Integrating dilution-based sequencing and population genotypes for single individual haplotyping

Hirotaka Matsumoto 0 Hisanori Kiryu 0 0 Department of Computational Biology, Faculty of Frontier Science, The University of Tokyo , 5-1-5, Kashiwanoha, Kashiwa, Chiba 277-8561 , Japan Background: Haplotype information is useful for many genetic analyses and haplotypes are usually inferred using computational approaches. Among such approaches, the importance of single individual haplotyping (SIH), which infers individual haplotypes from sequence fragments, has been increasing with the advent of novel sequencing techniques, such as dilution-based sequencing. These techniques could produce virtual long read fragments by separating DNA fragments into multiple low-concentration aliquots, sequencing and mapping each aliquot, and merging clustered short reads. Although these experimental techniques are sophisticated, they have the problem of producing chimeric fragments whose left and right parts match different chromosomes. In our previous research, we found that chimeric fragments significantly decrease the accuracy of SIH. Although chimeric fragments can be removed by using haplotypes which are determined from pedigree genotypes, pedigree genotypes are generally not available. The length of reads cluster and heterozygous calls were also used to detect chimeric fragments. Although some chimeric fragments will be removed with these features, considerable number of chimeric fragments will be undetected because of the dispersion of the length and the absence of SNPs in the overlapped regions. For these reasons, a general method to detect and remove chimeric fragments is needed. Results: In this paper, we propose a general method to detect chimeric fragments. The basis of our method is that a chimeric fragment would correspond to an artificial recombinant haplotype and would differ from biological haplotypes. To detect differences from biological haplotypes, we integrated statistical phasing, which is a haplotype inference approach from population genotypes, into our method. We applied our method to two datasets and detected chimeric fragments with high AUC. AUC values of our method are higher than those of just using cluster length and heterozygous calls. We then used multiple SIH algorithm to compare the accuracy of SIH before and after removing the chimeric fragment candidates. The accuracy of assembled haplotypes increased significantly after removing chimeric fragment candidates. Conclusions: Our method is useful for detecting chimeric fragments and improving SIH accuracy. The Ruby script is available at https://sites.google.com/site/hmatsu1226/software/csp. - Background Advances in experimental techniques for DNA sequencing and genotyping have made it possible to determine many individual human genomes and detect variations, such as single nucleotide polymorphisms (SNPs) [1,2]. This has brought about great progress in genome analyses, such as genome-wide association studies (GWAS) [3], inference of population structure [4], and expression phenotypes [5]. However, most technologies give only genotype information and most current research does not determine the haplotype origin of the variations. Haplotypes contain more detailed information than genotypes and are valuable for GWAS [6], and for analyzing genetic structures such as linkage disequilibrium, recombination patterns [1], and correlations between variations and diseases [7]. Determining haplotypes experimentally is difficult, and there are three main computational approaches for haplotype inference. The first approach is the statistical phasing method, which infers population haplotypes from population genotypes using statistical computation [8-12]. Algorithms for statistical phasing have been developed in response to technological advances for genotyping, and its basis is that the diversity of haplotypes is limited, and there are conserved haplotypes [13]. Because of the strategy, statistical phasing does not work well in chromosomal regions which exhibit several different haplotypes, particularly regions of low linkage disequilibrium. This approach is also weak in inferring long haplotypes because the complexity of population haplotypes increases exponentially according to the number of SNPs. In the second approach, haplotypes are inferred from genotypes of pedigrees. For example, a childs haplotypes are determined from the genotypes of child and its parents (trio-based haplotyping). The origin of childs alleles can be determined if only one of the parents has the same alleles. However, the haplotypes of sites at which all family members have the same genotype cannot be determined and, furthermore, family genotype data are not always available. In addition, neither the statistical phasing method nor this approach can identify spontaneous mutations. The third approach uses DNA sequencing data and is called single individual haplotyping (SIH) or haplotype assembly [14-22]. SIH utilizes the fact that each sequenced read is derived from only one of the haplotypes. If a read spans two or more heterozygous sites, the haplotype can be determined from the co-occurrence of alleles in the read. Two reads are determined to originate from the same chromosome if they overlap at a region that has at least one heterozygous site, and they have the same alleles at these sites. SIH did not attract much attention until recently, since it needed long DNA sequencing reads that spanned multiple heterozygous sites, and obtaining such reads quickly and economically was difficult. However, this situation is changing rapidly with the advent of new experimental techniques, such as fosmid pool-based next-generation sequencing [17,23,24], long read fragment technology [25], and dilution-amplification-based sequencing [26] that can produce virtual long reads. In these methods, long DNA fragments are separated into distinct lowconcentration aliquots which each contain less than one fragment per chromosomal region. After sequencing an aliquot with a next-generation sequencer and mapping short reads, clusters are formed in which the reads are close to each other. A cluster corresponds to a long DNA fragment and is supposed to be derived from a single haplotype. Thus, virtual long reads can be obtained by merging the short reads in a cluster (see Figure 1). Although such experimental techniques are sophisticated, they have the problem of producing chimeric fragments whose left and right parts match different chromosomes very well. Because long DNA fragments are separated into aliquots randomly, there are cases where an aliquot has some long DNA fragments derived from the same region of different chromosomes and, consequently, reads with different chromosomal origins are regarded as one cluster and merged into a single fragment (see Figure 1). In the process of developing MixSIH [22], which is the first SIH algorithm that can evaluate the reliability of a haplotype region, we have shown that such chimeric fragments significantly decrea (...truncated)