To analyze large pedigrees rapidly and to compare the linkage analysis results of different software packages, we developed a software tool called CompareLinkage to automate the following processes: (1) Converting of Affymetrix Mapping 10 K genotype data, pedigree files and marker information into the "Linkage" format 13, and detecting and fixing incompatibilities in pedigree genotypes. The input genotype text file for CompareLinkage can be a single text file containing genotypes for each sample or a combined text file as exported by the Affymetrix GDAS 3.0 software. (2) Automatically calling the software packages Merlin and Allegro for linkage analysis and converting the analysis results (LOD or non-parametric linkage (NPL) scores) into the input files for dChip to visualize the results in the context of genes and cytobands. (3) The SNP genotype data in the "Linkage" format can be converted into the dChip input files (genotype, pedigree and marker information files) to perform parametric linkage analysis by dChipLinkage. All steps are discussed in detail at the CompareLinkage software manual provided on the software website. All these functionalities are useful for cross-validation of linkage results and to identify concordance and discordances between different linkage analysis programs as well as between parametric and non-parametric linkage results.

A graphical user interface (GUI) for Windows was also implemented in Java. In this GUI users are allowed to set their own working directory and the location of the Perl interpreter through the "Setting" menu. CompareLinkage's functions of converting file formats and getting dChip input files are incorporated through the "Convert" and "GetCurve" menu (Figure 1). Since computing is usually time-consuming, the code of calling the Perl program is executed in separate thread to provide better interaction. The output of the Perl program can be viewed in the message window (Figure 2).

The Affymetrix Mapping 10 K array CEL files and genotype TXT files can be imported into dChip and visualized along cytobands and genes as previously reported 911. The information of the SNPs such as their genetic and physical distance and allele frequencies from three ethnic groups (Asian, African American and Caucasian) is obtained from the Affymetrix website 14 and converted into the genome information files for dChip. The information of the reference genes and cytobands is obtained from the UCSC genome bioinformatics database 15 for the matching human genome assembly (hg12 or hg15) of the SNP information, and is organized into the refGene and cytoband file provided with dChip.

We implemented a variant of the Lander-Green 5616 algorithm in the dChipLinkage module of dChip to perform multipoint parametric linkage analysis and compute a LOD score at each SNP position. Disease allele frequencies, penetrance information and phenocopy information for dominant and recessive disease models can be selected by the user through a dialog (Figure 3). The Mendelian genotype errors inconsistent with parental genotypes are detected and set to missing genotypes. To handle other genotyping errors or wrongly mapped SNP markers, we assume a conservative genotyping error rate of 0.01 1 (user adjustable) and regard observed genotypes as phenotypes in the likelihood computation 17. As a result, the computation of the probability of the observed genotype data at one marker given an inheritance vector v involves the summation over all the possible real genotypes (or equivalently the founder allele configurations):

where F_i represents the ith of all the possible founder allele configurations and is independent of v. P(real genotypes i | v, F_i) is 1 since an inheritance vector and founder allele configuration uniquely determines the real genotypes, and P(observed genotypes | real genotypes i) involves comparing the real genotype and observed genotype for all the individuals and multiplying the probability by the error rate of 0.01 (default value) for each disagreement and 0.99 for each agreement. We also use the matrix-vector multiplication algorithm and bit reduction due to founder phase symmetry described in 16, and the founder allele factoring technique reported in 617 to speed up the computation of single-locus and accumulative likelihood vectors as well as the likelihood vector of disease phenotypes.

We use the forward-backward computation in the Lander-Green algorithm to obtain the marginal probability distribution of inheritance vector at each SNP marker position given the data of all the markers on a chromosome. In addition the most likely inheritance vector at each marker given the genotype data of all the markers on this chromosome is calculated 6. Conditioned on the most likely inheritance vector at a marker and the observed genotype data, we can find the most likely founder allele configurations. When there are competing inheritance vectors with the same largest marginal probabilities at a marker, we select the one with fewer crossover events from the last marker since the distance between adjacent markers are small (average 300 kb) and it is therefore less likely to have multiple crossover events between two markers in a pedigree 7. Together these procedures give the haplotyping results of the pedigree data. dChipLinkage visualizes the haplotyping result in either the haplotype view or the ordered genotype view.

CompareLinkage can format Affymetrix Mapping 10 K SNP genotype output files and genotype files into the "Linkage" format and convert genome information and pedigree files into the formats suitable for Merlin (Version 0.10.2), GeneHunter (Version 2.1) and Allegro (Version 1.2). CompareLinkage removes all non-informative markers and calls the PedCheck software 18 to detect genotype incompatibilities using the pedigree information. A Mendelian genotype inconsistency at a SNP is handled by setting the genotype of this SNP in all the individuals to missing. For the analysis in GeneHunter, overlapping segments of large chromosomes are prepared, with each segment containing 150 or fewer markers with 75 markers in common between adjacent segments. Linkage scores are computed as the mean of two scores for the same marker from the two overlapping fragments. We ran genome-wide linkage analysis using all the three software packages and dChipLinkage for the 10 K SNP genotype data of three families: 5026.10 (Figure 4; autosomal recessive non-syndromic deafness disease, 13 bits, Asian), CR (Figure 5; recessive, 17 bits, Asian) and ER (Figure 6; dominant, 17 bits, Caucasian). For the parametric analysis, we use a disease frequency of 0.001, a penetrance value of 0.99 and a phenocopy of 0.01 for all the families and all the software packages. For GeneHunter and Allegro we ran both nonparametric and parametric analysis. For Merlin, the NPL_all statistic is computed. The allele frequencies are calculated based on the actual genotype data in each family. The LOD score or NPL score are computed at the position of the SNP makers. After running the analysis for all chromosomes, the two chromosomes with the largest LOD scores were selected from each pedigree and compared below.

Figures 7, 8, 9, 10, 11, 12 show the comparative LOD score and nonparametric score plots in dChip for these chromosomes analyzed with GeneHunter, Merlin, Allegro and dChipLinkage. The vertical red line in the figures indicates the significance threshold and is set to 3 for parametric analysis (LOD scores) and to 3.7 for non-parametric analysis (NPL score) based on statistical significance recommended by Lander and Kruglyak 19. The linkage scores largely agree with each other in the regions with significant LOD/NPL scores. GeneHunter, Merlin and Allegro detect the peaks in the chromosome 1 and 3 of the consanguineous family 5026.10 but compute lower LOD scores than dChipLinkage (indicated by arrows in Figure 7 and 8). For another consanguineous family CR with a recessive disease, all software packages detect similar peak regions in the two chromosomes denoted as A and B (Figure 9 and 10). For the family ER with a dominant disease, dChipLinkage computes similar overall patterns but reports a possible sporadic and non-significant peak (LOD < 1.6) in each chromosome (indicated by arrows in Figure 11 and 12).

To do parametric linkage analysis in dChipLinkage, a pedigree file is needed (Figure 13C). The file is similar to the standard pedigree file format but has an additional "Array" column matching each individual in the pedigree file to the corresponding genotype information in the genotype file through array names (the header line in Figure 13A). The data importing and analysis steps are:

1. Open dChip.

2. Select the Analysis menu and the Get External Data function to read in the genotype file in the text format (Figure 13A).

3. Select the genome information file downloaded from the dChip website (Figure 13B). This file is provided in three versions, each containing the SNP information like TSC SNP ID and genetic map locations but having different allele frequencies for each of the three ethnic groups (Asian, Caucasian and African Americans).

4. Select the Analysis menu and the Chromosome function to display the genotype calls, genes and cytobands along the chromosome

5. After the program has displayed the genotype data, select the Chromosome menu and the Linkage function to start the dChipLinkage module (Figure 3). Specify the pedigree file (Figure 13C) and other linkage parameters. Depending on whether the dChip "Chromosome View" displays one or all chromosomes, the linkage analysis will be performed for one or all chromosomes accordingly. For the analysis of the 5026.10 family, the recessive disease model is assumed, and a penetrance of 0.99, phenocopy of 0.01, disease allele frequency of 0.001 and a SNP marker error rate of 0.01 are used. The SNP allele frequencies in the genome information file are used and truncated to values between 0.001 and 0.999. This family has 13 bits and it takes about 20 minutes for the whole genome linkage analysis.

Using dChipLinkage to analyze the 5026.10 family, we were able to identify a region on the chromosome 1 (Cytogenetic region: 1p36.32 – 1p36.22) with LOD scores of greater than 2.3 (Figure 7, indicated by arrow). The most interesting gene in this region is ESPIN, which has previously been shown to be involved in deafness in mice 20 and two frameshift mutations in the gene have just recently been associated with deafness in two consanguineous families 21. Sequence analysis of the locus revealed that the parents (the individual 1 and 2 in Figure 4) and the unaffected child (the individual 6) are heterozygous for the insertion mutation and the affected children are homozygous (data not shown). In addition a novel locus with a maximum LOD score of 2.77 was identified on the chromosome 3 (Figure 8, indicated by arrow). The peak region on the chromosome 3 is about 2 Mb wide (Figure 14C). Using the GeneHunter software, we compute a maximum expected LOD score of 2.78 for this family under the specified parameters. Therefore we extract the most linkage information based on the dense SNP markers in this region. Figure 14 shows the LOD score curve together with genotype calls, inferred haplotypes and ordered genotypes based on haplotyping. In Figure 15 the results are presented in the context of cytobands and genes. The Chromosome/Export SNP data function can also export the text information of the SNPs, genes and cytobands in the region with linkage scores exceeding the threshold.

After the linkage computation is finished, the inferred haplotype information can be visualized. In the haplotype view (Figure 14 and 16), one can view the inference on how the founder chromosomes are crossed over and inherited by the descendants. The different colors represent distinct founder chromosomes, and for each individual, the father allele haplotype is displayed on the left and mother on the right. Since a pedigree contains no phase information of the founders 6, in the linkage computation we can assume that one child of each founder always inherits the whole grandfather-descent chromosome. This assumption does not affect the LOD score computation but reduces the number of bits in the Lander-Green algorithm by the number of founders and consequently reduces the analysis time. This is the reason that in Figure 14B the individual 1 has both father and mother haplotypes in pure color and individual 2 has only the father haplotype in pure color. By inspection of the observed genotype and the inferred haplotypes (Figure 16), one can see that only in the peak LOD score region all the affected children (individual 3, 4, 5 and 7) are homozygous and that the unaffected child (individual 6) is heterozygous. All the affected individuals share two copies of the identical chromosome segment (the pink color between the two arrows) presumably containing the disease locus. By two very close crossover events respectively in individual 6 (indicated by the black arrow) and individual 7 (indicated by the white arrow), the LOD score implicates the possible disease gene in a 2 Mb region and one can easily search the physical map for candidate disease genes in this region in the dChip chromosome view (Figure 15).