Utility of allozyme supernatant and formalin-fixed samples for phylogenomic and population genomic analyses
Formalin-fixed samples clustered with their corresponding specimen replicates in the phylogeny, despite having above-average levels of missing data (Fig. 4; S6). This suggests that these samples are amenable to phylogenetic analyses. By contrast, formalin-fixed samples did not cluster with corresponding specimen replicates in PC space, likely due to higher levels of missing data. Smith et al. (2020) found that missing data at informative sites biased phylogenomic estimation when using historical samples. Similar unevenness in missing data may partially explain some of the separation between the formalin-fixed and supernatant/pellet replicates in PC space in our study. In particular, when we pruned our data to only capture-based replicates and retained loci shared by 90% of samples (713 SNPs, n=25), the amount of missing data seems to load heavily on PC2 (12%), with the greatest separation between the formalin-fixed samples with the largest amounts of missing data and all other replicates. The two formalin-fixed samples with< 20% missing data clustered with their corresponding replicates, further indicating that missing data is likely driving this pattern. Rates of allelic mismatch (13–28%) could also explain some of the variation in PC space between capture-based replicates, though Ewart et al. (2019) found that simulated rates of allelic mismatch up to 10% did not impact PCA analyses in their study. Likewise, Stronen et al. (2018) report a similar pattern between historical and contemporary sample genotypes using Illumina BeadChips, where samples cluster along PC1 by sample-type (historical vs. contemporary), and along PC2 by geography, suggesting that this pattern may be common in studies using historical samples and not restricted to only target-capture approaches (but see van der Valk et al. 2019).
Finally, we inferred significantly different estimates of nucleotide diversity in formalin-fixed replicates compared with the supernatant and pellet replicates, although removing sites with high levels of missing data accounted for some, but not all, of this discrepancy. A similar pattern was reported by Ewart et al. (2019), who found a bias towards homozygosity in historical samples leading to lower estimates of genetic diversity, and a bias in missing data towards heterozygous sites, which may partially explain why removing sites with more missingness gave higher estimates of genetic diversity for formalin-fixed samples. We inferred a non-significant difference in heterozygosity among the capture-based replicates, suggesting that either allelic mismatch or biases in the types of sites that are missing may best explain variation in nucleotide diversity estimates. When sample sizes are large enough, estimating genetic diversity using only loci and individuals with low missingness will likely reduce these biases, although the appropriate threshold will likely vary by study. Based on the systematic biases we report here, we recommend future studies make thoughtful choices about data assembly and SNP filtering when using formalin-fixed samples for demographic parameter estimation.