainder of the paper, we refer to the data generated at BGI as S 1B (first individual) and S 2B (second individual), and the data generated at TÜBİTAK for the same individuals as S 1T and S 2T .

Alignment, coverage, GC content
To discover SNPs and short indels, we mapped the reads to the human reference genome (NCBI GRCh37) using the BWA aligner (version 0.6.2) [6], in paired-end mode ("sampe") and default options. We converted the mapping output to sorted, duplicate-removed, and indexed BAM files using SAMtools [7]. We calculate the expected coverage as: num mapped reads Â read length 2; 897; 310; 462 ðnumber of non À N bases in GRCh37Þ Next, SAMtools and BEDtools [8] were used to calculate the effective coverage: Finally, we used the FASTQC tool (version 0.10.1) [9] to collect basic statistics of the genomic sequence data (Table 1).

Variant calling
SNP and indel detection. After the initial alignment and the PCR-duplicate removal, we realigned the indel-containing reads to the reference genome using GATK Realigner tool. We then used the GATK UnifiedGenotyper tool to generate the SNP and indel call sets. We also used the GATK HaplotypeCaller as an alternative approach for variant calling. Next, we eliminated likely false positives using the GATK Variant Quality Score Recalibration (VQSR) tool with GATK resource bundle v2.5. Finally, we further filtered the call sets using the GATK VariantFiltration to remove low confidence calls (SnpCluster filter to remove SNPs if there are more than 3 SNPs in a 10 bp window). We applied the same variant calling pipeline to each of the four datasets separately: S 1B , S 1T , S 2B and S 2T . Pooled SNP and indel calling. As a second experiment, we tested whether pooling data from multiple sequencing runs for the same samples improve callset reproducibility. Our main question here was to understand if the slight differences in the coverage and depth of the datasets could be ameliorated by merging data for discovery, and if this would improve genotyping accuracy. For this purpose, we applied the SNP/indel detection pipeline to both samples by pooling two sequencing datasets (i.e. S 1BT , and S 2BT ) generated at BGI and TÜBİTAK.
However, we named the two datasets from the same sample as if they were generated from different genomes. In the remainder of the paper, we denote the SNP/indels genotyped within the BGI data from S 1 as B 1 , and the SNP/indels genotyped within the TÜBİTAK data from S 1 as T 1 for this experiment. Similarly, we have B 2 and T 2 for the sample S 2 .

Data Availability
We had previously deposited the sequence reads obtained from BGI to the SRA read archive (SRP021510). Primary run IDs relevant to this study are: SRR839600 for S 1B and SRR849493 for S 2B . Datasets generated at TÜBİTAK are also available as "secondary sequencing" data sets with sample IDs SRR2128004 and SRR2128088 respectively within the same SRA archive. We also released our scripts we used to map the reads and call the variants at https://github.com/ pinarkavak/robust, and the VCF files for the call sets are available at http://alkanlab.org/paperdata/Kavak_RobustNGS/.

Read length, coverage, GC content
We provide the basic analysis of the input data sets in Table 1. Briefly, we generated a total of more than 5.5 billion reads, equivalent to * 530 Gbps, where the effective sequence coverage per data set ranged from 29.5X to 49.2X. The reads sequenced at TÜBİTAK (S 1T and S 2T ) were 14bp longer than the reads sequenced at BGI (S 1B and S 2B ), and the GC contents were similar (Table 1).

Call Sets and comparisons
SNP and indel discovery. We generated SNP call sets for each sample and for pooled data sets (Methods). 4 SNP call sets: S 1T , S 1B , S 2T , S 2B ; and 2 pooled call sets for S 1 and S 2 , denoted as S 1BT , S 2BT were generated. 3 call sets per sample (i.e., S 1T , S 1B , and S 1BT for S 1 and S 2T , S 2B , and S 2BT for S 2 ) were compared with each other to quantify and characterize any differences. The SNP and indel statistics are summarized in Table 2. The SNP and indel statistics that were obtained by HaplotypeCaller are also shown in Table 3.
Separately generated call sets. Briefly, after potential false positive removal (Methods), we observed approximately 95% agreement between the pairs of SNP call sets generated from both genomes ( Table 4). The indel call sets showed a larger discrepancy, where only 18%-68% of each callset were shared with the other two call sets ( Table 4). The number of shared and discrepant SNP and indels of HaplotypeCaller are shown in Table 5. To understand the causes of different calls from the same genomes, we investigated the underlying sequence content of the discrepancies of novel SNP and indel calls in detail. First, we downloaded the human reference genome annotations (segmental duplications and common repeats) from the UCSC genome browser (http://genome.ucsc.edu), CNV call sets from the 1000 Genomes Project [11]. We then calculated and annotated the number of novel SNPs and indels (Fig 1, Fig 1A and 1C, Fig 1B and 1D). We found that 46%-59% of discrepant novel SNP calls intersected with common repeats, and 5%-28% intersected with segmental duplications. In addition, a 3%-5% of the discrepant calls were found within the CNV regions reported in the 1000 Genomes Project [11], and 0.3%-0.8% were discovered in low coverage areas (< 5X). Analysis of the discrepant indel calls yielded similar results (Fig 1B and 1D). Next, we investigated the types of common repeats for the discrepancies. The majority of discrepant calls were found to be within Alu and L1 repeats (Tables 6 and 7). The incongruent calls within satellites and low complexity repeats were negligible. In addition, a close look to Alu and L1 subfamilies revealed that the number of discrepant calls peaked at * 10% sequence divergence from consensus sequences, also showing negligible differences at recent and distant mobile element insertion loci (data not shown). Both of these observations can be explained by low mapping quality within these regions, causing the GATK VQSR algorithm to filter out such calls.
The significance of the discrepant SNP and indels in terms of predicted functionality was more closely investigated (Table 8). We found that 88%-95% of the discrepant SNP calls mapped to intergenic and intronic regions, where a 3.5%-4.5%-were predicted to be within coding exons, and ncRNAs. Indels showed similar properties, where only 0-3 of them were predicted to incur frameshifts.
Pooled BGI vs Pooled TÜBİTAK. The number of shared and discrepant SNP and indel calls are shown on Table 9. This strategy showed a better correspondence between the two datasets, reducing the contradicting call rate to 0.1%-0.8%. The number of shared and discrepant SNP and indel calls of pooled HaplotypeCaller are also shown on Table 10. Table 5. Comparisons of total and novel SNP and indel call sets generated from the genomes of S 1 and S 2 . S 1B , S 1T , S 1BT : S 1 calls from BGI, TÜBİTAK, and pooled datasets using HaplotypeCaller; S 2B , S 2T , S 2BT : S 2 calls from BGI, TÜBİTAK, and pooled datasets, respectively.

Discussion and Conclusion
With the improvements in cost efficiency, speed, and analysis algorithms, HTS platforms are now being considered to be used routinely as part of health care. This assumption prompted a pilot project called ClinSeq [12] that aims to investigate the strength and potential pitfalls of using HTS data in the clinic. However, the HTS technologies continue to evolve and new platforms are introduced almost every month. This, coupled with changes and updates of  algorithms to analyze HTS data raises questions about the maturity and robustness of HTS platforms for accurate discovery and genotyping of genomic variants.
In an effort to answer this question, we analyzed the genomes of two individuals, each sequenced twice using the same technology, albeit at different locations. Since our aim was to investigate the maturity of sequencing platforms in this study, we used the same tools to characterize both single nucleotide and short indel variants. Under the assumption of 100% robustness, one would expect to characterize the same set of variants in both sequencing datasets from the same genomes, however, this is not what we found.
We believe multiple factors contribute to this effect. First, since the library preparation is different, one may expect difference in GC% bias, as clearly seen in Table 1 of the manuscript. This leads to differences in read depth over different regions of the genome, which in turn causes discrepancies in variation calls. The GC% effect can also explain the over-representation of repeats and segmental duplications in terms of SNP discrepancies, as common repeats are high in GC content (41.45% GC within common repeats vs 40.33% GC in unique regions), together with difficulties in mapping to repeats and duplications. Second, although the make and model of the sequencing instruments are the same, they are individually different machines, which may account for slight differences in base calling errors. Third, mapping biases against repeats and duplications incur additional problems in terms of mapping and calling. We note that we used the same mapping and calling tools with the same parameters for all datasets in this study, therefore the tools should not be the reason for discrepancies. Although orthogonal methods are needed for definitive validations, we suggest that when there are more than one data set, one should use all the available data for higher accuracy. Sequencing machines, alignment and genomic variant discovery and genotyping algorithms change rapidly, and one must be careful when interpreting results. Here we demonstrated potential problems that may arise within HTS-based studies. Discrepancies between call sets generated from the same genomes may be complementary false positives and false negatives in each callset, in addition to common genotyping errors. Luckily, much of the differences were found within non-genic regions and common repeats, which are of less importance for most studies.