Reducing the exome search space for Mendelian diseases using genetic linkage analysis of exome genotypes

Many exome sequencing studies of Mendelian disorders fail to optimally exploit family information. Classical genetic linkage analysis is an effective method for eliminating a large fraction of the candidate causal variants discovered, even in small families that lack a unique linkage peak. We demonstrate that accurate genetic linkage mapping can be performed using SNP genotypes extracted from exome data, removing the need for separate array-based genotyping. We provide software to facilitate such analyses.


Background
Whole exome sequencing (WES) has recently become a popular strategy for discovering potential causal variants in individuals with inherited Mendelian disorders, providing a cost-effective, fast-track approach to variant discovery. However, a typical human genome differs from the reference genome at over 10,000 potentially functional sites [1]; identifying the disease-causing mutation among this plethora of variants can be a significant challenge. For this reason, exome sequencing is often preceded by genetic linkage analysis, which allows variants outside of linkage peaks to be excluded. The linkage peaks delineate tracts of identity by descent sharing that match the proposed genetic model. This combination strategy has been successfully used to identify variants causing autosomal dominant [2][3][4] and recessive [5][6][7][8][9][10][11] diseases, as well as those affecting quantitative traits [12][13][14]. Linkage analysis has also been used in conjunction with whole genome sequencing (WGS) [15].
Other WES studies have not performed formal linkage analysis, but have nonetheless considered inheritance information, such as searching for large regions of homozygosity shared by affected family members using genotypes obtained from genotyping arrays [16][17][18] or exome data [19,20]. This method does not incorporate genetic map or allele frequency information, which could help to eliminate regions from consideration, and is applicable only to recessive diseases resulting from consanguinity. Recently, it has been suggested that identity by descent regions be identified from exome data using a non-homogeneous hidden Markov model (HMM), allowing variants outside these regions to be eliminated [21,22]. This method incorporates genetic map information but not allele frequency information and requires a strict genetic model (recessive and fully penetrant) and sampling scheme (exomes of two or more affected siblings must be sequenced). It would be suboptimal for use with diseases resulting from consanguinity, for which filtering by homozygosity by descent would be more effective than filtering by identity by descent. Finally, several WES studies have been published that make no use of inheritance information whatsoever, despite the fact that DNA from other informative family members was available [23][24][25][26][27][28][29][30][31].
Classical linkage analysis using the multipoint Lander-Green algorithm [32], which is a HMM, incorporates genetic map and allele frequency information and allows for great flexibility in the disease model. Unlike the methods just mentioned, linkage analysis allows dominant, recessive or X-linked inheritance models, as well as permitting variable penetrances, non-parametric analysis and formal haplotype inference. There are few constraints upon the sampling design, with unaffected individuals able to contribute information to parametric linkage analyses. The Lander-Green algorithm has produced many important linkage results, which have facilitated the identification of the underlying disease-causing mutations.
We investigated whether linkage analysis using the Lander-Green algorithm could be performed using genotypes inferred from WES data, removing the need for the array-based genotyping step [33]. We inferred genotypes at the location of HapMap Phase II SNPs, [34] as this resource provides comprehensive annotation, including the population allele frequencies and genetic map positions required for linkage analysis. We adapted our existing software [35] to extract HapMap Phase II SNP genotypes from WES data and format them for linkage analysis.
We anticipated two potential disadvantages to this approach. Firstly, exome capture only targets exonic SNPs, resulting in gaps in marker coverage outside of exons. Secondly, genotypes obtained using massively parallel sequencing (MPS) technologies such as WES tend to have a higher error rate than those obtained from genotyping arrays [36]. The use of erroneous genotypes in linkage analyses may reduce power to detect linkage peaks or result in false positive linkage peaks [37].
We compared the results of linkage analysis using array-based and exome genotypes for three families with different neurological disorders showing Mendelian inheritance ( Figure 1). We sequenced the exomes of two affected siblings from family M, an Anglo-Saxon ancestry family showing autosomal dominant inheritance.
The exome of a single affected individual, the offspring of first cousins, from Iranian family A was sequenced, as was the exome of a single affected individual, the offspring of parents thought to be first cousins once removed, from the Pakistani family T. Families A and T showed recessive inheritance. Due to the consanguinity present in these families, we can perform linkage analysis using genotypes from a single affected individual, a method known as homozygosity mapping [33].

Results and discussion
Exome sequencing coverage of HapMap Phase II SNPs Allele frequencies and genetic map positions were available for 3,269,163 HapMap Phase II SNPs that could be translated to UCSC hg19 physical coordinates. The Illumina TruSeq platform used for exome capture targeted 61,647 of these SNPs (1.89%). After discarding indels and SNPs whose alleles did not match the HapMap annotations, a median 56,931 (92.3%) of targeted SNPs were covered by at least five high-quality reads ( Table  1). A median of 64,065 untargeted HapMap Phase II SNPs were covered by at least five reads; a median 78% of these untargeted SNPs were found to lie within 200 bp of a targeted feature, comprising a median 57% of all untargeted HapMap SNPs within 200 bp of a targeted feature.
In total, we obtained a minimum of 117,158 and a maximum of 133,072 SNP genotypes from the four exomes. The array-based genotyping interrogated 598,821 genotypes for A-7 and T-1 (Illumina Infinium HumanHap610W-Quad BeadChip) and 731,306 genotypes for M-3 and M-4 (Illumina OmniExpress Bead-Chip). Table 2 compares the inter-marker distances between exome genotypes for each sample to those for the genotyping array. The exome genotypes have much more variable inter-marker distances than the genotyping arrays, with a smaller median value.

Optimization of genotype concordance
We inferred genotypes at the positions of SNPs located on the genotyping array used for each individual so that we could investigate genotype concordance between the two technologies. We found that ambiguous (A/T or C/ G SNPs) comprised a high proportion of SNPs with discordant genotypes, despite being a small proportion of SNPs overall. For example, for A-7 at coverage ≥ 5 and t = 0.5 (see below), 77% (346 of 450) of discordant SNPs were ambiguous SNPs, while ambiguous SNPs composed just 2.7% of all SNPs (820 of 30,279). Such SNPs are prone to strand annotation errors, as the two alleles are the same on both strands of the SNP. We therefore discarded ambiguous SNPs, which left 29,459 to 52,892 SNPs available for comparison (Table 3). Several popular genotype-calling algorithms for MPS data require the prior probability of a heterozygous genotype to be specified [38,39]. We investigated the effect of varying this parameter, t, upon concordance of genotyping array and WES genotypes (given WES coverage ≥ 5; Table 3). Increasing this value from the default 0.001 results in a modest improvement in the percentage of WES genotypes being correctly classified, with most of the improvement occurring between t = 0.001 and t = 0.05. The highest concordance is achieved at t = 0.5, where all four samples achieve 99.7% concordance, compared to 98.7 to 98.9% concordance at the default t = 0.001.
We note that t = 0.5 may not be optimal for calling SNP genotypes on haploid chromosomes. At t = 0.5, the male M-4 had five × chromosome genotypes erroneously called as heterozygous out of 1,026 (0.49%), while the male T-1 had one such call out of 635 genotypes (0.16%). The same SNPs were not called as heterozygous by the genotyping arrays. No heterozygous × chromosome calls were observed at the default value of t = 0.001.

Linkage analysis and LOD score concordance
Prior to performing linkage analysis on exome and array SNP genotypes, we selected one SNP per 0.3 cM to ensure linkage equilibrium while retaining a set of SNPs dense enough to effectively infer inheritance. The resulting subsets of WES genotypes (Table 4) contained 8,016 to 8,402 SNPs with average heterozygosities of 0.40 or 0.41 among the CEPH HapMap genotypes, obtained from Utah residents with ancestry from northern and western Europe (CEU). The resulting subsets of array genotypes (Table 4) contained more SNPs (12,173 to 12,243), with higher average heterozygosities (0.48 or 0.49).
Despite this difference, there was good agreement between LOD scores achieved at linkage peaks using the different sets of genotypes ( Figure 2, Table 5). The median difference between the WES and array LOD scores across positions where either achieved the maximum score was close to zero for all three families (range -0.0003 to -0.002). The differences had a 95% empirical interval of (-0.572,0.092) for family A, with the other two families achieving narrower intervals ( Table 5).

Efficacy of filtering identified variants by location of linkage peaks
If our genetic model is correct, then variants lying outside of linkage peaks cannot be the causal mutation and can be discarded, thus reducing the number of candidate disease-causing variants. Table 6 lists the number of nonsynonymous exonic variants (single nucleotide variants or indels) identified in each exome, as well as the number lying with linkage peaks identified using WES genotypes. The percentage of variants eliminated depends upon the power of the pedigree being studied: 81.2% of variants are eliminated for the dominant family M, which is not very powerful; 94.5% of variants are The denominator for percentages is the total number of HapMap Phase II SNPs in that distance category.

Conclusions
Linkage analysis is of great potential benefit to WES studies that aim to discover genetic variants resulting in Mendelian disorders. As variants outside of linkage peaks can be eliminated, it reduces the number of identified variants that need to be investigated further. Linkage analysis of WES genotypes provides information regarding the location of the disease locus to be extracted from WES data even if the causal variant is not captured, suggesting regions of interest that may be targeted in follow-up studies. However, many such studies are being published that employ less sophisticated substitutes for linkage analysis or do not consider inheritance information at all. Anecdotal evidence suggests that a substantial proportion of MPS studies of individuals with Mendelian disorders fail to identify a causal variant, though an exact number is not known due to publication bias. We describe how to extract HapMap Phase II SNP genotypes from massively parallel sequencing data, providing software to facilitate this process and generate files ready to be analyzed by popular linkage programs. Our method allows linkage analysis to be performed without requiring genotyping arrays. The flexibility of linkage analysis means that our method can be applied to any disease model and a variety of sampling schemes, unlike existing methods of considering inheritance information for WES data. Linkage analysis incorporates population allele frequencies and genetic map positions, which allows superior identification of statistically unusual sharing of haplotypes between affected individuals in a family.
We demonstrate linkage using WES genotypes for three small nuclear families -a dominant family from which two exomes were sequenced and two consanguineous families from which a single exome was sequenced. As these families are not very powerful for linkage analysis, multiple linkage peaks with relatively low LOD scores were identified. Nonetheless, discarding variants outside of the linkage peaks eliminated between 81.2% and 99.43% of all nonsynonymous exonic variants detected in these families. The number of variants remaining could be reduced further by applying standard strategies, such as discarding known SNPs with minor allele frequencies above a certain threshold. Our work demonstrates the value of considering inheritance information, even in very small families that may consist, at the extreme, of a single inbred individual. As the price of exome sequencing falls, it will become feasible to sequence more individuals from each family, resulting in fewer linkage peaks with higher LOD scores.
Exome capture using current technologies yields large numbers of useful SNPs for linkage mapping. Over half of all SNPs covered by five or more reads were not targeted by the exome capture platform. Approximately 78% of these captured untargeted SNPs lay within 200 bp of a targeted feature. This reflects the fact that fragment lengths typically exceed probe lengths, resulting in flanking sequences at both ends of  a probe or bait being captured and sequenced. The serendipitous result is that a substantial number of non-exonic SNPs become available, which can and should be used for linkage analysis. We found that setting the prior probability of heterozygosity to 0.5 during genotype inference resulted in the best concordance between WES and array genotypes.
The authors of the MAQ SNP model recommend using t = 0.2 for inferring genotypes at known SNPs [38], while the default value used to detect variants is t = 0.001. Our results highlight the need to tailor this parameter to the specific application, either genotyping or rare variant detection. Although we anticipated WES genotypes being less accurate than array genotypes, all four samples achieved a high concordance of 99.7% for SNPs covered by five or more reads at t = 0. 5 We found that LOD scores obtained from WES genotypes agreed well with those obtained from array genotypes from the same individual(s) at the location of linkage peaks, with the median difference in LOD score zero to two or three decimal places for all three families. This was despite the fact that the array-based genotype sets used for analysis contained more markers and had higher average heterozygosities than the corresponding WES genotype sets, reflecting the fact that genotyping arrays are designed to interrogate SNPs with relatively high minor allele frequencies that are relatively evenly spaced throughout the genome. By contrast, genotypes extracted from WES data tend to be clustered around exons, resulting in fewer and less heterozygous markers after pruning to achieve linkage equilibrium. We conclude that if available, array-based genotypes from a high resolution SNP array are preferable to WES genotypes; but if not, linkage analysis of WES genotypes produces acceptable results.
Once WGS is more economical, we will be able to perform linkage analysis using genotypes extracted from WGS data, which will obviate the problem of gaps in SNP coverage outside of exons. The software tools we provide can accommodate WGS genotypes without requiring modification. In the future, initiatives such as the 1000 Genomes Project [1] may provide populationspecific allele frequencies for SNPs not currently included in HapMap, further increasing the number of SNPs available for analyses, as well as the number of populations studied.
The classic Lander-Green algorithm requires markers to be in linkage equilibrium [40]. Modeling linkage disequilibrium would allow incorporation of all markers without the need to select a subset of markers in linkage equilibrium. This would allow linkage mapping using distant relationships, such as distantly inbred individuals who would share a sub-linkage (< 1 cM) tract of DNA homozygous by descent. Methods that incorporate linkage disequilibrium have already been proposed, including a variable length HMM that can be applied to detect distantly related individuals [41]. Further work is being targeted towards approximations of distant relationships to connect sets of related pedigrees [42]. These methods will extract the maximum information from MPS data from individuals with inherited diseases.
We have integrated the relatively new field of MPS in families with classical linkage analysis. Where feasible, we strongly advocate the use of linkage mapping in combination with MPS studies that aim to discover variants causing Mendelian disorders. This approach does not require purpose-built HMMs, but can utilize existing software implementations of the Lander-Green algorithm. Where genotyping array genotypes are not available, we recommend utilizing MPS data to their full capacity by using MPS genotypes to perform linkage analysis. This will reduce the number of candidate disease-causing variants that need to be evaluated further. Should the causal variant not be identified by a WES study, linkage analysis will highlight regions of the genome where targeted resequencing is most likely to identify this variant.

Materials and methods
Informed consent, DNA extraction and array-based genotyping Written informed consent was provided by the four participants or their parents. Ethics approval was provided by the Royal Children's Hospital Research Ethics Committee (HREC reference number 28097) in Melbourne. Genomic DNA was extracted from participants' blood samples using the Nucleon™ BACC Genomic DNA Extraction Kit (GE Healthcare, Little Chalfont, Buckinghamshire, England).
All four individuals were genotyped using Illumina Infinium HumanHap610W-Quad BeadChip (A-7, T-1) or OmniExpress (M-3, M-4) genotyping arrays (fee for service, Australian Genome Research Facility, Melbourne, Victoria, Australia). These arrays interrogate  Individual T-1 was sequenced using one-quarter of a flow cell lane while the other three individuals were sequenced using one-eighth of a lane. Paired-end reads of 110 bp were generated. Reads were aligned to UCSC hg19 using Novoalign version 2.07.05 [43]. Quality score recalibration was performed during alignment, and reads that aligned to multiple locations were discarded. Following alignment, presumed PCR duplicates were removed using MarkDuplicates.jar from Picard [44]. Table S1 in Additional file 1 shows the number of reads at each stage of processing, while Tables S2 and S3 in the same file show coverage statistics for the four exomes.

WES genotype inference and linkage analysis
SNP genotypes were inferred from WES data using the samtools mpileup and bcftools view commands from release 916 of the SAMtools package [45], which infers genotypes using a revised version of the MAQ SNP model [38]. We required base quality and mapping quality ≥ 13. SAMtools produces a variant call format (VCF) file, from which we extracted genotypes using a Perl script.
These genotypes were formatted for linkage analysis using a modified version of the Perl script linkdatagen.pl [35] with an annotation file prepared for HapMap Phase II SNPs. This script chose one SNP per 0.3 cM to be used for analysis, with SNPs selected to maximize heterozygosity according to CEU HapMap genotypes [34]. Array-based genotypes were prepared for linkage analysis in the same way, using annotation files for the appropriate array.
The two Perl scripts used to extract genotypes from VCF files and format them for linkage analysis are freely available on our website [46], as is the annotation file for HapMap Phase II SNPs. Users may also download VCF files containing WES SNP genotypes for the four individuals described here (both for HapMap Phase II and genotyping array SNPs), as well as files containing genotyping array genotypes for comparison.
Multipoint parametric linkage analysis using WES and array genotypes was performed using MERLIN [47]. A population disease allele frequency of 0.00001 was specified, along with a fully penetrant recessive (family A, family T) or dominant (family M) genetic model. LOD scores were estimated at positions spaced 0.3 cM apart, and CEU allele frequencies were used.

WES variant detection
SAMtools mpileup/bcftools was also used to detect variants from the reference sequence with the default setting of t = 0.001. Variants were annotated by ANNOVAR [48] using the UCSC Known Gene annotation. For the purposes of filtering variants, linkage peaks were defined as the intervals in which the genome-wide maximum LOD score was obtained, plus 0.3 cM on either side.