Estimating divergence dates and evaluating dating methods using phylogenomic and mitochondrial data in squamate reptiles

Abstract

Recently, phylogenetics has expanded to routinely include estimation of clade ages in addition to their relationships. Various dating methods have been used, but their relative performance remains understudied. Here, we generate and assemble an extensive phylogenomic data set for squamate reptiles (lizards and snakes) and evaluate two widely used dating methods, penalized likelihood in r8s (r8s-PL) and Bayesian estimation with uncorrelated relaxed rates among lineages (BEAST). We obtained sequence data from 25 nuclear loci (∼500–1000bp per gene; 19,020bp total) for 64 squamate species and nine outgroup taxa, estimated the phylogeny, and estimated divergence dates using 14 fossil calibrations. We then evaluated how well each method approximated these dates using random subsets of the nuclear loci (2, 5, 10, 15, and 20; replicated 10 times each), and using ∼1kb of the mitochondrial ND2 gene. We find that estimates from r8s-PL based on 2, 5, or 10 loci can differ considerably from those based on 25 loci (mean absolute value of differences between 2-locus and 25-locus estimates were 9.0Myr). Estimates from BEAST are somewhat more consistent given limited sampling of loci (mean absolute value of differences between 2 and 25-locus estimates were 5.0Myr). Most strikingly, age estimates using r8s-PL for ND2 were ∼68–82Myr older (mean=73.1) than those using 25 nuclear loci with r8s-PL. These results show that dates from r8s-PL with a limited number of loci (and especially mitochondrial data) can differ considerably from estimates derived from a large number of nuclear loci, whereas estimates from BEAST derived from fewer nuclear loci or mitochondrial data alone can be surprisingly similar to those from many nuclear loci. However, estimates from BEAST using relatively few loci and mitochondrial data could still show substantial deviations from the full data set (>50Myr), suggesting the benefits of sampling many nuclear loci. Finally, we found that confidence intervals on ages from BEAST were not significantly different when sampling 2 vs. 25 loci, suggesting that adding loci decreased errors but did not increase confidence in those estimates.