2.1 Samples and DNA extraction

Chicken blood samples originated from experimental inbred lines kept at the Institute for Animal Health at Compton, UK (lines 72, C, WL and N), and the Basel Institute for Immunology in Basel, Switzerland (lines H.B15 and H.B19⁺), as detailed in Jacob et al. (2000), Shaw et al. (2007) and Wallny et al. (2006). These lines carry seven common B haplotypes: B2 (line 72), B4 and B12 (line C), B14 (line WL, sometimes referred as W), B15 (H.B15), B19 (H.B19) and B21 (line N). All the lines used in this study are homozygotes (NCBI Accession nos.: AJ248572 to AJ248586). In each haplotype are two class IIB loci: BLB1 (previously known as BLBI or BLBminor) and BLB2 (BLBII or BLBmajor), with alleles now designated as BLB1*02 and BLB2*02 from the B2 haplotype, etc. All alleles have different nucleotide sequences, except BLB1*12 and BLB1*19. DNA was isolated from blood cells by a salting out procedure (Miller et al., 1988).

2.2 Generating 43 artificial MHC genotypes

We artificially generated 43 genotypes of varying allelic complexity by combining equimolar amounts of DNA samples from the seven MHC haplotypes mentioned above (Table 1; created genotypes listed in Table S1).

2.3 Optimal primers for chicken MHC class IIB

We targeted 241 bp of the 270 bp exon 2 of MHC class IIB, the polymorphic region known to code for antigen binding sites, using the primers OL284BL (5-GTGCCCGCAGCGTTCTTC-3) and RV280BL (5-TCCTCTGCACCGTGAAGG-3; Goto et al., 2002). The primers are not locus-specific and bind to both loci of the chicken B complex.

2.4 Cross-species primer design for chicken MHC class IIB

To replicate designing primers without any a priori knowledge of the species’ MHC Class IIB structure or sequences, we downloaded 61 exon 2 MHC class IIB sequences from seven Galliform species (Coturnix japonica, Crossoptilon crossoptilon, Meleagris gallopavo, Numida meleagris, Pavo cristatus, Perdix perdix and Phasianus colchicus, all accession numbers are listed in the Table S2) from GenBank (https://www.ncbi.nlm.nih.gov/genbank/). We then used Primer3 (Rozen & Skaletsky, 1999; Untergasser et al., 2012) to design the forward primer GagaF1 (5-WTCTACAACCGGCAGCAGT-3) and the reverse primer GagaR2 (5-TCCTCTGCACCGTGAWGGAC-3) aiming at amplifying 151 bp of exon 2. No species were given more weight than others during primer design, and all default parameters of Primer3 (concerning melting temperatures and structural settings) were kept. The only exception is that we allowed up to two degenerate positions in the primer sequence. Note here that the design of primers was deliberately suboptimal to convey the challenges of primer design. Optimal primer design should include sequences from other avian groups than Galliform species and consider phylogenetic relationships to weigh variants (Babik, 2010; Burri et al., 2014; see Discussion for more details on primer design).

2.5 PCR amplification, library preparation and high-throughput sequencing

For both data sets, we replicated all individuals to estimate repeatability (n_individuals = 43 and n_amplicons = 86). Individual PCRs were tagged with a 10-base pair identifier, using a standardized Fluidigm protocol (Access Array™ System for Illumina Sequencing Systems, ©Fluidigm Corporation). We first performed a target-specific PCR with the CS1 adapter and the CS2 adapter appended. To enrich base pair diversity of our libraries during sequencing, we added four random bases to our forward primer. The CS1 and CS2 adapters were then used in a second PCR to add a 10-bp barcode sequence and the adapter sequences used by the Illumina instrument during sequencing.

The first PCR consisted of 3–5 ng of extracted DNA, 0.5 units FastStart Taq DNA Polymerase (Roche Applied Science), 1× PCR buffer, 4.5 mM MgCl₂, 250 M of each dNTP, 0.5 M primers and 5% dimethyl sulphoxide (DMSO). The PCR was carried out with an initial denaturation step at 95°C for 4 min followed by 30 cycles at 95°C for 30 s, 60°C for 30 s and 72°C for 45 s, and a final extension step at 72°C for 10 min. The second PCR contained 2 L of the product generated by the initial PCR, 80 nM per barcode primer, 0.5 units FastStart Taq DNA Polymerase, 1× PCR buffer, 4.5 mM MgCl₂ 250 M of each dNTP, and 5% dimethyl sulphoxide (DMSO) in a final volume of 20 L. Cycling conditions were the same as those outlined above, but the number of cycles was reduced to ten. Reducing the number of PCR cycles, the elongation time within PCR cycles and omitting the final extension step is recommended to reduce the number of chimeras when coamplifying multiple loci, because most incomplete primer extensions that generate chimeras are thought to be formed in the final cycles of PCRs and during the final extension step (see Discussion; Judo et al., 1998; Lenz & Becker, 2008; Smyth et al., 2010). However, we chose to process samples using conventional PCR conditions, because a high number of cycles may be necessary in some study systems and we wanted to replicate conditions used in most MHC wildlife studies. Thus, we purposefully wanted to evaluate the robustness of our pipeline in the more challenging setting where a high number of artefacts might be generated due to suboptimal PCR conditions.

PCR products were purified using an Agilent AMPure XP (Beckman Coulter) bead cleanup kit. The fragment size and DNA concentration of the cleaned PCR products were estimated with the QIAxcel Advanced System (Qiagen) and by UV/VIS spectroscopy on an Xpose instrument (Trinean, Gentbrugge, Belgium). Samples were then pooled to equimolar amounts of DNA. The library was prepared as recommended by Illumina (MiSeq System Denature and Dilute Libraries Guide 15039740 v05) and was loaded at 7.5 pM on a MiSeq flow cell with a 10% PhiX spike. Paired-end sequencing was performed over 2 × 251 cycles.

2.6 Data analysis with the ACACIA pipeline

ACACIA consists of 11 consecutive steps of data processing. The software requires two nonstandard python libraries (pandas (McKinney, 2010) and biopython (Cock et al., 2009)) and six third-party software (fastqc (www.bioinformatics.babraham.ac.uk/projects/fastqc/), flash (Magoč & Salzberg, 2011), vsearch (Rognes et al., 2016), blast (Altschul et al., 1990), mafft (Katoh & Standley, 2013) and oligotyping (Eren et al., 2013)), which can all be installed with one command. The input files are any number of FASTq files, which are the current canonical output of the Illumina platform. The step-by-step workflow is described below:

Subsequently, putative alleles with very low frequency (both at the individual and population level) are scrutinized again. If the proportion of reads of a putative allele within an individual amplicon is less than 10 times lower than the next higher ranking allele, and if it is very similar (one single different base) to another, more frequent allele present in the same individual amplicon, that putative allele is considered an artefact and removed. Finally, if an individual amplicon has fewer than 50 sequences following all of the allele calling validation steps, it is eliminated. Users are able to change all parameter values but ACACIA recommends settings based on our benchmarking. The output of this step consists of four files:

To evaluate the pipeline, we calculated both allele calling accuracy and allele calling repeatability. Allele calling accuracy was calculated as the percentage of alleles that have been correctly called across replicates. This was done by comparing the predicted genotype to the genotype generated by ACACIA. All alleles that were dropped out or false positives were marked as inaccurately called alleles. Allele calling repeatability on the other hand was calculated as the percentage of alleles called in both replicates (including false positives). Note here that allele calling accuracy and repeatability are not necessarily correlated if allele calling errors are highly repeatable, that is either false positives or allele dropout are consistent across replicates. However, allele calling accuracy can only be calculated if the genotype is known a priori, which will not be available to most wildlife studies. We have therefore calculated both measures to investigate the pitfalls of relying on allele repeatability to validate a genotyping pipeline.

We investigated the best abs_nor and low_por settings for our data sets by first looking at the allele calling accuracy and repeatability at varying abs_nor values (range: 0–40, with low_por set at 0) first, and at varying low_por values (range: 0–0.02, with the optimal abs_nor, in our case 10) second. The latter is how we recommend users to find their optimal settings, although the range of abs_nor and low_por values to be investigated may vary across different data sets, depending on where the ‘peak’ optimal setting lies.

The pipeline is supervised by a configuration text file (config.ini), which is appended every time users enter one of the settings mentioned above. Users can avoid running ACACIA interactively (and run the whole workflow in a ‘hands-free’ mode) by providing a complete config.ini file at the beginning of the workflow. A template of a config.ini file is given in ACACIA’s repository (https://gitlab.com/psc_santos/ACACIA/blob/master/config.ini).

2.7 Data analysis with the AmpliSAS pipeline

To compare how ACACIA performed relative to an existing relevant pipeline, we applied the web server AmpliSAS pipeline to our chicken data sets (Sebastian et al., 2016). The default AmpliSAS parameters of a substitution error rate of 1% and an indel error rate of 0.001% for Illumina data were used. We then tested for the optimal ‘minimum dominant frequency’ clustering threshold for a given filtering threshold (i.e. 0.5% for the ‘minimum amplicon frequency’), by testing a set of thresholds of 10%, 15%, 20% and 25%. All clustering parameters tested gave an allele calling accuracy of 97%, but we chose the 25% clustering threshold because it was the only parameter that resulted in no false positives.

Subsequently, AmpliSAS filters for clusters that are likely to be artefacts, including chimeras and other low-frequency artefacts that have filtered through the clustering step (Sebastian et al., 2016). The default setting for the filtering of low-frequency variants (i.e. ‘minimum amplicon frequency’) is 3%. However, this value was far too high for our data sets, and we tested a range of filtering threshold between 0% and 1% at 0.1% intervals (i.e. 0%, 0.1%, 0.2%, etc.). We assessed the optimal filtering threshold using both allele calling accuracy and repeatability.

2.8 Statistical analyses

To analyse the relationship between the number of alleles amplified and amplification bias/artefacts, we used generalized additive mixed models using the ‘mgcv’ package (Wood, 2006) in r version 3.6.3 (R Core Team, 2020). The three response variables that were explored using a binomial error distributions corrected for overdispersion (aka quasibinomial) were as follows: proportion of reads assigned to an allele, proportion of reads that were nonchimeric artefacts and proportion of reads that were chimeras. The response variables number of chimeric variant and number of parental variants that were generating chimeric variants were analysed using a Poisson error distribution, with the latter corrected for overdispersion (aka quasi-Poisson). The fixed term number of alleles amplified was entered as a smoother and was limited to 6 estimated degrees of freedom. To control for pseudoreplication, sample ID was entered as a random factor.

3.1 Sequencing depth for each data set and proportion of artefacts detected using ACACIA

A total of 530,101 paired-end reads were generated for the optimal primer dataset, which amounted to an average of 6,164 reads per amplicon (n = 86). For the cross-species primer data set, 994,338 paired-end reads were generated, amounting to an average of 11,562 reads per amplicon (n = 86). The proportion of artefacts identified at each step of the ACACIA pipeline for the chicken data sets combined is illustrated in Figure 1. Workflow filtering removed the highest proportion of reads when filtering for singletons (13.6%) and chimeras (14.2%). After all filters, 66.4% of the original raw reads were used for allele calling.

3.2 Optimal settings of different workflows

We compared allele calling repeatability across a range of different abs_nor and low_por settings when using the ACACIA workflow to identify the optimal settings according to genotyping accuracy for our data sets. We first fixed the abs_nor setting at 10 and tested different low_por values and found that the optimal setting (i.e. the highest accuracy values) was 0 across both data sets (Figure 2a.). Setting higher low_por values resulted in a higher allele dropout rate, which led to lower accuracy and repeatability scores. We then tested the optimal abs_nor setting for a fixed low_por value of 0 and found that the optimal setting was 10 across both data sets (Figure 2b). An abs_nor value of 0 increased the rate of false positives, while a value above 10 increased the rate of allele dropout.

For the AmpliSAS workflow, we investigated the optimal filtering threshold and found differing optimal values between data sets. For the optimal primer data set, we found that the optimal filtering threshold was 0.3, while 0.5 was found to be optimal for the cross-species primer data set (Figure 2c).

3.3 AmpliSAS versus ACACIA: optimal primer data set

When using the optimal settings of the ACACIA workflow, comparison of results with expected genotypes revealed that nine alleles dropped out, no false positives were found (Table 2), and as a result, allele calling accuracy was 98.5% (Figure 2a,b). All instances of allele dropout derived from the B21 haplotype. For two genotypes, both BLB1*21 and BLB2*21 dropped out. For four genotypes, only BLB2*21 dropped out, and for one genotype, only BLB1*21 dropped out (Table 2). Allele calling repeatability was 97.7%.

Using the optimal settings in AmpliSAS across 86 genotypes, a total of 17 alleles dropped out and one false positive was found (Table 2), which resulted in an allele calling accuracy of 97% (Figure 2c). As with ACACIA, most allele dropouts (16 of 17) derived from the B21 haplotype. For three genotypes, both BLB1*21 and BLB2*21 dropped out. For nine genotypes, only BLB1*21 alleles dropped out, and for one genotype, only BLB2*21 allele dropped out. Finally for one genotype, the allele dropout was BLB1*04 and the same genotype had a false-positive allele (Table 2). Allele calling repeatability was 95.3%. Therefore, the ACACIA workflow resulted in higher allele calling accuracy and repeatability than the AmpliSAS workflow.

3.4 AmpliSAS versus ACACIA: chicken cross-species primer dataset

Using the optimal settings of ACACIA, we found a total of 134 allele dropouts across the 86 genotypes and allele calling accuracy was 77.8% (Figure 2a,b). However, all dropouts were from the alleles BLB1*04, BLB1*15 or BLB1*21, which were never called in the genotypes they were predicted to occur. Further comparison between the allelic reads and the primers revealed two mismatches at the 1st bp and 16th bp within the forward primer. Across the whole data set, only 13 (0.001%), 114 (0.01%) and 11 (0.001%) reads prior to applying any downstream quality filtering steps after merging corresponded to BLB1*04, BLB1*15 or BLB1*21, respectively. By comparison, the range for all other alleles was between 25,812 (2.79%) and 115,489 (12.49%) and the range for all artefact reads were between 1 (0.0001%) and 5,535 (0.60%). Therefore, all allele dropouts in the cross-species data set when using the ACACIA workflow are explained by primer mismatch leading to very poor amplification and sequencing of these alleles, which were well within the lower range of artefact reads. Since BLB1*04, BLB1*15 and BLB1*21 dropped out in all genotypes, allele calling repeatability between both replicates was 100% when using the ACACIA workflow, which highlights that relying on allele calling repeatability when validating a genotyping workflow can be misleading.

Using the optimal settings of AmpliSAS, we found 152 allele dropouts across all genotypes and allele calling accuracy was 75% (Figure 2c.). As above, 134 dropouts were due to a mismatch with the forward primer. The remaining 17 alleles that dropped out were BLB1*12 or *19 (13 alleles) and BLB2*14 (4 alleles) (Table 2). Allele calling repeatability between both replicates was 96.1%. Therefore, as with the optimal primer data set the ACACIA workflow resulted in higher allele calling accuracy and repeatability than the AmpliSAS workflow.

3.5 Relationship between number of alleles amplified and artefacts

In the optimal primer data set, when amplifying two alleles within an amplicon (i.e. within haplotype), all alleles amplified and the proportion of reads assigned to alleles ranged from 0.24 to 0.36 (Figure 3). The latter confirms the suitability of the primer set design for this model system. In contrast, in the cross-species data set, primer mismatch and systematic allele dropout for the alleles BLB1*04, BLB1*15 or BLB1*21 meant that three haplotypes had a single allele instead of two (Figure 3). In both data sets, the contribution of allelic variants to the proportion of reads decreased sharply with increasing number of alleles when amplifying less than 4–6 alleles but starts to level when amplifying more than 4–6 alleles (Figure 3, optimal data set GAMM: F_{3.477, 542} = 237.3, p-value < .001; cross-species data set GAMM: F_{4.779, 420} = 99.73, p-value < .001). Amplification efficiency was significantly different between alleles in both data sets (optimal data set GAMM: F_{12, 542} = 10.63, p-value < .001; cross-species data set GAMM: F_{9, 420} = 35.53, p-value < .001). Both alleles from the B4 and B21 haplotypes in the optimal data set and the BLB2*04 allele in the cross-species primer data set consistently amplified poorly when coamplifying with alleles from other haplotypes (Figure 3; see Figure S1 for multiple comparison post hoc test of allele amplification). In the optimal primer data set, the low amplification efficiency of the B21 haplotype when coamplifying with other haplotypes explains the high allele dropout of alleles from this haplotype in more complex genotypes (i.e. when coamplifying 10 or more alleles; Figure 3). In contrast, the higher sequencing depth of the cross-species data set meant that BLB2*04 allele did not drop out. However, we identified a primer mismatch between BLB2*04 allele and the second base pair of the reverse primer, explaining the lower amplification efficiency of this allele when coamplified with other alleles.

The proportion of sequences classified as artefacts was much higher for PCRs using the optimal primer set than when using the cross-species primer set (Figure 4a,b; nonchimeric artefact GAMM: F_{1, 74} = 2,669.1, p-value < .001; chimera: F_{1, 74} = 180.4, p-value < .001), which is likely due to the fact that the fragment length of the optimal primer data set was longer relative to the cross-species primer data set (241 bp versus 151 bp, respectively; see Discussion). For both data sets in this study, when considering nonchimeric artefacts, there was a positive relationship between the proportion of artefacts and the number of alleles amplified (Figure 4a.; GAMM: F_1,74 = 207.3, p-value < .001). There is a logarithmic relationship between the proportion of chimeric artefacts and the number of alleles amplified whereby the proportion of chimeric reads no longer increased with number of alleles amplified when amplifying more than 4–6 alleles (Figure 4b, GAMM: F_4.857,74 = 35.77, p-value < .001). The total number of unique chimeric reads also tended to follow a logarithmic relationship, whereby the number of unique chimeric variants seemed to no longer increase with the number of alleles amplified when amplifying more than 10 alleles (Figure 3c, GAMM: F_4.06,74 = 117.5, p-value < .001). The relationship between the total number of parental variants generating chimeras and the number of alleles amplified also levelled when amplifying more than six alleles (Figure 4d.; GAMM: F_{4.06, 74} = 117.5, p-value < .001).

4.1 AmpliSAS versus ACACIA

Experimentally generating variation in the number alleles amplified with amplicon of known chicken MHC class IIB genotypes allowed us to validate our ACACIA pipeline to genotype systems with high complexity at high accuracy and repeatability across replicates in the optimal primer data set. While we achieved higher allele calling accuracy and repeatability using ACACIA than the AmpliSAS web server pipeline, we do not claim that ACACIA will necessarily perform better than AmpliSAS with all data sets. To demonstrate the latter, we would need to test both pipelines on a larger number of data sets and/or on simulated data sets. In addition, while our pipeline should suit data generated with any high-throughput sequencing technologies, we have only tested ACACIA with paired-end Illumina sequencing technology. Here, we focus on ACACIA versus AmpliSAS; however, further discussion on how ACACIA compares with other genotyping workflows can be found in the supplementary material.

Biedrzycka et al. (2017) reviewed extensively different genotyping workflows according to sequencing depth using AmpliSAS in highly complex MHC system (up to 42 alleles within an individual). The authors achieved an allele calling repeatability <90% when coverage <5,000 sequences regardless of the genotyping workflow used and reached 99% with a sequencing depth of 20,000. While our study does not allow to assess the relationship between the number of alleles amplified within amplicon and sequencing depth using ACACIA, our results were consistent with Biedrzycka et al. (2017) since allele calling repeatability was 97.7% for the optimal primer data set and 100% for the cross-species primer dataset, which had an average sequencing depth of 6,164 and 11,562 reads per amplicon, respectively. For the optimal primer data set, regardless of the genotyping pipeline used, allele dropout occurred in genotypes with a high number of alleles within amplicon (for ACACIA, 8 out of 9, and for AmpliSAS, 12 out of 14 genotypes with allele dropouts had 10 alleles or more). Our optimal primers amplified all alleles at a similar efficiency when amplified within single haplotypes suggesting that the primers are indeed optimally designed. For all instances, allele dropouts were alleles from the B4 and B21 haplotype that amplified poorly when coamplified with alleles from other haplotypes. Higher sequencing depth will reduce or even remove such allele dropout instances (Biedrzycka et al., 2017). Indeed for the cross-species primer data set, sequencing depth was nearly twice as high, and there were no instances of allele dropout due to the ACACIA pipeline (all allele dropouts were due to primer mismatch; see subsequent subsection of the discussion) and allele calling repeatability was 100%. Therefore, in order to reach allele calling repeatability values <99%, we advise researchers to aim for a sequencing depth of at least 10,000 reads per amplicon when amplifying more than 4 alleles per amplicon and of 20,000 reads when amplifying more than 15 alleles regardless of the bioinformatic workflow used (Biedrzycka et al., 2017).

The most apparent benefit of using the amplisas web server is that it is relatively easy to use for users with limited knowledge of scripting languages (such as python, perl, C++ or r). However, we have noticed that a number of studies report results using default settings when applying the amplisas pipeline to their data set. We find this concerning since, as our study demonstrates, the default clustering and filtering parameters are unlikely to be optimal for most data sets. Indeed, allele calling accuracy was much lower when using the default settings (81.8%) as compared to the optimal settings (97%) in the optimal primer data set in our study, due to high allele dropout. We therefore strongly discourage users from using default settings and advise to permutate between different filtering and clustering parameters to find the best settings for their data set when using the amplisas pipeline. As most wildlife studies cannot assess allele calling accuracy, duplicating samples and relying on repeatability is the only feasible method for most research to optimize their amplicon-based genotyping workflow. However, authors should bear in mind that due to the high recurrence of amplification and sequencing errors, high repeatability in allele calling between replicates does not necessarily entail an error-free workflow and may be misleading. Therefore, careful design of primers and PCR conditions to reduce artefacts during amplification is crucial to maximize amplicon-based genotyping accuracy regardless of the bioinformatic tools used (see further discussion on this below).

An important disadvantage of the AmpliSAS web server is that at the time of writing, sequencing depth per amplicon was limited to 5,000 reads. For data sets with sequencing depth above 5,000 reads, AmpliSAS can be run locally but we found that, unlike the web server, the local version of AmpliSAS had limited documentation and troubleshooting was time-consuming. Once installed, ACACIA does not require users to have experience with scripting languages, allows genotyping with virtually unlimited sequencing depth and provides output data reporting the number of reads kept at each step of the pipeline. The latter should aid users when deciding upon optimal parameters and thresholds. As for the AmpliSAS pipeline, we advise to not use default parameters of ACACIA without critically assessing different parameters for each data set. In particular, we urge users to permutate between different settings of abs_nor and low_por parameters. We advise to first search for the optimal abs_nor setting with a fixed low_por parameter of 0 because it is likely that it is only necessary to change the low_por parameter setting from 0 in data sets with ultra-deep sequencing depth. If it is subsequently found that the optimal low_por setting is greater than 0, users should repeat the permuting step of abs_nor until the optimal settings are found. Of course, finding optimal settings requires the inclusion of replicates for at least a subset of the data set. We therefore recommend that a sufficient number of replicates are always included in genotyping runs to obtain sufficiently accurate repeatability values.

4.2 The challenge of designing primers for nonmodel organisms

A common approach for primer design in complex genomic regions of nonmodel organisms includes aligning multiple sequences of phylogenetically related species. By building primers on consensus sequences, researchers assume that oligos will also amplify the target region in the species of interest. However, knowledge about related species is often limited to very few individuals. This means that primers can be designed in regions that are polymorphic in the target species. As a consequence, certain allelic variants are not amplified and homozygosity is overestimated. Indeed, this proved to be the case in our cross-species primer data set, whereby two mismatches (1st bp and 16th bp) within the forward primer (19 bp long) were sufficient to prevent the amplification of three alleles (out of 13). Interestingly, a single base pair mismatch between the second base pair of the reverse primer and the BLB2*04 allele did not prevent the amplification of this allele, although it did suffer severely from low amplification efficiency when in competition with other alleles. High sequencing depth for the cross-species primer data set prevented this allele from dropping out, regardless of the genotyping pipeline used. Our study therefore highlights the importance of carefully designed primers for amplicon-based genotyping.

Two nonmutually exclusive strategies can be used to decrease allele dropout in nonmodel organism with no a priori information on the target region. First, designing multiple primers and combining them within a PCR reaction are known to reduce allele dropout due to primer mismatch and allele amplification bias (Marmesat et al., 2016). In addition, combining multiple primers within PCRs reduces the need to sequence multiple primer sets separately, considerably reducing the cost of using multiple primer sets to genotype a novel target region. Second, isolating introns and exons flanking the highly polymorphic exons of interest can help inform primer design and even result in locus-specific primers (Babik et al., 2008, 2009; Biedrzycka & Radwan, 2008; Burri et al., 2008, 2014; Gaigher et al., 2016; Gillingham et al., 2016), which reduces allele dropout due to primer mismatch and allows assigning alleles to loci. Indeed, an important set-back of multilocus amplification is that assigning alleles to loci in complex systems is frequently impossible, even when using recent phasing algorithms (Huang et al., 2019), limiting the use of many population genetic analyses (Gillingham et al., 2017). In addition, the total number of loci is currently likely to be underestimated in many species since recent gene duplication means that different genes often carry identical alleles (Gaigher et al., 2016; Worley et al., 2008).

The characterization of flanking regions have typically relied on cloning followed by Sanger sequencing (Burri et al., 2008, 2014; Gaigher et al., 2016; Gillingham et al., 2016) or genome walking techniques such as vectorette PCRs (Babik et al., 2008, 2009; Biedrzycka & Radwan, 2008). Both of these approaches are time-consuming and thus only allow the characterization of a limited number of individuals, which may underrepresent sequence diversity during primer design. The development of long-read HTS of single molecules, such as Pacific Biosciences single-molecule real-time sequencing or Oxford Nanopore sequencing, offers much promise in characterizing longer regions of novel multigene families and consequently to enable more informed primer design (O’Connor et al., 2019). For instance, Fuselli et al. (2018) were recently able to develop an assembly pipeline that combined long-read HTS from a single sample with de novo short-read HTS assembly of six samples to characterize a 9Kb region of the MHC Class II (DRB) locus in Alpine chamois Rupicapra rupicapra. More complex multigene families have also been characterized in primate species (Larsen et al., 2014), including the MHC (Gordon et al., 2016; Westbrook et al., 2015), using long-read HTS. An important limitation with long-read HTS technologies is that they have higher sequencing error rates than short-read HTS (the reported error rates are 16% per base compared with 0.1% for Illumina HTS) and lower sequencing output (5–20 Gb/run compared with 200–600 for Illumina HTS) (Bansal & Boucher, 2019; Reinert et al., 2015). Therefore, while long-read HTS will undoubtedly improve our understanding of the MHC and other multigene complexes structure and consequently nonmodel species primer design, population studies genotyping a large number of individuals are likely to continue to rely on amplicon-based genotyping from short-read HTS for the foreseeable future.

4.3 Relationship between number of alleles amplified and artefacts

By knowing the exact expected alleles for the chicken genotypes, we were able to quantify chimeric artefacts precisely (Figure 1). There were a higher proportion of chimeric and nonchimeric artefacts in the optimal primer data set than in the cross-species primer data set. The most likely explanation for the latter is the shorter sequence for the cross-species primer data set (151 bp) compared with the optimal primer data set (241 bp). A shorter fragment reduces the number of base pairs that can be erroneously substituted/deleted and the number of breaking points for chimera formation. In addition, it is likely that the probability of incomplete elongation is inversely related to fragment length. Thus, fragment length appears to be the dominant factor predicting the proportion of artefactual reads.

As expected, the proportion of reads that were nonchimeric artefacts increased linearly as the number of alleles amplified with an amplicon increased, which can be explained simply by the fact that there are an increasing number of possible artefacts that can be generated as the number of initial template variants increases. An unexpected result was that the proportions of chimeras did not increase with increasing number of alleles amplified with an amplicon, when amplifying more than 4–6 alleles. Similarly, when amplifying more than 10 alleles, the number of chimeric variants no longer increased with increasing number of alleles amplified within an amplicon. Such saturation in chimera generation beyond a threshold of alleles amplified is likely to be a by-product of allele PCR competition. Indeed, as demonstrated by our own data, there is amplification bias whereby some gene variants are amplified preferentially relative to others (Marmesat et al., 2016; Sommer et al., 2013). Therefore, a few gene variants (~3–6 gene variants) are preferentially amplified and most chimeras originate from these dominantly amplified variants and few chimeras are generated from the poorly amplified variants. Indeed, we found that the number of parental variants generating chimeras in our data set did not increase with increasing number of alleles amplified when amplifying more than 4–6 alleles. The nonlinear relationship between chimera generation and number of alleles amplified has important implications when considering sequencing depth needed to accurately genotype complex multigene families, since it suggests that linearly increasing sequencing depth for increasing number of alleles coamplified is not necessarily the optimal strategy. The challenges of dealing with chimeras in genotyping pipelines are discussed below in detail.

4.4 Chimeras in genotyping pipelines

The formation of artificial chimeras during amplification is an important source of artefacts in amplicon sequencing projects (Lenz & Becker, 2008; Smyth et al., 2010), including those with newer sequencing technologies (Laver et al., 2016). Chimeras are challenging to identify as artefacts because they resemble real alleles generated by recombination, particularly in multigene families under high rates of interlocus genetic exchange (‘concerted evolution’), which is common in many MHC systems (Burri et al., 2008, 2010; Edwards et al., 1995; Gillingham et al., 2016; Hess & Edwards, 2002; Wittzell et al., 1999). Our results suggest that chimeras are more prevalent, harder to identify and potentially more reproducible across technical replicates than previously assumed. We expect the same to be true for similar projects with conserved, yet variable amplification targets such as the MHC.

One allele erroneously called as a real variant (i.e. a false positive) by the AmpliSAS pipeline in the optimal primer data set was actually a chimera between the BLB1*21 and BLB2*21 alleles. Furthermore, when using the AmpliSAS pipeline, 15 allele dropouts in the cross-species primer dataset were due to erroneous assignment of real allelic variants as chimera artefacts. Indeed, the BLB1*12 or *19 allele was identical to potential chimeric artefact sequences between BLB2*14 (85 possible breakpoints) and any of the following alleles: BLB1*04, BLB2*15, BLB2*19, BLB2*21 or BLB1*21 (Figure 5a.). In addition, BLB2*14 dropped out because it is identical to a chimera formed between the BLB1*02 and BLB1*12 or *19 alleles (33 breakpoints; Figure 5b.).

We have identified two factors that seemed to enhance chimera formation and challenge the distinction between artefact and real allelic variants. First, the combination of multiple real ‘parent’ sequences can yield the same chimeras, as illustrated in our examples in Figure 5a and Figure 5b, whereby any breakpoint in the shaded areas leads to the same chimeras. Second, peripheral breakpoints (Figure 5c) can generate chimeras that differ to parental sequences by as little as a single base pair. For instance, a chimera could be a product of the allele BLB2*21 combined with any of the other alleles shown in the alignment, with a breakpoint within the shaded area (Figure 5c.). Since the potential breakpoints are at the very end of the sequence, the chimera is very similar to one of its parents (in this example, it is different from BLB2*21 by only one base). In an attempt to deal with this issue as much as possible, we changed the default settings of VSEARCH so that chimeras can be detected even if they differ from one parent by one single base. Both the ‘multiple parents’ and the ‘peripheral breakpoints’ issues are likely to contribute to making chimeras reproducible across replicates.

Our study highlights the challenges of chimeras for amplicon-based genotyping. We purposefully used conventional PCR conditions to replicate methods used by most wildlife MHC studies. However, the formation of most chimeras is known to occur during the final cycles of PCR amplification when dNTP and primer concentrations are low and when incompletely elongated sequences are high (Judo et al., 1998; Lenz & Becker, 2008; Smyth et al., 2010). When target primers and dTNPs concentration are low during the latter stages of PCR cycles, incompletely elongated sequences act as primers and bind with the wrong sequences generating chimeras. Chimera formation during amplification can be simply reduced by adjusting the ratio of DNA template to dNTP and primer concentrations, reducing the number of cycles, increasing the extension step within PCR cycles and omitting the final extension step (which elongates high concentrations of incomplete chimeric sequences; Judo et al., 1998; Lenz & Becker, 2008; Smyth et al., 2010). Therefore, prior to any amplicon-based genotyping study, we advise researchers to reduce artefacts, including chimeras, during the wet laboratory stage of their workflow by applying carefully designed and optimal PCR conditions. Practices during the wet laboratory that reduce artefact generation in the first place are likely to be the most effective way of reducing genotyping errors regardless of the bioinformatic allele calling workflow used.