A comparison of pedigree, genetic and genomic estimates of relatedness for informing pairing decisions in two critically endangered birds: Implications for conservation breeding programmes worldwide

Abstract Conservation management strategies for many highly threatened species include conservation breeding to prevent extinction and enhance recovery. Pairing decisions for these conservation breeding programmes can be informed by pedigree data to minimize relatedness between individuals in an effort to avoid inbreeding, maximize diversity and maintain evolutionary potential. However, conservation breeding programmes struggle to use this approach when pedigrees are shallow or incomplete. While genetic data (i.e., microsatellites) can be used to estimate relatedness to inform pairing decisions, emerging evidence indicates this approach may lack precision in genetically depauperate species, and more effective estimates will likely be obtained from genomic data (i.e., thousands of genome‐wide single nucleotide polymorphisms, or SNPs). Here, we compare relatedness estimates and subsequent pairing decisions using pedigrees, microsatellites and SNPs from whole‐genome resequencing approaches in two critically endangered birds endemic to New Zealand: kakī/black stilt (Himantopus novaezelandiae) and kākāriki karaka/orange‐fronted parakeet (Cyanoramphus malherbi). Our findings indicate that SNPs provide more precise estimates of relatedness than microsatellites when assessing empirical parent–offspring and full sibling relationships. Further, our results show that relatedness estimates and subsequent pairing recommendations using PMx are most similar between pedigree‐ and SNP‐based approaches. These combined results indicate that in lieu of robust pedigrees, SNPs are an effective tool for informing pairing decisions, which has important implications for many poorly pedigreed conservation breeding programmes worldwide.

Recent studies argue that a better indication of genome-wide diversity can be obtained from genomic-based estimates of relatedness based on large numbers of genome-wide single nucleotide polymorphisms (i.e., SNPs; Knief et al., 2015;Taylor, 2015;Taylor et al., 2015).
To our knowledge, no study has compared pedigree-, geneticand genomic-based approaches for estimating relatedness to inform pairing decisions for conservation breeding programmes, despite there being over 350 vertebrates worldwide that are captive bred for release to the wild (Smith et al., 2011). Here, we evaluate these three approaches using two critically endangered birds endemic birds to Aotearoa New Zealand-the kakī/ black stilt (Himantopus novaezelandiae) and kākāriki karaka/orange-fronted parakeet (C. malherbi)-as proof of concept. Kakī and kākāriki karaka are excellent candidates for this research as both have active conservation breeding programmes, as well as multigenerational pedigrees (this study), microsatellite panels (Andrews, Hale, & Steeves, 2013;Steeves, Hale, & Gemmell, 2008) and genomic resources including species-specific reference genomes and whole-genome resequencing data (Galla et al., 2019; this study). In addition, because captive breeding pairs for both species are housed in separate enclosures and all offspring are intensively managed, kakī and kākāriki karaka present an excellent opportunity to examine relatedness in known family groups.
Once found on both main islands of Aotearoa, kakī experienced significant population declines throughout the 20th century due to introduced mammalian predators (e.g., feral cats, Felis catus; stoats, Mustela erminea; and hedgehogs, Erinaceus europaeus) along with braided river habitat loss and degradation (Sanders & Maloney, 2002). Today, an estimated 129 kakī are largely restricted to braided rivers of Te Manahuna/The Mackenzie Basin (Department of Conservation, personal comm.; Figure 1a) and recovery efforts include a conservation breeding programme that was initiated in the early 1980s (Reed, 1998). In addition to breeding birds in captivity, the kakī recovery programme also collects eggs from intensively monitored wild nests and rears them in captivity before wild release. Similar to kakī, kākāriki karaka were also once found on both main islands of Aotearoa and experienced population declines in the 19th and 20th centuries due to introduced mammalian predators (e.g., brushtail possums, Trichosurus vulpecula; rats, Rattus rattus and R. norvegicus; and stoats) and habitat loss (Kearvell & Legault, 2017 are relatively long-lived braided river specialists that breed predictably within the bounds of a spring and summer season (Pierce, 2013).
In contrast, kākāriki karaka belong to the Order Psittaciformes and are relatively short-lived beech forest specialists capable of breeding year-round, with prolific breeding periods associated with food abundance (e.g., beech forest masting events; Kearvell & Legault, 2017).
Here, we compare relatedness estimates from pedigree, microsatellites and genome-wide SNPs using known parent-offspring and full sibling relationships. We then compare pairing recommendations among these three approaches to assess how each translates to effective conservation management. Given that kakī and kākāriki karaka represent two taxonomically distinct bird species with different life history strategies, we anticipate the results of our research may be applicable to the wider conservation breeding community.

| Sample collection and DNA extraction
Animal ethics approval for this project has been granted by the New Familial relationships are known for all samples collected, as they were sampled from birds of known provenance in captive conditions. However, to verify that no sample was mislabelled during genetic and genomic processing, parent-offspring relationships were verified through an allele mismatch exclusion analysis (Jones & Ardren, 2003) using microsatellite panels previously developed for kakī (Steeves et al., 2008) and kākāriki karaka (Andrews, 2013), with a maximum allowed mismatch of one allele at one locus (see Microsatellite data F I G U R E 1 Current breeding distributions of wild kakī (a) and kākāriki karaka (b) in Aotearoa

| Pedigree-based relatedness
Multigenerational pedigrees were constructed for kakī and kākāriki karaka by entering studbook information (i.e., hatch date, sex, parentage and status) into the programme PopLink v. 2.5.1 (Faust, Bergström, Thompson, & Bier, 2018). Sex for all individuals was verified using molecular markers 2550F/2718R (Fridolfsson & Ellegren, 1999) for kakī and P2/P8 (Griffiths, Double, Orr, & Dawson, 1998) for kākāriki karaka, with PCR products run on a 3% agarose gel for visual characterization, with positive controls included. Due to the short distance between P2/P8 alleles on the Z and W chromosomes in kākāriki karaka (Robertson & Gemmell, 2006), 2 µl of PCR products using a tagged forward primer was combined with 11.7 µl for- 1.6.20190628 (Lacy, Ballou, & Pollak, 2012), selecting only the individuals used in this study (n = 36 in kakī and kākāriki karaka) and treating all unknown individuals in the pedigree as wild founders. In order to produce direct comparisons of pairwise relatedness coefficients (R) between pedigree, genetic and genomic data, R was calculated from pedigree kinship data using R(xy) = 2 * f(xy)/√{(1 + Fx) (1 + Fy)}. In this formula, f(xy) is the kinship between two individuals (x and y) and Fx and Fy are the inbreeding coefficients of individuals x and y (Crow & Kimura, 1970).
Samples were prepared for genotyping by adding 0.5 µl of PCR product to 11.7 µl formamide and 0.3 µl GeneScan™ LIZ® 500 size standard (Applied Biosystems) and were genotyped on either a 3130xl or 3730xl Genetic Analyser (Applied Biosystems). Chromatograms were visualized using GeneMarker v. 2.2. To avoid bias by potential dye shifts (Sutton, Robertson, & Jamieson, 2011), peaks were scored manually. The number of alleles and standard estimates of per locus diversity-including expected heterozygosity (H O ) and observed heterozygosity (H E )-were produced using GenAlEx v. 6.5 (Peakall & Smouse, 2006;Smouse & Peakall, 2012). Tests for deviations from Hardy-Weinberg and linkage disequilibrium for these loci using samples that are representative of larger kakī and kākāriki karaka populations can be found in Steeves et al. (2008), Steeves, Maloney, Hale, Tylianakis, and Gemmell (2010) and Andrews (2013), respectively.
For kākāriki karaka, only eight of the 18 microsatellite markers previously described were polymorphic in this study and these eight loci were used in all downstream analyses.
Genetic-based R estimates were produced in the programme COANCESTRY v. 1.0.1.9 (Wang, 2011). COANCESTRY offers seven different estimators of relatedness, and to choose the most appropriate estimator for the kakī and kākāriki karaka microsatellite data sets, we employed the simulation module within COANCESTRY using allele frequencies, missing data and error rates from our microsatellite data sets. To produce dyads that represent the relationships and degree of inbreeding found within kakī and kākāriki karaka, we used R package "identity" (Li, 2010) to generate 10,879 dyads for kakī and 1,484 dyads for kākāriki karaka based on the pedigrees of both species. The frequency of each unique dyad in the kakī and kākāriki karaka data sets was scaled to create 1,000 dyads for each set that are representative of relationships between individuals used in this study. The COANCESTRY simulations were conducted using allele frequencies, error rates and missing data rates from each microsatellite data set, with settings changed to account for inbreeding. The triadic likelihood approach (Wang, 2007) was selected given it had the highest Pearson's correlation with "true" relatedness for both data sets (see Supporting Information for details). This approach is also preferred, as it is one of the few estimators that accounts for instances of inbreeding (Wang, 2007).
To estimate R with our genetic data set, COANCESTRY programme parameters were set to account for inbreeding, with the number of reference individuals and bootstrapping samples set to 100.

Reference genomes
A reference genome for kakī has already been assembled (Galla et al., 2019) and was used in this study. To assemble a de novo reference genome for kākāriki karaka, a paired-end library was prepared at the FastQC v. 0.11.8 (Andrews, 2010) was used to evaluate the quality of the raw Illumina data and assess potential sample contamination. Initial read trimming was performed using TrimGalore v. 0.6.2 (Krueger, 2019) and Cutadapt v. 2.1 (Martin, 2011) with an end trim quality of 30, a minimum length of 54 and using the --nextseq two-colour chemistry option. A median Phred score of 20 was also used for initial read trimming to remove obvious data errors; it should be noted that the assembly programmes used here (i.e., Meraculous-2D to assembly to assess heterozygosity and contamination. Two genome assembly programmes were tested for assembly performance: Meraculous-2D v. 2.2.5.1 (Chapman et al., 2011) and MaSuRCA v.
3.2.9 (Zimin et al., 2013). Meraculous was run using trimmed reads in diploid mode 1, with all other assembly parameters set to default.
MaSuRCA was run using untrimmed reads, as it incorporates its own error correction pipeline. MaSuRCA parameter adjustments include a grid batch size of 300,000,000, the longest read coverage of 30, a Jellyfish hash size of 14,000,000,000 and the inclusion of scaffold gap closing; all other parameters were set to default for nonbacterial Illumina assemblies. The final assembly using the Meraculous pipeline was more fragmented (i.e., an N50 of 28.5 kb with 67,046 scaffolds > 1 kb), while the MaSuRCA genome was less fragmented (i.e., an N50 of 107.4 kb with 66,212 scaffolds > 1 kb) but contained possible artefacts due to heterozygosity (i.e., tandem repeats flanking short stretches of "N"s). To correct for these issues, the Meraculous assembly was first aligned to the MaSuRCA assembly using Last v.
959 (Kielbasa, Wan, Sato, Horton, & Frith, 2011); then, alignments were filtered to find matches where the Meraculous assembly spans the entirety of the tandem repeat in the MaSuRCA scaffolds, but lacking the tandem repeat or stretch of "N"s (i.e., gaps). In those cases, the aligned sequence in the MaSuRCA scaffold was replaced with the Meraculous match. All compute requirements needed to assemble the kākāriki karaka genome are available in Supporting Information.

Whole-genome resequencing
Kakī resequencing libraries were prepared at IKMB using a TruSeqⓇ Nano DNA Library Prep Kit following the manufacturer's protocol and were sequenced across 34 lanes of an Illumina HiSeq 4,000. FastQC v. 0.11.4 and 0.11.8 (Andrews, 2010) were used to evaluate the quality of the raw Illumina data for kakī and kākāriki karaka, respectively. Kakī resequencing reads were subsequently trimmed for the Illumina barcode, a minimum Phred quality score of 20 and a minimum length of 50 bp using Trimmomatic v. 0.38 (Bolger, Lohse, & Usadel, 2014). Because kākāriki karaka libraries were produced using different library preparation protocols and nextera chemistry, reads were trimmed using TrimGalore v. 0.6.2 (Krueger, 2019) for nextera barcodes and two-colour chemistry, using a median Phred score of 20, end trim quality of 30 and a minimum length of 54. Prior to mapping, the kakī reference genome was concatenated to a single chromosome using the custom perl script "concatenate_genome.
pl" (Moraga, 2018a) for use in an aligned project that used both resequencing and genotyping-by-sequencing reads (see Galla et al., 2019). The kakī and kākāriki karaka reference genomes were indexed, and resequencing reads were mapped using Bowtie2 v. 2.2.6 and v. 2.3.4.1 (Langmead & Salzberg, 2012), respectively, with the setting --very-sensitive. Resulting SAM files were converted to BAM and were sorted using Samtools v. 1.9 (Li et al., 2009). Read coverage and variant calling were performed using mpileup in BCFtools v.
1.9 (Li et al., 2009). The custom perl script "split_bamfile_tasks.pl" (Moraga, 2018b) was used to reduce the computational time needed for mpileup by increasing parallelization. SNPs were detected, filtered and reported using BCFtools. Filtering settings were set to retain biallelic SNPs with a minor allele frequency (MAF) greater than 0.05, a quality score greater than 20 and a maximum of 10% missing data per site. After a series of filtering trials for each species (see Supporting Information for details), depth for kakī was set to have an average mean depth greater than 10, while kākāriki karaka depth was set so that each site had a minimum depth of 5 and a maximum depth of 200. Resulting SNPs for both data sets were pruned for linkage disequilibrium using BCFtools with the r 2 set to 0.6 and a window size of 1,000 sites. Sites were not filtered for Hardy-Weinberg equilibrium, as the nature of these data sets (mostly family groups) violates the assumptions of random mating. Per site missingness, depth and diversity-including proportion of observed and expected heterozygous SNP sites per individual, nucleotide diversity and SNP density per kb-were evaluated in the final sets using VCFtools v.

SNP-based relatedness
To produce estimates of R using whole-genome SNPs, the programme KGD (Dodds et al., 2015)  We evaluated the scaled KGD approach with other maker-based relatedness estimators, including the triadic likelihood approach (Wang, 2007), the KING estimator (Waples, Albrechtsen, & Moltke, 2019) and the r xy method (Hedrick & Lacy, 2015), using parent-offspring relatedness as a benchmark for precision. We found that the

| Pairing recommendations
We used two complementary methods in PMx v. 1.6.20190628  to determine whether pairing recommendations change using pedigree-, microsatellite-and SNP-based approaches for estimating R. First, we used mate suitability index (MSI), which scores how valuable offspring of a potential pair would be by taking into account four factors: deltaGD (i.e., the net positive or negative genetic diversity provided to the population), the difference of mean kinship values of the pair, the inbreeding coefficients of resulting offspring and the extent of unknown ancestry (Ballou, Earnhardt, & Thompson, 2001;Lacy et al. 2012). MSI scores scale from 1 to 6, with 1 being "very beneficial," and 6 being "very detrimental." An additional category denoted with a "-" indicates "very highly detrimental" pairings. Here, we assign this category with a numerical MSI score of 7. MSI scores provide a standardized approach for comparing pairing recommendations within and among species, including those based on the three approaches used in this study. However, Ballou et al. (2001) recommend caution when using automated pairing recommendations such as MSI in small and inbred populations.
Thus, we also used mean kinship (MK) rank, which is an approach known to perform well in small and inbred populations (Ballou & Lacy, 1995;Rudnick & Lacy, 2008

| Pedigree-based relatedness
This study has produced the first multigenerational pedigrees for two critically endangered endemic birds from Aotearoa. The kakī pedigree includes 2,481 wild and captive individuals recorded from 1977 to present, with a pedigree depth ranging from 1 to 8 generations (3.35 average). The number of founders and founder genome equivalents in the kakī pedigree (94 and 12.9, respectively) is high relative to the kākāriki karaka pedigree (16 and 12, respectively), and the % known ancestry is lower (55% and 100%, respectively; 0.02), with averaged full sibling R of 0.52 ± SD 0.02 (Figure 2). The kākāriki karaka pedigree includes 624 captive individuals from 2003 to present, with an a pedigree depth ranging from 1 to 5 generations (2.48 average). Pedigree-based R for the 36 focal kākāriki karaka ranged from 0 to 0.67, with an average R of 0.19 ± SD 0.18.
Average R between all parent-offspring was 0.52 ± SD 0.03, with averaged full sibling R being 0.51 ± SD 0.02 (Figure 2).

| Microsatellite diversity and relatedness
All eight microsatellite markers for kakī successfully amplified in all individuals used in this study. The number of alleles present across kakī loci ranged from 2 to 4 (average 3.13 ± SD 0.64; Table 3), with overall fewer alleles found here than reported in previous studies utilizing these loci with more individuals (Hagen, Hale, Maloney, & Steeves, 2011;Steeves et al., 2010). While 18 microsatellite markers were amplified in kākāriki karaka, one was removed from this study for not successfully amplifying in more than 50% of individuals (locus OFK56) and nine were removed for being monomorphic (Table 3).
The number of alleles among polymorphic loci ranged from 2 to 4 (average 3.0 ± SD 0.93), with overall fewer alleles found here than reported in previous studies (Andrews, 2013;Andrews et al., 2013).  Table 3).
Microsatellite-based R between all kakī used in this study ranged from 0 to 0.85, with an average R of 0.16 ± SD 0.19. Average R between all known kakī parent-offspring (0.44 ± SD 0.13) was below the minimum expected relatedness value of 0.5, with a larger standard deviation of R values compared to pedigree-based estimates.
Averaged microsatellite-based full sibling R (0.41 ± SD 0.20) also had a larger deviation around the mean compared to the parent-offspring estimates (Figure 2).
Microsatellite-based R between all kākāriki karaka used in this study ranged from 0 to 0.84, with an average R of 0.18 ± SD 0.22. Similar to kakī, average R between all known kākāriki karaka parent-offspring relationships (0.47 ± SD 0.19) was below the minimum expected R value of 0.5, with a larger standard deviation of R values compared to pedigree-based estimates. Averaged full sibling R (0.49 ± SD 0.21) also had a larger deviation around the mean compared to microsatellite-based parent-offspring estimates ( Figure 2).

| Kākāriki karaka reference genome assembly
Reference genome library preparation and Illumina NovaSeq™ sequencing resulted in 584.47 million total reads for the kākāriki karaka genome. The final kākāriki karaka genome assembly was 1.15 GB in length, which is within the range of most assembled avian genomes (e.g., Zhang et al., 2014

| SNP discovery and diversity
Library preparation and Illumina sequencing resulted in 6.07 billion total reads for kakī (168.69 ± SD 65.32 million reads). In addition to the individual used for the reference assembly, 3.64 billion total reads (average = 103.92 ± SD 29.76 million reads) were produced for the additional 35 kākāriki karaka in this study. More SNPs were discovered during initial SNP discovery using kākāriki karaka than kakī, and more remained postfiltering (Table 4). These filtered SNPs were used for all downstream analyses. Average missingness was low for both data sets (Table 4), but lower for kākāriki karaka than kakī, as kākāriki karaka had a hard minimum cut-off for depth during filtering that resulted in no missing data. Average depth for both data sets was relatively high (Table 4), with kakī having slightly higher average depth. Average diversity statistics (nucleotide diversity, and the average observed and expected SNP heterozygosity per individual postfiltering) were similar in both species, with diversity in kakī being slightly higher. SNP density using the kakī data set was higher than the kākāriki karaka data set (Table 4).

| SNP-based relatedness
SNP-based R between all kakī used in this study ranged from 0.13 to 0.61, with an average R of 0.27 ± SD 0.09. Similar to pedigree-based estimates, average R between all known kakī parent-offspring was slightly higher than the expected relatedness value of 0.5 with a small standard deviation relative to microsatellite-based estimates (0.54 ± SD 0.03). Averaged full sibling R also had a larger deviation around the mean (0.52 ± SD 0.05) than parent-offspring relationships (Figure 2).

SNP-based R between all kākāriki karaka used in this study
ranged from 0.08 to 0.67, with an average R of 0.30 ± SD 0.12.
Similar to pedigree-based estimates, average R between all known kākāriki karaka parent-offspring was slightly above the expected R value of 0.5 with a small standard deviation relative to genetic-based

| Comparison of relatedness estimates and pairing recommendations
All kakī and kākāriki karaka R estimates using pedigree-, microsatellite-and SNP-based approaches correlated with one another with high statistical significance (Mantel test, p < .001; Pearson's correlation, p < .001; Figure 3). Of all the approaches, the correlation coefficient between pedigree-and SNP-based approaches was markedly higher than between other approaches, indicating that they are the most concordant (Figure 3).  (Figure 4). While the distributions of MSI scores between each approach were different, each approach produced scores that correlated significantly with one another (Pearson's correlation, p < .01-0.001). Similar to correlations between R estimates, correlation coefficients between pedigree-and SNP-based MSI scores were highest (Pearson's r = 0.5, see Figure S1 for details). Of the 320 possible kakī pairings, 38% did not experience a change in MSI score value between pedigree-and SNP-based approaches; however, 20% of pairings experienced an MSI score change that was 3+ categories different. In 2% of pairings, pedigree-based MSI scores were categorized as a 1 (i.e., preferred pairing), while SNP-based MSI scores were categorized as a 7 (i.e., very highly detrimental). Correlations between MK ranks provided by the three approaches were significant between pedigree-and SNP-based approaches only (Pearson's r = 0.75, p < .001; see Figure S2 for details  Figure 4). Each approach also produced scores that correlated significantly with one another (Pearson's correlation, p < .001), with the highest correlation coefficients seen between pedigree-and SNP-based MSI scores (Pearson's r = 0.65, see Figure S1 for details).
Of the 324 possible pairings for kākāriki karaka, 59% did not experience a change in MSI score value between pedigree-and SNPbased approaches; however, 9% of pairings experienced an MSI score change that was 3+ categories different. In 2% of pairings, pedigree-based MSI scores were categorized as a 1 (i.e., very beneficial), while SNP-based MSI scores were categorized as a 7 (i.e., very highly detrimental). Correlations between MK rank were significant between pedigree-and SNP-based approaches (Pearson's r = 0.64, p < .001) and microsatellite-and SNP-based approaches (Pearson's r = 0.51, p = .002, see Figure S2 for details). Among pedigree-and SNP-based MK ranks, 53% of individuals experienced a minimal rank shift of 0-3 categories, 31% experienced a moderate rank shift of 4-7 categories, and 8% experienced a high rank shift of ≤8 categories.

| D ISCUSS I ON
This study is the first to compare pedigree-, microsatellite-and SNP-based estimates of relatedness and subsequent pairing

| Relatedness comparisons
When producing empirical estimates of relatedness, simulations were performed to choose the most suitable estimator for microsatellites, and various relatedness estimators and filtering schemes for SNPs were evaluated to find the approach that best approximated known parent-offspring relationships as a biologically relevant benchmark. While different relatedness estimators and filtering schemes will result in different point estimates of relatedness, this study demonstrates an approach for producing relatedness estimates that are well suited for our particular data set and research question.
Pedigree-based estimates of parent-offspring and full sibling relatedness approximated 0.5 for both kakī and kākāriki karaka ( Figure 2), with some measures being slightly higher, which likely reflects intergenerational inbreeding and/or higher pedigree depth from the baseline (reference) population. These results are consistent with expectations, as pedigrees are based on the probability of Mendelian inheritance, which postulates that first-order relationships (i.e., parents and offspring, and siblings) share 50% of their genomic information (Lacy, 1995;Wright, 1922). We expect realized (i.e., empirical) parent-offspring relationships to also approximate 0.5, but a broader range of realized relatedness estimates among full siblings, as they may receive different genetic material from each parent due to recombination and independent assortment during meiosis and random fertilization (Hill & Weir, 2011, 2012Speed & Balding, 2015). Even when pedigrees are robust, this study highlights an unavoidable shortcoming as pedigrees do not adequately capture true relatedness between full siblings. We anticipate this uncaptured diversity may prove useful for maximizing existing diversity, especially in conservation breeding programmes with relatively few founders (Ballou & Lacy, 1995).
Compared to the pedigree-based approach, our empirical data sets (i.e., microsatellites and SNPs) capture more variation between siblings than parents and offspring (Figure 2). A broad range of microsatellite-based relatedness estimates were observed in both parent-offspring and sibling relationships, compared to the SNPbased approach. In some instances, even parent-offspring pairings appeared relatively unrelated using microsatellites (e.g., minimum parent-offspring R = 0.14 in kakī and R = 0 in kākāriki karaka), which underscores the lack of precision in this approach and how it could inadvertently lead to poorly informed pairing recommendations.
These large ranges of relatedness values using microsatellites can be explained because genetic-based relatedness values between parent-offspring and full siblings are based on allele frequencies, and relatedness between individuals that share common alleles will be lower than individuals that share rare alleles (Speed & Balding, 2015;Wang, 2011). This bias in relatedness values can be exacerbated when samples sizes are small (Wang, 2017), which is typical of conservation breeding programmes. Furthermore, the lack of precision using microsatellites shown here is consistent with studies that suggest relatively few markers with low allelic diversity are insufficient for estimating relatedness and inbreeding, especially in genetically depauperate species (e.g., Attard et al., 2018;Escoda et al., 2017;Hellmann et al., 2016;Taylor, 2015;Taylor et al., 2015).
Compared to microsatellite-based relatedness, SNP-based relatedness showed a relatively small range with parent-offspring and full sibling relatedness estimates approximating 0.5, and full siblings showing a wider range of values than parent-offspring relationships ( Figure 2). Not only is this pattern consistent with expectations given the behaviour of chromosomes during meiosis and random fertilization, but it also shows more precision than the microsatellite data sets. Other researchers have found similar results in a diverse range of wild taxa, indicating that thousands of genome-wide SNPs show more precision than microsatellites when measuring relatedness and inbreeding (e.g., Attard et al., 2018;Hellmann et al., 2016;Hoffman et al., 2014;Lemopoulos et al., 2019;Thrasher, Butcher, Campagna, Webster, & Lovette, 2018).
Beyond parent-offspring and full sibling relationships, pedigree-and SNP-based relatedness estimates showed the highest concordance with one another among the three approaches used ( Figure 3). In kakī, the data sets used here include non-captivebred individuals with intensively monitored wild parents. These results provide more credibility to the semi-wild kakī pedigree, where socially monogamous wild pairs of kakī are assumed to be the genetic parents of offspring at nests (but see also Overbeek et al., 2017). Still, it should be noted that many pairs with pedigree-based relatedness values of 0 had SNP-based relatedness values ranging upwards of 0.40 in kakī 0.33 and in kākāriki karaka, which approximates first-and second-order relationships in both species (Figure 2). This indicates that pedigree-based R between these individuals may be downwardly biased by the assumption that no variance in relatedness exists among founders, missing information and/or low pedigree depth (Balloux et al., 2004;Bérénos et al., 2014;Hammerly et al., 2016;Hogg et al., 2019;Kardos et al., 2015;Lacy, 1995;Pemberton, 2008;Rudnick & Lacy, 2008;Tzika et al., 2009).

| Pairing recommendations
When these relatedness values are translated into pairing recommendations using MSI scores and MK rank, there is a high concordance between pedigree-and SNP-based approaches, with SNP-based MSI scores being significantly higher than pedigreeand microsatellite-based approaches. The latter result is somewhat expected, given that average relatedness estimates using SNPs was highest among the approaches used here, and empirical estimates of relatedness and inbreeding are usually higher than pedigrees as they more effectively capture relatedness between founders or misassigned individuals (Hammerly et al., 2016;Hogg et al., 2019). With that said, when making pairing recommendations using kinship-based pairing decisions (e.g., Ballou & Lacy, 1995), it is often the relative kinships between individuals that are more important than absolute values (Galla et al., 2019;McLennan, Wright, Belov, Hogg, & Grueber, 2019). This suggests that pedigree-and SNP-based approaches both yield similar results for pairing recommendations, with some important differences. For example, while correlation coefficients between these two sets of MSI scores are high relative to other comparisons, there are instances where pairings are considered "highly beneficial" (i.e., MSI category 1) when using the pedigree and "very highly detrimental" (i.e., MSI category 7) when using SNPs. When comparing MK ranks between pedigree-and SNP-based approaches, some kakī and kākāriki karaka experienced large shifts in rank (i.e., ≥8 positions difference) depending on the approach used. Although we expect some differences between pedigree-and SNP-based MSI scores and MK ranks, we attribute these very large differences to errors in the pedigree (e.g., Hammerly et al., 2016) or violations of the assumption that there is no variance in founder relationships (e.g., Hogg et al., 2019). Of all kakī and kākāriki karaka pairings that experienced a large shift between pedigree-and SNP-based MSI scores, most feature recurring individuals with wild parentage (i.e., founders), and in one recurring occasion, a wild individual (kakī) with high pedigree depth that likely represents an entry error in the pedigree.

| Management implications
Pedigree-, genetic-and genomic-based tools each have their advantages to inform conservation management. For example, pedigrees capture both genetic and demographic considerations dating back to the founding of the population, while empirical estimates of relatedness can circumvent pedigree errors and issues surrounding founder relationships by expressing realized relatedness between all sampled individuals. From the results shown here, we recommend that when conservation breeding programmes are poorly pedigreed (i.e., pedigrees of low depth or containing missing data), SNPs should be incorporated to provide a precise indicator of relatedness to genetically inform pairing decisions. The microsatellite panels used here have shown low precision in estimating relatedness, with demonstrated downstream effects for pairing recommendations compared to pedigree-and SNP-based approaches. More microsatellites could be developed to mitigate this shortcoming; however, other studies indicate that a larger number microsatellites (e.g., 20-40 markers) may not equate to higher precision for relatedness estimates and inbreeding coefficients, especially in threatened species with low allelic diversity (Nietlisbach et al., 2017;Robinson et al., 2013;Taylor, 2015;Taylor et al., 2015). Further, the time and cost associated with building larger microsatellite panels and generating microsatellite data will likely be surpassed by the production of genomewide SNPs, either by a whole-genome resequencing approach as shown in this study or by a reduced representation sequencing approaches (e.g., RAD sequencing, or genotyping-by-sequencing; Galla et al., 2016;Narum, Buerkle, Davey, Miller, & Hohenlohe, 2013). Currently, for kakī and kākāriki karaka, reduced representation sequencing is more cost-effective than whole-genome resequencing (i.e., approximately one-third of the price, depending on the genome size, as of 2019)-but we foresee more people shifting towards whole-genome resequencing in the near future, given the decreasing cost of high-throughput sequencing (Hayden, 2014) and the ability to ask more research questions using wholegenome resequencing data sets (see Future Directions below for details). This is particularly true for birds, whose genomes are small (e.g., 1.05-1.26G) relative to many vertebrates (Zhang et al., 2014).
We anticipate SNPs will be particularly applicable in circumstances when pedigrees are the least reliable. For instance, when the founders of a conservation breeding population have no ancestry data available and are likely to be related, SNP-based relatedness estimates between individuals can be used to avoid highly related matings . This situation may not only coincide with the original founding event of a captive population, but iteratively when individuals are sourced from wild or translocated populations to augment the captive population, as suggested in Frankham (2008) and Hogg et al. (2019). For example, in kākāriki karaka, whole-genome resequencing has been made available for all current breeding individuals in the conservation breeding programme, including individuals who are founders themselves. Because birds of unknown ancestry are being routinely sourced from highly endangered wild populations, and will also be founders, we anticipate the need for resequencing these birds as they are incorporated into the breeding programme to assess their relatedness to other individuals. In addition to traditional captive-bred populations (i.e., ex situ management), this approach is applicable to intensively managed wild populations (i.e., in situ management), where robust pedigrees are lacking, but conservation translocations can be informed by relatedness between individuals in a managed landscape (e.g., kākāpō, Strigops habroptilus, Elliott, Merton, & Jansen, 2001;scimitar-horned oryx, Oryx dammah, Wildt et al., 2019).
While we expect SNPs will be important for pairing recommendations moving forward, we do not expect they will eclipse well-established pedigrees, as both approaches have advantages for conservation breeding. Instead, we envision a combined approach where realized relatedness from SNPs can be used to augment data-rich pedigrees. With that said, there are relatively few studies to date that effectively combine existing pedigree data with genomic estimates of relatedness to inform pairing recommendations (but see Hogg et al., 2019;Ivy, Putnam, Navarro, Gurr, & Ryder, 2016)  , this approach requires caution, as the calculation of pedigree-based identity by descent for subsequent generations-including kinship and gene diversity-will be affected by the addition of empirical data . We acknowledge this approach requires further investigation and validation, particularly for species that receive periodic influx of wild individuals of unknown ancestry in their conservation breeding programme.

| Future directions and concluding remarks
This study has produced pedigrees and whole-genome sequences for two critically endangered species. Beyond estimating relatedness, these tools provide an exciting opportunity to explore other questions relevant to conservation, such as characterizing the genomic basis of fitness traits, including those associated with inbreeding depression (Kardos, Taylor, Ellegren, Luikart, & Allendorf, 2016; but see also Kardos & Shafer, 2019) or adaptation to captivity (e.g., Grueber et al., 2017). We also envision using the genomic resources developed here to further investigate best practice for making pairing recommendations, for example, agent-based, multigenerational simulations can be used to evaluate whether genome-wide diversity is best maximized using pedigrees, SNPs or a combination approach.
Given that SNPs have been successfully used to estimate relatedness for different purposes across a wide diversity of taxonomic groups outside of this study (as reviewed in Attard et al., 2018), we anticipate a SNP-based approach for estimating relatedness and making subsequent pairing recommendations will be applicable beyond birds. In the meantime, for poorly pedigreed populations worldwide, we recommend a SNP-based approach to estimate relatedness for subsequent pairing recommendations. It should be noted that many approaches used to date have used de novo reduced representation approaches (e.g., genotyping-by-sequencing, RADseq; Narum et al., 2013) for SNP discovery, which typically have more missing data, lower depth and fewer SNPs than the reference-guided whole-genome resequencing approach used here. While these factors may contribute to bias in relatedness estimates (but see Dodds et al., 2015), research still indicates that fewer SNPs, with more missing data and lower depth, than those presented here provide more precision than microsatellites (Attard et al., 2018). We expect reduced representation approaches will persist in the short term, especially for species with large and complex genomes (e.g., some fish, amphibians and invertebrates) that otherwise cannot yet be affordably resequenced across entire conservation breeding programmes. With that said, we also expect whole-genome resequencing projects like ours will gain momentum in the years to come, as these data can be better leveraged to address multiple questions related to conservation genetic management (Harrisson, Pavlova, Telonis-Scott, & Sunnucks, 2014; see also above). In the meantime, we look forward to seeing more poorly pedigreed conservation breeding programmes for taxonomically diverse species from around the world incorporate SNPs for estimating relatedness to inform pairing decisions.

ACK N OWLED G EM ENTS
We are grateful for the support of Te Rūnanga o Ngāi Tahu Endeavour Fund (UOCX1602 awarded to TES), the Brian Mason Scientific and Technical Trust (awarded to SJG and TES), and the Mohua Charitable Trust (awarded to TES). We thank the three anonymous reviewers for their thoughtful and constructive feedback to this manuscript.

CO N FLI C T O F I NTE R E S T
None declared.

DATA AVA I L A B I L I T Y S TAT E M E N T
Genomic data provided in this manuscript are available through a password protected server on the Conservation, Systematics and Evolution Research Team's website (http://www.uccon sert. org/data/). Kakī and kākāriki karaka are taonga (treasured) species. For Māori (the indigenous people of Aotearoa), all genomic data obtained from taonga species have whakapapa (genealogy that includes people, plants and animals, mountains, rivers and winds) and are therefore taonga in their own right. Thus, these data are tapu (sacred) and tikanga (customary practices, protocols and ethics) determine how people interact with it. To this end, the passwords for the genomic data in this manuscript will be made