Selection and evaluation of bi-allelic autosomal SNP markers for paternity testing in Koreans

We used SNP genotyping data from the Ansan-Ansung cohort and the Twin-Family cohort, which were part of the Korean Genome and Epidemiology Study (KoGES) [23]. DNA was extracted from blood samples collected from individuals. The participants of the Ansan-Ansung cohort study were 10,030 adults aged 40 to 69, who live in Ansan or Ansung. Among them, 8840 individuals were genotyped with the Affymetrix Genome-Wide Human SNP Array 5.0. The Twin-Family cohort study, consisting of 3202 twins and their families, collected SNP data from 1716 individuals using the Affymetrix Genome-Wide Human SNP Array 6.0.

In this study, the selection of candidate SNP markers and calculation of allele frequencies were based on the Ansan-Ansung cohort data, and the performance of markers was evaluated using the Twin-Family cohort data. The Ansan-Ansung cohort data included 352,228 bi-allelic autosomal SNPs. Among them, SNPs not included in the Twin-Family cohort data were excluded. To avoid the influence of selection pressure, SNPs within the range of a gene list (hg19) were discarded. Quality control (QC) steps were performed as follows: (1) Samples with individual missing rates higher than 0.05 were filtered. (2) SNPs with missing genotype rates higher than 0.01 were removed. (3) SNPs that deviated from the Hardy-Weinberg equilibrium (p value < 10⁻⁵) were removed. Next, we estimated kinship coefficients to identify potential relatives. Samples were filtered until kinship coefficients of all pairs of individuals were lower than 0.0884, meaning that all pairs were treated as third-degree or more distant relationships. The fixation index (F_ST) was calculated between Ansan and Ansung populations. Then, we performed linkage disequilibrium (LD)-based SNP pruning with the following parameters: window size = 500, step size = 50, and r² threshold = 0.01. Finally, 200 candidate SNPs with the highest minor allele frequencies (MAFs) were selected among the retained SNPs. The minimum distance between candidate SNPs in each chromosome was 10 Mbp. PLINK v1.90 was used to conduct QC steps and LD pruning and to calculate F_ST and MAF [24]. Kinship coefficient was estimated using KING v2.2.4 [25].

In order to select the appropriate number of markers needed for the paternity test, we simulated 10,000 pedigrees using MERLIN v1.1.2 [26]. Since this program requires centimorgan (cM) position for each SNP, we obtained this information from the genetic map of the CHB (Han Chinese in Beijing, China) population of the 1000 Genomes Project (available at ftp://​ftp.​1000genomes.​ebi.​ac.​uk/​vol1/​ftp/​technical/​working/​20130507_​omni_​recombination_​rates/​). If there was no information about the genetic position of the SNP, the genetic position of the nearest SNP was used. The structure of simulated pedigrees is shown in Fig. 1. Founder genotypes were randomly generated based on the previously calculated allele frequencies of candidate SNPs. Then, it was assumed that each parent contributes one allele to the offspring. Alleles spaced less than 25 cM were clustered and passed to the offspring based on the estimated haplotype frequencies.

In paternity testing, we evaluated genetic evidence by comparing likelihoods of hypotheses using the equation: LR = Pr(G|H_parent-child)/Pr(G|H_unrelated), where G is the observed genotype data, H_parent-child is the hypothesis that two tested people are in a parent-child relationship, and H_unrelated is the hypothesis that two tested people are unrelated [22]. We calculated LRs for each marker and multiplied them all to obtain final LRs for a set of markers using a “likelihoodMerlin” function of “pedprobr” package in R [26, 27]. This function can take into account linkage when calculating LRs. Ayres [8] estimated that at least 70 SNP loci were required for paternity testing in the motherless duo case, so the number of markers used in the analysis was increased from 70 to 200 by 10. When the log₁₀LR was greater than or equal to 5, which provides very strong support for H_parent-child [9], AF and child were judged to be in a parent-child relationship. The accuracy is defined as the percentage by which the relationship of two individuals is correctly determined as parent-child or non-parent-child. The false positive is an error in which a person who is not the TF is falsely included as the TF. In contrast, the false negative means that the TF is falsely excluded.

The Twin-Family cohort data was used to examine the performance of selected markers. Individuals with missing genotype data of these SNPs were filtered out. We collected all pairs of two individuals who were in parent-child or unrelated relationships. We also used second-degree relative pairs, such as uncle-nephew or grandparent-grandchild, to check LR values when the AF is the close relative of the TF. The methods used to decide paternity and calculate the accuracy, false positive, and false negative were written above. The method of maternity testing was the same as that of paternity testing, except that only the gender of the typed person is female. Thus, we calculated LRs for all pairs of individuals, ignoring their gender.

The number of bi-allelic autosomal SNPs included in both data was 280,905. Since functional markers have a possibility that the selection pressure affects the allele frequency [9], 125,485 SNPs in the gene region were excluded to avoid this. A total of 12,238 SNPs with a genotyping rate lower than 0.99 were removed. The total genotyping rate of the retained SNPs from the Ansan-Ansung cohort was 0.9987. One hundred ninety-three SNPs were removed due to failure to pass the Hardy-Weinberg exact test. Among 8840 participants with a genotyping rate per individual higher than 0.95, 8621 unrelated samples were selected. All SNPs had F_ST values lower than 0.01, so it could be considered that there was no significant genetic difference between the populations in these regions. Finally, 10,538 independent (pairwise r² < 0.01) SNPs were retained by the LD pruning method.

To select highly informative SNPs, we calculated MAFs of 10,538 SNPs from 8621 unrelated samples and selected 200 SNPs located far from each other (> 10 Mbp) with a high MAF. These SNPs had an MAF in the range [0.49, 0.5].

A total of 10,000 families were generated to determine the appropriate number of markers for paternity testing. Within each family, we were able to collect four types of AF-child pairs as shown in Fig. 1: The AF was the TF (AF 1), the AF was the close relative of TF (AF 2 and 3), and the AF was the unrelated person (AF 4). LRs were calculated for each pair, using 70, 80, …, and 200 SNPs.

Table 1 shows the accuracy, false-positive rate, and false-negative rate in simulated duo cases. Since we used a high value of LR cutoff, the false-negative rate was very high with 70-80 loci, which were suggested in a previous study [8]. When the number of loci was 150, no TF was falsely excluded. However, there were some cases that the first-degree relative of the TF was judged to be the TF, so the false-positive rate was 0.0033%. One hundred percent accuracy was achieved when 160 or more SNPs were used. As a result, we selected a set of 160 SNPs for paternity testing without errors. The details of the selected SNPs are shown in Supplementary Table 1.

Figure 2 shows the distribution of log₁₀LR values for true parent-child pairs. The average, minimum, and maximum values of the log₁₀LR for parent-child pairs were 12.05, 6.02, and 18.98, respectively. In non-parent-child pairs, LRs were all zero due to SNP mismatches between the two tested people.

The family genotype data included 1716 samples. Among them, 816 samples, whose genotypes of finally selected SNPs were observed, were used for validation. Out of a total of 332,520 pairs, those with unknown or uncertain relationships were excluded from the results. Finally, there were 295, 19, and 331,778 pairs of parent-child, second-degree relative, and unrelated individuals, respectively. When paternity testing was performed using 160 selected SNPs, the accuracy reached 100%. The LR values for non-parent-child pairs were zero, and the average, minimum, and maximum values of the log₁₀LR for parent-child pairs were 12.18, 8.42, and 19.26, respectively (Fig. 3). We also identified the number of opposite homozygosity for 160 SNPs, which means the child and the AF are homozygous and have different alleles; for example, the child has allele AA and the AF has allele BB [15, 28]. The more mismatches are observed, the less likely the two individuals are in parent-child relationships. No mismatch was found in any of the 160 loci in all true parent-child relationships. When the AF was the first-degree relative of the TF and the unrelated man, the average number of mismatches between the AF and the child was 10.95 and 20, respectively.