Experimental validation of in silico predicted RAD locus frequencies using genomic resources and short read data from a model marine mammal

BACKGROUND

Restriction site-associated DNA sequencing (RADseq) has revolutionized the study of wild organisms by allowing cost-effective genotyping of thousands of loci. However, for species lacking reference genomes, it can be challenging to select the restriction enzyme that offers the best balance between the number of obtained RAD loci and depth of coverage, which is crucial for a successful outcome. To address this issue, PredRAD was recently developed, which uses probabilistic models to predict restriction site frequencies from a transcriptome assembly or other sequence resource based on either GC content or mono-, di- or trinucleotide composition. This program generates predictions that are broadly consistent with estimates of the true number of restriction sites obtained through in silico digestion of available reference genome assemblies. However, in practice the actual number of loci obtained could potentially differ as incomplete enzymatic digestion or patchy sequence coverage across the genome might lead to some loci not being represented in a RAD dataset, while erroneous assembly could potentially inflate the number of loci. To investigate this, we used genome and transcriptome assemblies together with RADseq data from the Antarctic fur seal (Arctocephalus gazella) to compare PredRAD predictions with empirical estimates of the number of loci obtained via in silico digestion and from de novo assemblies.

RESULTS

PredRAD yielded consistently higher predicted numbers of restriction sites for the transcriptome assembly relative to the genome assembly. The trinucleotide and dinucleotide models also predicted higher frequencies than the mononucleotide or GC content models. Overall, the dinucleotide and trinucleotide models applied to the transcriptome and the genome assemblies respectively generated predictions that were closest to the number of restriction sites estimated by in silico digestion. Furthermore, the number of de novo assembled RAD loci mapping to restriction sites was similar to the expectation based on in silico digestion.

CONCLUSIONS

Our study reveals generally high concordance between PredRAD predictions and empirical estimates of the number of RAD loci. This further supports the utility of PredRAD, while also suggesting that it may be feasible to sequence and assemble the majority of RAD loci present in an organism's genome.

Predicting RAD-seq Marker Numbers across the Eukaryotic Tree of Life

High-throughput sequencing of reduced representation libraries obtained through digestion with restriction enzymes--generically known as restriction site associated DNA sequencing (RAD-seq)--is a common strategy to generate genome-wide genotypic and sequence data from eukaryotes. A critical design element of any RAD-seq study is knowledge of the approximate number of genetic markers that can be obtained for a taxon using different restriction enzymes, as this number determines the scope of a project, and ultimately defines its success. This number can only be directly determined if a reference genome sequence is available, or it can be estimated if the genome size and restriction recognition sequence probabilities are known. However, both scenarios are uncommon for nonmodel species. Here, we performed systematic in silico surveys of recognition sequences, for diverse and commonly used type II restriction enzymes across the eukaryotic tree of life. Our observations reveal that recognition sequence frequencies for a given restriction enzyme are strikingly variable among broad eukaryotic taxonomic groups, being largely determined by phylogenetic relatedness. We demonstrate that genome sizes can be predicted from cleavage frequency data obtained with restriction enzymes targeting "neutral" elements. Models based on genomic compositions are also effective tools to accurately calculate probabilities of recognition sequences across taxa, and can be applied to species for which reduced representation data are available (including transcriptomes and neutral RAD-seq data sets). The analytical pipeline developed in this study, PredRAD (https://github.com/phrh/PredRAD), and the resulting databases constitute valuable resources that will help guide the design of any study using RAD-seq or related methods.