An empirical pipeline for choosing the optimal clustering threshold in RADseq studies

Abstract

Genomic data are increasingly used for high resolution population genetic studies including those at the forefront of biological conservation. A key methodological challenge is determining sequence similarity clustering thresholds for RADseq data when no reference genome is available. These thresholds define the maximum permitted divergence among allelic variants and the minimum divergence among putative paralogues and are central to downstream population genomic analyses. Here we develop a novel set of metrics to determine sequence similarity thresholds that maximize the correct separation of paralogous regions and minimize oversplitting naturally occurring allelic variation within loci. These metrics empirically identify the threshold value at which true alleles at opposite ends of several major axes of genetic variation begin to incorrectly separate into distinct clusters, allowing researchers to choose thresholds just below this value. We test our approach on a recently published data set for the protected foothill yellow-legged frog (Rana boylii). The metrics recover a consistent pattern of roughly 96% similarity as a threshold above which genetic divergence and data missingness become increasingly correlated. We provide scripts for assessing different clustering thresholds and discuss how this approach can be applied across a wide range of empirical data sets.