Facilitating population genomics of non-model organisms through optimized experimental design for reduced representation sequencing

Background Genome-wide data are invaluable to characterize differentiation and adaptation of natural populations. Reduced representation sequencing (RRS) subsamples a genome repeatedly across many individuals. However, RRS requires careful optimization and fine-tuning to deliver high marker density while being cost-efficient. The number of genomic fragments created through restriction enzyme digestion and the sequencing library setup must match to achieve sufficient sequencing coverage per locus. Here, we present a workflow based on published information and computational and experimental procedures to investigate and streamline the applicability of RRS. Results In an iterative process genome size estimates, restriction enzymes and size selection windows were tested and scaled in six classes of Antarctic animals (Ostracoda, Malacostraca, Bivalvia, Asteroidea, Actinopterygii, Aves). Achieving high marker density would be expensive in amphipods, the malacostracan target taxon, due to the large genome size. We propose alternative approaches such as mitogenome or target capture sequencing for this group. Pilot libraries were sequenced for all other target taxa. Ostracods, bivalves, sea stars, and fish showed overall good coverage and marker numbers for downstream population genomic analyses. In contrast, the bird test library produced low coverage and few polymorphic loci, likely due to degraded DNA. Conclusions Prior testing and optimization are important to identify which groups are amenable for RRS and where alternative methods may currently offer better cost-benefit ratios. The steps outlined here are easy to follow for other non-model taxa with little genomic resources, thus stimulating efficient resource use for the many pressing research questions in molecular ecology. Supplementary Information The online version contains supplementary material available at 10.1186/s12864-021-07917-3.


Background
Evolutionary and ecological population genetic studies are important to understand how the diversity of life on earth is distributed, has evolved and may respond to future environmental changes [1]. A grand challenge has been to document this biodiversity and understand its role in maintaining ecosystem functionality, particularly in the ocean [2] and even more so in frontier areas such as the deep-sea and polar regions [3]. Molecular data collection has benefitted from a revolution in sequencing technologies such that genomics, where billions of nucleotides are screened simultaneously, is now an integral part of the biological toolbox [4][5][6]. Genome-wide data open new avenues of ecological and evolutionary research, especially to study local adaptation [7,8]. Given ever-increasing rates of anthropogenic disturbance, it is crucial to assess spatio-temporal genomic diversity, adaptation patterns and resilience of non-model organisms [9,10].
Similar to previous methodology shifts in population genetics (e.g. from Amplified Fragment Length Polymorphisms [AFLP] to microsatellites), the transition to novel methods requires detailed understanding of the new technology, its potential as well as its pitfalls, and careful experimental planning. While some study systems are moving towards population-specific shallow resequencing of whole genomes (e.g. important commercial fish species) [11,12], many species of interest with less extensive genomic resources rely on reduced representation sequencing (RRS) techniques to subsample the genome. Among the most popular RRS techniques are Restriction site-Associated DNA sequencing (RADseq) [13] and Genotyping by Sequencing (GBS) [14]. A plethora of similar methods with unique names have been developed [5,[15][16][17][18]. Here, we follow the reasoning of Campbell et al. [18] and use the term RRS [19] to refer to all of these methods, which are attractive because they make more frugal use of sequencing volume than whole genome sequencing.
In RRS, one or several restriction endonuclease enzymes are used to first fragment the target genome into smaller portions, thus reducing sequencing costs. Millions of reads from high-throughput sequencing platforms are then aligned against either a reference genome or, alternatively, a de novo reference catalog of loci [20,21]. Subsequently, genetic variants, most commonly single nucleotide polymorphisms (SNPs) are determined. In addition, approaches have been developed to use RRS data to create microhaplotypes [22,23], or identify microsatellites [24] or copy number variants (CNV). The latter relies on summary statistics of the frequency of heterozygotes and the read ratio, which should differ between singleton and duplicated SNPs [25,26]. RRS has provided many important insights across a wide range of taxa from different ecosystems, e.g. with respect to population structure and demography, as well as hybridization, landscape or seascape genomics, QTL mapping, phylogeography, and shallow phylogenies (e.g. [5,[27][28][29][30][31][32]). Limitations and problems of RRS include the potential for allele dropout, PCR duplicates, genotyping error, as well as insufficient coverage and/or low marker density (the number of genetic markers that are sequenced in relation to genome size) [5,33,34]. Unnecessary costs, inability to answer the research, or, in the worst case, incorrect conclusions may be the consequence. Good experimental design, however, can help avoid or mitigate some of these issues.
Effective and cost-efficient RRS experiments must be well designed. First, one should establish whether the species of interest truly represents one species or if cryptic species are present. This can be problematic in nonmodel taxa and has potentially large downstream implications for RRS such as high divergence but few shared loci [35,36]. A useful complement is therefore DNA barcoding to screen for cryptic species [37,38]. Alternatively, RRS can be specifically employed for species delimitation purposes [30,39,40], but this should be a deliberate choice before designing the RRS setup. For such a scenario it would be especially important to sequence many fragments thereby increasing the likelihood of capturing genetic markers that are conserved across, yet discriminatory between species. In general, the research question fundamentally determines whether the application of RRS is appropriate. For example, providing evidence for significant, evolutionary neutral genetic population structure may be easier and less expensive with a good number (> 10) of multi-allelic microsatellites [41]. However, RRS may be better suited to identify loci that are putatively affected by spatially variable selection and therefore involved in local adaptation. To this end, the density of markers that can be realized for a given species, which depends on genome size and complexity, as well as research budget, should be considered.
With low marker density one may run the risk of accepting unreasonably high rates of false positives (outliers that are not based on biological reality) in genome scans leading to biased or erroneous inferences [42,43]. Consequently, there is debate about the usefulness of RRS (or RADseq in particular), especially for inferring local adaptation patterns [34,44]. The genomic characteristics of a target species, most importantly its genome size and the level of linkage disequilibrium (LD), are crucial to design a RRS experiment. With little genomic information, a priori calculations may be inaccurate. Therefore, it is vital to assess, optimize and critically ponder the advantages and limitations of RRS for a given research project to avoid the creation of sequence data that are unsuitable to answer the study question and/or inefficient use of resources. A most critical point is to properly strike a balance between sequencing depth (coverage) and number of fragments, which is roughly proportional to the number of genetic markers. The estimated number of fragments generated from a genome determines the marker density (as the number of fragments translates approximately into the number of SNPs), while avoiding unnecessary "over"-sequencing of the genomic fragments, i.e. loci or RADtags, to save sequencing costs. Both excessive (> 100×) and uneven or too low (< 10×) coverage is detrimental for accurate locus reconstruction and SNP calling, particularly in de novo approaches [45]. Hence, RRS experimental procedures may benefit from thorough optimization. In this context, we used the framework of a large research project ("Refugia and Ecosystem Tolerance in the Southern Ocean") to optimize RRS for a diverse set of taxa in parallel. The Southern Ocean hosts a unique marine fauna with high levels of endemism [46,47], but is increasingly subject to external pressures, such as warming, pollution and living resource exploitation [48][49][50][51]. Population genomic approaches are needed to understand the genetic structure and connectivity of Antarctic fauna, so that appropriate management and conservation actions can be developed (e.g. [52][53][54]).
In this molecular pilot experiment, we seek to investigate and optimize the applicability of RRS to a range of Antarctic non-model taxa across the animal kingdom. The target organisms are ecologically important, abundant, and widely distributed in the Southern Ocean and cover a variety of habitatsfrom benthos to pelagic birds. Specifically, we aim to develop economic and robust experimental setups for RRS population genomic studies in an ostracod group, two amphipod species, two bivalve species, two sea star species, two fish species, and two bird subspecies ( Table 1). The outlined approach should be readily adoptable for other taxa of interest. We lay out a clear and concise protocol to follow a priori for any RRS experiment on non-model species that will help researchers to evaluate the costs, benefits, and risks of such projects.
We specifically aim to (i) collate information about the genomic properties of the target taxa; (ii) assess in silico which restriction enzymes are likely to yield the desired number of fragments; (iii) test selected restriction enzyme digestions in the laboratory; (iv) optimize restriction enzyme choice, size selection window and the number of individuals to be pooled per sequencing library (based on the previous results); and (iv) sequence and analyze test RRS libraries of promising experimental setups. These extensive pilot analysesincluding literature research, computational analyses, and laboratory workare designed to comprehensively evaluate all information for each target species or species complex. In the workflow of optimizing the setup for each target taxon, we strive to use the same restriction enzymes (or combinations) for several taxa whenever possible to reduce the costs for specifically designed barcodes and adaptors. Results shall ultimately facilitate informed decisions about whether and how RRS for each taxon could be conducted. We critically discuss these considerations and suggest alternative approaches in two cases.

Results
The optimization process of RRS experimental setups for non-model species is iterative and includes many deliberate choices that must be made based on the best available knowledge. Relatively constant variables, i.e. the number and quality of samples, the research budget and the main research question, should be considered during the entire process and flexible variables, such as restriction enzymes, size selection window and the number of Table 1 Target taxa for a molecular pilot experiment to test and optimize the experimental setup for reduced representation sequencing (RRS) individuals to be pooled, should be adjusted to reach the desired outcome.

Genome characteristics
First, published genomic resources of our target taxa were collected. Available information is highly variable across target taxa, with typically more genomic resources available for vertebrate groups ( Table 2). Genome size among ostracods varies considerably, with Macrocyprididae estimated at approximately 166 Mb (or 0.17 C) [68]. One published ostracod genome (Cyprideis torosa) with a genome size comparable to our target species was available [57]. Amphipods show very large variability in genome size [72] with extreme cases that dramatically exceed the size estimates of all other target taxa studied here (up to 63,198 Mb or 64.62 C, Table 2). Two amphipod reference genomes were available (Hyalella azteca, Parhyale hawaiensis) [58,59]. In addition to these reference genomes we simulated a large (10,000 Mb) and a very large (30,000 Mb) genome for amphipods, according to genome size estimates from species of the same family (Eusiridae: 7.16 C, Lysianassidae: 27 C) [73]. For bivalves and sea stars more reference genomes were available, but not from species closely related to the target species. In both cases, we selected three reference genomes of varying size ( Table 2). The Antarctic fish target species have genome size estimates available as well as a reference genome from a species from the same family (Notothenia coriiceps) [66]. In birds, no genome size estimates for our target species were available, but bird genome size appears to be relatively constrained between approximately 1 and 2 Gb and a reference genome from the same family has been published (Fulmarus glacialis) [67]. We decided to aim at 50,000 fragments as initial targets for our optimizations in all taxa, except for fish and sea stars ( Table 2). In the latter target taxa, we aimed at 20,000 fragments initially, because we had more samples available and thus were interested in covering more individual samples from a wider geographic range at the expense of marker density. Note that these targets are highly study specific and depend on the budget, number of samples to be sequenced and, most importantly, exact research question of a given RRS project.

In silico digestions
We estimated how many RRS fragments twelve restriction enzymes and enzyme combinations (listed in Table 3) would produce. These estimates were conducted using various reference genomes and simulated genomes. We estimated the fragment number in total as well as in various size selection windows (see below and Additional File 2). As expected, the fragment number is influenced primarily by the type of enzyme and the genome size. The tested combinations showed promising results with fragment numbers close to our defined targets in all species, but there was not one single enzyme or enzyme combination that produced promising results in all species. We aimed at using as few different enzyme setups across species as possible and in addition opted for enzymes or enzyme combinations that were previously used in our laboratory as much as possible. Using the same setup for several RRS experiments reduces costs as the same adaptor sets can be reused multiple times. Therefore, we kept five initial setups that yielded promising fragment numbers: EcoRI, PstI, ApeKI, MspI and a double digest with EcoRI and MspI.

Empirical digestions
Based on preliminary in silico results, we tested the genome digestion by several enzymes and enzyme combinations in the laboratory. High quality bird DNA was not available, preventing empirical digestion tests for this group. Ostracod DNA was whole genome amplified and this proved problematic for the Bioanalyzer instrument, because the results indicated overloading even after multiple dilutions. In total, 75 empirical digestions were conducted, several of which produced unusable results even after repeating the experiment. The sensitivity of the Bioanalyzer to small irregularities especially in the size range below 500 bp made it impossible to infer sensible patterns in many cases (Additional File 3). We therefore must caution that Bioanalyzer results only sometimes provide useful additional information that increase the confidence in estimates obtained in silico. Nevertheless, from the successful runs it appeared that the empirical results are usually more similar to in silico digestions with genomes from related species than of simulated genomes ( Fig. 1 and Additional File 3). For example, in fishes ApeKI was estimated to produce significantly more small than large fragments using the N. coriiceps reference genome, which was at least roughly confirmed through empirical digestion (Fig. 1). Here, using EcoRI provides few fragments overall, which  [17] proved difficult to accurately depict using the Bioanalyzer. In contrast, the tested double digest provided a consistent picture in four out of six replicates for the two fish species (Fig. 1). Here, we also noted a pronounced spike at around 650 bp and the size window was therefore deliberately kept lower (see below and Tables 4 and  5).

RRS setup
With all information gathered thus far, we proceeded to optimize the RRS experimental setup for each of the target taxa. We planned the same setup for species from the same class, when the genomic differences between those species were unknown (in Bivalvia and Asteroidea), or when they were related and therefore likely to have similar genomic properties (Actinopterygii and Aves). In contrast, we designed two different setups in Malacostraca, because the genomes of C. obesa and E. pontomedon may have very different sizes (Tables 2 &  4). Experimental setups, i.e. restriction enzymes and size selection window, were furthermore tuned to suit a sequencing experiment with the HiSeq 2500 or 4000 platforms, respectively. The choice of the sequencing platform can be modified based on instrument availability and budget. In the following, results for use with a HiSeq 2500 platform are listed (Table 4), the same results for a HiSeq 4000 platform can be found in Additional File 4 (further calculations e.g. aimed at using a NovaSeq platform can be obtained by adjusting the R code; see: https://github.com/notothen/radpilot). The setup for optimizing results as listed here also includes the consideration that it would be cost-efficient to use the same enzyme or enzyme combinations for several species whenever possible, because adaptors can then be reused. Therefore, when several enzymes (or combinations) seemed promising according to in silico digestion, we attempted to choose setups that were also promising in other target species. For ostracods, we assumed a genome size of 250 Mb and 500 Mb as worst-case scenario. Using the C. torosa reference genome, a digest with ApeKI and size selection of 200-350 bp would yield 31.9× coverage (or half of that in the worst-case scenario). With this setup and genome size, we would achieve an estimated marker density of approximately one SNP every 1.5 kb. In amphipods, different setups per species are required. Given the highly uncertain genome size of 27,000 Mb for C. obesa and 7000 Mb for E. pontomedon (based on same family estimates) [72], double digest RADseq experiments with SbfI and MspI and EcoRI and SphI, respectively, would yield the desired coverage.
Marker density in both cases is expected to be low, due to the large genome size ( Table 4). Because of uncertainty with respect to genome size and an anticipated low marker density, we stopped RRS optimization in amphipods and instead explored alternatives. For both bivalve species, a genome digestion with ApeKI and size selection of 200-260 bp seemed promising with all three reference genomes and would yield around one SNP per 20 kb. Similarly, in sea stars we found setups with ApeKI and a slightly wider size selection that should yield good results, although results varied depending on the reference genome used. For the Antarctic fishes of the genus Trematomus, a double digest setup with EcoRI and MspI in a size window of 250-450 bp should yield desired coverage and marker density. Regarding the snow petrels, a setup with PstI and 200-300 bp size selection seemed appropriate, yielding one SNP every 9 kb. Overall, results indicate that with only three enzyme choices, it should be possible to achieve the desired coverage and marker density in five of our six target classes (excluding Malacostraca as discussed above) ( Table 4).

RRS test libraries
Pilot libraries with optimized setups were sequenced, yielding a total of 531 million (M) reads. After demultiplexing and quality control, 422 M reads were retained. These reads were spread relatively evenly across libraries, species, and individuals (average and standard deviation across all taxa and libraries: 4.5 ± 2.1 M reads). All but five individuals received more than 1 M reads and most individuals received more than 3 M reads. We created de novo catalogs from these reads using Stacks [21,98,99] with varying M and n parameters [45]. Optimal parameters varied (M = n = 3-6) among taxa (   Restriction enzyme(s) and size window was optimized for number of fragments and coverage as in Table 4 optimization also revealed varying levels of diversity, e.g. sea stars showed relatively high levels of polymorphism, while the bird library produced many loci but few SNPs (Additional File 7). Comparing the unfiltered numbers of loci and coverage across individuals underlined the inverse relation of these two variables (Table 5, Fig. 2). In ostracods, our target estimates were matched best. In bivalves and sea stars, more loci than expected were . Libraries were prepared as listed in Table 5 Fig . 3 The iterative process of reduced representation sequencing (RRS) optimization. Empirical (in vitro, left of the dotted line) and computational (in silico, right of dotted line) analyses are part of this process. Core procedures to identify suitable experimental setups are in silico and empirical enzyme digestion and sequencing of a pilot RRS library (blue boxes). The coverage and marker density that can be achieved with a given setup needs to be repeatedly checked and fine-tuned (red box). We used 30 × coverage and a density of at least one SNP per 10,000 bp as target values but emphasize that these values need to be defined depending on the specific goals of a research project sequenced at the expense of coverage, although coverage was still reasonable. Two individuals of B. loripes had low coverage due to low initial numbers of reads, indicating errors during library preparation or degraded input DNA. The fish libraries contained considerably less loci than expected at high coverage, while the opposite was true for the bird library. The latter also showed very uniform low coverage at approximately 10×. Overall, these results show promise for full scale RRS libraries with sufficiently high coverage in four of five libraries.

Discussion
High-throughput sequencing methods promise new avenues of ecological and evolutionary research in nonmodel organisms. We provide a detailed workflow to evaluate and optimize reduced representation sequencing (RRS) techniques for any animal species of interest (Fig. 3). This approach is reproducible and ensures that researchers are well-informed about the advantages and drawbacks of RRS for their research question. Different RRS setups (i.e. various species and libraries constructed via different protocols, enzymes and size selection windows) were successfully sequenced together on one HiSeq lane. Most individuals included in this multilibrary-multiplex received adequate sequencing effort, which has been problematic in other studies that pooled individuals directly after ligation [33]. From our experience (including this and previous studies in our laboratory; see e.g. [33,83,87]) it seems that careful, repeated quantification and standardization of DNA from every individual before and after PCR are key to achieve equivalent sequencing effort across individuals. A pilot sequencing experiment can then yield valuable insights before proceeding with sequencing at a larger scale.
Here, more loci than expected were assembled in most taxa (ostracods, bivalves, sea stars) at sufficiently high per locus coverage. This highlights the value of choosing parameters conservatively, e.g. under-rather than overestimating the number of sequencing reads. The fish library yielded fewer loci than expected at higher coverage. Pooling more individuals, increasing the size window, or changing the restriction enzyme setup altogether including new optimization are future options to further optimize this project, although the current setup also yields useful data. The bird library produced coverage that is directly at the advised limit of 10× [45]. This may be partly related to low quality input DNA, which was mostly extracted from feathers. Alternative sampling and/or DNA extraction protocols and further testing are needed before sequencing full scale libraries for snow petrels. Overall, a few key properties determine the feasibility and cost of RRS in non-model organisms.

Predictability of reduced representation experiments
Planning a genome reduction through restriction enzyme digestion starts with an imperative question: how large is the target genome? Non-model species often lack information on genome size, which complicates RRS optimization [100]. If genome size appears relatively conserved across species within a taxonomic class (e.g. Asteroidea), it can be assumed that the species of interest from this class has similar genome size. Some imprecisions regarding the exact size have only limited effects on overall accuracy. Alternatively, in other groups, such as amphipods, genome size is highly variable, spanning two orders of magnitude [72,73,101]. In this case, using an inaccurate genome size estimate has the potential to dramatically impact the parameters one aims to optimize. In addition, very large genomes are often highly repetitive, which significantly hampers downstream bioinformatics and population genomics [74,102,103]. Therefore, with the current state of knowledge, we opted to exclude amphipods from our trial RRS libraries.
Estimating genome size with flow cytometry or conducting a series of test libraries could be alternative ways forward. For ostracods, bivalves, sea stars, and birds more loci were found than expected. This might indicate that genome sizes were consistently larger than expected. Another, likely explanation is that the enzymes used (ApeKI, PstI) produce more fragments than in silico digestions predicted. For example, the number of fragments resulting from four base cutters may be more difficult to predict as they sometimes produce so many fragments that effectively the entire genome would be sequenced [100]. The five-base recognition site of ApeKI features a degenerate base, which may have a similar effect. The methylation sensitivity of ApeKI may also provide more genomic markers in genic regions [104]. It is unclear, however, how general this prediction holds across metazoans. Finally, some of the excess loci recovered may be artefacts from library preparation, PCR duplicates, or incorrect locus assembly [21]. Rigorous downstream filtering and/or comparison of several, differently filtered datasets may help determine the true biological signal. Whatever the reason, the higher-thanexpected number of loci still lead to sufficient coverage, except in the bird library. The latter is likely related to low quality/quantity of input DNA. Few bird samples were available, some only as feathers, which yielded very little DNA. Whole genome amplification (WGA) could be an option to increase yield for RRS as successfully applied in ostracods (this study) and insects [105].
Finally, even with reliable genome size estimates and well-tested enzymes, the empirical results may differ from in silico expectations. In Trematomus fishes, approximately half of the expected sites were found, despite well-known genome size [69][70][71]. Genomic architecture may play an important role in affecting the number of cut sites per restriction enzyme. We used the draft genome of a related species from the same family to estimate the number of fragments. The endemic Antarctic notothenioid fishes, however, are characterized by frequent chromosomal rearrangements and large numbers of transposable elements [69,106,107]. The genus Trematomus constitutes an example of a relatively recent marine adaptive radiation [108,109]. Therefore, in this particular case, the genome of a closely related species may provide relatively poor accuracy for cut site estimations.
We have tested various enzymes and enzyme combinations that have been successfully used in RRS studies (Table 3). Yet, many previous studies achieved overall relatively little marker density, which is problematic if looking for genome-wide adaptation patterns [34]. With increasing output of sequencers, aiming at higher marker density is not an unachievable goal. Genome size, restriction enzyme characteristics and genomic complexity influence the predictability. Altogether, our results highlight the importance of conducting test libraries before embarking on larger, multi-library sequencing projects. In our case, ApeKI together with a narrow size window seems robust and powerful to create many genomic fragments (and thus sufficiently high marker density) across taxa with small to medium genome size. Using the same restriction enzyme for several projects drastically reduces cost as the same custom-made barcodes and adaptors can be used.

Decision making for population genomics
As we illustrated here, there are many experimental choices that may lead to inefficient or "broken" [34] RRS experiments. Given the publication bias towards successful applications [110], it is likely that a large number of unsuccessful applications of this technology to nonmodel species exist. It is crucially important that researchers actively engage in the decision-making process when choosing restriction enzymes, size selection windows, and the number of individuals to be pooled per sequencing lane. Furthermore, the research objectives and budget should be critically evaluated and matched. In other words, investigating genome-wide polygenic adaptation patterns in a non-model species with large, complex genome may simply not be feasible on a small budget. The number of individuals to be included is another aspect that weighs in on these considerations and latest developments in the field enable the inclusion of this parameter in in silico simulations [111]. In situations where sampling is not restricted, inferences of spatial genetic structure for example may benefit more from wider geographic sampling coverage than from higher marker density. If sampling more localities is unfeasible as may be the case in the Antarctic realm, it can be beneficial to instead invest in high density sequencing (as in several markers per linkage group). With sufficient genome coverage even advanced coalescent modeling is possible using RRS data [112].
We recommend following a few guiding principles when planning RRS for population genomics (but see also e.g. [5,45,100]). First, clear targets with respect to the number of individuals to be screened in a project (and/or in follow-up projects) and the marker density needed for the research objective should be defined. Determining the necessary marker density is difficult and depends on the degree of linkage disequilibrium [34]. We aimed for and achieved in three out of five cases a marker density of at least one SNP per 10,000 bp or 100 SNPs per Mbp. How valuable these marker densities are will only become apparent after full scale sequencing projects and depends also on factors that cannot be controlled through the experimental setup. However, our optimization approach yielded marker densities considerably higher (median 68 SNPs per Mbp) than in the survey of Lowry et al. (2017) [34] (median 4.08 tags per Mbp). Second, in silico estimations of how these targets can be reached and approximations of the associated costs should be obtained. The number of markers and individuals must be matched to reach a certain coverage (e.g. an average target of 30×). Subsequently, it is useful to briefly evaluate the trade-offs and benefits of RRS and other methods. If a promising combination of RRS method, enzyme, size selection, and sequencing effort is found, it is often worthwhile to conduct a pilot experiment before running the full sequencing experiment (Fig. 3). However, it is also advisable to stick to one approach afterwards and not change for example the sequencing platform, the size window or other properties of the setup that will otherwise reduce comparability between datasets. Finally, it is also important to thoroughly test and optimize the bioinformatic processing and data filtering to obtain a robust population genomic dataset [21,113].

Alternative approaches
In some cases, RRS might not be the right choice for molecular ecological research (Fig. 3). A plethora of other genomic or genetic methods exists, which may offer more appropriate cost-benefit ratios. SNP genotyping arrays are a common and highly reproducible alternative, but usually only for species with more genomic resources (which exist for some Antarctic taxa; see e.g. [114]). Similarly, whole genome resequencing is providing the most extensive datasets which can be used for a wide range of analyses [11,12,115]. However, this is still too costly for many research projects, especially if information across many individuals and/or localities is needed. Another option is to focus on the expressed part of the genome and use a form of sequence capture enrichment (e.g. [103,116,117]) or RNAseq [118], or both [119,120]. These approaches are versatile and can provide valuable information, even for museum samples [121,122]. However, substantial expertise and prior investment in the development of custom methods is necessary for species that have not been investigated yet. With a limited budget and research objectives that do not depend on whole genome scans for selection, more classical molecular approaches are sometimes a good alternative. Nuclear microsatellite markers remain powerful to describe population structure and can be multiplexed and screened in large numbers. These markers can also benefit from high-throughput sequencing [123,124]. Mitogenome sequencing and assembly using long-range PCR is another useful approach, particularly for phylogeographic applications [125,126]. The amphipod and bird species evaluated here may currently be more amenable to such methods instead of RRS.

Conclusions
An extensive evaluation and optimization protocol allowed us to identify whether RRS is a suitable option for population genomics in a range of Antarctic animals. We have achieved promising results in some classes (ostracods, bivalves, sea stars, and fishes) that will be further developed soon. In other cases (amphipods and birds/degraded samples) alternative strategies such as mitogenome, capture sequencing or microsatellites seem more appropriate. The detailed considerations outlined here are a guideline for researchers to make informed decisions about the use of RRS or alternative methods. This is particularly important for species where genomic information remains scarce.

Specimen sampling
Samples of all target species were available from recent expeditions to the Southern Ocean (Additional File 1). For ostracods, we used existing DNA extractions of Macrocyprididae from the Southern Ocean that were already taxonomically identified and described [55,127]. The amphipod target species were collected during RV Polarstern [128] expedition ANTXXIX-3 PS81. More details on Eusirus pontomedon (note that we initially included these specimens tentatively as Eusirus aff. perdentatus, but the taxonomy was updated during the course of this project) are provided in [56], while details of investigated Charcotia obesa are given in [129]. The bivalves Laternula elliptica and Aequiyoldia eightsii were sampled by scuba diving in the shallow water of Potter Cove (King George Island, western Antarctic Peninsula; by F. Pasotti) and Rothera station (Adelaide Island, West Antarctic Peninsula; courtesy of the British Antarctic Survey) in 2016. Two sea star species (Bathybiaster loripes and Psilaster charcoti) were collected during international expeditions with RRS James Clark Ross and RV Polarstern to the South Orkney Islands (JR15005 in 2016, PS77 in 2011), the Weddell Sea (PS81 in 2013), West Antarctic Peninsula (PS77 in 2011), and with RV L'Astrolabe to Adélie Land (REVOLTA 1 in 2010). Emerald rockcods (Trematomus bernacchii) were sampled in 2014 around James Ross Island with gill nets [130]. Scaly rockcods (Trematomus loennbergii) were sampled in the Ross Sea as bycatch of the exploratory Antarctic toothfish (Dissostichus mawsoni) longline fishery. Dead birds and feathers of snow petrels (Pagodroma spp.) were sampled during the BELARE 2017-2018 expedition in the vicinity of the Princess Elisabeth Station, and additional samples were obtained from Signy and Adelaide Islands as courtesy of the British Antarctic Survey. Samples were stored frozen, dried, or in > 90% ethanol until DNA extraction.

Genomic resources
Prior to computational analyses, genomic information was collated for all target species or, if such information was not available, from the closest related species. Published reference genomes were collected from the literature and online resources, such as GenBank and Ensembl [131]. In addition, genome size estimates were retrieved from genomesize.com [132] and other published estimates based on flow cytometry (e.g. [69]). Genome size estimates as C values were transformed to Mb for comparison (1 pg = 978 Mb) [133].

In silico genome digestion analyses
We used SimRAD to computationally digest genomic DNA at sites matching a restriction enzyme recognition site [134]. In total, seven restriction enzymes and combinations thereof were tested (Table 3). These were chosen based on what is commonly used in comparable studies and to cover a variety of enzymes ranging from very common (MseI, MspI, ApeKI) to medium (EcoRI, SphI, PstI) and rare cutters (SbfI). Reference genomes from related species as well as two simulated genomes per taxonomic class were used for these in silico digestions. Simulated genomes were generated randomly using Sim-RAD, but with GC content as in the available reference genome(s) and with two different fixed sizes per taxonomic class to cover the approximate range of genome sizes known for this class ( Table 2). The total number of fragments that these enzymes (or enzyme combinations for double digest setups) produced were estimated, as well as the number of fragments in various size selection windows (between 210 and 260, 240-340, 0-100, 100-200, 200-300, 300-400, 400-500, 500-600, 600-700, 700-800, and 800-900 bp). Approximate targets for the number of fragments in each species of interest were defined (Table 2) and restriction enzyme and size selection combinations that provided fragment numbers close to our target numbers (50,000 ± 10,000 or 20,000 ± 10,000) were retained for downstream testing. After narrowing down the enzyme choice and conducting empirical digestion analyses, we ran additional in silico digestions for a final optimization of the size window and thus number of fragments for each specific case. During these fine-tuning analyses we tested as many different size selection windows as needed (in some cases > 20 additional size windows between 50 and 250 bp width) to find a suitable estimate of the number of fragments.

Empirical genome digestion analyses
Laboratory experiments were conducted with promising restriction enzymes to complement results from computational analyses. For each species, DNA from three individuals was used to test two or three restriction enzymes or enzyme combinations. Genomic DNA was extracted using either the commercial DNA extraction kits NucleoSpin Tissue (Macherey-Nagel) or DNeasy Blood & Tissue (Qiagen) and following the manufacturer's guidelines, or with a standard salting out protocol [86], or, for the bivalves, with a standard cetyl trimethylammonium bromide (CTAB) protocol. Subsequently, DNA quality and quantity were checked using the fluorescence assay Quant-iT PicoGreen dsDNA (Thermo Fisher Scientific Inc.), an Infinite M200 microplate reader (Tecan Group Ltd.) and 1% agarose gel electrophoresis. Whenever possible, only high-quality DNA extractions were used. Because of their small size, extractions from individual ostracods yielded insufficient quantities of DNA for downstream protocols, and sample numbers per locality were very low. Hence, the entire genomic DNA of ostracods was amplified using the REPLI-G kit (Qiagen) for whole genome amplification of 1 μL extracted DNA with high-fidelity polymerase Phi 20 and multiple displacement amplification following the manufacturer's protocol. For this purpose, extractions with the highest DNA concentrations from different species of Macrocyprididae, mainly of the Macroscapha tensa-opaca species complex, were selected [127]. For all target species, 100 ng genomic DNA of three biological replicates per species was digested with 10 units of a selected restriction enzyme at 37°C (EcoRI, MspI and PstI) or 75°C (ApeKI) for 2 h in a total volume of 10 μL. Reactions were purified with CleanPCR (GC Biotech) according to the manufacturer's protocol. Between 1 and 5 ng of the purified digested DNA was loaded on a High Sensitivity DNA chip (Agilent Technologies) and run on an Agilent 2100 Bioanalyzer System. Results were exported from the 2100 expert software (Agilent) as XML files and read into R v4.0.4 [135] using the bioanalyzeR package v0.5.1 [136]. Additional R packages used in this project were here v1.0.1 [137], seqinR v1.0-2 [138], the tidyverse packages [139], ggsci v2.9 [140], and gridExtra v2.3 [141] (see also more details under: https://github.com/notothen/radpilot). Because it is not possible to accurately standardize the number of fragments in an empirical digest without knowledge of the true genome size, we compared the shape of the curves of produced fragments (number of loci or DNA concentration vs. locus size or length) between in silico and empirical digests ( Fig. 1 and Additional File 3).

RRS setup optimization
In order to choose a promising restriction enzyme and size selection combination, we calculated the sequencing coverage per fragment as follows: We conservatively aimed at a coverage of approximately 30× for each fragment per individual, higher than other minimum recommendations [15,36,45]. Given that the accuracy of our genome size estimates is unknown, we aimed for relatively high coverage, so that in a "worst-case scenario", where the genome size is actually twice as large as we estimated (or any other factor leads to twice as many fragments as assumed), we would still reach a coverage of approximately 15×. The number of individuals per sequencing library was set to 96, corresponding to one PCR plate. Sequencing with a HiSeq 4000 platform (Illumina) should conservatively yield approximately 300 million reads per sequencing lane, while on a HiSeq 2500, we expect approximately 200 million reads. These coverage calculations were applied to fragment numbers from in silico results based on available reference genomes and extrapolated to a final, conservative estimate of genome size based on the best available knowledge (Table 4). This extrapolation is likely not biologically accurate but serves as a conservative correction factor. We then used in silico estimates again to further tweak the size window of a chosen restriction enzyme or enzyme combination in each target species to achieve the desired coverage, while considering the size range in which the two HiSeq machines work best. Finally, we estimated the number of SNPs across the genome as a measure of marker density (analogous to [34]) for a chosen enzyme and size selection setup and sequencing machine, assuming one SNP every other 100 bp. The latter estimate is based on our own experience, predominantly from fish genomes (but see also e.g. [142]). If an estimate of the naturally occurring SNP density across the genome is known for the target species or a related species, then this should be used. We provide an R function where any estimate can be used as input for marker density calculations. In general, all our calculations and plots should be reproducible with our spreadsheet tables and R scripts available at https://doi. org/10.5281/zenodo.5045574 and at https://github.com/ notothen/radpilot.

RRS library preparation and sequencing
The information collected so far convinced us not to pursue RRS in amphipods (see discussion); they were therefore not included in the test libraries. In addition, not enough high molecular weight DNA samples of P. nivea confusa (one of the snow petrel subspecies) were available. Eventually, five RRS test libraries for eight target species were constructed using 6, 8, 10, or 14 individuals and two controls per species and sequenced on one lane of a HiSeq 2500 unit (see Table 5 in results section). With this setup, we attempted to realize the previously estimated fixed variables for our coverage calculations, i.e. an estimated 250 million reads spread over 94 individuals and between 53,399 and 81,605 fragments. We originally aimed at 96 individuals, but too many samples of low-quality DNA dropped out during sample preparation. In addition, the estimated number of fragments varied between target species, but the conservative estimates in all other aspects should allow for some flexibility here. The libraries were all prepared by the same person at the KU Leuven laboratory using custom protocols that are based on two main references: the original ddRAD protocol by Peterson et al. (2012) [15] and the original GBS protocol by Elshire et al. (2011) [14]. We adjusted these protocols slightly and provide a full-length description of the laboratory procedure in Additional Files 5 & 6. In both cases, the standardized high-quality DNA was first digested with restriction enzyme(s), followed by adaptor and barcode ligation, purification, PCR, another purification and finally quantification and pooling. The libraries were then s e n t t o t h e K U L e u v e n G e n o m i c s C o re ( w w w . genomicscore.be), where all five libraries were individually size selected on a Pippin Prep unit (Sage Science), checked for quantity using qPCR, pooled, and pairedend sequenced on one lane of a HiSeq 2500 platform (Illumina).

Sequence analyses
Sequencing data were checked using FastQC v0.11.5 [143] and then demultiplexed and cleaned (options -c and -q) using the process_radtags module of Stacks v2.4 [98,99]. Because some of our multiplexing barcodes for the PstI library were contained in longer ApeKI barcodes, we demultiplexed the ApeKI libraries first and captured reads that were discarded in the process. These reads were subsequently used for demultiplexing of the PstI library. All demultiplexing runs were conducted without barcode rescue to avoid cross-contamination between libraries. The Stacks pipeline was also used for each target species independently to create a de novo assembly and call genotypes. Building contigs from pairedend reads is not possible with GBS data in Stacks [21], because the orientation of the reads is ambiguous. In this case (libraries 1, 2, 3, 5), we concatenated the four output files per individual of process_radtags to run the pipeline as if it was single-end data. Our size selection windows were designed to avoid overlap between the two reads of one fragment, so this approach should work well, albeit creating shorter haplotypes. We used Stacks' default value for m, i.e., a minimum stack coverage of 3, which generally produces consistent results at typical coverage rates [36]. Choosing parameters M and n to control the formation of loci within and across individuals on the other hand is study dependent. We explored a parameter range of n = M = [1 .. 9] following Rochette and Catchen [45] to strike a balance between over-and undermerging alleles and loci. To compare results from the different parameters only loci present in 80% of the samples (50% in the case of ostracods) were retained. Further detailed filtering would be required for downstream population genomic analyses.
Additional file 1. Samples used for reduced representation sequencing (RRS) optimization. DNA from these samples was used for empirical restriction enzyme digestions with different enzymes (single digest EcoRI, PstI, MspI, or double digest EcoRI-MspI) and for RRS pilot libraries. Some samples were extracted twice as replicates (marked as _rep in sample ID). Three samples per species (family in the case of ostracods) were used for empirical digestions. The amphipod (C. obesa and E. pontomedon) samples and one T. loennbergii were used for empirical digestions, but not included in any RRS library.
Additional file 2. In silico estimates of the number of fragments. Estimates were produced through in silico restriction enzyme digestions for reduced representation sequencing (RRS) optimized for approximately 30× coverage. The number of fragments depends on the restriction enzyme/combination, the size window, the assumed genome size, and the reference genome used for in silico digestion. Reference genomes of related species were used as well as simulated genomes; in this case the size and GC content used to simulate the genomes are listed. The number of fragments were extrapolated to the assumed genome size. Only two different enzyme and size selection setups per target species are listed here (for RRS setups optimized for HiSeq 2500 or HiSeq 4000 sequencing runs, respectively; the same as in Table 4, Table 5, Additional File 4); further estimates can be found in spreadsheets available at https://doi.org/10.5281/zenodo.5045574.
Additional file 3. Comparisons of empirical and in silico restriction enzyme digestions. Empirical Bioanalyzer results (left figure panels) with digested DNA are shown as concentration over fragment size and estimated loci numbers over locus size from in silico digestions (right figure panels) for all target taxa except fish (these are shown in Fig. 1).
Additional file 4. Reduced representation sequencing (RRS) setups for seven individually optimized protocols. These setups were optimized in order to be run on a HiSeq 4000 platform (Illumina). The choice of restriction enzyme(s) and size window was optimized to obtain approximately 30× coverage (or half that value in a worst-case scenario) with the assumed genome size (conservatively estimated based on available information, see Table 2). Marker density was estimated as a comparable measure to the metastudy by Lowry et al. (2017) [34].  [14]. The protocol is scaled for use with 192 samples and with restriction enzymes PstI or ApeKI; the reagent volumes can be scaled down/up to suit other sample numbers; if other enzymes are used, the respective reaction conditions must be adjusted.
Additional file 7. Results from parameter optimization for de novo assembly and genotyping. Eight parameter optimization series were conducted following Rochette & Catchen (2017) [45] to identify optimal parameters to genotype reduced representation sequencing (RRS) data with Stacks v2.4 (Rochette et al. 2019) [21]; one test series for each species/species complex. The Stacks parameter m was kept constant (m = 3), while parameters M and n were varied together from 1 to 9. Subsequently, only loci present in 80% of the samples were retained and for each M = n parameter the number of loci and polymorphic loci was plotted, as well as the proportion of these loci containing 0 to 10 or > 10 SNPs. In ostracods, the library contained DNA from a species-complex, resulting in very few shared loci across 80% of the samples. Therefore, in this case results based on loci shared by 50% of samples are shown. Optimal M = n values were decided in all cases with this information (and reported in Table 5). Note, however, that it is impossible to make absolute calls regarding the ideal value.