Harnessing the power of RADseq for ecological and evolutionary genomics

Abstract

A revolution is occurring in ecological and evolutionary genetics, driven by the development of techniques such as Restriction-site-Associated DNA sequencing (RADseq) that allow relatively low-cost discovery and genotyping of thousands of genetic markers for any species, including non-model species. Here we provide an overview of the diverse RADseq techniques that have been developed and highlight some of the research questions these powerful methods can be used to answer. We discuss how technical differences among the many variant methods lead to trade-offs in experimental design and analysis, and describe general considerations for designing a RADseq study.

Introduction

The development of Restriction site-Associated DNA Sequencing (RADseq) was deemed among the most significant scientific breakthroughs within the last decade¹. RADseq has fueled studies in ecological, evolutionary, and conservation genomics by harnessing the massive throughput of next-generation sequencing to uncover hundreds or thousands of polymorphic genetic markers across the genome in a single, simple, and cost-effective experiment^2,3. Like other reduced-representation sequencing approaches, RADseq targets a subset of the genome, thus providing advantages over whole-genome sequencing including greater depth of coverage per locus (and therefore improved confidence in genotype calls) and sequencing of greater numbers of samples for a given budget. Unlike many other methods for generating genome-wide data, RADseq does not require any prior genomic information for the taxa being studied. Consequently, RADseq has become the most widely used genomic approach for high-throughput SNP discovery and genotyping in ecological and evolutionary studies of non-model organisms.

The widespread adoption of RADseq has sparked further innovation of the core RADseq technique (BOX 1). Numerous variations promise to increase flexibility (e.g., in the number of loci assayed) and decrease cost and effort in ecological and evolutionary genomics studies. However, technical differences among the methods lead to important considerations for all steps of genomic studies, from costs of library preparation and sequencing, to the types of bias and error inherent in the resulting data, to the types of scientific questions that can be addressed. A comprehensive review of RADseq methods is thus critically needed to aid researchers in choosing an approach and avoiding erroneous scientific conclusions from RADseq data, a problem that has plagued other new marker types in the past^4–6.

Box 1. Common RADseq-related techniques.

I. Sequence adjacent to single restriction enzyme cut sites

Original RAD^10,64 digests genomic DNA with one restriction enzyme, followed by mechanical shearing to reduce fragments to the appropriate length for sequencing, which (unlike other methods) creates variance in fragment sizes at each locus.

2bRAD^15,65 uses type IIB restriction enzymes, which cleave DNA upstream and downstream of the recognition site, resulting in short fragments of uniform length (33–36bp).

II. Sequence fragments flanked by two restriction enzyme cut sites

a. Single enzyme, indirect size selection

Genotyping by Sequencing (GBS)⁶⁶ uses a common-cutter enzyme and PCR preferentially amplifies short fragments.

Sequence-based Genotyping (SBG)⁶⁷ uses a rare-cutter and one or two common-cutters and PCR preferentially amplifies short fragments

b. Double enzyme, indirect size selection

Complexity Reduction of Polymorphic Sequences (CRoPS)⁶⁸ uses two enzymes and a proprietary library preparation kit (originally developed for 454 pyrosequencing).

c. Single enzyme, direct size selection

Reduced Representation Libraries (RRL)^11,69 are unique in using a blunt-end common-cutter enzyme, followed by a size selection step and a proprietary Illumina library preparation kit.

Multiplexed shotgun genotyping (MSG)⁵⁵ uses one common-cutter enzyme and a size selection step.

ezRAD¹⁶ uses one or more common-cutter enzymes, and a proprietary kit for Illumina library preparation.

d. Double enzyme, direct size selection

Double-digest RAD (ddRAD)¹⁴ uses two restriction enzymes, with adaptors specific to each enzyme, and size selection by automated gel cut.

Variations on the above techniques include using methylation-sensitive enzymes⁷⁰; adding more restriction enzymes to existing protocols to further reduce the set of loci^67,71; adding a second digestion to eliminate adaptor dimers¹⁸; adapting RADseq techniques to other sequencing platforms such as Ion Torrent^71–73; and other minor technical modifications^56,74.

Figure within BOX 1.

Numbers of articles citing the original papers describing each RADseq protocol over time. Data for 2015 are extrapolated using numbers of articles cited from January through September 2015. Protocols are arranged by order of first appearance in the literature. Data generated using Web of Science.

The core feature of RADseq techniques is the use of restriction enzymes to obtain DNA sequence at a genome-wide set of loci. Restriction enzymes have long been utilized to sample loci across the genome and generate information on population-level variation^7,8, including genome-wide surveys for genetic variation in humans⁹. Whereas these previous techniques focused on polymorphisms within restriction cut sites or used Sanger sequencing, RADseq uses next-generation sequencing to generate sequence data adjacent to a large number of restriction cut sites^10–12. RADseq loci can occur in all areas of the genome (i.e. coding and non-coding regions), and individuals within or between closely related species generally share most loci due to conservation of cut sites. All RADseq methods are broadly applicable across a wide range of taxa and scientific questions (BOX 2). However, some techniques have been used more widely in certain systems, largely due to historical contingencies rather than relative suitability of the various approaches to different species (e.g. CRoPS, GBS, and RRL have been used primarily in agricultural species¹³).

Box 2. Ecological and evolutionary insights from RADseq data.

Exemplar studies illustrate ecological, evolutionary, and conservation related questions that can be answered using RADseq.

Inbreeding and genomic diversity

A study investigating heterozygosity-fitness correlations in seals found that genome-wide heterozygosity estimated using RADseq had a nearly fivefold higher correlation with fitness than did 27 microsatellite loci⁵². RADseq genomic diversity estimates were also used to characterize the influence of social structure on autosome vs. sex chromosome diversity in Tonkean macaque monkeys⁷⁵.

Effective population size (N_e)

Thousands of SNPs generated using RADseq were used to estimate N_e in salmon and smelt from western North America^76,77.

Population structure, phylogeography, and conservation units

RADseq was used to develop a population-informative SNP panel to monitor stock composition in salmon and delineate population units to harvest as discrete rather than mixed stocks^76,78; see also Emerson et al. 2010⁷⁹, Combosch & Vollmer 2015⁸⁰, Gaither et al. 2015⁸¹.

Introgression

Hohenlohe et al.²⁰ used RADseq to identify 3180 species-diagnostic SNPs and calculate admixture between a native and an invasive trout species; see also Eaton & Ree 2013⁸², Ford et al. 2015⁸³.

Genomics of adaptation

A study using GWAS and F_ST outlier tests of RADseq data from two butterfly species in four parallel hybrid zones found that genomic regions harboring genes controlling color pattern were the most divergent between the two species⁴⁹; see also Hohenlohe et al. 2010²⁵, Chutimanitsakun et al. 2011⁸⁴, Ruegg et al. 2014⁵⁰.

Phylogenomics

RADseq data generated a highly resolved tree for 16 species of Lake Victoria cichlid fish, whereas previous analyses using AFLP, microsatellites, or a handful of sequence-based markers failed to resolve species level relationships for these species⁸⁵.

Note that we use the term RADseq here to refer to any of several related methods that rely on restriction enzymes to determine the set of loci that will be sequenced in a reduced-representation library. The term “RAD” was originally used to describe one particular method¹⁰, but has subsequently been used to describe a range of methods^14–16. “Genotyping-by-sequencing” (GBS) has also been used to describe these methods¹⁷, but we use the term “RADseq” here because the methods we review are united by their use of restriction enzymes, which is captured in the RADseq acronym. Although we focus primarily on RADseq applications to ecological and evolutionary genetics in natural populations (BOX 2), much of our discussion is also relevant to other RADseq applications, such as trait-mapping in agricultural species¹³.

The RADseq family of methods

RADseq techniques share several basic steps (FIG. 1). All methods start with relatively high molecular weight genomic DNA¹⁸ and begin by digesting it with one or more restriction enzymes. All methods add specific sequencing adapters, or double-stranded oligonucleotides, that are required by all next-generation sequencing platforms. Adapters added during RADseq protocols may contain barcodes – short unique sequences generally 6–12bp – that are used to identify individual samples that are sequenced together (multiplexed) in a single library. Depending on the enzyme(s) used, RADseq protocols also reduce and/or select sizes of DNA fragments optimal for next-generation sequencing.

Figure 1.

Step-by-step illustration of five RADseq library prep protocols. All protocols begin by digesting relatively high quality genomic DNA with one or more restriction enzymes. For most protocols, the sequencing adapters (oligos) are added in two stages, with one set of oligos added during a ligation step early in the protocol, and a second set of oligos incorporated during a final PCR step. The second set of oligos extends the length of the total fragment to produce the entire Illumina adapter sequences. In contrast, the original RADseq adds adapters in three stages. For Illumina sequencing, the adapters on either end of each DNA fragment must differ, and therefore some protocols (e.g. original RADseq, ddRAD, ezRAD) use “Y-adapters” that are structured to ensure that only fragments with different adapters on either end are PCR-amplified (illustrated here as Y-shaped adapters). Other protocols (e.g. GBS) simply rely on the fact that fragments without the correct adaptors will not be sequenced. To generate fragments of an ideal length for sequencing, most methods use common-cutter enzymes (e.g. 4–6bp cutters) to generate a wide range of fragment sizes, followed by a direct size selection (gel-cutting or magnetic beads, e.g. ddRAD, ezRAD) or an indirect size selection (as a consequence of PCR amplification or sequencing efficiency, e.g. GBS). In contrast, the original RADseq uses a mechanical shearing step to produce fragments of an appropriate size, and incorporates a size selection step only to increase Illumina sequencing efficiency and remove adapter dimers. 2bRAD uses IIB restriction enzymes to produce small fragments of equal size across all loci (33–36bp).

RADseq methods differ in the order and details of enzyme digestion, adapter ligation, barcoding, and size selection, as well as type of sequence data that can be produced at each locus. These differences can be used to place techniques into major groups (BOX 1). Below we discuss important variations among methods at each step and some of the consequences for library preparation, the resulting data, and subsequent bioinformatic analyses.

Starting genomic DNA

RADseq techniques have been optimized based on starting material comprised of high molecular weight genomic DNA, and thus these techniques may perform poorly with highly degraded genomic DNA¹⁸. For example, in methods without enzyme-specific adaptors (e.g., ezRAD, CRoPS), smaller fragments of starting genomic DNA not adjacent to cut sites may end up in the sequencing library, thus wasting sequencing effort on non-RAD loci. The original RADseq technique¹⁰ also requires higher-molecular-weight DNA than other methods, because the mechanical shearing step is most consistent and efficient with relatively large fragments remaining after enzyme digestion (see below).

In general, more starting DNA is often beneficial, as it may reduce the number of PCR cycles required and thus reduce the problem of PCR duplicates (see below). Some of the protocols originally recommended fairly large amounts of DNA (up to 1 μg per sample for original RAD¹⁹ or 5.5 μg for RRL¹¹); however, most RADseq methods are somewhat flexible in the total amount of DNA required per sample, and can often be implemented with 50–100 ng of DNA per sample. One exception would be when using a PCR-free library preparation method that requires larger amounts of starting DNA, as in one implementation of ezRAD¹⁶.

Restriction enzyme digestion

RADseq protocols differ in the number of restriction enzymes used and the frequency with which these enzymes cut the genome, with “common-cutters” defined as restriction enzymes that cut more frequently than “rare-cutters,” generally a result of the length of their recognition sequence. Techniques also fall into two major groups depending on how the set of loci sequenced relates to the distribution of enzyme cut sites across the genome. The original RADseq protocol and 2bRAD aim to produce sequence data at all cut sites for the restriction enzyme. In contrast, all other techniques depend on sequencing of genomic fragments produced by two enzyme cut sites separated by a specified genomic distance (typically 300–600bp apart, with the distance determined by direct or indirect size selection; see below). These cut sites may be from the same or different enzymes, depending on whether the method uses one or two enzymes. For each method, common-cutter or rare-cutter enzymes can be used to tailor the number of loci produced. For example, for the original RADseq protocol, a very rough estimate is that an 8-cutter will cut every 4ˆ8 = 65,536bp, while a 6-cutter will cut every 4ˆ6 = 4,096bp; this calculation can be adjusted to account for the GC content of the recognition sequence and the genome under study (Davey et al. 2011).

Adapter ligation

RADseq techniques differ in how adapters are constructed and ligated to DNA fragments, and also how they are designed to ensure that only the target genomic DNA fragments (i.e. those adjacent to restriction cut sites) are sequenced. In some cases, adapters are designed to ligate only at the characteristic single-stranded sticky end that is left at restriction cut sites after digestion. Many Illumina sequencing-based RADseq protocols also use “Y-adapters” that are structured to ensure that only fragments with the adapter combinations required for sequencing are PCR amplified (FIG 1). Some techniques adopt proprietary library preparation kits for adapter ligation (e.g. ezRAD, CRoPS, RRL), which may increase the reliability as well as the cost of reagents for library construction. Using adapters from proprietary kits can also lead to lower specificity in ligation, because these adapters do not ligate to the sticky ends, and therefore sequence data could be generated from fragments of degraded DNA not adjacent to restriction cut sites¹⁶.

Size selection

For most protocols, the restriction digest reduces genomic DNA to a wide range of fragment lengths, and then a size selection step is used to isolate fragments of ideal lengths for sequencing. This leads to key distinctions among RADseq protocols (BOX 1): for all the methods that sequence DNA fragments flanked by two cut sites, the set of loci to be genotyped is further reduced by this size selection, because each potential locus has a characteristic fragment size determined by the distance between cut sites. In these techniques, size selection is done either indirectly, as a consequence of PCR amplification or sequencing efficiency (e.g. GBS, CRoPS), or directly, using manual or automated gel cutting techniques or magnetic beads (e.g. RRL, MSG, ezRAD, ddRAD). For these methods, consistency of size selection across libraries is critical for producing data on a comparable set of loci across samples; inconsistency can lead to different sets of loci appearing in different libraries, resulting in wasted sequencing effort and high levels of missing genotypes.

In contrast, original RADseq and 2bRAD do not use size selection to reduce the set of loci to be sequenced; instead, all loci adjacent to restriction cut sites are targeted by these two methods. The original RADseq follows digestion by a single enzyme with a mechanical shearing step to produce fragments appropriate for Illumina sequencing. This means that each sequenced fragment has a cut site on one end and a randomly sheared end on the other, and a distribution of fragment sizes is produced at each locus. As a result, the size selection step does not further reduce the set of loci, but is used only to optimize Illumina sequencing efficiency and remove adapter dimers. The 2bRAD method is unique among the RADseq protocols in that it uses IIB restriction enzymes to produce short fragments that are of equal size across all loci (33–36 bp).

Barcoding

Use of barcodes built into the adapters allows multiplexing of individual samples early in library preparation for some of the protocols (this is sometimes called “pooling,” but should not be confused with pooling of individuals into one barcode; BOX 3). During library preparation, as soon as barcoded adapters are ligated to each sample, the samples can be multiplexed, which can greatly reduce the time and expense of subsequent steps in studies with large numbers of samples. Multiplexing of samples early in the library preparation requires the use of “in-line” barcodes, which are short sequences (typically 6–12bp) immediately adjacent to the genomic DNA. Adapters from proprietary kits do not have in-line barcodes, and therefore custom-made adapters are required for in-line barcoding. Many techniques can also be used with combinatorial barcoding, in which DNA fragments from each sample are identified by a unique combination of two different identifiers, typically one in-line barcode, and one Illumina index (6–8bp located near the middle of the adapter) added at the PCR stage to the opposite end of the DNA fragment (e.g. Peterson et al.¹⁴). An alternative combinatorial barcoding strategy would be to use two Illumina indexes, one on each side of the DNA fragment. However, this strategy would not allow multiplexing of samples early in the library preparation. Another alternative would be to use in-line barcodes on both sides of the DNA fragment; however, this strategy would be redundant because all Illumina libraries have at least one index, and would also waste sequencing effort on a redundant inline barcode. Combinatorial barcoding decreases the total number of adapters required to distinguish individual samples, so for instance a set of 24 barcoded adapters and 16 indexes can uniquely identify 384 samples in a sequencing lane.

Box 3. Pooling.

Pooling of samples without individual barcoding during RADseq library prep can allow estimation of population allele frequencies at reduced cost^86–88. However, several sources of error are unique or magnified for pooled sequencing. Unequal representation of DNA from individual samples could lead to inaccurate allele frequency estimates^89,90, and PCR duplicates will amplify this problem⁸⁸. In addition, identification of allele dropout, paralogs, mapping errors, and hidden population structure is more difficult or impossible for pooled data^86,88. Similarly, distinguishing sequencing error from low-frequency alleles is more difficult for pooled data.

Errors caused by unequal representation of individual samples in pooled sequencing libraries can be substantially reduced by using large per-pool sample sizes and depth of coverage, and removal of PCR duplicates^88,91,92. The prevalence of PCR duplicates can be reduced by using a small number of PCR cycles, which should be feasible for pooled sequencing with a large starting amount of genomic DNA. Generating and comparing sequence data for replicate pools for each population can also help identify and correct for unequal representation of individual samples⁸⁸. However, this does not mitigate problems with identifying paralogs or allele drop-out.

Researchers should also be aware of restrictions in analyses that can be conducted with pooled sequence data. Analyses requiring individual genotypes are not possible with this type of data, such as assignment tests (e.g. Bayesian clustering analyses with STRUCTURE⁹³), relatedness tests, or estimates of inbreeding coefficients. Several approaches for inferring population history or detecting selection depend on accurate estimates of linkage disequilibrium (LD)^94,95, and while there is limited power to estimate LD with the unphased data that results from individually-barcoded RADseq data, it is not possible at all with pooled data. More fundamentally, pooling assumes that all samples in a pool are from a single well-mixed population, and cryptic population structure will be obscured if multiple groups are unknowingly combined within a pool.

Type of sequence data

Most RADseq techniques currently use Illumina sequencing. Illumina machines offer a range of sequence read lengths (currently 50 to 300bp, and likely to increase further) and also the option of either single-end sequencing, which produces one “forward” read per DNA fragment, or paired-end sequencing, which produces one “forward” read and one “reverse” read per fragment. These options can be applied to all RADseq libraries, although paired-end sequencing would not be beneficial for 2bRAD, which produces very short fragments (33–36bp). For all other methods, forward reads begin from the restriction enzyme cut site, and longer reads typically capture more genomic sequence. For all methods that target loci flanked by two cut sites, reverse reads begin at the second cut site, and therefore these reads will align at identical locations in the genome for each locus. In contrast, paired-end sequencing using the original RADseq protocol produces a very different type of data. While the forward reads begin at the cut site, the reverse reads start from the randomly sheared end, typically 400–700bp away. Therefore the reverse reads at any given locus are staggered²⁰, and these data can be used to assemble long contigs, for example as long as 1kb if library fragments are tailored to be this length^19,21. These RAD contigs allow for better identification of paralogs²², provide more sequence for BLAST searching of functionally important loci²⁰, and could provide haplotype data for genealogical or phylogenetic analysis. Longer contig sequences also allow for the design of PCR primers or sequence capture probes to target loci of interest for further study^23,24.

For all methods, read pairs produced by paired-end sequencing may overlap depending on read length and fragment size range, so that if fragments are less than 200–300bp long (e.g., some fragments produced using GBS with a common-cutter enzyme), increasing read lengths or using paired-end sequencing may not gain any genomic sequence information. However, overlapping read pairs may be used to improve genotyping accuracy by increasing depth of coverage toward the ends of the reads, which tend to have higher rates of sequencing error.

Bioinformatic analyses

Post-sequencing analyses will generally share several basic steps for data generated using all RADseq methods. Initial analyses will include de-multiplexing and trimming of barcodes (if present), filtering reads based on the presence of the expected restriction enzyme cut site and sequence quality, and possibly trimming if quality declines toward the end of reads. For some RADseq methods, PCR duplicates can be removed during initial analyses to improve downstream genotyping accuracy (see below). If a reference genome is available, loci can then be identified by alignment of sequence reads to the reference. Alternatively, loci can be assembled de novo by clustering similar sequence reads together and assuming that variation among reads at a locus represents either sequencing error or allelic variation. After locus discovery, long contigs can be generated for paired-end data generated using the original RADseq (see above). Genotyping can be conducted using maximum likelihood²⁵ or Bayesian approaches^26,27; maximum likelihood methods may require higher depth of coverage than Bayesian methods, particularly if Bayesian approaches make use of population-level allele frequencies to set prior probabilities on genotypes.

Several programs designed specifically for analyzing RADseq data are available (e.g., Stacks²⁸, pyRAD²⁹, UNEAK³⁰, in addition to other publicly available scripts and pipelines). Stacks contains a number of flexible modules to conduct all parts of the analysis, from quality filtering to locus identification (either reference-aligned or de novo) to genotyping and calculating population genetic statistics. pyRAD, designed specifically for phylogenetic applications, conducts quality filtering and de novo locus identification and genotyping, with the advantage that it can handle insertion-deletion variation among alleles and may thus be better suited to studies with a broader taxonomic scale. UNEAK is part of the TASSEL pipeline for association mapping with GBS data³¹ and applies a network-based SNP detection algorithm, but is somewhat less flexible than other software in aspects such as read trimming and parameters for de novo locus identification. RADseq data can also be analyzed using more generic software tools for quality filtering, alignment to a reference genome, and genotyping. Following genotyping, further filtering is typically recommended to remove loci and/or individual samples with large proportions of missing data. The appropriate level of filtering at this stage depends on the study goals and the subsequent analyses to be conducted, as these vary in their sensitivity to missing data and sample size of individuals and loci. Several recent publications have highlighted how the details of RADseq data analysis, particularly the parameters used in de novo locus identification, can profoundly affect analytical results^32–34. Some of this work provides explicit recommendations for how to apply bioinformatic tools to RADseq data. Overall, it is critical for researchers to vary the parameters used in all steps of the analysis, from quality filtering to locus identification and genotyping, to critically evaluate the sensitivity of the results and optimize the analysis depending on the study goals.

Sources of Error and Bias

RADseq methods share some sources of sequencing and genotyping errors with all next-generation sequencing methods³⁵. However, there are also several unique potential sources of error and bias in RADseq methods, the impact of which can vary across library preparation protocols and statistical analyses.

Allele dropout and null alleles

Allele dropout manifests in RADseq when a polymorphism occurs at a restriction enzyme recognition site, resulting in a failure to cut the genomic DNA at that location. Alleles lacking the complete recognition site will not be sequenced and are therefore “null alleles.” If a SNP occurs within a null allele, failure to sequence the allele could cause genotyping errors, with individuals heterozygous for the null allele appearing as homozygotes. The absence of a restriction cut site could also drive allele dropout for loci at neighboring cut sites, because the post-digestion fragment lengths may fall outside the selected size range for methods that use size selection to reduce the set of loci (FIG. 3A).

The frequency of allele dropout will increase with the cumulative length of the restriction enzyme recognition sites, simply due to an increase in the probability of mutations in longer sequences³⁶. Simulation studies also indicate that allele dropout will increase with overall levels of polymorphism in the study system, and will have a greater impact on data generated by ddRAD than original-RAD because loci depend on the presence of two cut sites rather than one^36,37.

Genotyping errors caused by allele dropout can bias population genetic statistics through underestimation of genomic diversity, overestimation of F_ST, and an increase in false positives and negatives in F_ST outlier tests^36,37. However, there is evidence that the impact of these biases may be limited unless effective population sizes are large (N_e>10⁵)³⁶. F_ST biases can be largely compensated by removing loci with null alleles from the dataset. In theory, loci with null alleles should be identifiable by high variance in depth of coverage across individual samples, as some individuals will lack one or both copies at the locus. However, many other factors cause variance in depth of coverage (see below), so it is not always a reliable indicator of null alleles. Nevertheless, loci with a high prevalence of null alleles will be removed by many standard filtering practices that retain only loci that are successfully genotyped across some minimum percentage of individual samples. Although removal of loci with null alleles should largely compensate for biased F_ST estimates, it may do little to compensate for biased diversity estimates. Loci with null alleles are expected to occur more frequently in genomic regions with higher mutation rates and/or levels of standing genetic diversity, and thus the absence of these loci from the dataset will tend to lead to systematic underestimation of overall genomic diversity³⁷.

PCR duplicates and genotyping errors

Most next-generation sequencing library preparation protocols have a PCR step during which clonal DNA fragments (“PCR duplicates”) are generated from the original genomic DNA fragments (“parent fragments”)^38,39. During PCR, stochastic processes can cause one allele to amplify more than the other at a given locus in an individual sample. This potential skew can lead to downstream genotyping errors because heterozygotes can appear as homozygotes, or alleles containing PCR errors can appear as true alleles (FIG 3B). Studies report that PCR duplicates can occur at high frequencies in RADseq data (e.g. 20–60% of reads^20,38,39). In theory, PCR should not systematically favor one allele over another at a given locus, and therefore parameters estimated from a large number of loci are unlikely to be substantially biased. However, analyses requiring high genotyping accuracy at individual loci, such as outlier tests or parentage assignments, could produce erroneous results if PCR duplicates are present.

For sequence data generated using most next-generation sequencing protocols, PCR duplicates can be identified and removed bioinformatically to improve genotyping accuracy. This is possible in protocols that have a mechanical or random enzymatic fragmentation step, so that PCR duplicates can be identified as fragments that start and end at identical positions in the genome. This method can also be used to identify PCR duplicates in sequence data generated using original RADseq with paired-end sequencing, because of the mechanical shearing step (FIG. 3B). In some circumstances (when the distance between forward and reverse reads is very short or local coverage is very high), this filter will remove fragments that are not duplicates but that, by chance, have the same start and end points. However, this should occur rarely and should be conservative with respect to genotyping accuracy. This method cannot be used to identify PCR duplicates in any other RADseq protocols, because all fragments for a given locus will have identical start and stop positions for these protocols².

Another recently developed method shows promise for identifying PCR duplicates through the use of degenerate base regions within the sequencing adapters to tag parent fragments prior to PCR^39–41. This method could be incorporated into any protocol that uses custom-designed adaptors. An alternative method for dealing with PCR duplicates is to eliminate the PCR step of library prep altogether, as in ezRAD with Illumina PCR-free kits¹⁶. However, PCR-free kits are currently much more expensive and require much more genomic DNA (1μg) than other RADseq protocols.

Variance in depth of coverage among loci

Whereas PCR duplicates and allele dropout can cause genotyping errors as a result of preferential sequencing of certain alleles within RADseq loci, several other phenomena can cause preferential sequencing of certain loci over other loci. These phenomena should not cause genotyping errors, but will require greater overall sequencing effort to obtain sufficient depth for the loci that sequence less commonly. One well-known phenomenon is preferential amplification of fragments based on GC content during PCR^2,42–44, and this bias should equally affect all RADseq methods that include a PCR step. Another is the preferential amplification of shorter fragments over longer fragments. This issue will affect all RADseq methods that sequence fragments flanked by two cut sites (BOX 1), because each locus has a characteristic fragment length. This issue will not affect 2bRAD because all loci are uniform in length, and will not affect the original RADseq because each locus is represented by a variety of fragment lengths (see above).

Another phenomenon influencing variance in depth of coverage among loci is driven by the mechanical shearing step in the original RAD. Fragments <10kb shear with lower efficiency, and therefore loci originating from shorter restriction fragments will yield fewer reads than loci originating from longer fragments⁴². However, this phenomenon should have less influence on the majority of original RADseq studies, which typically use rare-cutters that digest genomic DNA to fragments >10kb.

When coverage varies widely among loci, obtaining sufficient numbers of reads to accurately genotype the low-coverage loci will require an increase in the average depth of coverage across all loci. To accomplish this, the number of individuals multiplexed per sequence lane must be decreased, and this will increase the cost of the research project or decrease the number of individual samples that can be analyzed. Alternatively, low-coverage loci could simply be removed from the dataset if sufficient data can be obtained from high-coverage markers, and this is commonly done in practice.

How to design a RADseq study

Designing a RADseq study for a particular application requires several major considerations regarding the most appropriate RADseq method, sampling and sequencing strategies, budget, and other methodological details. Trade-offs among selected methods are summarized in Table 1.

TABLE 1.

Summary of trade-offs among five RADseq methods.

	Original RAD	2bRAD	GBS	ddRAD	ezRAD
Options for tailoring number of loci	Change restriction enzyme	Change restriction enzyme	Change restriction enzyme	Change restriction enzyme or size selection window	Change restriction enzyme or size selection window
Number of loci per 1 Mb of genome size*	30–500	50–1000	5–40	0.3–200	10–800
Length of single-end loci	≤1kb if building contigs; otherwise ≤300bp**	33–36bp	<300bp**	≤300bp**	≤300bp**
Cost per barcoded/indexed sample	Low	Low	Low	Low	High
Effort per barcoded/indexed sample	Medium	Low	Low	Low	High
Uses proprietary kit?	No	No	No	No	Yes
Can identify PCR duplicates?	with paired-end sequencing	No	with degenerate barcodes	with degenerate barcodes	No
Specialized equipment needed	Sonicator	None	None	Pippin Prep***	Pippin Prep***
Suitability for large or complex genomes****	good	poor	moderate	good	good
Suitability for de novo locus identification (no reference genome)*****	good	poor	moderate	moderate	moderate
Available from commercial companies (in 2015)	Yes	No	Yes	Yes	No

^*Estimated as follows: original RAD, assuming either a 6-cutter or 8-cutter; 2bRAD, assuming type IIB enzymes with recognition sites containing 5–7 specific nucleotides; GBS, values from Elshire et al.⁶⁶; ddRAD, from Table 1 in Peterson et al¹⁴ and allowing for up to double the size range; ezRAD, values from Toonen et al¹⁶ for species with reference genomes.

^**Based on current limits in sequencing technology

^***Can alternatively be used with standard gel equipment

^****Based on ability to reduce total number of loci and lengths of loci

^*****Based on lengths of loci to distinguish paralogs and duplicate sequence

Number of loci

The number of loci identified and genotyped by RADseq methods depends on the genome size, the frequency of the restriction cut sites in the genome, and the number of cut sites targeted for sequencing. Computational tools are available to estimate the number of loci expected for each protocol^43,45. RADseq methods that target all cut sites (original RAD and 2bRAD) or use common-cutter enzymes (GBS) without a direct size selection step generally provide more loci, but the number can be adjusted by the choice of enzyme. In contrast, protocols involving an explicit size-selection step (e.g. ddRAD, ezRAD) can not only adjust the number of loci by choice of enzyme(s), but also by changing the size range selected, and thus they typically have more flexibility to provide a smaller number of loci. Alternatively, another way to reduce the number of loci in any RADseq protocol is to design probes for a subset of RADseq loci and use these to capture and sequence selected loci (i.e., “RAD Capture” or “Rapture,” unpublished data, see Further Information).

The optimal number of loci depends on the goals of the study. Studies focused on estimating neutral or genome-wide processes, such as phylogenetic relationships, geographic population structure, gene flow, introgression, or individual inbreeding (identity by descent) often require only several hundred to a few thousand SNP-containing RADseq loci to adequately sample the genome^20,46–48. In contrast, studies seeking to characterize functionally important regions across the entire genome, such as those exhibiting signatures of selection, require a larger set of markers (e.g., up to tens or even hundreds of thousands of RADseq loci)^25,49,50. In mapping studies the optimal number of RADseq loci depends on the expected extent of linkage disequilibrium along chromosomes and recombination patterns. For instance, a laboratory F2 cross or very recently admixed population would require fewer loci, although statistical power may be increased with large numbers of progeny and more markers. For association mapping in an outbred population, many more markers would be required. Quantifying diversity patterns along chromosomal stretches (e.g. runs of homozygosity) to estimate recent and historical effective population size and inbreeding also requires tens of thousands of loci^48,51,52.

Some biological factors may also increase the number of loci that should be targeted. Bottlenecked or small populations with low genomic variation may require sequencing more loci to accurately quantify levels of variation. Genomes with a history of whole-genome or gene duplication, or genomes with high levels of transposable elements or other repeat sequence, may also require large numbers of loci to compensate for stringent filtering (removal) of problematic loci.

Type of sequence reads

Longer sequence reads and/or paired-end sequencing reads provide many advantages, including improved locus identification, discrimination of paralogous or repetitive sequence, and BLAST searching for functionally important loci. For most RADseq protocols, sequence length is limited primarily by sequencing technology (e.g. typically up to 150bp reads with Illumina, but up to 300bp in some cases). Many research questions can be sufficiently addressed with relatively short reads (e.g. 100bp) and single-end sequencing. However, as described above, longer RADseq loci can be obtained by assembling contigs from paired-end sequence reads with the original RAD (up to 1kb²⁰), and this method can be particularly advantageous for complex genomes in the absence of a reference genome. 2bRAD produces the shortest reads of all methods (33–36bp), so is not recommended for de novo locus identification or in the case of large and complex genomes (e.g. the human genome⁵³).

Prior genomic resources

Prior reference sequence can provide numerous advantages for RADseq studies. A reference genome sequence, a poorly assembled set of genomic scaffolds, or even a set of previously identified RAD loci can greatly improve the ability to filter paralogous or repetitive sequences, identify insertion-deletion variation, and remove non-target DNA sequence (e.g. bacterial contamination)⁵⁴. A well-assembled reference genome provides further advantages. For instance, mapping studies can use information on physical positions of loci to infer haplotypes across larger chromosomal regions covering multiple loci⁵⁵. The GBS and MSG methods have been used in this way for trait mapping in model species, where chromosomal blocks of parental ancestry are relatively large. Population genomic studies can use a reference genome assembly to conduct sliding window analyses and increase statistical power to detect genomic regions of interest, such as regions under divergent selection between populations^25,50. In the absence of a reference genome, long contigs generated with the original RAD should provide the greatest ability to distinguish paralogous or repetitive sequences (see above)^19–21.

Depth of sequencing coverage

Libraries from all RADseq methods can be sequenced to produce different depths of coverage, and the ideal depth for individually barcoded samples varies widely across studies. At one extreme, laboratory mapping studies with a well-assembled reference genome may be most efficient with very low coverage (<1×). Much higher coverage is required (e.g. 20–30×) for confident de novo locus discovery and genotyping in diploids, and even higher coverage would be required in polyploid taxa. Alternatively, in some cases individuals may be pooled into single barcodes (BOX 3), with much lower coverage per individual because individual genotypes are not assigned.

Budget

Often the major expense in producing RADseq data is the sequencing itself. The total sequencing effort is divided among the number of loci, the number of samples and populations, and the desired coverage per locus per individual. However, the different protocols can also differ considerably in the expense of library preparation, and in the way in which library preparation costs scale with the number of samples. For instance, while the original RADseq protocol has a relatively large number of steps, samples are multiplexed early in the protocol and the subsequent steps are conducted on mixtures of up to 96 or more barcoded samples, so the marginal cost of increasing samples is minimized both in terms of time and money (FIG 2). In contrast, the cost of ezRAD scales roughly linearly with samples because multiplexing does not occur until the end, so this method may be most appropriate for small numbers of samples or pools of samples¹⁶. Some RADseq protocols also require an initial financial investment in specialized barcoded adapters, although a single set of such oligonucleotides is often sufficient for a large number of libraries. In addition, some RAD protocols may require the purchase of specialized laboratory equipment. The original RAD requires the use of a DNA sonicator, and RADseq protocols that use a direct size selection (e.g. ddRAD, ezRAD) can increase precision and consistency of size selection, and decrease the possibility of cross-contamination, by using a Pippin Prep¹⁴ (Sage Science, Beverly, MA).

Figure 2.

Sources of error and bias in RADseq data. (a) Example of allele dropout for a RADseq protocol that uses size selection to reduce the number of loci to be sequenced. Gray lines represent chromosomes within one individual, red squares represent restriction cut sites, colored squares represent heterozygous SNPs, and brackets represent genomic regions that are sequenced. Mutation in Restriction Cut Site B for Haplotype 1 makes the post-digestion fragment containing the SNP too long to be retained during size selection for Haplotype 1, eliminating the possibility of sequencing of any loci on that fragment, and causing the individual to appear homozygous at the heterozygous SNP. (b) See Figure 1 of Andrews et al. 2014³⁸. Example of fragments produced after PCR for one heterozygous locus for different RADseq protocols, and the reads retained after bioinformatic analyses. PCR duplicates are shown with the same symbol (circle, square, asterisk or triangle) as the parent fragment from the original template DNA. By chance, some alleles will amplify more than others during PCR. For all protocols, PCR duplicates will be identical in sequence composition and length to the original template molecule. For the original RADseq, this feature (i.e., identical length) can be used to identify and remove PCR duplicates bioinformatically, because original template molecules for a given locus will not be identical in length. For alternative RADseq methods, this feature cannot be used to identify PCR duplicates, because all original template molecules for a given locus are identical in length. High frequencies of PCR duplicates can cause heterozygotes to appear as homozygotes or can cause PCR errors to appear as true diversity.

Comparability of data

A final consideration when designing a RADseq study is the consistency of data across sequencing runs and across laboratories. Inconsistency in size selection could produce variation among libraries for methods that use size selection to reduce the set of loci. The consistency of different size selection techniques (automated or manual gel extraction vs. bead-based selection) has not been rigorously quantified, but magnetic beads are likely much less consistent⁵⁶. Methods that target every cut site (original RAD, 2bRAD) are generally expected to be more consistent across libraries (but see Sources of Error). There can be some consistency in the loci genotyped even across methods, depending on the choice of restriction enzymes. For instance, the loci sequenced using SbfI and EcoRI in a ddRAD protocol should be a subset of those sequenced using SbfI with original RAD.

Complementary approaches

Although RADseq has many benefits as a tool for SNP genotyping and discovery, it is not the best method of choice for every ecological and evolutionary study. Two major alternative reduced representation approaches that take advantage of next generation sequencing are transcriptome sequencing (RNAseq)⁵⁷ and targeted (probe-based) capture⁵⁸ (BOX 4). Whole genome re-sequencing and whole genome pooled sequencing are other alternatives that provide much more genomic information than reduced representation techniques^59–61. However, despite the increasing feasibility of whole-genome re-sequencing for population studies, many ecological and evolutionary questions stand to gain little from such an increase in genome-wide data. For example, a RADseq study using several thousands of markers to detect selection based on allele frequency or linkage disequilibrium is more likely to be limited by the number of individuals sampled than the density of markers.

Box 4. Alternatives to RADseq.

Two major alternative reduced representation next-generation sequencing methods are transcriptome sequencing (RNAseq) and targeted (probe-based) capture.

Transcriptome sequencing (RNAseq): sequencing transcribed regions of the genome using RNA as a starting point in library preparation.

Advantages

RNAseq can be used to quickly sequence thousands of functional genomic regions in virtually any species with limited or no genomic resources⁹⁶. Most transcripts can be annotated against existing genome databases⁹⁷, providing a much stronger functional context when compared to anonymous RADseq loci.

Disadvantages

RNAseq provides limited opportunity to dynamically scale sequencing effort based on question or experimental design. Individual transcripts may differ by several orders of magnitude in relative abundances⁹⁸, complicating genotyping⁹⁹ and increasing sequencing costs. Functional annotation may be limited in taxonomic groups with poor database representation. RNAseq requires high quality samples, which can limit its feasibility for many studies.

Targeted (Probe-based) capture: sequencing of pre-selected genomic regions using a DNA probe to isolate regions of interest

Advantages

Targeted capture is highly scalable, able to sequence a single locus¹⁰⁰ or hundreds of thousands of loci^101,102. Technical performance is typically very high¹⁰³, with low variance in sequencing coverage across regions and individuals^36,42,104. Capture can be applied across moderate to deep evolutionary timescales^105–107 and on degraded DNA samples, making it popular for phylogenetic^34,108,109 and ancient DNA studies^110–115.

Disadvantages

Primary limitations for capture are the availability of genomic resources for designing probes, and generally higher cost than RADseq or RNAseq⁵⁸.

Alternative genomic approaches can also be used to complement RADseq for more comprehensive or flexible investigation in a particular system. For instance, the development of de novo reference genomes for non-model species is becoming increasingly feasible as sequencing and assembly technologies continue to improve^62,63, and such a reference provides numerous advantages for analysis of RADseq data from population-level sampling^25,49,50,54. Transcriptome sequencing can also complement RADSeq data by targeting coding (and presumably functional) sequence, whereas RADseq interrogates both coding and non-coding loci. RADseq can also be used as the first step in a larger study to focus on significant loci. For instance, RADseq can provide a genome-wide scan to identify candidate loci of interest, and sequence data at these loci can then be used to design probes for sequence capture. Subsequent targeted sequencing could then be conducted on a large number of samples at greatly reduced cost per sample, and with poorer quality DNA.

Conclusions

RADseq techniques have enormous power and versatility for SNP discovery and genotyping in ecological and evolutionary genomics, but researchers should employ careful consideration in choosing and applying these methods. Numerous RADseq protocols have been developed that differ not only in the technical details and cost of the library prep, but also in the types of data produced and the sources of genotyping error and bias. Therefore different protocols will be better suited to different study systems, budgets, and research questions. Despite rapid changes in sequencing technology and costs, we anticipate that reduced representation sequencing approaches like RADseq will continue to be a critical tool for genomic studies of natural populations into the foreseeable future. When implemented appropriately, RADseq approaches provide efficient, flexible, and cost-effective avenues to unleash the power of next-generation sequencing technologies for gaining new insights into ecological, evolutionary, and conservation-related questions.

KEY POINTS.

1. RADseq has fueled studies in ecological, evolutionary, and conservation genomics by using next-generation sequencing to uncover hundreds or thousands of polymorphic loci across the genome in a single, simple, and cost-effective experiment. RADseq does not require any prior genomic information for the taxa being studied, and is therefore particularly advantageous for studies of non-model organisms.

2. Numerous technical variations on RADseq have been developed, promising to increase flexibility and decrease cost and effort in genomics studies. Differences among the methods lead to important considerations for all steps of genomic studies, from costs of library preparation and sequencing, to the types of bias and error inherent in the resulting data, to the types of scientific questions that can be addressed.

3. Allele dropout, PCR duplicates, and variance in depth of coverage among loci are important sources of error and bias in RADseq studies, and the prevalence of these phenomena will vary across RADseq methods.

4. Other important considerations when designing a RADseq study include the number, length, and coverage of loci needed to address the research question; availability of prior genomic resources; budget; and consistency of data across sequencing runs and laboratories.

5. There is no single best or most flexible RADseq method. Researchers must weigh the trade-offs of the different methods, and choose the approach best suited to their study goals.

GLOSSARY

Allele dropout

failure of an allele present in a sample to be detected by sequencing

Adapters

double-stranded oligos that must be ligated to DNA fragments prior to next-generation sequencing. Illumina adapters contain regions that anneal to the flow cell, an “index” sequence that act as a barcode to identify individual samples, and primer binding sites for bridge amplification and sequencing of the DNA fragment and indexes

Barcode (also called “inline barcode”)

a short unique sequence (typically 6–12bp) used to identify individual samples. Inline barcodes occur on the end of the adapter that is immediately adjacent to the genomic DNA fragment after adapter ligation. The barcode is sequenced immediately prior to sequencing of the DNA fragment, and thus the barcode sequence will appear at the beginning of sequence reads

Combinatorial barcoding

using two different barcoding methods, usually a standard Illumina index and an inline barcode. This method can reduce the number of adapters that must be purchased, thus reducing library prep cost

Contig

a group of overlapping sequence reads assembled to form a longer sequence

Depth of coverage

the number of sequence reads for a given locus or nucleotide site

Filtering

removing unwanted sequence reads from a dataset due to low sequence quality, low depth of coverage, evidence for paralogy, or other reasons

Illumina index

a unique 6bp or 8bp sequence incorporated into Illumina adapters that functions as a barcode to identify individual samples

Sequencing library

DNA prepared for next-generation sequencing. The DNA must be an appropriate length for sequencing and must have sequencing adapters ligated

Next-generation sequencing (“Massively parallel sequencing”)

Technologies first emerging around 2005 that sequence millions of DNA molecules simultaneously

Null allele

An allele present in a sample that fails to be identified by genotyping. The presence of a null allele leads to “allele dropout.”

Paired-end sequencing

Illumina sequencing of both ends of each DNA fragment

Paralog

sequence originating through duplication within the genome

Pooling

combining multiple individual samples into a DNA library with only one unique identifier (e.g. one barcode or one index)

Reduced-representation library

DNA library comprised of a subset of loci, rather than the entire genome

Restriction site Associated DNA (RADseq)

a method for sequencing thousands of genetic loci adjacent to restriction cut sites across the genome using massively parallel (“next generation”) sequencing. Sometimes referred to as “Genotyping by Sequencing.”

Single-end sequencing

Illumina sequencing of only one end of each DNA fragment

Sticky end (also called “DNA overhang”)

the string of single-stranded DNA remaining on the end of a DNA fragment that has been digested with a restriction enzyme. Also called a “DNA overhang.” Some restriction enzymes produce “blunt ends” (double-stranded ends) rather than sticky ends

Biographies

Kimberly R. Andrews is a postdoctoral research fellow in Lisette Waits’ group (Department of Fish & Wildlife Sciences) collaborating with Paul Hohenlohe’s group (Department of Biological Sciences) at the University of Idaho. Her work focuses on investigating the ecological and evolutionary factors driving dispersal, gene flow, adaptation, and speciation, and applying this knowledge to conservation and management issues. She received her PhD from the University of Hawai‘i at Mānoa and has conducted postdoctoral research at the University of Hawai‘i and as a Marie Curie Research Fellow at Durham University.

Jeffrey Good is an assistant professor in the Division of Biological Sciences at the University of Montana, Missoula, USA. His laboratory combines population, comparative, and functional genomics to understand the genetic basis of speciation and adaptation in model and non-model systems.

Paul A. Hohenlohe is an Assistant Professor in the Institute for Bioinformatics and Evolutionary Studies, and the Departments of Biological Sciences and Statistical Science at the University of Idaho. His research group focuses on evolutionary and conservation genomics in a wide variety of organisms, using RADseq and other tools. He received his Ph.D. from the University of Washington, worked as a conservation biologist and conducted postdoctoral research at Oregon State University and the University of Oregon before joining UI in 2011.