Joint Variant and De Novo Mutation Identification on Pedigrees from High-Throughput Sequencing Data

Abstract

The analysis of whole-genome or exome sequencing data from trios and pedigrees has been successfully applied to the identification of disease-causing mutations. However, most methods used to identify and genotype genetic variants from next-generation sequencing data ignore the relationships between samples, resulting in significant Mendelian errors, false positives and negatives. Here we present a Bayesian network framework that jointly analyzes data from all members of a pedigree simultaneously using Mendelian segregation priors, yet providing the ability to detect de novo mutations in offspring, and is scalable to large pedigrees. We evaluated our method by simulations and analysis of whole-genome sequencing (WGS) data from a 17-individual, 3-generation CEPH pedigree sequenced to 50× average depth. Compared with singleton calling, our family caller produced more high-quality variants and eliminated spurious calls as judged by common quality metrics such as Ti/Tv, Het/Hom ratios, and dbSNP/SNP array data concordance, and by comparing to ground truth variant sets available for this sample. We identify all previously validated de novo mutations in NA12878, concurrent with a 7× precision improvement. Our results show that our method is scalable to large genomics and human disease studies.

1. Introduction

Whole-genome and exome sequencing has been successful in the elucidation of highly penetrant genes in early childhood diseases and is making inroads in complex trait studies entailing thousands of samples (Gilissen et al., 2012). Because of its shotgun nature, mis-mapping of short reads in complex genomic regions, and relatively high sequencing error rates, calling variants from human high-throughput sequencing (HTS) data still results in substantial false positives and false negatives (Ajay et al., 2011). The problem is magnified when looking for de novo mutations in parent–offspring trios or larger families, as this enriches for sequencing artifacts (Veltman and Brunner, 2012). This is problematic since de novo mutations are thought to be responsible for about half of all early neurodevelopmental childhood disorders (Veltman and Brunner, 2012) and likely a similar fraction of neonatal/prenatal cases (Saunders et al., 2012; Talkowski et al., 2012).

Numerous methods have been proposed for the identification of variants from HTS data. Initially, methods were based on frequentist approaches that used heuristic filters (McKernan et al., 2009) but soon were displaced by Bayesian approaches (DePristo et al., 2011; Garrison and Marth, 2012) that leveraged prior information and were able to deal with the uncertainty in the data (Marth et al., 1999). Results from the 1000 Genomes Project provided large-scale empirical validation of these approaches and identified the major sources of artifacts in variant calling: mapping ambiguities caused by the complexity of the human genome that result in spurious variants, inflated or nonlinear base quality values from the sequencing platforms, and library artifacts and sequencing chemistry systematic errors that result in false positives and compositional biases (1000 Genomes Project Consortium, 2010, 2012). Nevertheless, because of difficulty of modeling all sources of errors and artifacts, variant-calling methods evolved into complex, multistep pipelines addressing separately different problems in the data, including (1) mapping and alignment, (2) realignment for indels or complex regions, (3) base quality recalibration, (4) initial variant calling, and (5) variant quality score recalibration. These methods are complex to use, are slow, and cannot deal effectively with related individuals who are present in family studies and the emerging clinical applications of HTS.

In order to improve upon these problems, we developed a novel Bayesian network framework that calls variants simultaneously across a pedigree implicitly leveraging shared haplotypes in its members and incorporating a Mendelian segregation model. Here we present how our Bayesian framework escapes combinatorial explosion (as compared with more simplistic approaches), is highly scalable to large pedigrees, can deal with low coverage and missing data, and scores de novo mutations. This approach can be applied to single samples, groups of unrelated individuals, pedigrees, or a combination thereof. Coupled with a fast read mapping and aligning algorithm, our approach can process reads-to-variants in a matter of minutes (whole exome) to a few hours (whole genome) using commodity hardware and scaling with O(n³) complexity with the number of samples.

2. Methods

2.1. Mapping and alignment

We mapped and aligned paired-end sequencing reads from FastQ files to the hg19 human reference genome with decoys, as constructed by the 1000 Genomes Project (1000 Genomes Project Consortium, 2012). Briefly, reads were mapped by an aligner that used a 2-bit encoded index of hashes and then querying this index using hashes computed on the reference genome. We simultaneously mapped paired reads, incrementally analyzing each chromosome through a sliding-window process exceeding the size of the expected mate-pair insert size, and performing read lookup on the first w nucleotides (default word size and step was 20). Hits were aligned with in a multistage hierarchical alignment process, going from perfect string matching, to simple substitution, to seeded matching, and to a banded Gotoh alignment algorithm (Gotoh, 1982). For all recoded mapping locations, MAPQ scores were calculated (Li et al., 2008) and were output as a BAM file (Li et al., 2009). Base quality values were also recalibrated using reported base quality and read group ID as covariates, and provided in a separate recalibration file (1000 Genomes Project Consortium, 2010). For more details, see Supplementary Methods (Supplementary Material is available online at www.liebertpub.com/cmb).

2.2. Haploid and diploid variant calling

In what follows, we will use the following random variables for a particular locus:

• T, reference nucleotide

• S, read set, set of nucleotides mapped to the locus

• R, read nucleotide, a single nucleotide mapped to the locus

• H, haplotype, a single nucleotide variant

• G, genotype, the one (haploid) or two (diploid) haplotypes of a variant

• N, de novo, true iff a genotype is possible only as a result of a de novo mutation

What we want to do is to determine the probability of the genotype G given the set of nucleotides mapped to a locus and other ancillary information such as the quality scores and variant probabilities. That is, we want to compute P(G|S). Using Bayes formula, this can be computed as follows: \documentclass{aastex}\usepackage{amsbsy}\usepackage{amsfonts}\usepackage{amssymb}\usepackage{bm}\usepackage{mathrsfs}\usepackage{pifont}\usepackage{stmaryrd}\usepackage{textcomp}\usepackage{portland,xspace}\usepackage{amsmath,amsxtra}\pagestyle{empty}\DeclareMathSizes{10}{9}{7}{6} \begin{document} \begin{align*}P ( G \mid S ) = \frac { P ( S \mid G ) P ( G ) } { \sum_g P ( S \mid g ) P ( g ) } \end{align*} \end{document}

This requires that we compute the conditional probability P(G|S) and the prior P(G).

Each of the reads in S can be considered to be conditionally independent of each other. That is, if we know the value of the genotype G, then the errors in the different reads are independent of each other. This allows us to decompose the conditional probability distribution (CPD) P(s|G), \documentclass{aastex}\usepackage{amsbsy}\usepackage{amsfonts}\usepackage{amssymb}\usepackage{bm}\usepackage{mathrsfs}\usepackage{pifont}\usepackage{stmaryrd}\usepackage{textcomp}\usepackage{portland,xspace}\usepackage{amsmath,amsxtra}\pagestyle{empty}\DeclareMathSizes{10}{9}{7}{6} \begin{document} \begin{align*}P ( s \mid G ) = \prod_{r \in s} P ( r \mid G )\end{align*} \end{document}

That is, the conditional probability of the whole set of reads is obtained by taking the product of the conditional probabilities of the individual reads. This now requires computation of the probability of a particular haplotype (say the nucleotide T) given that we have seen a particular nucleotide in the mapped read (say C). If sequencing was error-free, then this probability would be zero, and the only possible explanation would be the nucleotide actually seen in the read. However, there is certainly a nonzero probability of an error. An estimate of this error rate is taken from the quality score supplied in the original sequencing files (e.g., FastQ files). If calibration information is available, then this probability will be adjusted by this calibration information (see above). Let the probability of an error be ε and consider the case where haploid genotypes are being called, then: \documentclass{aastex}\usepackage{amsbsy}\usepackage{amsfonts}\usepackage{amssymb}\usepackage{bm}\usepackage{mathrsfs}\usepackage{pifont}\usepackage{stmaryrd}\usepackage{textcomp}\usepackage{portland,xspace}\usepackage{amsmath,amsxtra}\pagestyle{empty}\DeclareMathSizes{10}{9}{7}{6} \begin{document} \begin{align*}P ( r \mid h ) = \begin{cases}1 - \epsilon \qquad r = h \\ \; \epsilon / 3 \qquad r \neq h\end{cases}\end{align*} \end{document}

The expression ε/3 comes from the fact that ε is the overall probability of an error and there are three nucleotides other than the correct one each of which is considered equally likely. Such haploid calling is used for the non-PAR regions of the sex chromosomes in males.

For diploid regions of the genome, the Bayesian calculation is only slightly more complex. Let the genotype g consist of two haplotypes \documentclass{aastex}\usepackage{amsbsy}\usepackage{amsfonts}\usepackage{amssymb}\usepackage{bm}\usepackage{mathrsfs}\usepackage{pifont}\usepackage{stmaryrd}\usepackage{textcomp}\usepackage{portland, xspace}\usepackage{amsmath, amsxtra}\pagestyle{empty}\DeclareMathSizes {10} {9} {7} {6} \begin{document} $$\langle h_1 , h_2 \rangle$$ \end{document} , then \documentclass{aastex}\usepackage{amsbsy}\usepackage{amsfonts}\usepackage{amssymb}\usepackage{bm}\usepackage{mathrsfs}\usepackage{pifont}\usepackage{stmaryrd}\usepackage{textcomp}\usepackage{portland,xspace}\usepackage{amsmath,amsxtra}\pagestyle{empty}\DeclareMathSizes{10}{9}{7}{6} \begin{document} \begin{align*}P ( r \mid \langle h_1 , h_2 \rangle ) = \frac { P ( r \mid h_1 ) + P ( r \mid h_2 ) } { 2 } \end{align*} \end{document}

That is the conditional probability is the average of the probabilities for the two constituent haplotypes of the genotype.

2.2.1. Priors

In addition to the conditional probabilities described above, we also use the priors for the different possible variants, P(G). This is done using a CPD of transition probabilities P(G|t), that is, P(G) = P(G|t). The table can be generated from prior data for humans (Levy et al., 2007) or from empirical data from previous calls in similar samples.

Consider the simple case when we are estimating prior probabilities for haploid genotypes. Let n_h be the total number of times allele h has been observed at this locus and n = ∑_h h be the total number of observations. Then P(G) is crudely estimated by \documentclass{aastex}\usepackage{amsbsy}\usepackage{amsfonts}\usepackage{amssymb}\usepackage{bm}\usepackage{mathrsfs}\usepackage{pifont}\usepackage{stmaryrd}\usepackage{textcomp}\usepackage{portland,xspace}\usepackage{amsmath,amsxtra}\pagestyle{empty}\DeclareMathSizes{10}{9}{7}{6} \begin{document} \begin{align*}P ( h ) = \frac { n_h } { n } \end{align*} \end{document}

However, if n_h = 0, then P(h) = 0, which is impermissible. A common way of dealing with this is to add a Laplace correction (Koller and Friedman, 2009): \documentclass{aastex}\usepackage{amsbsy}\usepackage{amsfonts}\usepackage{amssymb}\usepackage{bm}\usepackage{mathrsfs}\usepackage{pifont}\usepackage{stmaryrd}\usepackage{textcomp}\usepackage{portland,xspace}\usepackage{amsmath,amsxtra}\pagestyle{empty}\DeclareMathSizes{10}{9}{7}{6} \begin{document} \begin{align*}P ( h ) = \frac { n_h + p_n } { n + 1 } \end{align*} \end{document}

where p_h ≥ 0 and∑_h p_h = 1. In practice, this is done by using the default values from above to generate the p_h, that is, \documentclass{aastex}\usepackage{amsbsy}\usepackage{amsfonts}\usepackage{amssymb}\usepackage{bm}\usepackage{mathrsfs}\usepackage{pifont}\usepackage{stmaryrd}\usepackage{textcomp}\usepackage{portland,xspace}\usepackage{amsmath,amsxtra}\pagestyle{empty}\DeclareMathSizes{10}{9}{7}{6} \begin{document} \begin{align*}p_h = P ( h \mid t )\end{align*} \end{document}

For diploid genotypes, Hardy–Weinberg equilibrium is assumed and the diploid priors are computed from the frequencies of the haploid genotypes. The crude estimates are given by \documentclass{aastex}\usepackage{amsbsy}\usepackage{amsfonts}\usepackage{amssymb}\usepackage{bm}\usepackage{mathrsfs}\usepackage{pifont}\usepackage{stmaryrd}\usepackage{textcomp}\usepackage{portland,xspace}\usepackage{amsmath,amsxtra}\pagestyle{empty}\DeclareMathSizes{10}{9}{7}{6} \begin{document} \begin{align*}P ( \langle h_1 , h_2 \rangle ) = \alpha_ { h_1 , h_2 } \frac { n_ { h_ { 1 } } } { n } \frac { n_ { h_ { 2 } } } { n } \end{align*} \end{document}

where \documentclass{aastex}\usepackage{amsbsy}\usepackage{amsfonts}\usepackage{amssymb}\usepackage{bm}\usepackage{mathrsfs}\usepackage{pifont}\usepackage{stmaryrd}\usepackage{textcomp}\usepackage{portland,xspace}\usepackage{amsmath,amsxtra}\pagestyle{empty}\DeclareMathSizes{10}{9}{7}{6} \begin{document} \begin{align*}\alpha_{a , b} = \begin{cases}1 \qquad a = b \\ 2 \qquad a \neq b\end{cases}\end{align*} \end{document}

But as above, this has a problem when either \documentclass{aastex}\usepackage{amsbsy}\usepackage{amsfonts}\usepackage{amssymb}\usepackage{bm}\usepackage{mathrsfs}\usepackage{pifont}\usepackage{stmaryrd}\usepackage{textcomp}\usepackage{portland, xspace}\usepackage{amsmath, amsxtra}\pagestyle{empty}\DeclareMathSizes {10} {9} {7} {6} \begin{document} $$n_{h_{1}} = 0$$ \end{document} or \documentclass{aastex}\usepackage{amsbsy}\usepackage{amsfonts}\usepackage{amssymb}\usepackage{bm}\usepackage{mathrsfs}\usepackage{pifont}\usepackage{stmaryrd}\usepackage{textcomp}\usepackage{portland, xspace}\usepackage{amsmath, amsxtra}\pagestyle{empty}\DeclareMathSizes {10} {9} {7} {6} \begin{document} $$n_{h_{2}} = 0$$ \end{document} . An analogous Laplace correction is given by \documentclass{aastex}\usepackage{amsbsy}\usepackage{amsfonts}\usepackage{amssymb}\usepackage{bm}\usepackage{mathrsfs}\usepackage{pifont}\usepackage{stmaryrd}\usepackage{textcomp}\usepackage{portland,xspace}\usepackage{amsmath,amsxtra}\pagestyle{empty}\DeclareMathSizes{10}{9}{7}{6} \begin{document} \begin{align*}P ( \langle h_1 , h_2 \rangle ) = \frac { \alpha_ { h_1 , h_2 } ( n_ { h_ { 1 } } n_ { h_ { 2 } } + n_ { h_ { 1 } } p_ { h_ { 2 } } + p_ { h_ { 1 } } n_ { h_ { 2 } } ) + p_ { h_ { 1 } } , h_2 } { ( n + 1 ) ^2 } \end{align*} \end{document}

where p_h = P (h|t) and \documentclass{aastex}\usepackage{amsbsy}\usepackage{amsfonts}\usepackage{amssymb}\usepackage{bm}\usepackage{mathrsfs}\usepackage{pifont}\usepackage{stmaryrd}\usepackage{textcomp}\usepackage{portland, xspace}\usepackage{amsmath, amsxtra}\pagestyle{empty}\DeclareMathSizes {10} {9} {7} {6} \begin{document} $$p_{h_{1} , h_{2}} = P ( \langle h_1 , h_2 \rangle \mid t )$$ \end{document} .

2.2.2. Mapping errors

The mapper provides an estimate of the probability that a mapping is incorrect in the form of the MAPQ score for each read (Li et al., 2008). Let η be the probability that the read is mapped incorrectly as indicated by the MAPQ score, then we can compute a new CPD P′(r|h) in terms of the original P(r|h): \documentclass{aastex}\usepackage{amsbsy}\usepackage{amsfonts}\usepackage{amssymb}\usepackage{bm}\usepackage{mathrsfs}\usepackage{pifont}\usepackage{stmaryrd}\usepackage{textcomp}\usepackage{portland,xspace}\usepackage{amsmath,amsxtra}\pagestyle{empty}\DeclareMathSizes{10}{9}{7}{6} \begin{document} \begin{align*}{ \rm P}^{ \prime} ( r \mid h ) = ( 1 - \eta ) { \rm P} ( r \mid h ) + \eta { \rm P} ( r )\end{align*} \end{document}

where \documentclass{aastex}\usepackage{amsbsy}\usepackage{amsfonts}\usepackage{amssymb}\usepackage{bm}\usepackage{mathrsfs}\usepackage{pifont}\usepackage{stmaryrd}\usepackage{textcomp}\usepackage{portland, xspace}\usepackage{amsmath, amsxtra}\pagestyle{empty}\DeclareMathSizes {10} {9} {7} {6} \begin{document} $${ \rm P} ( r ) = \sum \nolimits_h { \rm P} ( r \mid h ) { \rm P} ( h )$$ \end{document}

2.3. Joint Bayesian calling in pedigrees

Given the genotypes G_f, G_m, and G₁ and the sets of mapped reads S_f, S_m, and S₁ for the father, mother, and one offspring, respectively, the joint distribution for the combination is \documentclass{aastex}\usepackage{amsbsy}\usepackage{amsfonts}\usepackage{amssymb}\usepackage{bm}\usepackage{mathrsfs}\usepackage{pifont}\usepackage{stmaryrd}\usepackage{textcomp}\usepackage{portland,xspace}\usepackage{amsmath,amsxtra}\pagestyle{empty}\DeclareMathSizes{10}{9}{7}{6} \begin{document} \begin{align*}P ( G_f ) P ( S_f \mid G_f ) P ( G_m ) P ( S_m \mid G_m ) P ( G_1 \mid G_f , G_m ) P ( S_1 \mid G_1 )\end{align*} \end{document}

Later it will be clearer if we write this expression for the joint distribution somewhat more abstractly as \documentclass{aastex}\usepackage{amsbsy}\usepackage{amsfonts}\usepackage{amssymb}\usepackage{bm}\usepackage{mathrsfs}\usepackage{pifont}\usepackage{stmaryrd}\usepackage{textcomp}\usepackage{portland,xspace}\usepackage{amsmath,amsxtra}\pagestyle{empty}\DeclareMathSizes{10}{9}{7}{6} \begin{document} \begin{align*}P ( \chi ) = P ( G_f ) & P ( S_f \mid G_f ) P ( G_m ) P ( S_m \mid G_m ) P ( G_1 \mid G_f , G_m ) P ( S_1 \mid G_1 ) \\ & = P ( { S \mid G} ) P ( { G} ) P ( G_1 \mid G_f , G_m ) \end{align*} \end{document}

P(χ) is shorthand for the joint distribution over all variables, and S and G indicate the random variables for all sets of reads and all genotypes, respectively (Fig. 1).

FIG. 1.

Bayesian network diagram for a parent–offspring trio.

Using Bayesian rules, it is possible to extract from this expression the posterior distributions that we are looking for—in this case, P(G_f| s ), P(G_m| s ), and P(G₁| s )are the posterior distributions, respectively, for the father's, mother's, and offspring's genotypes given all the mappings of reads.

All of the factors in the joint distribution except one are provided by the single-sample calling. P(G_f) and P(G_m) are the prior probabilities for the genotypes. These are computed from tables of known nucleotide frequencies, and if site-specific priors are supplied, from the allele counts there. The factors P (S_i|G_i) link the nucleotides seen in the mapped reads to the genotypes.

What is new here is the factor P (G₁|G_f, G_m), which is a CPD linking the genotypes of the parents to the offspring's. This factor takes into account Mendelian inheritance, mutation, and the type of chromosome, that is, whether it is an autosome or a sex chromosome. Supplementary Table S3 shows part of the CPD for SNPs in an autosome.

Similar tables are used for the sex chromosomes depending on the sex of the offspring. The three cases are as follows: Y chromosome in a son that is inherited only from the father; X chromosome in a son that is inherited only from one of the mother's two alleles; and the X chromosome of a daughter that is inherited from the father's only X chromosome and one of the mother's two X chromosomes. The details of these inheritance patterns are specified as part of the configuration of the reference sequence for a genome and can be configured for other patterns in nonmammalian species.

2.3.1. De novo mutation scoring

In the strictly Mendelian inheritance of Supplementary Table S3, an instance such as \documentclass{aastex}\usepackage{amsbsy}\usepackage{amsfonts}\usepackage{amssymb}\usepackage{bm}\usepackage{mathrsfs}\usepackage{pifont}\usepackage{stmaryrd}\usepackage{textcomp}\usepackage{portland, xspace}\usepackage{amsmath, amsxtra}\pagestyle{empty}\DeclareMathSizes {10} {9} {7} {6} \begin{document} $$P ( \langle { \rm A , T} \rangle \mid \langle { \rm A , C} \rangle , \langle { \rm A , C} \rangle )$$ \end{document} will have a probability of 0. However, allowance is made for a de novo mutation between the parents and the offspring with a probability of 10⁻⁹ (Conrad et al., 2011). In order to ensure that the sums of the probabilities in the CPD are still 1, the other cases all have a slightly reduced probability. This small probability means that significant evidence is needed from the mapped reads before such a de novo call will be made.

Whenever a call is made that can only occur as a result of a de novo mutation, a special DNP score is calculated. This is the posterior score P (N₁| S ), where N₁ is a true false value, indicating that the offspring's genotype can only occur as a result of a de novo mutation and S is all the variables for the parents' and offspring's sets of reads. As usual in VCF files, it is output as a Phred score (Danecek et al., 2011). The value of N₁ is computed deterministically from the genotypes G , so the deterministic CPD P (N_i| G ) can be included in the joint distribution without disturbing the other variables \documentclass{aastex}\usepackage{amsbsy}\usepackage{amsfonts}\usepackage{amssymb}\usepackage{bm}\usepackage{mathrsfs}\usepackage{pifont}\usepackage{stmaryrd}\usepackage{textcomp}\usepackage{portland,xspace}\usepackage{amsmath,amsxtra}\pagestyle{empty}\DeclareMathSizes{10}{9}{7}{6} \begin{document} \begin{align*}P ( \chi ) = P ( N_1 \mid { G} ) P ( { S \mid G} ) P ( { G} ) P ( G_1 \mid G_f , G_m )\end{align*} \end{document}

P (N₁| S ) is then computed in the normal way by summing over all the values of G and applying Bayes theorem.

The Bayesian analysis above can be extended easily to families with more offspring. The joint distribution then becomes \documentclass{aastex}\usepackage{amsbsy}\usepackage{amsfonts}\usepackage{amssymb}\usepackage{bm}\usepackage{mathrsfs}\usepackage{pifont}\usepackage{stmaryrd}\usepackage{textcomp}\usepackage{portland,xspace}\usepackage{amsmath,amsxtra}\pagestyle{empty}\DeclareMathSizes{10}{9}{7}{6} \begin{document} \begin{align*}P ( G_f ) P ( S_f \mid G_f ) P ( G_m ) P ( S_m \mid G_m ) \prod_i P ( G_i \mid G_f , G_m ) P ( S_i \mid G_i )\end{align*} \end{document}

That is, the factors for all the offspring are multiplied in. In strictly theoretical terms, there is no problem with this formulation. Nevertheless, if the calculations are implemented naïvely, the execution time will increase exponentially with the number of offspring. However, the offspring are conditionally independent of each other given the parents' genotypes, which permits the calculations to be done in time 0 (n³), where n is the number of potential genotypes for a single sample.

2.3.2. Multigenerational pedigrees

Pedigrees can be much more complex than just a simple family with two parents and some number of offspring. Supplementary Figure S1 shows an extended pedigree and Figure 2 the corresponding Bayesian network.

FIG. 2.

Bayesian network diagram for an extended pedigree.

To express the joint distribution for such a pedigree, it is necessary to introduce some notation. Let i^↑ be the (unique) father of sample i and i^† be the (unique) mother of sample i. i is a root if it has no father or mother. It is assumed that if one parent is present, then the other is also. This can be achieved by adding a sample that contains no reads (S_i = ∅) if sequencing data from such a sample is absent.

The factor for a root node contains the prior for the genotype and the CPD for the reads given the genotype, P (s_i|G_i) P (G_i). The factor for a nonroot node contains the CPD for the genotype given the genotype of the parents and the CPD for the reads given the genotype, P (G_i|G_i^↑, G_i^†) P (S_i|G_i). All of these factors have been met already in the nuclear family case.

When all these factors are combined, the joint distribution is \documentclass{aastex}\usepackage{amsbsy}\usepackage{amsfonts}\usepackage{amssymb}\usepackage{bm}\usepackage{mathrsfs}\usepackage{pifont}\usepackage{stmaryrd}\usepackage{textcomp}\usepackage{portland,xspace}\usepackage{amsmath,amsxtra}\pagestyle{empty}\DeclareMathSizes{10}{9}{7}{6} \begin{document} \begin{align*}\prod_{i \ \hbox{is root}} P ( s_i \mid G_i ) P ( G_i ) \times \prod_{i \ \hbox{not root}} P ( G_i \mid G_{i^{ \uparrow}} , G_{i^{\dag}} ) P ( s_i \mid G_i )\end{align*} \end{document}

As before, there is a danger that a naïve calculation of the posterior distributions P (G_i| s ) will explode exponentially with the size of the pedigree. If the pedigree contains no inbreeding, then standard techniques for variable elimination in Bayesian networks are guaranteed to converge to the required posterior. For example, using the message passing algorithm of Shafer and Shenoy (Shafer and Shenoy, 1990), the computation can be arranged as two passes. First, a backward pass proceeds from the leaves toward the roots. It incrementally computes a partial joint distribution that includes all the samples that are descended from each node (this calculation for a parent requires only the result of the backward pass for each of the offspring). Second, a forward pass from the root nodes to the leaves calculates a partial joint distribution that includes each sample and all samples that are not descended from it (this calculation requires the result of the forward pass for each offspring as well as the backward pass for each of its siblings and half-siblings). Finally, the results of the forward and backward pass for each sample i are combined to compute P (G_i| s ). Practical techniques for implementations are discussed by Koller and Friedman (2009) and Pearl (1988).

If there is some inbreeding, then this algorithm breaks down (there will come a point where it is not possible to proceed with the forward pass of the algorithm). This can be solved at the cost of further computation by iterating the backward and forward passes. Whenever a result is required in a pass that has not already been computed, then an old value from a previous pass (or a default prior) is used. The iterations continue until there are small changes in the calls and their scores.

2.4. Complex variant calling

Any region that included indel alignments, complex variants such as MNPs, or clusters of SNPs in close proximity suggesting spurious alignments was handled by a novel haplotype-aware “complex region” calling algorithm. This aims to achieve one or two consistent haplotypes that explain all the reads that overlap the complex region. The caller proceeds as follows: (1) determining (small) complex regions that are candidates for the haplotype caller; (2) complex regions are extended by merging adjacent complex regions and interesting loci; (3) extracting hypotheses from the alignments for potential haplotypes; (4) Bayesian calling is performed similarly to simple variants, except that P (r|h) is calculated by a dynamic programming alignment technique that realigns each read against all the hypotheses; (5) presenting the resulting calls as simple SNPs wherever possible. When calling multiple samples together, all samples can contribute potential haplotypes to the calling, but their number is pruned to include at most six alternative haplotypes ranked on how often they occur. Usually, 10–20% of whole-genome sequencing (WGS) calls are processed by the complex caller and they are labeled as such in the output VCF. For more details, see Supplementary Methods, Supplementary Table 1, and Supplementary Figures 2 and 3.

2.5. Variant scores

The variant caller outputs its results in a VCF file. The GQ and QUAL fields are provided using the Bayesian posterior probabilities for the genotype and following the standard definition for VCF files (Danecek et al., 2011). In addition, we rescored variant quality values to deal with technology and mapping artifacts that are not captured in the Bayesian model. Our adaptive variant rescoring (AVR) model is based on Breiman's random forest ensemble learning technique (Breiman, 2001). We use attributes derived, or directly extracted, from the info and format fields of the VCF (Supplementary Table S2). We trained the model with data from the 1000 Genomes and the HapMap projects (DePristo et al., 2011). The final AVR score is the average of the scores returned by the individual trees. For more details, see Supplementary Methods.

2.6. Benchmarking

2.6.1. Simulations

To evaluate our method, we simulated a WGS dataset for a parent–offspring trio by constructing the four haplotypes of the founders from the human reference sequence (hg19) and applying polymorphism using as a template the data from the 1000 Genomes Project to a rate of 10⁻⁶. We then recombined the parents' haplotypes with a crossover rate of 1 in 100 Mb, ensuring that at least one crossover per chromosome arm existed. From this template, we simulated 2×100 bp paired-end reads with an average insert size of 350 bp and a sequencing error rate similar to that of the Illumina platform with a higher probability of error at the 3′-end of the reads and an average error rate of 1%. Reads were simulated randomly across the genome to an average depth of 5×.

2.6.2. Gold-standard data

In addition to simulations, we analyzed a set of gold-standard WGS data from a three-generation CEPH/Utah family of 17 members (Supplementary Fig. S4) produced by Illumina Inc. (San Diego, CA) as part of their Platinum Genomes resource. DNA from lymphoblastoid cell lines obtained from the Coriell Institute were sequenced with the HiSeq 2000 system to 50× average depth using 2×100 bp libraries of ∼350 bp insert size. Raw data were downloaded from the European Nucleotide Archive (Acc. Nos. ERA172924 and ERA185981). We aligned reads and performed calls in three nuclear family subsets and the entire pedigree for comparison (see Supplementary Fig. S4). See Supplementary Materials for source of orthogonal data used for comparisons. To compare ground truth and called VCFs, we utilized vcfeval, a command of rtgTools that includes a novel algorithm for comparing variant call sets that deals with indel/MNP representation discrepancies, minimizing false positives and negatives through a dynamic programming approach, for accurate performance evaluation (Cleary et al., submitted). rtgTools is freely available online from the Genome in a Bottle consortium repository.

2.6.3. HTS data analysis

We analyzed simulated or real HTS data with a multithreaded Java implementation of our methods that is part of the commercially available rtgVariant v1.2 software (Real Time Genomics, Inc., San Bruno, CA). Software was run on commodity dual-quad Intel servers running the Linux operating system (Amazon EC2, m3.2 × large instances running CENTOS, Intel Xeon E5-2670 eight core, 34 GB RAM, launched by StarCluster). Reads were mapped in groups of up to 50 million reads jobs using 8 cores to the human genome reference hg19 with decoys, as described by the 1000 Genomes Project (1000 Genomes Project Consortium, 2012). BAM files were merged as needed and passed to the variant caller providing, when necessary, a PED file describing the pedigree relationships and sex of the sample. Variant calling was parallelized by splitting jobs by chromosomes and each job was multithreaded in eight cores. As described in the results, samples were either called independently or in groups. The output VCFs were analyzed either in its entirety, or after filtering with an AVR score of ≥0.15 for whole-genome data. Compressed VCF files with the different calls sets produced for this study are available at the NCBI trace archive online.

For comparison purposes we also aligned the gold-standard data with BWA v0.6.1 and called variants with two other methods: GATK Unified Genotyper v1.7 (DePristo et al., 2011), and Samtools-hybrid v0.1.7 (Li et al., 2009). These tools were used with their default parameters, including filtering down to the first tranche of VQSR for GATK. In addition, to compare with another approach for family-aware variant calling, we reanalyzed the VCFs from Samtools-hybrid with PolyMutt v0.16 (Li et al., 2012).

3. Results

3.1. Simulations

For an initial evaluation of our method, we simulated HTS paired-end reads of 100 bp length with an average insert size of 350 bp, to an average depth of 5× for a parent–offspring trio. Figure 3 shows a receiver operator characteristic (ROC) curve showing a comparison between calling the offspring independently (singleton), jointly with her parents using the population calling without considering the family relationships, and jointly but using the family calling, sorted by GQ score. These results demonstrate that the family caller performs much better that the other two cases having a greater area under the curve (AUC) and being able to retrieve more true positives at the expense of fewer false positives at any given GQ score threshold. This improvement is particularly notable at reduced depth of coverage versus high-depth coverage (see below; data not shown).

FIG. 3.

Receiver operator characteristic (ROC) curve of call sets from a simulated parent–offspring trio. Data are for the offspring called as either a single sample (offspring; red line), with its parents using the population caller (green line), or jointly as a trio with the family caller (blue line).

3.2. Analysis of gold-standard samples

To better assess the performance of our method with real HTS data and its associated artifacts, we analyzed the data from a set of gold-standard samples from a widely used CEPH/Utah pedigree sequenced with 2×100 bp PE Illumina reads to about 50× (see Methods). We focus our analysis on NA12878, a female in the second generation, for which extensive orthogonal validation data exist, including fosmid-end Sanger sequence data (Kidd et al., 2008), Complete Genomics WGS data, Illumina OMNI SNP-array genotype data (1000 Genomes Project Consortium, 2012), and experimentally validated germline and cell-line somatic de novo mutation data (Conrad et al., 2011; Ramu et al., 2013).

Table 1 shows that compared with singleton calling, the family caller produced more SNV/indel/MNP calls while maintaining good quality, as judged by commonly used quality metrics such as Ti/Tv, Het/Hom ratios, and dbSNP concordance. Note that the data were filtered using an AVR ≤0.15 cutoff for our caller, whereas for GATK we used the first tranche of VQSR, as recommended (DePristo et al., 2011).

Table 1.

Statistics and Quality Metrics for Different Call Sets for NA12878

Call set	SNVs	Indels	MNPs	Het/Hom	Ti/Tv	% dbSNP (r129)
RTG single	3,329,797	558,242	31,070	1.55	2.11	90.80
RTG trio	3,363,619	595,030	33,686	1.57	2.11	90.40
GATK/VQSR	3,263,289	610,837	N/A	1.51	2.09	91.70

SNV, single-nucleotide variants; MNP, multiple-nucleotide variants; Het, heterozygote; Hom, homozygote; Ti, transitions; Tv, transversions.

We then compared the calls sets with reference datasets obtained by orthogonal methods for the sample NA12878, including SNP OMNI array data generated by the 1000 Genomes Project (1000 Genomes Project OMNI Poly, for polymorphic sites), the Get-RM high-confidence calls (Ball et al., 2012), and the NIST genome-in-a-bottle arbitration calls (Zook et al., 2014) (Table 2). Although all call sets show good concordance with the orthogonal data, the family caller showed slightly better agreement with the arbitration data sets of the NIST, in particular in the high-confidence regions (GiaB-BED). As a proxy for false positives, we report the variants discovered across ∼50,000 sites in the OMNI SNP array that appear monomorphic in the 1000 Genomes Project samples (DePristo et al., 2011; 1000 Genomes Project Consortium, 2012). All numbers are very low, and although the family caller appears to have higher call rate in these sites, manual inspection of a random set of alignments suggests that many of these calls are real, but either small indels or MNPs, which cannot be typed in the microarray platform.

Table 2.

Concordance with Orthogonal Datasets for NA12878 Call Sets

Call set	1000 Genomes Project OMNI Poly (TP %)	1000 Genomes Project OMNI Mono (FP %)	Get-RM (TP %)	GiaB (TP %)	GiaB-BED ^a (TP %)
RTG single	97.50	0.10	97.40	89.80	93.40
RTG trio	97.50	0.24	97.00	90.50	94.10
GATK/VQSR	97.80	0.17	87.80	88.40	92.50

High-confidence Browsen Extensible Data (BED) region. TP, true positives; FP, false positives.

While these comparisons suggest that joint trio calling performs better, they are filtered at the default quality score threshold and do not reveal the performance across the entire scoring range. We thus constructed ROC curves using as a baseline (or ground truth) the arbitration dataset developed by the NIST (Fig. 4A) (Zook et al., 2014), the independently obtained variants calls produced by Complete Genomics (Fig. 4B) (Drmanac et al., 2009), or a pedigree-based gold standard derived from phasing the large family of the CEPH pedigree and testing whether variant calls segregate in a consistent fashion with the haplotype phases (Segregation Phasing or SP standard; Fig. 5) (De La Vega et al., in preparation). As can be seen in Figures 4 and 5, the family caller outperforms the single-sample calling (greater AUC), allowing making a better trade-off between true positives and false positives at realistic thresholds. As compared with GATK-UG, our family caller in general shows greater AUC in the range up to the end of the VQSR first tranche, but there is variation depending on the baseline used. Fewer differences are seen with the NIST GiaB data set; this could be the result of biases to GATK calls and/or Illumina platform data during the construction of this version of the GiaB baseline, whereas the CGI data are completely independent. The best results are observed with the SP gold standard (De La Vega et al., in preparation), were the trio calls show better AUC than the single-sample calls, and significantly better than GATK/VQSR (note that the ROC curve for the latter is truncated at the end of the first tranche of VQSR, as recommended) (DePristo et al., 2011). Notably, when the trio is called in the context of the data of remaining 14 members of the pedigree, even when their relationships are not used, the AUC improves further (trio-cohort line, Fig. 5).

FIG. 4.

ROC curve of different calls sets for NA12878 vs. the CGI calls (A) and NIST GiaB calls (B). Calls are for the following methods: single sample (offspring; green); trio (blue); and GATK/VQSR single sample (offspring red). rtgVariant calls are sorted by AVR score (see Methods); GATK are sorted by VQSLOD score.

FIG. 5.

ROC curve of different calls sets for NA12878 vs. the segregation phasing (SP) standard as baseline. Calls are for the following methods: single sample (offspring; green); trio (blue); trio concurrently with the remaining samples of the pedigree samples but without considering their relationships; and GATK/VQSR single sample (offspring red). rtgVariant calls are sorted by AVR score (see Methods); GATK are sorted by VQSLOD score.

3.3. Mendelian errors

We enumerated Mendelian inheritance errors (MIEs) across the trio samples. Table 3 shows that the family caller reduces the number of MIEs down to just 0.09%, dramatically smaller than the single-sample approaches. It should be noted that this is accomplished without resorting to filtering and indeed the family caller returns more calls than single-sample calling (cf. Table 1). We compared the results of our method with those from PolyMutt (Li et al., 2012), a likelihood-based method aimed at improving variant calls in a trio-aware fashion and identify de novo mutations. PolyMutt requires as input a call set from another variant calling, so in this case we used the output from SamTools-hybrid (Li et al., 2009). As can be seen in Table 3, PolyMutt reduces the number of MIEs to 0.32%, still 3.5-fold more than our method.

Table 3.

Mendelian Errors Across Trio Samples for Different Callers

	Mendelian inconsistent calls
Calling method	n	%
Singleton caller	335,626	5.62
Family caller	4,870	0.09
GATK/VQSR	351,904	9.2
PolyMutt	9,283	0.32

3.4. Identification of de novo mutations

Without joint calling, there are tens of thousands of offspring heterozygote calls whose variant alleles are not called in either parent, thus appearing as candidate de novo mutations (Table 4). However, we know from prior studies that the number of true de novo mutations that appear in the germline should be closer to 100 (Conrad et al., 2011; Li et al., 2012). Even considering that this data is obtained from cell lines where somatic de novo mutations exist (about 1000) (Conrad et al., 2011), the de novo candidates in single-sample call sets are hugely contaminated with false positives. However, the family caller identifies just under 3,000 candidates—a reduction of 60×.

Table 4.

De Novo Mutation Identification in NA12878

	De novo candidates		Validated de novo concordance
Call set	All	Filtered ^a	Germline ^b	Somatic ^b	Indel ^c
Singleton	16,902	N/A	48 (100%)	941 (99%)	46 (84%)
Trio	3,857	2,208	48 (100%)	941 (95.6%)	45 (81.8%)
GATK/VQSR	N/A	74,727	47 (98%)	859 (96.7%)	48 (87%)
PolyMutt	6,851	N/A	45 (93.8%)	844 (95%)	0 (0%)

DNP ≥50.

Conrad et al. (2011).

Ramu et al. (2013); autosomes only.

The 1000 Genomes Project experimentally validated a subset of the de novo germline and somatic mutations in the NA12878 trio that we can use as a reference (Conrad et al., 2011; Ball et al., 2012). As compared with the validated de novo mutations set, we observed 100% and 99% sensitivity in detecting the reported germline and somatic de novo single-nucleotide mutations, respectively (note that the cell line batch may be different and may have slightly different somatic mutations). Through the analysis of variant segregation to the third generation, we confirmed 99% of the Conrad et al.'s (Conrad et al., 2011; Zook et al., 2014) germline mutations (somatic variants do not segregate, as expected). Importantly, this high de novo sensitivity was achieved while reducing the number of candidate de novo mutations by greater than sixfold without using ad hoc filters. We derived a specific quality score for de novo mutations (DNP, see Methods). Filtering with this score (e.g., DNP ≥50), the number of candidate mutations can be reduced further (Supplementary Fig. S5).

We also analyzed the de novo mutations identified by PolyMutt (Li et al., 2012) in the same data. As shown in Table 4, PolyMutt identifies twice the number of candidate de novo mutations (i.e., half the specificity) and is less sensitive in identifying the 1000 Genomes Project–validated de novos, as compared with our joint trio calling method. In addition, PolyMutt identifies only single-nucleotide de novo mutations.

3.5. Multigenerational pedigrees

Expanding the analysis to multigenerational pedigrees requires new methods that avoid combinatorial explosion. We implemented a belief propagation algorithm in our Bayesian network model that allows conveying priors beyond the nuclear family. To evaluate the advantage of this method, we simulated data for a multigenerational pedigree at different depths of coverage. Supplementary Figure S6 shows the pedigree structure we used to simulate data. On a first instance, all individuals were simulated at a depth of 5×, and in a second example, we simulated the grandparents to 30× and the rest to 5×. Supplementary Figure S7 shows that the AUC of the ROC for either the mother or the first father (F1) increases in the call set from the family caller as compared with single-sample calls (simulated to a depth of 5×). To better understand the impact of prior propagation from one generation to another, we simulated 30× coverage for the grandparents and 5× for the rest of the pedigree. In Supplementary Figure S8, we observe that while the benefit of family calling for the grandparents is small (they have already high coverage), there is a substantial benefit for the other samples, such as F1 and even M. The impact in distal samples is smaller (data not shown).

We also tested our method in the complete CEPH/Utah pedigree data. Supplementary Table S4 shows that calls have excellent quality metrics, similar to the results for nuclear families, while maintaining an MIE rate of 3.6% across the entire pedigree.

4. Discussion

Whole-exome and genome sequencing has been very successful in elucidating causal genes for highly penetrant Mendelian diseases and is now being used in diagnostic. Recent reports suggest that de novo mutations could account for over 50% of early neurodevelopmental diseases in outbred populations (Veltman and Brunner, 2012). The identification of such de novo mutations greatly benefits from sequencing and concurrent analysis of the parents. In spite of this, variant identification and genotyping methods from HTS data often ignore the family relationships between the samples and the fact that the same haplotypes are being redundantly sequenced. This is an important piece of information that can be leveraged to reduce the inevitable errors that occur in variant calling because of the nature of HTS. There have been some attempts to perform “family-aware” variant rescoring for improving the calls sets or the identification of de novo mutations (Cartwright et al., 2012; Li et al., 2012; Peng et al., 2013; Ramu et al., 2013). However, these methods are post hoc, not principled, and slow. In contrast, our method utilizes an extensible Bayesian network framework that permits fast joint family calling leveraging information across related samples using Mendelian inheritance priors, scores de novo mutations, and is scalable to multigenerational pedigrees. Our results with data from gold-standard samples demonstrate that family calling outperforms single-sample approaches, producing high-quality calls and reducing significantly MIEs (Tables 1 and 2). This is accomplished without filtering, but instead correcting the undercalled heterozygotes in parents using otherwise weak evidence. In fact, the family caller returns additional variants at any given quality threshold, that with the additional information are now scored higher (Table 1; Figs. 4 and 5).

We observed a high concordance with orthogonal reference data for sample NA12878, suggesting excellent sensitivity and specificity. Nevertheless, these reference data are not infallible and have pitfalls. For example, invariant sites from the 1000 Genomes Project OMNI array have been widely used as a proxy for false positives. We found a slight increase in OMNI monomorphic sites in the trio calls. When we analyzed alignments by hand for such variants, we found reasonable support for about half of these, which frequently are MNPs or small indel calls, and are unlikely to be typed correctly in the SNP array.

The ROC curve analysis presents a more complete view of the performance of the variant calling methods and the recalibration scores ability to discriminate between true and false positives. Simulation data perform as expected, trio calling significantly improving over simple population calling and singe sample, in particular in low-depth situations. ROC curves of calls from high-depth WGS data of NA12878 versus baselines from orthogonal (CGI) or arbitration reference sets (NIST) show a greater AUC when the sample is called as a trio with her parents. While the former baseline call sets could have inherent problems, a baseline derived from variants that inherit in the proper haplotype context in the 11 offspring of NA12878 provided a cleaner set of variants to be used as ground truth (De La Vega et al., in preparation). The ROC curves versus the SP standard show a clear improvement of our Bayesian caller over GATK, and greater AUC of trio calls, in particular when called in the context of the complete pedigree data. This improvement comes partly in the form of additional complex calls (indels and MNPs). We explain this as the result of the complex caller using all alignments across the samples to prune the most unlikely haplotype hypothesis, and highlight the benefits of calling groups of samples together whether they are related or not.

We could have completely eliminated MIEs; however, we allow for a small probability of de novo mutations (a prior of 10⁻⁹, consistent with estimated mutation rates per generation) (Conrad et al., 2011). This allows capturing true de novo variants in offspring while reducing significantly the background of false-positive de novos (Table 4). While high sensitivity can be achieved by simply reporting variants that pass less stringent filtering thresholds (and in so doing increasing substantially the MIEs), our family calling achieves high sensitivity, concurrently with a 10× reduction in MIEs (Table 4). Further reduction of false positives can be achieved by filtering with the DNP score, improving precision even further (Supplementary Fig. S5). We evaluated the detected de novos comparing to validated mutations from the 1000 Genomes Project (Conrad et al., 2011). Comparatively, our method outperforms PolyMutt, another method we tested that aims to identify de novo mutations, as reflected in our higher true-positive and lower false-positive rates. Further, our method detect de novo MNPs and indels (cf Table 4), and is much faster (data not shown).

A possible pitfall of our method is that in regions of copy number variation (CNV) simple Mendelian inheritance is not expected. Nevertheless, our Bayesian network model includes a factor for copy number of the template. This is normally set to diploid, but is changed to accommodate haploid calls in sex chromosomes in males (except for the PAR region in X), and can certainly be changed for regions of CNV if those regions are known a priori. In addition, one may want to infer the copy number from the data and indeed our model allows for estimating the CPD of the copy number if desired. Such implementation is aimed for future developments.

Another area of further improvements is in the identification of indels and MNPs (complex variants). Our analysis with the SP standard showed that indels greater than 20 bp, in particular insertions, and MNP exhibit higher false positives than small indels and SNVs (1–4%; data not shown). Currently, our default banding allows initial alignments of up to ∼25 bp, although the complex caller could extend these further during variant calling (see Supplementary Figure S9 for a histogram of the length distribution of detected indels). If additional indel sensitivity is desired, the banding can be increased at the expense of slower alignment time. Alternatively, additional hypotheses for complex variant calling can be obtained by other means, such as local assembly of reads, either those aligned directly during mapping or those placed in the approximate location because of the mapped position of their mate.

Our methods are very efficient and fast (Supplementary Fig. S10); coupled with our hash-based fast mapping and hierarchical alignment method, whole genome sequence data of a parent–offspring trio sequenced to 50× depth (2×100 bp) can collectively be analyzed from reads to variant calls in ∼30 hours on a single-commodity server and is amenable to large-scale parallelization for further speed improvements. In summary, our results show that joint pedigree calling outperforms singleton and population variant calling in pedigrees, allows for the identification of de novo mutations with greater specificity, and is scalable to large genomics and human disease studies. We believe that analytical advances such as this are crucial for the adoption of genomes and exomes in large family disease studies and clinical settings.

Footnotes

Acknowledgments

We would like to thank Chad Harland (Livestock Improvement Corp.), George Weinstock (Washington University), and Carlos Bustamante (Stanford University) for fruitful discussions. We are grateful to Michael Eberle (Illumina, Inc.) and Steve Lincoln (formerly Complete Genomics, Inc.) for releasing publicly the CEPH/Utah pedigree WGS data we utilized and for providing related information. We also thank the 1000 Genomes Project Consortium and Justin Zook (NIST) for the reference data used in this work. We dedicate this work to the memory of John G. Cleary.

Authors' Contributions

J.G.C., S.I., S.A.I., L.T., and F.M.V. conceived methods. J.G.C., R.B., K.G., S.I., S.A.I., A.J., R.L., M.R., D.W., and L.T. developed and implemented algorithms. F.M.V. and J.G.C. conceived the validation study and analysis strategy. S.N.-M., R.B., B.S.H., M.R., R.L., and L.T. performed data analysis. S.N.-M. and R.L. performed production runs. J.G.C. and F.M.V. wrote the article.

Author Disclosure Statement

The authors declare that during the time this work was performed they were employed by Real Time Genomics, Inc., or its subsidiaries, and received salaries and stock options. The work presented here is covered by U.S. Patent 7,640,256 and additional pending patents.

References

1000 Genomes Project Consortium. 2010. A map of human genome variation from population-scale sequencing. Nature, 467, 1061–1073.

1000 Genomes Project Consortium. 2012. An integrated map of genetic variation from 1,092 human genomes. Nature, 491, 56–65.

Ajay

S.S.

, Parker

S.C.J.

, Ozel Abaan

, et al. 2011. Accurate and comprehensive sequencing of personal genomes. Genome Res. 21, 1498–1505.

Ball

M.P.

, Thakuria

J.V.

, Zaranek

A.W.

, et al. 2012. A public resource facilitating clinical use of genomes. Proc. Natl. Acad. Sci. USA, 109, 11920–11927.

Breiman

2001. Random forests. Mach. Learn., 45, 5–32.

Cartwright

R.A.

, Hussin

, Keebler

J.E.M.

, et al. 2012. A family-based probabilistic method for capturing de novo mutations from high-throughput short-read sequencing data. Stat. Appl. Genet. Mol. Biol., 11.

Conrad

D.F.

, Keebler

J.E.M.

, DePristo

M.A.

, et al. 2011. Variation in genome-wide mutation rates within and between human families. Nat. Genet., 43, 712–714.

Danecek

, Auton

, Abecasis

, et al. 2011. The variant call format and VCFtools. Bioinformatics, 27, 2156–2158.

DePristo

M.A.

, Banks

, Poplin

, et al. 2011. A framework for variation discovery and genotyping using next-generation DNA sequencing data. Nat. Genet., 43, 491–498.

10.

Drmanac

, Sparks

A.B.

, Callow

M.J.

, et al. 2009. Human genome sequencing using unchained base reads on self-assembling DNA nanoarrays. Science, 327, 78–81.

11.

Garrison

, and Marth

2012. Haplotype-based variant detection from short-read sequencing. arXiv 1207.3907v2.

12.

Gilissen

, Hoischen

, Brunner

H.G.

, et al. 2012. Disease gene identification strategies for exome sequencing. Eur. J. Hum. Genet., 20, 490–497.

13.

Gotoh

1982. An improved algorithm for matching biological sequences. J. Mol. Biol., 162, 705–708.

14.

Kidd

J.M.

, Cooper

G.M.

, Donahue

W.F.

, et al. 2008. Mapping and sequencing of structural variation from eight human genomes. Nature, 453, 56–64.

15.

Koller

, and Friedman

2009. Probabilistic Graphical Models. MIT Press, Cambridge, MA.

16.

Levy

, Sutton

, Ng

P.C.

, et al. 2007. The diploid genome sequence of an individual human. PLoS Biol., 5, e254.

17.

, Chen

, Zhan

, et al. 2012. A likelihood-based framework for variant calling and de novo mutation detection in families. PLoS Genet., 8, e1002944.

18.

, Handsaker

, Wysoker

, et al. 2009. The sequence alignment/map format and SAMtools. Bioinformatics, 25, 2078–2079.

19.

, Ruan

, and Durbin

2008. Mapping short DNA sequencing reads and calling variants using mapping quality scores. Genome Res. 18, 1851–1858.

20.

Marth

G.T.

, Korf

, Yandell

M.D.

, et al. 1999. A general approach to single-nucleotide polymorphism discovery. Nat. Genet. 23, 452–456.

21.

McKernan

K.J.

, Peckham

H.E.

, Costa

G.L.

, et al. 2009. Sequence and structural variation in a human genome uncovered by short-read, massively parallel ligation sequencing using two-base encoding. Genome Res. 19, 1527–1541.

22.

Pearl

1988. Probabilistic Reasoning in Intelligent Systems. Morgan Kaufmann, Burlington, MA.

23.

Peng

, Fan

, Palculict

T.B.

, et al. 2013. Rare variant detection using family-based sequencing analysis. Proc. Natl. Acad. Sci. USA, 110, 3985–3990.

24.

Ramu

, Noordam

M.J.

, Schwartz

R.S.

, et al. 2013. denovoGear:. Nat. Meth., 10, 985–987.

25.

Saunders

C.J.

, Miller

N.A.

, Soden

S.E.

, et al. 2012. Rapid whole-genome sequencing for genetic disease diagnosis in neonatal intensive care units. Sci. Transl. Med., 4, 154ra135.

26.

Shafer

G.R.

, and Shenoy

P.P.

1990. Probability propagation. Ann. Math. Artif. Intell. 2, 327–351.

27.

Talkowski

M.E.

, Ordulu

, Pillalamarri

, et al. 2012. Clinical diagnosis by whole-genome sequencing of a prenatal sample. N. Engl. J. Med. 367, 2226–2232.

28.

Veltman

J.A.

, and Brunner

H.G.

2012. De novo mutations in human genetic disease. Nat. Rev. Genet. 13, 565–575.

29.

Zook

J.M.

, Chapman

, Wang

, et al. 2014. Integrating sequencing datasets to form highly confident SNP and indel genotype calls for a whole human genome. Nat. Biotech. 32, 246–251.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

2.74 MB