Advances in genomic analyses, particularly massively parallel sequencing (MPS), have enabled unprecedented high throughput SNP analyses which have fostered the development of Forensic Genetic Genealogy (FGG). This subdiscipline of forensic genetics exploits genome-wide dense single nucleotide polymorphism (SNP) data generated by whole genome sequencing (WGS) or microarrays to determine near and distant kinship relationships. While targeted sequencing, in theory, can be used in forensic contexts to interrogate selected SNPs at lower cost (although current commercial forensic options are relatively costly), whole genome sequencing (WGS) offers a more comprehensive approach, capturing significantly more genetic variation and enabling deeper kinship inference. As sequencing costs continue to decline, WGS is becoming increasingly preferred in FGG for its broader utility and richer dataset (Mandape, et al., 2024). The kinship relationships can facilitate investigative leads for identification of human remains and source attribution of donors of crime scene DNA evidence more effectively than current forensic genetic methods (Dowdeswell, 2023; Kling, 2019; Kling and Tillmar, 2019; Mandape et al., 2024; Budowle et al., 2024). Kinship associations typically are determined using either IBS (identity by State) or IBD (identity by Descent) segmented-based methods (Henn et al., 2012; Browning and Browning, 2011; Browning and Browning, 2013; Huff et al., 2011; Lipatov et al., 2015; Staples et al., 2016; Kaplanis et al., 2018; Ramstetter et al., 2018; Morimoto et al., 2018; Seidman et al., 2020). While these approaches provide high accuracy, they differ from the likelihood ratio (LR)-based framework traditionally used in kinship analysis, which underpins the validity and acceptance of short tandem repeat (STR) interpretation methods in forensic applications. LR-based statistics are routinely employed to support identifications, making them a critical component of forensic practice. To enable the forensic community to leverage SNP data, an analogous LR-based framework is needed that offers the flexibility, rigor, interpretability, and validity required for broad acceptance. Accordingly, an LR-based interpretation method tailored to whole genome sequencing (WGS) data for SNP selection and relationship inference was developed, with a focus on pairwise comparisons up to second degree relatives.

Like traditional kinship analysis, the approach described herein calculates the likelihoods of each selected SNP for specific relationships, with LRs obtained by comparing the likelihoods of genetic data under alternative relationships. Assuming independence among SNPs, the cumulative LR is calculated by multiplying the LR values for each individual SNP. Therefore, selecting the most informative SNPs is critical to the success of the method. Our approach is unique in that it enables dynamic SNP selection in tandem with likelihood ratio (LR) calculations compared with traditional kinship software that relies on fixed, pre-selected markers. This dynamic integration allows for greater flexibility and improved performance when working with WGS data. A straightforward approach was used to refine a subset of SNPs from the hundreds of thousands to millions identified by WGS, prioritizing high minor allele frequency (MAF) markers observed in each case. These markers offer high discrimination power for relationship inference. Subsequently, the data are further subset to include SNPs located in genomic regions identified by Genome-in-a-Bottle as being easy to sequence or genotype (Dwarshuis et al., 2024). This subset is then further refined by selecting only SNPs that exhibit nominal or no linkage and linkage disequilibrium (LD). This approach provides a curated panel of SNPs with high MAF for robust LR support for true close kinship relationships, particularly up to the second degree.

Indeed, Yousefi et al. (2018) demonstrated in a Dutch sample population that 50 independent SNPs (MAF > 0.2; average MAF = 0.35) are sufficient to yield low probabilities of identity (or population match probabilities) of 6.9 × 10⁻²⁰ and 1.2 × 10⁻¹⁰ for unrelated individuals and siblings, respectively. Chakraborty et al. (1999) estimated that approximately 40 SNPs (MAF ∼ 0.5) could achieve random match probabilities around 10⁻¹⁵ and 33 SNPs (MAF ∼ 0.5) would be required to reach an exclusion probability of 99.9% in a typical trio paternity case. Furthermore, the higher the MAF for a suite of binary SNPs (i.e., approaching the highest heterozygosity of 0.5) is globally, the lower are the effects of population substructure. Moreover, with high heterozygosity any positive predictive power with binary SNPs for associations with private genetic information would be negligible (Budowle and van Daal, 2008).

Here, the validation of a statistical approach, KinSNP-LR (version 1.1), is described for computing LRs based on WGS generated SNP data. Instead of a priori selecting a fixed panel of informative SNPs, the first SNP on an end of a chromosome passing the MAF threshold is selected from a large candidate panel and the next SNP at a specified genetic distance (e.g., 30–50 centimorgans (cM)) and meeting the MAF criterion is selected and so on across the genome. The LRs for multiple relationships are calculated based on the methods described in Thompson (1975), Ge et al. (2010), and Ge et al. (2011). This approach maximizes the number of SNPs with little to no linkage and LD selected in a case-specific manner.

This paper describes the development and validation of a LR-based methodology designed to conduct close kinship relationship testing by dynamically selecting informative SNPs based on MAF and MGD from genome-wide SNP data. Our approach is unique in that it integrates dynamic SNP selection based on output data, rather than relying on a fixed, pre-selected SNP set. The described approach bridges modern genomic datasets with the traditional LR-based framework, allowing kinship testing laboratories to leverage comprehensive genomic information while maintaining established LR-based interpretive methodologies. KinSNP-LR, as implemented in this study, does not incorporate comprehensive population parameters, such as Fst or linkage between markers; these are features that can be considered in future iterations. This initial implementation is intended to demonstrate the core logic and workflow of dynamically integrating SNP selection with LR calculations for close kinship analysis. The system is evaluated for accuracy using empirical and simulated datasets.

However, the traditional LR framework assumes marker independence, necessitating the selection of only a limited subset of SNPs from the total typed data to minimize the confounding effects of linkage and LD³¹. The current implementation of the methodology supports parent-child, full-sibling, and second degree relationships, as well as distinguishing unrelated individuals. These close relationships are most commonly encountered in missing and unidentified persons cases handled by medical examiner offices, where the goal is to confirm identity rather than generate investigative leads. While more distant relationships could be incorporated in future iterations by including linked SNPs, such comparisons generally are less fruitful for identity confirmation due to the higher number of potential matches and reduced specificity.

This LR methodology assumes access to high-quality genotyping data. When genotypes are derived from low-quality or low-quantity DNA samples, rigorous quality control measures are necessary to ensure accurate relationship testing, which may include the use of metrics such as genotype quality scores, read depth, or other laboratory-specific thresholds for SNP quality. The default mutation rate in the methodology is set at 0.001, corresponding to the Q30 threshold, as genotyping errors are significantly more frequent than the actual biological mutation rate (∼10⁻⁸). For datasets of lower quality, the mutation rate can be adjusted to higher values (e.g., 0.01) to account for increased error rates. In this study, the mutation rate (or better stated typing error rate) was performed only on parent-child relationships as is typically done in such cases. However, one could consider applying the typing error rate to all relationships tested herein since typing error will impact all sequence data.

Additionally, targeted panels may be considered, as they can be enriched to increase sensitivity of detection. The LR approach would remain similar, though the number of candidate SNPs will be fewer unless the DNA is highly degraded. For close relationships up to second degree, the number of SNPs may be sufficient, as only a limited number are needed. However, enrichment approaches still may suffer from differential detection of SNP states, and overall data quality will impact success. While the choice between targeted panels and WGS has historically been influenced by cost and laboratory preference, declining sequencing costs are making WGS increasingly attractive due to its broader genomic coverage and greater utility in distant kinship inference. This LR tool was developed to accommodate WGS data, but if one would like to use it with a targeted panel, the panel should be evaluated for its ability to include a maximum number of unlinked (or loosely linked) SNPs.

Substructure, which likely exists at some level, was not included in this iteration of the software because these SNPs likely exhibit low Fst values, particularly for each individual population. The impact of Fst on calculations involving high frequency SNP alleles should minimize their impact of substructure on LR calculations and support robust relationship testing across populations. Future versions of KinSNP-LR could explore the impact of a Fst correction based as described by Balding and Nichols (1994).

The selection of an appropriate large set of candidate SNPs for relationship testing was derived from a preselected gnomAD panel containing 222,366 SNPs. These SNPs were filtered based on a MAF threshold of 0.3. Subsequently, the data were further subset to include SNPs located in genomic regions identified by Genome-in-a-Bottle as being easy to sequence or genotype. One might consider a much lower MAF, such 0.01, as rare alleles may increase the LR; however, such SNPs would not apply to all kinship comparisons as they are quite uncommon and have little value in most cases. A high MAF ensures that the SNPs from the large, preselected candidate panel of 222,366 SNPs would be applicable routinely. The high MAF threshold enhances the informativeness of selected SNPs, while reducing risk of unintended medical interpretations. This filtering strategy ensures a robust and unbiased dataset, providing a solid foundation for FGG as well as for other forensic applications, such as mixture interpretation. These selection steps provided SNPs that are amenable to validation studies performed herein, supporting reliability and reproducibility in forensic applications.

There have been reports by Mostad et al. (2023) and Andersen et al. (2025) that also provide approaches to use LRs with kinship analyses based on SNP data. They both have some positive features and limitations, as does KinSNP-LR. Mostad et al. (2023) use the Lander-Green algorithm, basically a Hidden Markov model (HMM), like the methods described by Epstein et al. (2000) and Boehnke and Cox (1997). Their method can exploit larger amounts of SNP data with the assumption of linkage equilibrium and no genetic interference, in which the likelihood of individuals given a defined relationship is calculated using the allele frequencies, the recombination fractions between the markers, and transition probabilities of the IBD values of linked markers. Thus, instead of using a maximum of a few hundred SNPs with the assumption of little or no linkage between the markers, HMM-based methods in theory could use all available SNPs to increase the accuracy of the relationship tests. However, the Mostad et al. (2023) approach does not address mutation or genotyping error which could be a problem with low quality data, although mutation rates could be added to their model. The benefits of HMM may be limited for close relationships, as shown herein and the accuracy is already very high for close relationships with a few hundred SNPs. However, if LRs are desired for more distant relationships, HMM-based methods can be utilized in subsequent versions of KinSNP-LR. Andersen et al. (2025) used a preselected panel of 43 SNPs with relatively high MAF for LR calculations accounting for genotype errors for direct comparison of single source samples to persons of interest. Our approach is in some ways similar to that of Andersen et al. (2025) but extended to indirect kinship comparisons and makes use of more SNPs. KinSNP-LR provides a model for possible solutions to address LR/SNP-based calculations by dynamically selecting high MAF SNPs common to several population groups, MGD to account for linkage effects, incorporation of a basic mutation (or genotyping error) rate model, and being flexible on a case-by-case perspective.

In addition, including multiple reference family members instead of pairwise comparisons (i.e., joint probabilities), which are not described herein because the current genetic genealogy applications are mostly with pairwise relationship tests, can be considered in future studies. For cases with more than one family reference sample, a pedigree LR can be calculated capturing all genetic information from all the references. Ge et al. (2010), based on the Elston-Stewart algorithm (Elston and Stewart, 1971), provide the methods for calculating a pedigree likelihood which incorporated genetic mutation, population substructure and accommodations for missing genetic data.

There are other ways to implement algorithms other than the one used herein which is a traditional, case-specific approach for LR calculation based on genotype combinations, where the formulae differ according to the genotype combination of the participants. For example, unified LR calculation formulae could be considered to address all genotype combinations with a single, more generalizable equation (Egeland, et al., 2017; Ma, et al., 2024). This would reduce reliance on complex code with numerous conditional statements which in turn potentially could reduce computational efficiency and increase risk of implementation errors, especially when dealing with high-density SNP data. KinSNP-LR is designed for close relationship testing for a specific targeted audience, such as medical examiners. The code for this software is only a few hundred lines. Therefore, these concerns that can be obviated to some degree with a unified LR calculation formulation are less likely an issue with KinSNP-LR which has been tested under the conditions shown in this study. In future iterations, where more distant and complex relationships would be addressed, the unified approach will be considered.

The validation results from both empirical and simulation data show high accuracies of close relationships, particularly for first degree relationships. One potential limitation is that the sample size of 50 pedigrees may be considered small. However, the data shown in Figures 6–9 and Table 4 support that the sample size is sufficient to evaluate the performance of the software tool for the tested relationships. On another point, the ratios for selecting relationships with maximum and second maximum likelihoods differed for empirical data (>100) and simulated data (>1,000). Sampling is a potential explanation for the observed differences between empirical and simulation results. The empirical data have less samples, which tend to result in lower ratios overall. With more samples, as with the simulated data, larger ratios will be observed. These differences between the maximum and the second maximum likelihoods are merely observations and may not be construed as a recommended threshold(s) for casework. Threshold selection will be done at the laboratory level and based on the risk profile that a laboratory or jurisdiction defines operationally.

Comparisons with other kinship software were not conducted. The reasons were the input data format and requirements, or even assumptions and methods, are different for different software programs. To seamlessly integrate with WGS-based genetic genealogy application, KinSNP-LR can accept any VCF file without pre-defined panels, which is unique among all LR-based relationship testing software programs. Due to the different input data, assumptions, methods, and/or software implementations, different software programs may use different sets of variants, population features, and methods in relationship testing, which will lead to different LR results or even different conclusions but not provide insight into accuracy of any given software. Thus, the evaluation (e.g., accuracy) of any software program should be conducted with ground truth data, which is what was done in this study, instead of comparing with other software programs.

In addition, the WGS data generated for some forensic samples may yield 0.5X to 5X coverage. If, for example, DeepVariant (Poplin et al., 2018) were used to call variants with additional high-quality filters, such as Genotype Quality (GD) ≥ 30 (equivalent to 99.9% genotyping accuracy) and a minimum read depth ≥5, such samples may yield limited data. However, for higher quality samples, typically ∼1,600 to ∼350,000 high quality variants could be detected for relationship testing. Even with further MAF and MGD filtering, there should be enough variants for close relationship tests, which usually only require dozens to a few hundred SNPs. On the other hand, relationship test software programs that require pre-selected panels may not work well for WGS data generated from typical forensic samples, as there may not be enough SNPs recovered with smaller pre-selected SNP panels after stringent quality filtering. A pre-selected panel may start with upwards of ∼3,500 to ∼10,000 compared with millions or tens of millions of SNPs by WGS.

In summary, the methodology presented in this study marks a significant step forward in bridging traditional forensic relationship testing with modern genomic technologies. By applying a LR-based framework with dynamically selected high-MAF SNPs, forensic laboratories can harness the data generated by WGS while maintaining compliance with accredited relationship testing standards. The use of unlinked markers and case-specific SNP selection distinguishes this method from existing LR-based kinship tools, making it especially applicable for FGG and other applications requiring statistically robust assessments of close relationships. This approach is particularly important for medical examiners, who are focused on confirming identity, which is a process that typically relies on comparisons of close relatives. In cases where STR testing is not possible due to degraded or limited DNA, this SNP-based LR approach offers a reliable and scientifically defensible alternative. As the field evolves, these advancements provide a strong foundation for broader adoption of relationship testing approaches that combine the rigor of traditional statistical models with the scalability of modern genomics.