Alignment of Reads to Target

Enrichment specificity can be assessed by comparing the proportion of sequencing reads that map to the target regions. The lower the specificity, the higher the sequencing capacity required to achieve the desired target coverage. We observed large variability in the total number of reads produced by each of the three enrichment methods (Table 1, Table S4). This variability is also evident for the PCR and aHC technical replicates we conducted (for the Pool of 20; Tables S5, S6). It is common practice in whole genome sequencing (WGS) to remove potential duplicate reads to avoid biases in coverage analyses as well as downstream analyses, but applying this practice in pooled targeted sequencing of a relatively small target region with a high depth of coverage is still a matter of debate. Therefore we calculated alignment statistics both before and after removing potential duplicate reads. PCR showed the highest percentage of sequencing reads that map to the target region both before and after duplicate read removal (Table 1, Table S4). Conversely, both aHC and sHC showed higher proportions of mapped on-target reads with good mapping quality scores (20) both before and after duplicate read removal (Table 1, Table S4). The mapping quality score of reads is an important factor in accurate variant detection and the specificity of target enrichment impacts directly on target coverage.

Download:

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

Table 1. Target sequence enrichment success before duplicate removal.

https://doi.org/10.1371/journal.pone.0026279.t001

Target Coverage Depth and Uniformity

Target coverage depth directly affects the ability to detect variants, and depth is affected by the removal of potential duplicate reads. The higher enrichment specificity of PCR resulted in a higher overall mean read depth for target bases as compared to aHC and sHC, taking pool size and number of lanes sequenced into account regardless of duplicate read removal (Figure 1; Figures S1, S2, S3; Tables S7, S8). PCR yielded a higher percentage of target bases covered at 20× per individual across all pool sizes (Figure 1; Figure S1). However, target regions were not covered in a uniform way. For example, we found different coverage of protein coding versus non-coding target regions with duplicate read removal affecting the depth by approximately 100–200 reads but not the overall trend (Tables S9, S10; Figures S4, S5, S6, S7). Both aHC and sHC preferentially covered protein coding regions over non-coding regions across all pool sizes, whereas PCR demonstrated a bias in the opposite direction (Tables S9, S10; Figures S4, S5, S6, S7, S8, S9, S10, S11; t-test p-value0.05 in all pools, for all methods). The same trends were observed in the technical replicates conducted (Figures S8, S9, S10, S11). An analysis of %GC, repeat and low complexity regions in the protein coding and non-coding target regions (Table S11) showed that non-coding DNA contains a higher proportion of repeat elements, thereby making it difficult to design highly specific oligonucleotide probes, affecting coverage for the hybrid capture methods. PCR experiments tended to favour the overall lower GC content of non-coding regions (Figures S12, S13, S14).

Download:

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

Figure 1. Target coverage per individual in pool before duplicate removal.

This shows a cumulative relative frequency plot of the percentage of target bases with X coverage depth normalized by the number of individuals sequenced for: (A) Pool of 2, (B) Pool of 10, (C) Pool of 20 and (D) Pool of 50 individuals. The x-axis is in increments of 10× coverage. The black squares/lines illustrate the data for PCR enrichment, the blue squares/lines illustrate the data for aHC enrichment and the orange squares/lines illustrate the data for sHC enrichment. The first square represents the percentage of target bases with 10× coverage per individual in the pool, and so on for each square in increments of 10×. This analysis assumes equal representation of each individual in the pool of DNA.

https://doi.org/10.1371/journal.pone.0026279.g001

Variant Detection Sensitivity and Specificity

Variant discovery is linked with coverage depth, but study design power importantly also depends on a balance between false positive and false negative variant discovery rates. A major reason for the removal of duplicate reads is to remove biases in variant detection and calling. To address issues related to removing duplicate reads in variant detection and frequency estimation in pooled targeted resequencing we analyzed all pools with the removal of duplicate reads before variant calling, and pools of 1, 10 and 50 individuals for the PCR and sHC enrichment without the removal of duplicate reads. We found the total number of called variants to increase with pool size, in keeping with the variants known to be present in each pool (Tables S12, S13). The removal of potential duplicate reads reduces the total number of variants called, with the effect being largest for PCR enrichment and for larger pools (Tables S12, S13). As the number of sequence-identified variants increased, the proportion present in dbSNP129 decreased regardless of duplicate read removal (Tables S14, S15). This trend could either be due to a higher false positive rate in larger pools, or to the fact that deep sequencing identified variants not present in dbSNP. We utilized HapMap, Illumina chip and 1KG data available for the pooled individuals to directly address questions of false positive and false negative rates (Table S1). sHC demonstrated the highest sensitivity to detect HapMap variants across all pool sizes and for both removing and not removing duplicate reads, except in the case of enriching a single individual after duplicate read removal (in which case aHC performed best; Table 2, Table S16). The removal of duplicate reads has a dramatic effect on the sensitivity in the pool of 1 enriched by PCR. Although the pre-duplicate read removal sensitivity is higher overall the difference in sensitivity is only approximately 1–3%. The same trend was observed when considering 1KG variants and the union of all known variants (Tables S19, S20, S21, S22).

Download:

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

Table 2. HapMap variation detection sensitivity after duplicate removal.

https://doi.org/10.1371/journal.pone.0026279.t002

We found that PCR had overall lower sensitivity to detect known singleton HapMap variants compared to HC methods (Table S23). Similarly, HC methods showed higher sensitivity to detect the variants identified in the single-individual pool particularly after duplicate read removal (Tables S24, S25), and sHC generally performed better than aHC. The ability to accurately call variants depends on sequence coverage, and the depth is affected by duplicate read removal. The read depth of false negative HapMap variants was significantly different to that of true positives, for both HC methods across pools of 2–50 individuals (Figures S15, S16, S17; data not shown pool of 2 and 50) (t-test p-value0.05 in all cases). A similar trend was observed for PCR (Figures S15, S16, S17). For both hybrid capture methods there was a trend towards a lower GC content in 200 base-pair regions around false negative HapMap variants compared to true positive variants, and the pattern was similar before and after duplicate read removal (Figures S18, S19, S20). This trend was not as prominent for the PCR experiments. The ability to call variants is also tied to the frequency of the variant in the pool. The false negative HapMap variants tended to have lower allele frequencies in the pools compared to true positives, and this trend was accentuated before duplicate read removal (Figures S21. S22, S23, S24, S25). This is in keeping with the fact that false negatives have lower depth coverage, making low frequency variant detection more difficult.

We found specificity (true negative rate), calculated on the basis of HapMap loci monomorphic in the pooled samples, to decrease as the complexity of the pool increased, and for a given pool the specificity was higher after duplicate read removal (Table 3, Table S17). False positives could be ascribed to genotype misclassification in HapMap or to sequencing error in our experiment. To resolve this, we examined data across 22 of the pooled samples present in both HapMap and 1KG. 1KG data corroborate the pooled sequencing findings across over 92% of overlapping loci for pools of more than one sample after duplicate reads are removed. For sHC, the concordance is 100% regardless of pool size when duplicate reads are removed, but is reduced to 95% when duplicates are included for the pool of 1 individual (Table 4, Table S18). The inclusion of duplicate reads uniformly increases the proportion of calls corroborated by 1KG for PCR. We examined the rate of genotype discordance between HapMap and 1KG at all sites in the regions examined for the 22 samples and found it to be 1.8%. Given the deep coverage of target bases in our experiment and concordance with 1KG we infer that the calculated false positive rates are likely to be overestimates.

Download:

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

Table 3. HapMap variation detection specificity after duplicate removal.

https://doi.org/10.1371/journal.pone.0026279.t003

Download:

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

Table 4. 1KG support for HapMap false positive loci after duplicate removal.

https://doi.org/10.1371/journal.pone.0026279.t004

Variant Frequency Estimation

The usefulness of pooled sequencing approaches in complex trait studies is primarily encapsulated by the ability to perform association tests through allele frequency estimate comparisons between pools of disease cases and controls. We compared estimated allele frequencies from the resequenced pools with those from HapMap and 58BC data and found that the sHC designs achieve the highest accuracy (Figures 2–3, Figures S26, S27, S28). The accuracy of frequency estimates improved with increasing pool size and was higher after duplicate read removal. The correlation between estimated allele frequency from sequencing the pool of 50 and from known genotypes was 95.8%, 97.9%, and 99.0% for PCR, aHC, and sHC respectively when duplicate reads were removed (Figure 2). However, when duplicate reads were included in the analysis the correlation in the same pool increased slightly for the PCR enrichment and dropped slightly for the HC methods (Figure 3). The decrease in correlation between true and estimated allele frequency pre-duplicate read removal was also seen for the pool of 10 individuals (Figures S26, S27). The allele frequency estimates appear to be stable and robust. For example, frequency estimates from the technical replicates of the Pool of 20 have a correlation of 98.59% for PCR and 99.31% for aHC (Figures S29, S30). Overall, pooled sequencing resulted in under-estimates of the true allele frequency regardless of duplicate read removal (Tables S26, S27).

Download:

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

Figure 2. Accuracy of non-reference allele frequency estimation at HapMap/58C intersection variants for the Pool of 50 after duplicate removal.

An analysis of the correlation between the non-reference allele frequency estimates from the sequencing based variant caller and the allele frequency calculated from the reference genotypes. The analysis includes the true positive variants called by the sequencing based variant caller for which there were missing genotypes in the reference genotypes. The correlation coefficient is the Pearson's correlation coefficient. The figure shows the analysis for: (A) PCR enrichment, (B) aHC enrichment and (C) sHC enrichment.

https://doi.org/10.1371/journal.pone.0026279.g002

Download:

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

Figure 3. Accuracy of non-reference allele frequency estimation at HapMap/58C intersection variants for the Pool of 50 before duplicate removal.

An analysis of the correlation between the non-reference allele frequency estimates from the sequencing based variant caller and the allele frequency calculated from the reference genotypes. The analysis includes the true positive variants called by the sequencing based variant caller for which there were missing genotypes in the reference genotypes. The correlation coefficient is the Pearson's correlation coefficient. The figure shows the analysis for: (A) PCR, (B) aHC and (C) sHC enrichment.

https://doi.org/10.1371/journal.pone.0026279.g003

We found the per-individual read depth at called variants to be weakly correlated with frequency estimate accuracy, and to vary across enrichment methods (Figures S31, S32, S33, S34). The inclusion of potential duplicate reads before the analysis increased this correlation (Figures S32, S35). There was a stronger correlation between the number of variant alleles in the pool and the accuracy of the allele frequency estimates (Figures S36, S37, S38, S39, S40). This correlation was also higher when potential duplicate reads were included in the analysis (Figures S37, S40). Interestingly, the higher the number of variant alleles in the pool, the worse the allele frequency estimates, a trend consistently observed across all enrichment methods and pool sizes. Specifically, we observed that low frequency variants tended to be more accurately estimated (Figures 2–3; Figures S26, S27, S28, S36, S37, S38, S39, S40).

Reproducibility of Results

Reproducibility was assessed by performing technical replicates for PCR and aHC for the Pool of 20 individuals as a representative example. The HC replicates yielded more consistent results in terms of the number of reads produced and median coverage of target bases (Tables S5, S6). The sensitivity of HapMap variant detection varied by 4% between PCR replicates, and 2% between aHC replicates (Table S28). We next considered the number of variants that overlap between replicates as a function of the total number of unique variants called across replicates. The overlap rates of called variants across pairs of replicates were low (59%) for both PCR and aHC (Table S29). For variants called in both technical replicates the correlation between estimated allele frequencies was found to be high (98.6% and 99.3% for PCR and aHC respectively) (Figures S29, S30). When comparing allele frequencies for these overlapping variants (i.e. expecting identical estimates under an ideal experimental scenario), we found an average absolute allele frequency difference of 2.7% for PCR (across 7,233 overlapping variants) and 2.1% for aHC (6,713 variants) (Table S29).

Cost

We compared the relative cost implications of the different study designs considered here. Considering the results after duplicate reads were removed, the Pool of 10 individuals had the highest sensitivity and specificity for pools greater than 1 individual but they were only 2% higher than the Pool of 50 which provided better allele frequency estimates and was more cost-effective. For example, for a pooling experiment involving 1000 cases and 1000 controls the Pool of 50 would be associated with 30% lower costs based on the number of sequencing lanes required as compared to the Pool of 10 and 86% lower costs than sequencing each individual on a single lane. Within each pool size, the cost of PCR was 3-fold more expensive than either of the hybrid-capture enrichment methods.

Ethics Statement

This study has been approved by the ethics committee of the Wellcome Trust Sanger Institute (WTSI). This study only used extracted DNA from cell-lines, which falls outside of the UK Human Tissue Act. The use of the 1958BC samples is covered by a material transfer agreement (MTA) with the ALSPAC Laboratory, University of Bristol (the 1958BC sample custodian), which stated that the 1958BC had been collected under UK NHS Research Ethics Committee approval from SouthEast MREC, in Aug. 2002. REC Ref. MREC 01/1/44. The HapMap Populations/ELSI Group made recommendations for the HapMap project during the initial planning phase, and developed an informed consent form template (http://hapmap.ncbi.nlm.nih.gov/consent.html). The use of the HapMap CEU DNA is governed by these individually signed informed consent forms that grant permission for the use of the DNA in future studies approved by relevant ethics committees. The use of the HapMap DNAs were approved by the HapMap Repository (Coriell).

DNA Samples

The samples sequenced consisted of 31 HapMap CEU individuals and 19 individuals from the 1958 British Birth Cohort (58BC). The HapMap DNA samples were obtained from Coriell Repositories and the sample IDs are: NA12249, NA12156, NA12004, NA11831, NA12716, NA11832, NA11993, NA12057, NA11995, NA12006, NA12144, NA12802, NA12146, NA12005, NA12003, NA07000, NA12043, NA12044, NA11992, NA11881, NA11994, NA07345, NA12154, NA06994, NA06985, NA12239, NA07022, NA07034, NA12155, NA07056, NA06993. Individuals with a 1KG superscript were sequenced as part of pilot 1 of the 1,000 Genomes Project [4].

Region Selection

The genomic regions selected for sequencing (Table S3) had shown suggestive evidence for association with type 2 diabetes following cumulative analysis of low frequency/rare variants directly typed on GWAS chips using a collapsing method [19]. Association in these regions did not replicate when further sample sets were tested. The targets for enrichment span 1.6 Mb in total and include entire genic regions that encompass 3′ and 5′ UTRs, introns, and exons, and have been defined as 50 Kb either side of the transcriptional start and stop sites.

Array and Solution Oligonucleotide pool design

Genomic coordinates for the regions of interest were submitted to Nimblegen for the design of custom 385K arrays covering the target regions. Oligonucleotide pools for hybridization in solution phase were generated by Nimblegen to cover the same target regions. To cover real-estate on the array, three further regions were added on the hybrid-capture arm of the experiment (for a total of 1.96 Mb). These additional regions were excluded from the analysis presented here. This exclusion results in an under-estimation of the percentage of reads mapping back to target for the aHC and sHC experiments in Table 1.

Preparation of the pools

Each DNA sample was quantified using standard picogreen protocols and normalized to 50 ng/l. The pools were generated by mixing the required volumes of the appropriate number of samples to give a final concentration of each pool of 50 ng/l. The concentration of the resulting pool was checked using picogreen. Aliquots of the same pool were used for both PCR and hybrid-capture.

PCR

Primers were designed automatically using Primer 3 to achieve a 5-fold depth of 5- and 10 kb amplicons across the target regions. Where necessary, manual primer design of 5 kb amplicons using Primer 3 was used to fill any gaps in the coverage following the automatic design. In total 462×10 kb STSs and 737×5 kb STSs were designed automatically and 88×5 kb STSs manually. All primers were pre-screened on a set of four genomic DNAs. Products were separated on an 0.8% agarose gel, visualised with ethidium bromide staining and scored as pass/weak/fail. Based on the prescreening results a final set of STSs were chosen to give 3-fold coverage over the target regions which consisted of 256×10 kb STSs and 256×5 kb STSs. Aliquots of the same DNA pools used for hybrid capture were used as template for PCR amplification with each STS. 5 kb amplicons were amplified as follows: Primers were pre-aliquoted at a concentration of 10 ng/l, 4 l per well into 384-well PCR plates. A premix was made consisting of 2 µl of 10× Buffer (as supplied with the enzyme), 0.4 l 10 mM dNTPs, 0.8 l 50 mM MgSO4 (as supplied with the enzyme), 0.12 l Platinum Hi-Fi Taq, 11.8 l DDW and 30 ng of pooled DNA per reaction and added to the pre-aliquoted primers. PCR cycling conditions were as follows: 98°C for 3 minutes, followed by 15 cycles of 94°C for 30 seconds, 68°C for 30 seconds, with the annealing temperature decreasing by 1oC per cycle, 68°C for 5 minutes followed by 19 cycles of 94°C for 30 seconds, 58°C for 30 seconds, 68°C for 5 minutes followed by 68°C for 10 minutes. 10 kb amplicons were amplified as follows: Primers were pre-aliquoted at a concentration of 10 ng/l, 4 l per well into 384-well PCR plates. A premix was made consisting of 2 l of 10× Buffer (as supplied with the enzyme), 0.4 l 10 mM dNTPs, 0.8 l 50 mM MgSO4 (as supplied with the enzyme), 0.16 l Platinum Hi-Fi Taq, 11.14 l DDW and 90 ng of pooled DNA per reaction and added to the pre-aliquoted primers. PCR cycling conditions for were as follows: 98°C for 3 minutes, followed by 15 cycles of 94°C for 30 seconds, 68°C for 30 seconds, with the annealing temperature decreasing by 1oC per cycle, 68°C for 10 minutes followed by 19 cycles of 94°C for 30 seconds, 58°C for 30 seconds, 68°C for 10 minutes followed by 68°C for 10 minutes. Products were visualised using ethidium bromide staining. PCR products from each DNA pool for all STSs were pooled together in equimolar amounts and used to construct an Illumina library prior to sequencing as described below.

Illumina Library Construction

20 g of DNA were sheared to 100–400 bp using a Covaris S2 following manufacturer's protocols and the settings Duty Cycle, 20%; Intensity, 5.0; Cycles/burst, 200; Duration, 90; Mode, Freq Sweeping. Sheared samples were quantitated on a Bioanalyzer (Agilent, Santa Clara, USA). 10–15 g of sheared DNA were end-repaired, A-tailed and Illumina sequencing adapters ligated to the resulting fragments using the Illumina Paired-End DNA Sample Prep protocol with the slight modification that the gel size selection step was replaced with a SPRI bead purification (following manufacturer's protocol).

Array Hybridization

5 g of each library were hybridized to a custom Nimblegen 385-K array following manufacturer's protocols (Roche/Nimblegen) with the modification that no pre-hybridization PCR was performed. Captured samples were washed and eluted in 50 l of PCR-Grade water following manufacturer's protocols. Eluted samples were amplified using a master-mix containing 2 mM MgCl2, 0.2 mM dNTPs, 0.5 M PE.1. 0.5 M PE.2 and 3 units of Platinum® Pfx DNA Polymerase per sample. Samples were aliquoted into 3 individual wells of a plate and amplified using the following conditions: 94°C for 5 minutes followed by 20 cycles of 94°C for 15 seconds, 58°C for 30 seconds, 72°C for 30 seconds and a final extension of 72°C for 5 minutes. PCR products were purified using SPRI beads prior to sequencing.

Solution Hybridization

1 g of each library was hybridized to an oligo pool following manufacturer's protocols with the modifications that 14 cycles of pre-hybridization PCR were performed and 50× COT1DNA was used in the hybridization. Following hybridization the captured samples bound to the Streptavidin beads were washed following manufacturer's protocols. Post-capture PCR was performed on the captured samples bound to the beads as described above.

Sequencing

Captured libraries were sequenced on the Illumina Genome Analyzer II (GAII) platform as paired-end 37-bp or 54-bp reads, following manufacturer's protocols. The raw sequencing reads are available through the European Genome-Phenome Archive (http://www.ebi.ac.uk/ega, accession EGAS00001000134) and the European Nucleotide Archive (http://www.ebi.ac.uk/ena, accession ERP000770).

Read Mapping and Sequence Analysis

The reference human genome used in these analyses was UCSC assembly hg18 (NCBI Build 36), including unordered sequence. Each lane of sequencing was mapped to the reference genome using Maq (v0.7.1) with default parameters [20]. For pools with multiple lanes of sequencing, the individual lane mappings were merged with the Maq utility mapmerge. The phred-scaled base quality scores from the GAII were recalibrated using the Quality Score Recalibration tool in the Genome Analysis Toolkit (v1.0 build January 21, 2010) [21]. Duplicate reads were identified and marked using Picard (v1.17; http://picard.sourceforge.net/), and for a subset of the analyses duplicates were removed with SamTools (v0.1.7) [22]. The number of reads mapped and mapped to target regions was calculated using the view utility in SamTools. The %GC versus coverage analysis was performed using the CollectGcBiasMetrics utility in Picard. The analysis of the repeat and low-complexity content of the coding and non-coding target regions were performed with the RepeatMasker software (v. open-3.2.9) [23].

Variant Calling and Frequency Estimation

Variants were called on the merged BAM file from all lanes for a pool. The BAM file used to call variants had recalibrated base quality scores, reads mapping off the end of the reference soft-clipped, and either duplicate reads marked or removed. The variant calling and frequency estimation was performed by Syzygy (v0.9.5.39) using the default parameters. Syzygy calls single nucleotide variants and single base insertion/deletions [7] (http://www.broadinstitute.org/software/syzygy/). This analysis only considered Syzygy single nucleotide variant calls. Variants are defined as a locus having 1 non-reference allele, an allele different than the reference genome used for mapping, present in the pool. Syzygy assigns a confidence score to all variant calls (high, medium and low). We analyzed all the called variants regardless of confidence.

Comparison Genotypes

The sensitivity, specificity and frequency estimation analyses were conducted by comparing the variants and frequency estimates from the Syzygy calls to the known variant content in the pool using existing genotype data for each pooled individual. We used the non-redundant release 27 HapMap genotypes for the 31 HapMap individuals used in the pooling experiments. The genotypes were mapped to the forward strand of Build 36 of the reference genome and sensitivity analysis included all loci where the HapMap genotypes indicated that there was at least one non-reference base in the pool, whereas the specificity and allele frequency estimation analysis only included loci where all individuals in the pool had non-missing genotype data. Twenty-two of the HapMap individuals used in our pooling experiments were sequenced in Pilot 1 of the 1,000 Genomes Project. We used 1KG genotypes4 for these individuals from the final pilot 1 call set released March 28, 2010. Due to the fact that no pool consisted solely of individuals sequenced in 1KG, we are unable to perform specificity analysis for the 1KG loci. The 1958 Birth Cohort (58BC) genotypes came from 2 sources. Sixteen of the pooled individuals were genotyped as part of the Wellcome Trust Case Control Consortium 2 (WTCCC2; Illumina 1.2 M Duo platform) [24] and 6 individuals were genotyped as part of this project at the Wellcome Trust Sanger Institute (Illumina 670K platform). The WTCCC2 genotypes were downloaded from the European Genotype Archive (http://www.ebi.ac.uk/ega/). The frequency estimation and variant discovery specificity analysis for the pool of 50 was based on the intersection of variants that occurred in both the HapMap and 58BC genotype sets. The variant discovery sensitivity analysis for the pool of 50 was carried out by taking the union of variants in 1KG, HapMap and 58BC genotype sets. The dbSNP variants used were dbSNP129 variants downloaded from the UCSC genome browser, with all rsIDs that mapped 2 locations in the genome removed (referred to as the non-redundant dbSNP129). The coding/non-coding analysis was performed by defining coding intervals for each gene as per the March 27, 2009 release of the consensus coding sequence (CCDS) project [25].

Statistical Sequence and Variant Analysis

All statistical analyses were performed with the R statistical software package [26]. The target regions and called variants were separated into different subsets and two-sided, two-sample t-tests with unequal variances were performed to assess differences in the means of the distributions. An obtained t-test p-value of 0 indicates that the p-value of the test was more significant than the statistical software R would calculate (the highest exponent on the machine used for calculation is 1024). The correlation coefficients reported in Figure 2 and Figures S15, S16, S17, S18, S19, S20, S21, S22, S23, S24 are Pearson's correlation coefficients. Figures S19, S20, S21 further investigate the relationship between individual read depth and allele frequency accuracy, defined as the HapMap frequency minus the Syzygy estimated frequency, by a least squares fitting of the model, , and the red lines in these figures shows the resulting estimate of the intercept and . Figures S22, S23, S24 further investigate the relationship between allele count and allele frequency accuracy, as defined above, by a least squares fitting of the model, , and the red lines in these figures shows the resulting estimate of the intercept and .