QTL mapping in autotetraploids using SNP dosage information
 2k Downloads
 26 Citations
Abstract
Key message
Dense linkage maps derived by analysing SNP dosage in autotetraploids provide detailed information about the location of, and genetic model at, quantitative trait loci.
Abstract
Recent developments in sequencing and genotyping technologies enable researchers to generate highdensity single nucleotide polymorphism (SNP) genotype data for mapping studies. For polyploid species, the SNP genotypes are informative about allele dosage, and Hackett et al. (PLoS ONE 8:e63939, 2013) presented theory about how dosage information can be used in linkage map construction and quantitative trait locus (QTL) mapping for an F_{1} population in an autotetraploid species. Here, QTL mapping using dosage information is explored for simulated phenotypic traits of moderate heritability and possibly nonadditive effects. Different mapping strategies are compared, looking at additive and more complicated models, and model fitting as a single step or by iteratively reweighted modelling. We recommend fitting an additive model without iterative reweighting, and then exploring nonadditive models for the genotype means estimated at the most likely position. We apply this strategy to reanalyse traits of high heritability from a potato population of 190 F_{1} individuals: flower colour, maturity, height and resistance to late blight (Phytophthora infestans (Mont.) de Bary) and potato cyst nematode (Globodera pallida), using a map of 3839 SNPs. The approximate confidence intervals for QTL locations have been improved by the detailed linkage map, and more information about the genetic model at each QTL has been revealed. For several of the reported QTLs, candidate SNPs can be identified, and used to propose candidate trait genes. We conclude that the high marker density is informative about the genetic model at loci of large effects, but that larger populations are needed to detect smaller QTLs.
Keywords
Quantitative Trait Locus Quantitative Trait Locus Mapping Single Nucleotide Polymorphism Marker Quantitative Trait Locus Location Potato Cyst NematodeIntroduction
Much progress has been made in recent years in developing methods for linkage analysis and QTL mapping in autotetraploid species. New genotyping technologies, such as the Illumina Infinium platform and sequencingbased methods, such as RAD sequencing (Baird et al. 2008) or genotyping by sequencing (Elshire et al. 2011), are enabling the construction of highdensity single nucleotide polymorphism (SNP) maps. Some of these technologies carry information not simply on presence/absence of an allele at a SNP locus, but about the actual allele dosage, which provides important information in polyploid species.
Hackett et al. (2013) have developed the statistical theory for using SNP dosage information to estimate recombination frequencies between SNPs in an autotetraploid population, and have applied this to a fullsib potato (2n = 4x = 48) population derived from a cross between processing clone 12601ab1 and the cultivar Stirling. This paper also developed methodology for interval mapping of quantitative trait loci (QTL), using the dosage information. The methodology used a hidden Markov model (HMM) (Rabiner 1989) to estimate genotype probabilities for each offspring along the chromosomes, and then modelled trait values as an additive function of the homologue effects, weighted by the genotype probabilities. This builds on previous work by Hackett et al. (2001) on QTL mapping in autotetraploids, which extended the mixture model approach to mapping QTL of Jansen (1992). The QTL model was applied to analyse allele intensity ratios of SNPs from the Illumina Infinium platform as a check on the position and phase of the SNP markers.
The allele intensity ratios are uncomplicated traits for QTL mapping, in that they are expected to have a very high ratio of genetic to random variance, to map to a single chromosomal location and to follow an additive function of the homologue effects. If a SNP has parental genotypes AABB × AABB, the offspring are expected to have genotypes AAAA, AAAB, AABB, ABBB and BBBB in a 1:8:18:8:1 ratio and the offspring allele intensity ratios are expected to cluster around values of 0.0, 0.25, 0.5, 0.75 and 1.0 (i.e. an additive function of the number of B alleles present). In general, we are interested in mapping traits with lower heritabilities, which may be affected by more than one QTL and where there may be dominance or other nonadditive effects.
In this paper, we present a simulation study on the efficacy of QTL interval mapping based on SNP dosages to model more general phenotypic traits of low to moderate heritability. We investigate the threshold LOD score for declaring a QTL present, and compare the additive model with a more complicated model for the 36 possible QTL genotypes. We then apply the model to reanalyse QTL data on this population for flower colour from Bradshaw et al. (2008), for maturity, height and resistance to late blight (Phytophthora infestans (Mont.) de Bary) from Bradshaw et al. (2004) and on resistance to potato cyst nematode (PCN, Globodera pallida) from Bryan et al. (2004).
Materials and methods
Potato mapping population and linkage map
The cross between the potato processing clone 12601ab1 and the cultivar Stirling has been studied extensively, and different linkage maps have been published. An early AFLP map based on 78 F_{1} offspring was published by Meyer et al. (1998) and a more detailed AFLP and SSR map, based on the full population of 227 F_{1} offspring was published by Bradshaw et al. (2008). This map contained 453 mapped markers and identified some, but not all, of the chromosomal groups based on the locations of SSR markers together with shared markers from the ultra high density (UHD) mapping population SH × RH (van Os et al. 2006). Stirling, 12601ab1 and 190 F_{1} offspring have recently been genotyped using the Infinium 8303 potato SNP array (Felcher et al. 2012). A SNP linkage map, using the dosage information, was generated by Hackett et al. (2013). This map assigned locations to 3839 SNPs, and identified all the chromosomal groups. Simplex SNPs identified 46 of the 48 chromosomes for Stirling and 45 chromosomes for 12601ab, giving a very comprehensive linkage map. SNP marker positions in the potato reference genome provided for QTL locations are obtained using the version 4.03 potato genome pseudomolecules (Potato Genome Sequencing Consortium 2011; Sharma et al. 2013).
Phenotypic traits
In this paper, we reanalyse some phenotypic traits from previous publications on this population by Bradshaw et al. (2004, 2008), Bryan et al. (2004). These papers describe the data collection and preliminary analysis including the estimation of heritability in detail. The following traits were analysed:
Flower colour (Fc), from Bradshaw et al. (2008), scored as a qualitative trait: blue like 12601ab1 or white like Stirling, for the clones that flowered over 3 years of trials (178 out of 190) from 1999 to 2001.
Maturity (Mat), from Bradshaw et al. (2004), scored on a one (early, all plants dead) to nine (late, all plants still green) scale in 1999 and 2000 and analysed as clone means over the 2 years. The heritability was estimated as 0.916.
Canopy height (Ht), from Bradshaw et al. (2004), measured in centimetres (cm) from the top of the drill in 1999, 2000 and 2001 and analysed as clone means over the 3 years. The heritability was estimated as 0.876.
 1.
Fourth field assessment of foliage blight (Fb4), on a 1–9 scale of increasing resistance, on 11 August 1998, which proved the optimal date for discriminating between clones. The heritability was estimated as 0.878.
 2.
Glasshouse assessment of tuber blight (Tb %) as a percentage of infected tubers in 1999. The heritability was estimated as 0.870.
 3.
Wholeplant glasshouse assessment of the presence/absence of Stirling’s major Rgene (Rgene). It was possible to categorise all 190 clones. Further details of this assessment using a simple and complex race are given in Stewart et al. (2003).
Resistance to the white potato cyst nematode Globodera pallida (PCN), from Bryan et al. (2004), scored as the count of cysts and analysed after a square root transformation. The heritability was estimated as 0.937.
Method for QTL mapping
 1.
Estimate the QTL genotype probabilities at each SNP from the parental genotypes and phases and the offspring dosages. This uses a Hidden Markov Model (HMM) and gives a 36 × s matrix P _{G} for each offspring G, G = 1…n with the probability of each of the 36 possible QTL genotypes π at the s SNPs on a linkage group.
 2.
Interpolate QTL genotype probabilities between the SNPs at a 1 centiMorgan (cM) spacing along each chromosome, using a cubic smoothing spline.
 3.
Model the trait values as a function of the QTL genotype, using a normal mixture model with a constant variance.
Fitting a normal mixture model to the trait values
Jansen (1992) developed a general mixture model for QTL mapping and showed how the model fitting can be separated into two steps, a weighted regression step where the trait values are regressed on the QTL genotypes, weighted by their probabilities P _{G} (derived initially from the HMM in step 1), and an updating step where the QTL genotype probabilities are updated from marker data and the current estimates of the QTL model parameters.
The conditional probabilities \(p(\pi_{G} Y_{G},x_{G}) = p(\pi_{G} x_{G})f(Y_{G} \pi_{G})/f(Y_{G} x_{G}) = P_{G} f(Y_{G} \pi_{G})/f(Y_{G} x_{G})\) where P _{ G } are the QTL genotype probabilities from the HMM. The likelihood equation can be solved (Jansen 1992) using an iterative approach based on the EM algorithm, alternating an Expectation step updating the conditional probability with a Maximisation step of a weighted regression to calculate the parameters of the QTL model \(f(Y_{G} \pi_{G})\) until the likelihood converges. The conditional probabilities are the QTL genotype probabilities P _{G} from the HMM for the first step, and are then updated using the probability distribution function \(f(Y_{G} \pi_{G})\) and rescaled to sum to one. Our experience with diploid QTL mapping software such as MapQTL 5 (van Ooijen 2004) is that occasionally the number of iterations can be high and spurious QTLs can be inferred in regions where the marker information is sparse: to avoid this here, a maximum of ten iterative steps was imposed. The LOD score is calculated as log_{10}(likelihood ratio) = log_{10}(L) − log_{10}(L _{0}), where L _{0} is the likelihood of the trait data in the absence of a segregating QTL i.e. that the trait has a normal distribution with a single mean for the population.
An alternative model fitting approach omits the iterative updating of the weights and uses a single regression step, weighted by the QTL genotype probabilities P _{G} from the HMM. This approach is used in some diploid QTL mapping software such as the procedures in GenStat 15 for Windows (Payne et al. 2012) and some of the options in R/QTL (Broman et al. 2003). For a dense marker map, this should be a good approximation to the iterative procedure. The simulation study compares these two model fitting procedures.
The form of the QTL model \(f(Y\pi)\)
In this model, the constant µ _{ C } is harder to interpret because of the constraints and does not correspond either to the trait mean or to a specific genotype mean.
Exploration of simpler models

Simplex QTL Qqqq × qqqq, which segregates in a 1:1 ratio

Duplex QTL QQqq × qqqq, either segregating in a 1:4:1 ratio qqqq:Qqqq:QQqq where there is additivity or partial dominance, or 5:1 Qqq:qqqq where there is complete dominance

Doublesimplex QTL Qqqq × Qqqq, either segregating in a 1:2:1 ratio qqqq:Qqqq:QQqq where there is additivity or partial dominance, or 3:1 Qqq:qqqq where there is complete dominance
The best model among these can be identified using an information criterion such as the minimum Schwarz information criterion (SIC) (Schwarz 1978): \(SIC =  2\log L + p\log m_{o}\) where L is the likelihood for the simple model, p is the number of parameters in the simple model and m _{ o } is the number of observations (i.e. the 36 genotype means).
This strategy fits an additive model, but then tests for nonadditive effects at the most likely QTL location.
QTL simulation study
Hackett et al. (2013) carried out a small study simulating offspring data from the estimated linkage maps to look at the proportion of the offspring genotypes that were estimated correctly by a HMM analysis of the parental genotypes and phases and the offspring dosages. From this study, chromosome XII was picked as a group of short length (87 cM), average SNP density (118 SNPs in the linkage group, or 1.35 markers per cM) and among the best reconstructed (the mean proportion of genotypes estimated correctly was 0.877) to form the basis of the simulation study here. Sets of 200 or 400 offspring were simulated from the parental genotypes and phases for the map of chromosome XII, with an additional marker (the QTL) included at position 27 cM. This marker was only used to simulate trait data and was not used for QTL interval mapping. Different configurations of trait data were estimated for QTL mapping, using the formulae in “Appendix” to estimate the size of allele effect that results in a QTL explaining a given proportion of the trait variance. A single set of SNP data was used for each configuration, and multiple traits were generated within each configuration as a constant m (set equal to 10.0), plus the genotype effects, plus a normally distributed term for environmental variation.
 Set 1a

Random data (100 traits)
 Set 1b

Random data (100 traits and 400 offspring)
 Set 2a

Simplex QTL, expected to explain 15 % of the trait variance, on homologue h1 from parent 1
 Set 2b

Simplex QTL, expected to explain 10 % of the trait variance, on homologue h2 from parent 1
 Set 2c

Simplex QTL, expected to explain 5 % of the trait variance, on homologue h2 from parent 1
 Set 2d

Simplex QTL, expected to explain 5 % of the trait variance, on homologue h2 from parent 1, using 400 offspring
 Set 3a

Duplex additive QTL, expected to explain 10 % of the trait variance, on homologues h6 and h8 from parent 2
 Set 3b

Duplex dominant QTL, expected to explain 10 % of the trait variance, on homologues h1 and h2 from parent 1
The traits in each of these sets were analysed using the additive model (1) to find the location with the highest LOD score on chromosome XII and to see how well the QTL configuration was reconstructed. The random data (set 1) and the three sets with a QTL explaining 10 % of the trait variance were also analysed using the complete model (2). In both cases, the results from fitting the model as a single weighted regression were compared to fitting the model by an iteratively reweighted regression process. The three sets with a QTL explaining 10 % of the trait variance were analysed further to identify whether there were twoallele models compatible with the genotype means.
Here, we ran a minimum of 200 permutations for each scenario. From the above equations, the approximate 95 % confidence interval is then given by the L = 184th and the U = 197th ordered maximum LOD scores. A further 300 permutations were analysed if necessary, and the 465th and the 484th ordered values of the combined sets were used as the 95 % confidence interval. For the simulated sets, the permutations were obtained as four permutations of each of the 50 traits in each set, which were analysed to give a single threshold for each set.
Results
Simulation study
Analysis of simulation set 1: random data
Simulation set 1a
Parameter  Additive model  Complete model  

With iteration  Without iteration  With iteration  Without iteration  
Mean  SD  Mean  SD  Mean  SD  Mean  SD  
Position  41.0  29.44  39.9  26.84  39.8  28.54  35.9  16.3 
µ _{ C }  10.0  0.44  10.0  0.39  10.2  0.62  10.2  0.49 
R ^{2}  3.9  2.34  2.1  1.67  15.9  6.67  2.9  3.12 
LOD  2.5  0.81  2.4  0.79  11.3  1.92  9.5  1.74 
rms  0.96  0.095  0.98  0.095  0.84  0.114  0.98  0.106 
Simulation set 1b
Parameter  Additive model  Complete model  

With iteration  Without iteration  With iteration  Without iteration  
Mean  SD  Mean  SD  Mean  SD  Mean  SD  
Position  41.2  26.46  40.4  23.41  43.5  28.8  38.4  17.66 
µ _{ C }  10.0  0.28  10.0  0.26  10.0  0.39  10.0  0.31 
R ^{2}  1.6  1.08  0.94  0.775  8.1  4.21  1.3  1.27 
LOD  2.3  0.78  2.3  0.77  10.7  1.63  9.5  1.41 
rms  0.98  0.066  0.99  0.066  0.91  0.066  0.98  0.068 
Analysis of simulation set 2: simplex QTLs, using the additive model
Simulation set 2a
Parameter  With iteration  Without iteration  

Mean  SD  Mean  SD  
Position  28.5  4.80  28.4  4.75 
µ _{ C }  11.64  0.291  11.59  0.266 
α_{2}  −0.87  0.172  −0.83  0.162 
α_{3}  −0.85  0.172  −0.82  0.160 
α_{4}  −0.84  0.164  −0.80  0.153 
α_{6}  0.02  0.182  0.02  0.158 
α_{7}  0.07  0.195  0.06  0.183 
α_{8}  0.00  0.180  0.01  0.163 
R ^{2}  15.8  4.64  14.0  4.17 
LOD  8.4  2.25  8.3  2.23 
rms  0.99  0.094  1.01  0.094 
One (two) LOD support interval  47 (50)  47 (50)  
LOD permutation threshold (N = 200)  3.85 (3.48, 4.15)  3.54 (3.25, 4.01) 
Simulation set 2b
Parameter  With iteration  Without iteration  

Mean  SD  Mean  SD  
Position  30.0  10.85  31.1  9.03 
µ _{ C }  9.93  0.365  9.95  0.327 
α_{2}  0.70  0.205  0.67  0.192 
α_{3}  0.02  0.230  0.02  0.217 
α_{4}  0.00  0.186  0.00  0.173 
α_{6}  0.02  0.231  0.02  0.213 
α_{7}  −0.01  0.223  −0.01  0.208 
α_{8}  0.02  0.222  0.02  0.200 
R ^{2}  12.0  4.91  10.2  4.06 
LOD  6.5  2.16  6.4  2.12 
rms  0.98  0.118  1.00  0.119 
One (two) LOD support interval  40 (48)  40 (47)  
LOD permutation threshold (N = 200)  3.92 (3.61, 4.57)  3.87 (3.47, 4.47)  
LOD permutation threshold (N = 500)  3.96 (3.70, 4.30)  3.87 (3.57, 4.13)  
No. significant traits  42  43  
Position  29.4  11.37  30.8  9.22 
α_{2}  0.74  0.195  0.70  0.182 
R ^{2}  13.3  4.12  11.2  3.30 
rms  0.97  0.120  0.99  0.120 
One (two) LOD support interval  32 (40)  33 (40) 
Two hundred permutations were run initially for this data set. The LOD scores were above the upper bound of the permutation threshold for 39 traits, with and without iteration, and below the lower threshold for four traits without iteration and three traits with iteration. A further 300 permutations were run to try to resolve the remaining traits. Table 4 shows the thresholds and confidence limits for both the original 200 permutations and the combined set of 500 permutations, which decreases the size of the confidence interval. Based on 500 permutations, 43 were above the upper threshold without iteration, four were below the lower threshold and three were still unresolved. The corresponding figures with iteration were 42 traits above the upper threshold, four below and four unresolved. We focus here on the traits above the upper threshold. α_{2} was the only significant coefficient for 30 of the 43 traits above the threshold without iteration, and for 29 of the 42 traits above this threshold with iteration: the remainder identified additional homologues as making a significant contribution. The means for the significant traits, at the bottom of Table 4, showed higher values of α_{2} and R ^{2} and a lower rms, especially using the model with iteration.
Simulation set 2c
Parameter  With iteration  Without iteration  

Mean  SD  Mean  SD  
Position  31.1  18.73  31.3  18.28 
µ _{ C }  10.0  0.507  10.0  0.453 
α_{2}  0.445  0.269  0.424  0.241 
R ^{2}  7.4  2.85  5.8  2.47 
LOD  4.3  1.22  4.2  1.22 
rms  0.97  0.088  0.99  0.086 
One (two) LOD support interval  32 (48)  34 (48)  
LOD permutation threshold (N = 500)  4.11 (3.85, 4.29)  3.95 (3.77, 4.11)  
No. significant traits  28  29  
Position  28.5  17.65  28.5  17.34 
α_{2}  0.520  0.215  0.483  0.203 
R ^{2}  9.3  1.72  7.4  1.53 
rms  0.96  0.094  0.98  0.093 
One (two) LOD support interval  17 (26)  19 (27) 
Simulation set 2d
Parameter  With iteration  Without iteration  

Mean  SD  Mean  SD  
Position  26.3  11.59  27.0  10.62 
µ _{ C }  9.9  0.277  9.9  0.235 
α_{2}  0.488  0.149  0.461  0.136 
R ^{2}  6.1  2.41  5.1  2.25 
LOD  6.3  2.10  6.2  2.12 
rms  0.98  0.063  0.99  0.061 
One (two) LOD support interval  42 (49)  43 (50)  
LOD permutation threshold (N = 500)  3.87 (3.72, 4.25)  3.74 (3.60, 4.18)  
No. significant traits  41  41  
Position  24.5  7.58  25.4  6.86 
α_{2}  0.513  0.149  0.488  0.130 
R ^{2}  6.8  1.98  5.8  1.88 
rms  0.98  0.063  0.99  0.061 
One (two) LOD support interval  33 (40)  35 (41) 
Analysis of simulation set 3: duplex QTLs, using the additive model
Simulation set 3a
Parameter  With iteration  Without iteration  

Mean  SD  Mean  SD  
Position  26.9  13.10  27.0  12.25 
µ _{ C }  10.0  0.316  10.0  0.289 
α_{2}  0.022  0.198  0.016  0.168 
α_{3}  −0.015  0.204  −0.011  0.189 
α_{4}  −0.008  0.196  −0.009  0.180 
α_{6}  0.607  0.195  0.564  0.174 
α_{7}  −0.009  0.217  0.001  0.192 
α_{8}  0.628  0.170  0.576  0.158 
R ^{2}  12.4  5.23  9.7  3.86 
LOD  6.4  2.07  6.2  2.00 
rms  1.00  0.135  1.030  0.131 
One (two) LOD support interval  44 (48)  46 (48)  
LOD permutation threshold (N = 500)  4.08 (3.96, 4.28)  3.99 (3.84, 4.12)  
No. significant traits  40  42  
Position  27.2  13.59  27.2  12.26 
α_{6}  0.649  0.188  0.591  0.170 
α_{8}  0.658  0.167  0.602  0.154 
R ^{2}  13.9  4.58  10.7  3.339 
rms  0.98  0.135  1.02  0.129 
One (two) LOD support interval  34 (38)  38 (40) 
Simulation set 3b
Parameter  With iteration  Without iteration  

Mean  SD  Mean  SD  
Position  29.8  16.55  29.4  15.24 
µ _{ C }  11.5  0.318  11.4  0.302 
α_{2}  −0.031  0.193  −0.039  0.183 
α_{3}  −0.592  0.175  −0.555  0.162 
α_{4}  −0.574  0.172  −0.542  0.160 
α_{6}  −0.041  0.180  −0.024  0.172 
α_{7}  −0.062  0.156  −0.041  0.155 
α_{8}  −0.053  0.207  −0.042  0.197 
R ^{2}  8.5  3.36  7.1  3.24 
LOD  4.8  1.58  4.7  1.59 
rms  1.03  0.114  1.05  0.115 
One (two) LOD support interval  38 (47)  37 (46)  
LOD permutation threshold (N = 500)  4.10 (3.85, 4.52)  3.96 (3.76, 4.43)  
No. significant traits  26  27  
Position  25.0  12.66  26.1  11.48 
R ^{2}  10.93  2.572  9.27  2.542 
rms  0.999  0.097  1.017  0.099 
One (two) LOD support interval  18 (22)  20 (24) 
Analysis using the complete model
Analysis of the traits where a QTL explains 10 %, using the complete model
Parameter  With iteration  Without iteration  

Mean  SD  Mean  SD  
Set 2b: simplex QTL  
Position  32.9  20.15  33.5  12.80 
R ^{2}  19.8  6.09  10.1  4.53 
LOD  14.5  2.36  13.2  2.27 
rms  0.891  0.120  1.00  0.129 
One (two) LOD support interval  14 (25)  24 (32)  
LOD permutation threshold (N = 500)  14.72 (14.44, 15.50)  12.89 (12.36, 13.23)  
Set 3a: additive duplex QTL  
Position  33.5  20.99  32.8  11.85 
R ^{2}  19.3  6.36  9.7  4.92 
LOD  14.3  2.61  13.0  2.50 
rms  0.920  0.132  1.03  0.130 
One (two) LOD support interval  15 (28)  23 (33)  
LOD permutation threshold (N = 500)  14.95 (14.49, 15.43)  13.01 (12.64, 13.44)  
Set 3b: dominant duplex QTL  
Position  31.2  19.87  30.7  9.63 
R ^{2}  18.2  5.93  9.2  3.51 
LOD  13.8  1.79  12.7  1.78 
rms  0.920  0.109  1.02  0.111 
One (two) LOD support interval  20 (34)  19 (29)  
LOD permutation threshold (N = 500)  14.41 (14.07, 14.60)  12.68 (12.22, 12.90) 
Testing for simpler models using the 36 genotype means from the complete model and the additive model
Testing for simpler models for the simulation sets explaining 10 % of the trait variance
Complete model  Additive model  

Set 2b: simplex QTL  
Number significant  27  43 
Correct model selected (qQqq × qqqq)  16  29 
Correct parent selected  4  1 
Other simple model selected  6  11 
Additive model selected  1  2 
Set 3a: additive duplex QTL  
Number significant  19  42 
Correct model selected (qqqq × qQqQ, with 1:4:1 ratio)  14  31 
Correct parent selected  4  9 
Other simple model selected  1  0 
Additive model selected  0  2 
Set 3b: dominant duplex QTL  
Number significant  24  27 
Correct model selected (QQqq × qqqq, with 5:1 ratio)  18  21 
Correct parent selected  5  6 
Other simple model selected  1  0 
Additive model selected  0  0 
Comparison of model fitting with and without iteration
Some further summary measures were investigated to establish the differences between model fitting with and without iteration, and are shown in Online Resource 1. The ratio of the mean LODs for each simulation set is only slightly above one, being in the range from 1.00 to 1.04 for the additive models and 1.10–1.19 for the complete models. There is more change in the mean R ^{2}, with the ratio varying from 1.13 to 1.86 for the additive models and 1.96–6.23 for the complete models. Plotting the ratio of the maximum LOD with iteration to that without iteration against the position of the maximum LOD with iteration shows that the LOD ratio is close to one when the maximum position is close to the middle of the chromosome but that there is a clear trend for the ratio to increase when the maximum position approaches the ends of the chromosome (Online Resource 2). There are only a few instances of this in Online Resource 2 part (b), which is based on simulation 2d (a simplex QTL explaining 5 % of the variance in a population of 400, detected for 41/50 traits), but more in Online Resource 2 part (a), with no true QTL present, and in Online Resource 2 part (c), where the complete model is fitted. Near the ends of the chromosome, there is inevitably less information about the QTL genotype provided by the marker data, and the weights will be more sensitive to the trait data and prone to overfit it. The iterative model fitting procedure was always terminated after ten iterations, and a check was made to see how often there was a failure to converge. As detailed in Online Resource 1, for the additive model at most 2 % failed to converge, and these were when testing a position close to the ends of the chromosome (0–4 and 75–87 cM). For the complete model fitted to 200 observations, more than half of the iterative procedures failed to converge. Bartlett’s test of homogeneity of variance was used to compare the variance of position estimates under the two fitting procedures. For the additive models, there was no significant difference, but for the complete models the variance was significantly larger when fitted with iteration, as detailed in Online Resource 1, due to QTLs being incorrectly detected near the ends of the chromosomes [Online Resource 2 part (c)]. We conclude that the iterative fitting method overfits in situations of low information: close to the ends of the chromosomes, when there is no strong QTL present, or when the number of parameters to be estimated is high as in the complete model. It is a particular problem in the latter case, where particularly few of the twoLOD support intervals contained the true QTL location.
Conclusions from the simulation study
We conclude from the simulation study that the optimal strategy is to fit the additive model to each trait as a weighted regression on the genotype probabilities without iteration. This aims to minimise overfitting to the data. TwoLOD support intervals were found to be a reasonable approximation to a 95 % confidence interval for the true position, while oneLOD intervals were too small to be a reliable indication. The mean trait values associated with the 36 QTL genotypes can then be calculated from the genotype probabilities at the most likely location, and a range of simple twoallele models fitted to identify which models minimise the SIC. This strategy is used in the analysis of the experimental data below.
Experimental study
The experimental study was based on 190 F_{1} offspring from the cross between Stirling and 12601ab1. To establish a genomewide threshold, the sequential permutation procedure of Nettleton and Doerge (2000) was used as described above. A minimum of 200 permutations were analysed for each of the experimental traits, using the additive model without iteration for all 12 chromosomes, and the maximum LOD over the 12 chromosomes was calculated. If the LOD score for a trait was within the approximate 95 % confidence interval for the threshold, a further 300 permutations were analysed.
Estimated QTL locations from analysis without iteration for the phenotypic traits scored on the experimental population
Trait  Chr.  Position (cM) and twoLOD support interval  LOD  LOD threshold 95 % CI  No. permutations  R ^{2}  Previously detected in this cross?  Simple model and candidate SNPs (if identified) 

Flower colour  X  44 (43–46)  37.9  (4.79, 5.09)  500  51.9  Yes (B2008)  Simplex, h6 
Flower colour  XI  84 (53–93)  4.9  (4.79, 5.09)  500  7.9  No  Dominant duplex, h6 and h8. Candidate c1_4947 
Flower_colour (residual)  XI  69 (51–91)  5.4  (4.79, 5.15)  500  7.9  No  Dominant duplex, h6 and h8. Candidate c1_4947 
Mat  V  15 (14–22)  42.3  (4.89, 6.25)  200  57.7  Yes (B2004, B2008)  Simplex on h1, candidate c2_47609 but possible further effect of h5 
Ht  V  21 (14–26)  20.0  (4.67, 5.35)  200  35.6  Yes (B2004)  As for Mat 
Rgene  XI  84 (81–87)  40.8  (4.72, 5.43)  200  57.7  Yes (B2004)  Simplex, h4 Candidate c2_37342 
Fb4  IV  26 (22–30)  19.7  (4.82, 6.01)  200  33.4  Yes (B2004)  Additive duplex for resistance on h1 and h4 Candidate c2_7756 
Fb4  V  21 (13–27)  9.6  (4.82, 6.01)  200  17.8  Yes (B2004)  Simplex, h1 Candidate c2_47609 
Tb%  V  22 (13–26)  10.6  (4.91, 5.53)  200  20.7  Yes (B2004)  Simplex, h1 Candidate c2_47609 
Tb%  IV  25 (20–34)  8.3  (4.91, 5.53)  200  16.3  Yes (B2004)  Additive duplex for resistance on h1 and h4 Candidate c2_7756 
PCN  IV  28 (22–34)  16.6  (4.84, 5.83)  200  29.8  Yes (Br2004)  Additive duplex for resistance on h6 and h8 
PCN (residual)  XI  22 (14–35)  5.4  (4.86, 5.21)  500  8.6  Yes (Br2004)  Additive duplex for resistance on h7 and h8. Candidate c2_33657 
Flower colour
Among the 190 offspring, 100 were scored as 0 = white (like Stirling), 78 as 1 = blue (like 12601ab1) and the remaining 12 were missing (no flowers). When the 0/1 scores were mapped as a quantitative trait, the most significant association was on chromosome X, with the peak LOD of 37.9 at 44 cM, well above the upper permutation threshold for this trait of 5.09. The only significant homologue effect was that of homologue h6, from 12601ab1, suggesting that this QTL is a simplex effect on h6. When twoallele models were fitted to the genotype means, the model of a simplex allele on h6 had the lowest SIC of −17.6, although this was only slightly lower than that of the full (additive) model where SIC = −16.4. A further association was also investigated on chromosome XI, with the peak LOD of 4.9. This lies within the 95 % confidence interval for the LOD threshold, based on 500 permutations (4.69, 5.09). Analysis of the residual flower colour after regression on the effect of LG X increased the LOD to 5.4, above the upper permutation threshold for the flower colour residual of 5.15. In both models, homologues h6 and h8 (from 12601ab1) showed significant, similarsized effects. A simple model with dominant effects of h6 or h8 had the minimum SIC of −33.7 for the residual flower colour, with an additive effect of h6 and h8 being a slightly worse fit with SIC = −30.3.
Twogene model for the probability of blue flowers
Chromosome XI, c1_4947 = AAAA  Chromosome XI, c1_4947 = AAB  

Count  Observed proportion  Predicted proportion  Count  Observed proportion  Predicted proportion  
Chromosome X, h6 pseudoSNP = AAAA  16  0.000  0.002  80  0.108  0.108 
Chromosome X, h6 pseudoSNP = AAAB  14  0.143  0.141  80  0.907  0.907 
The candidate SNPs on chromosome X, c2_27827 and c2_27806, both map to the same potato genome superscaffold (PGSC0003DMB000000106), at positions chr10:50697563 and chr10:50615253, respectively. This is in fairly close proximity to superscaffold PGSC0003DMB000000008 which contains DArT markers associated with violet flower colour reported by Śliwka et al. (2012). The candidate SNP on chromosome XI (c1_4947, chr11:41448860) maps to genome superscaffold PGSC0003DMB000000017. The inheritance of flower colour is reviewed by van Eck et al. (1993), who mapped the locus P involved in blue anthocyanin production to chromosome XI, and the locus F for flowerspecific expression of the colour to chromosome X in diploid potato. Bradshaw (2006) reviewed the inheritance in tetraploid potato. We postulate that the SNPs detected here are close to the F locus on chromosome X and the P locus on chromosome XI (although there are some recombinants), that Stirling has the recessive genotype pppp ffff and consequently white flowers, that 12601ab1 has genotype ppPP fffF and blue flowers and that only the offspring with genotype ppP fffF have blue flowers, as in Table 2.2 of Bradshaw (2006). The probability of an offspring inheriting the dominant P allele and the F allele and therefore having blue flowers is 5/6 × 1/2 = 5/12, giving expected frequencies here of 74 offspring with blue flowers and 104 with white flowers, which agrees well with the observed figures of 78 with blue flowers and 100 with white.
Maturity
Bradshaw et al. (2004, 2008) reported a large QTL affecting maturity on chromosome V, with a simplex allele from Stirling explaining 54 % of the phenotypic variance. This QTL was also detected in the current study, with the peak LOD score of 42.3 at 15 cM for the additive model, which explained 57.7 % of the variance. The chromosomal effects suggested that the most significant effect is that of simplex allele Q_{S}qqq on homologue h1 of Stirling. Analysis of the QTL genotype means using simple twoallele models showed that a simplex allele on h1 had the lowest SIC of 84.4, compared to SIC = 85.5 for the full (additive) model. The closest candidate SNP with this simplex configuration is c2_47609 at 18 cM, and regression on this SNP genotype explained 55.0 % of the phenotypic variance. The presence of the Q_{S} allele was associated with earlier maturity (mean difference 2.3, SE 0.15). However, the analysis showed further significant effects associated with 12601ab1, suggesting a simplex allele Q_{X} associated with earliness on homologue h5 of 12601ab1 with a smaller effect. Including this in the analysis of the QTL genotype means gave a lower SIC of 73.9. The nearest SNP with this configuration is c1_15292 at 27 cM. This was also significant (p < 0.001) when included in a regression model of the maturity scores and increased the percentage variance explained to 61.0 %, with a mean maturity difference of 0.77 (SE 0.14). Marker c2_47609 (chr05:5972404) maps to genome superscaffold PGSC0003DMB000000243 and is ~1.5 Mb from the recently published CDF1 gene (PGSC0003DMG400018408, chr5:45388804541736) coding for plant maturity (Kloosterman et al. 2013).
Canopy height
Bradshaw et al. (2004) reported that canopy height showed a similar genetic configuration to maturity and this was confirmed here. The peak of the LOD profile was at 21 cM with a LOD of 20.0, and explained 35.6 % of the phenotypic variance. The simplex SNP c2_47609 from h1 of Stirling explained 30.1 % of the variance in height, and the Q_{S} allele was associated with shorter height (mean difference 8.7 cm, SE 0.96). Inclusion of the simplex SNP c1_15292 from h5 of 12601ab1 increased the percentage variance explained to 35.2 %. This allele was also associated with shorter height, with a mean difference of 3.7 cm (SE 0.93). Analysis of the QTL genotype means also confirmed that a model with separate allele effects for h1 and h5 had the minimum SIC. These results suggest that early maturing potatoes have a lower canopy height and begin to senesce, while later maturing ones continue to grow.
Blight resistance
Among the 190 offspring, 113 were classified by Stewart et al. (2003) as having Stirling’s major Rgene, and 77 as lacking it. When the presence/absence scores were mapped as a quantitative trait, the most significant association was on chromosome XI, with the peak LOD of 40.8 at 84 cM. The only significant chromosomal effect was that of homologue h4, from Stirling. Analysis of the QTL genotype means showed that a simplex QTL on h4 has SIC = −32.8, considerably lower than the next best model, the full (additive) model, with the SIC = −16.5. There is a candidate simplex SNP qqqQ at this position on h4, c2_37342, which also has the highest association with the Rgene scores using a chi square test of independence (χ ^{2} = 106.4 with one degree of freedom). This Q allele is absent in 73 of the 77 offspring without the Rgene and present in 92 of the 113 with the Rgene. SNP marker c2_37342 maps to genome superscaffold PGSC0003DMB000000575 which is on chromosome VI. However, further analysis of the genome sequence flanking this marker shows a very strong BLAST hit to chromosome XI superscaffold PGSC0003DMB000000623 (P value 5.7e11) at position chr11:44363050, only slightly less significant than the value for the ‘correct’ location (4.9e–15). This indicates that for an as yet unexplained reason in the Stirling × 12601ab1 cross, this marker is segregating as a chromosome XI marker. Superscaffold PGSC0003DMB000000623 (chr11: 44,275,52644,528,585) is adjacent to superscaffold PGSC0003DMB000000017 (chr11: 40,928,095–44,225,525), which harbours a large cluster of NBLRR genes at the distal end of chromosome XI including closely related homologues of the late blight resistance gene R3a (Huang et al. 2004; Jupe et al. 2012).
Both foliage blight (Fb4) and tuber blight (Tb %) mapped to the same region of chromosome V as maturity and height, as shown by Bradshaw et al. (2004). For Fb4, the peak of the LOD profile was at 21 cM, with a LOD of 9.6, and the additive model explained 17.8 % of the phenotypic variance and for Tb % the peak of the LOD profile was at 22 cM with a LOD of 10.6 and explained 20.7 % of the trait variance. The best genetic model was a simplex allele Qqqq on homologue h1 of Stirling, with the Q allele associated with susceptibility to blight. There was no evidence that alleles from 12601ab1 had a significant effect on these traits. Regression on the candidate simplex SNP c2_47609 at 18 cM explained 18.0 % of the variance of Fb4 with a mean effect of 1.7 (SE 0.27), and 21.6 % of the variance of Tb %, with a mean effect of 27.2 % (SE 3.9).
A further QTL for Fb4 and Tb % was mapped to chromosome IV, as found by Bradshaw et al. (2004). For Fb4, the peak of the LOD profile for the additive model was at 26 cM, with a LOD of 19.7, and explained 33.4 % of the phenotypic variance and for Tb % the peak of the LOD profile was at 25 cM with a LOD of 8.3 and explained 16.3 % of the trait variance. Exploration of different genetic models indicated that the best model was for Stirling carrying a duplex allele QqqQ with the Q alleles associated with resistance on homologues h1 and h4 and the Qqqq and qqqQ offspring having intermediate resistance. This model had SIC = 96.7 for Fb4, compared to the second best SIC of 110.5 for a dominant duplex model, and SIC = 291.6 for Tb %, compared to the second best SIC of 305.3 for the full (additive) model. A candidate SNP with this configuration is the duplex SNP c2_7756 at 25 cM. Regression of the trait values on the three genotypes of this SNP explained 43.6 % of the variance in Fb4 and 19.2 % of the variance for Tb %. The means for Fb4 are qqqq = 1.7 (SE 0.29), qqqQ = 5.1 (SE 0.13) and qqQQ = 6.3 (SE 0.28) (where 9 is resistant). The corresponding means for Tb % are 84.1 % (SE 4.97), 57.1 % (SE 2.35) and 38.2 % (SE 5.06), where a low Tb % indicates resistance. SNP marker c2_7756 (chr04:4023794) locates to potato genome superscaffold PGSC0003DMB000000330, which along with adjacent superscaffold PGSC0003DMB000000296 harbours NBLRR genes, including close homologues of the functional late blight resistance gene R2 (Jupe et al. 2012).
PCN resistance
Bryan et al. (2004) reported a major QTL for PCN resistance on chromosome IV for this population, and this was confirmed here. The peak of the LOD profile for the additive model of the trait √(cyst counts) was at 28 cM, with a LOD of 16.6, and explained 29.8 % of the phenotypic variance. Exploration of different genetic models indicated that the best model is for 12601ab1 carrying a duplex allele qQqQ with the Q alleles associated with resistance (i.e. low cyst counts) on homologues h6 and h8 and the qQqq and qqqQ offspring having intermediate resistance. This had SIC = 74.9, compared with the second best model, full (additive) with SIC = 89.1. There is no candidate SNP with this configuration nearby, and so a pseudoSNP was constructed from the genotypes of two simplex SNPs c1_16358 (chr04:3240076) at 25.9 and c2_21847 (chr04:5030164) at 26.8 cM. This explained 31.2 % of the phenotypic variance, and the mean √(cyst counts) were qqQQ = 2.9 (SE 0.25), qqqQ = 4.0 (SE 0.11), qqqq = 5.7 (SE 0.21). These SNPs both locate to a region of chromosome IV that contains several homologues of the functional tomato Hero nematode resistance gene and the R2 late blight resistance gene (Ernst et al. 2002; Jupe et al. 2012).
No further QTLs were significant for PCN resistance, but again following Bryan et al. (2004), the residuals after regressing the PCN scores on the pseudoSNP from chromosome IV were constructed and the QTL analysis was repeated. A further QTL was detected on chromosome XI, as found by Bryan et al. (2004). The peak of the LOD profile for the additive model was at 22 cM, with a LOD of 5.4, and explained 8.6 % of the phenotypic variance. Exploration of different genetic models indicated that the best model was 12601ab1 carrying an additive duplex allele qqQQ with the Q alleles associated with resistance on homologues h7 and h8 and the qqQq and qqqQ offspring having scores midway between the qqqq and the qqQQ individuals (SIC = 81.1, compared with SIC = 87.8 for the second best model, a doublesimplex effect). There is a candidate SNP c2_33657 at 21 cM with this duplex configuration on h7 and h8. Regression on this as an additive trait explained 9.7 % of the trait variance, and possession of each Q allele was associated with a mean decrease of 0.7 (SE 0.16) in the PCN score. SNP marker c2_33657 (chr11:2274063) is located in genome superscaffold PGSC0003DMB000000152, which along with adjacent scaffold PGSC0003DMB000000505 contains a small cluster of NBLRR resistance gene homologues.
Discussion
In this paper, we have explored and extended the QTL mapping methodology proposed by Hackett et al. (2013) to incorporate information on SNP dosages for QTL mapping. This method is generally applicable to autotetraploid species, provided that the model of random chromosomal segregation is reasonable. Previous work on mapping in autotetraploids, such as Bradshaw et al. (2008), has not been able to align linkage maps from the two parents due to a lack of informative markers segregating in both parents, and consequently QTL mapping has been carried out for each parent separately. Using SNP dosage information, there are far more configurations that enable alignment of the parental maps, and so the effects of alleles from both parents can be studied simultaneously. This gives extra insights, for example that the wellreported maturity effect on chromosome V is not simply derived from the Stirling parent, but that the alleles from the 12601ab1 parent also contribute significantly to this trait. Many QTL mapping methods with sparser maps fit mixture models by iteratively updating the QTL genotype probabilities along the chromosome to allow for the uncertainty between marker positions, but here we have showed that with such a dense map there is no need for such an iterative process, and that using such a process is likely to overfit the data.
The simulation study also showed that it is better to use an additive model of chromosomal effects to locate the positions most associated with a quantitative trait than to fit a complete model with 36 genotype means at each position. The latter approach again leads to overfitting to the data. When random traits are analysed by such a flexible model fitted at a large number of positions, high LOD scores are obtained. The LOD threshold therefore has to be set at a high level to avoid false positives, with the consequence that the power to detect QTLs is low. Once the best position has been located using the additive model, the mean trait values for the 36 possible QTL genotypes can be estimated, and explored by fitting a range of twoallele models to see if the data are compatible with a simpler model. We have shown that this approach correctly identifies a dominant duplex QTL in more than half of the significant simulations. The denser maps, with simplex SNPs on almost all the homologous chromosomes and a high density of higher dosage markers, give increased confidence in modelling the allelic effects. It would be possible to recalculate the LOD profile based on a simpler model, but this would require specific programming for each detected QTL, and the simulation study suggests that it would be unlikely to change the QTL location to any large extent.
Despite the advantages of the denser map, the size of the mapping population limits the size of QTLs that can be detected. A population of 180–200 offspring is typical of many experimental studies seen in practice but the simulation study has shown the limitations of this. A single QTL explaining 5 % of the variance in a population of 200 was detected in only 29 of 50 simulations, and the size of the effect was on average overestimated. If a population of size 400 was used, a QTL explaining 5 % was detected for 41 of 50 simulations, and the magnitude of the effect was estimated more accurately. Larger QTLs can be detected in populations of 180–200, and candidate SNPs for markerassisted selection can be identified, but to detect smaller QTLs reliably, larger population sizes will be needed.
The simulation study here considered a single QTL together with an environmental effect. In reality, there are likely to be several QTLs of different sizes involved. If a large QTL is detected, its effect can be removed by regression on a candidate SNP and further QTL mapping conducted on the residuals, as was done here for PCN resistance. This should increase the power to detect smaller QTLs. In a clonal crop such as potato, increased replication of the population will increase the heritability and the power to detect smaller QTLs.
The QTL analysis was applied to investigate flower colour, maturity and resistance traits in autotetraploid potato. Most of the QTLs detected here were previously detected using an earlier AFLP and SSR map. The confidence intervals for locations have been improved by the more detailed linkage map, and more information about the genetic model at each QTL becomes available. For several of the QTLs reported here, candidate SNPs can be identified. A major advantage of the SNP map is the ability to link directly to the potato genome sequence for the vast majority of SNP markers used in this study. In several cases, there is strong evidence to suggest that the SNP explaining most of the phenotypic variation is quite close to the causal gene for the trait. For example, for plant maturity, marker c2_47609 maps at most ~1.5 Mb from the gene known to control the trait (Kloosterman et al. 2013). For most of the resistance traits analysed, the best SNP marker maps within NBLRR gene clusters located at welldocumented resistance ‘hotspots’ on chromosomes IV and XI, allowing the possibility of using candidate gene approaches for targeted gene isolation. Moreover, these findings are strongly indicative that the resistances, whether major R gene or partial resistance due to largeeffect QTLs, are likely to be due to the action of NBLRR genes in the potato genome.
Notes
Acknowledgments
Support for this work from the Scottish Government’s Rural and Environment Science and Analytical Services Division (RESAS) is gratefully acknowledged. We thank two anonymous reviewers for their suggestions, which have improved the paper.
Conflict of interest
The authors declare that they have no conflict of interest.
Supplementary material
References
 Baird NA, Etter PD, Atwood TS, Currey MC, Shiver AL et al (2008) Rapid SNP discovery and genetic mapping using sequenced RAD markers. PLoS ONE 3:e3376PubMedCentralPubMedCrossRefGoogle Scholar
 Beavis WD (1994) The power and deceit of QTL experiments: lessons from comparative QTL studies. American Seed Trade Association. 49th Annual Corn and Sorghum Research Conference. Washington, pp 252–268Google Scholar
 Bradshaw JE (2006) Genetics of agrihorticultural traits. In: Gopal J, Khurana SMP (eds) Handbook of potato production, improvement and postharvest management. The Haworth Press, Binghamton, pp 41–75Google Scholar
 Bradshaw JE, Pande B, Bryan GJ, Hackett CA, McLean K et al (2004) Interval mapping of quantitative trait loci for resistance to late blight [Phytophthora infestans (Mont.) de Bary], height and maturity in a tetraploid population of potato (Solanum tuberosum subsp tuberosum). Genetics 168:983–995PubMedCentralPubMedCrossRefGoogle Scholar
 Bradshaw JE, Hackett CA, Pande B, Waugh R, Bryan GJ (2008) QTL mapping of yield, agronomic and quality traits in tetraploid potato (Solanum tuberosum subsp. tuberosum). Theor Appl Genet 116:193–211PubMedCrossRefGoogle Scholar
 Broman KW, Wu H, Sen Ś, Churchill GA (2003) R/QTL: QTL mapping in experimental crosses. Bioinformatics 19:889–890PubMedCrossRefGoogle Scholar
 Bryan GJ, McLean K, Pande B, Purvis A, Hackett CA et al (2004) Genetical dissection of H3mediated polygenic PCN resistance in a heterozygous autotetraploid potato population. Mol Breed 14:105–116CrossRefGoogle Scholar
 Churchill GA, Doerge RW (1994) Empirical threshold values for quantitative trait mapping. Genetics 138:963–971PubMedCentralPubMedGoogle Scholar
 Elshire RJ, Glaubitz JC, Sun Q, Poland JE, Kawamoto K et al (2011) A robust, simple genotypingbysequencing approach for high diversity species. PLoS ONE 6:e19379PubMedCentralPubMedCrossRefGoogle Scholar
 Ernst K, Kumar A, Kriseleit D, Kloos DU, Phillips MS, Ganal MW, Ganal MW (2002) The broadspectrum potato cyst nematode resistance gene (Hero) from tomato is the only member of a large gene family of NBSLRR genes with an unusual amino acid repeat in the LRR region. Plant J 31:127–136PubMedCrossRefGoogle Scholar
 Felcher KJ, Coombs JJ, Massa AN, Hansey CN, Hamilton JP et al (2012) Integration of two diploid potato linkage maps with the potato genome sequence. PLoS ONE 7:e36347PubMedCentralPubMedCrossRefGoogle Scholar
 Hackett CA, Bradshaw JE, McNicol JW (2001) Interval mapping of QTLs in autotetraploid species. Genetics 159:1819–1832PubMedCentralPubMedGoogle Scholar
 Hackett CA, McLean K, Bryan GJ (2013) Linkage analysis and QTL mapping using SNP dosage data in a tetraploid potato mapping population. PLoS ONE 8:e63939PubMedCentralPubMedCrossRefGoogle Scholar
 Huang S, Vleeshouwers VGAA, Werij JS, Hutten RCB, van Eck HJ, Visser RGF, Jacobsen E (2004) The R3 resistance to Phytophthora infestans in potato is conferred by two closely linked R genes with distinct specificities. Mol PlantMicrobe Int 17:428–435CrossRefGoogle Scholar
 Jansen RC (1992) A general mixture model for mapping quantitative trait loci by using molecular markers. Theor Appl Genet 85:252–260PubMedGoogle Scholar
 Jupe F, Pritchard L, Etherington G, MacKenzie K, Cock PJA, Wright F, Sharma SK, Bolser D, Bryan GJ, Jones JDG, Hein I (2012) Identification and localisation of the NBLRR gene family within the potato genome. BMC Genom 13:75CrossRefGoogle Scholar
 Kempthorne O (1957) An introduction to genetic statistics. John Wiley & Sons, New YorkGoogle Scholar
 Kloosterman B, Abelenda JA, Gomez MMC, Oortwijn M, de Boer JM et al (2013) Naturally occurring allele diversity allows potato cultivation in northern latitudes. Nature 495:246–250PubMedCrossRefGoogle Scholar
 MacKenzie K, Hackett CA (2012) Association mapping in a simulated barley population. Euphytica 183:337–347CrossRefGoogle Scholar
 Meyer RC, Milbourne DM, Hackett CA, Bradshaw JE, McNicol JW et al (1998) Linkage analysis in tetraploid potato and association of markers with quantitative resistance to late blight (Phytophthora infestans). Mol Gen Genet 259:150–160PubMedCrossRefGoogle Scholar
 Nettleton D, Doerge RW (2000) Accounting for variability in the use of permutation testing to detect quantitative trait loci. Biometrics 56:52–58PubMedCrossRefGoogle Scholar
 Payne R, Murray D, Harding S, Baird D, Soutar D (2012) Introduction to GenStat for Windows, 15th edn. VSN International, Hemel HempsteadGoogle Scholar
 Potato Genome Sequencing Consortium (2011) Genome sequence and analysis of the tuber crop potato. Nature 475:189–197CrossRefGoogle Scholar
 Rabiner LR (1989) A tutorial on Hidden Markov Models and selected applications in speech recognition. Proc IEEE 77:257–286CrossRefGoogle Scholar
 Schwarz GE (1978) Estimating the dimension of a model. Ann Stat 6:461–464CrossRefGoogle Scholar
 Sharma SK, Bolser D, de Boer J, Sonderkaer M, Amoros W et al (2013) Construction of reference chromosomescale pseudomolecules for potato: integrating the potato genome with genetic and physical maps. Genes Genomes Genetics 3:2031–2047PubMedCentralPubMedGoogle Scholar
 Śliwka J, Jakuczun H, Chmielarz M, HaraSkrzypiec A, Tomczyńska I, Kilian A, GuzowskaZimnoch E (2012) Late blight resistance gene from Solanum ruizceballosi is located on potato chromosome X and linked to violet flower colour. BMC Genet 13:11PubMedCentralPubMedGoogle Scholar
 Stewart HE, Bradshaw JE, Pande B (2003) The presence of Rgenes for resistance to late blight (Phytophthora infestans) of potato (Solanum tuberosum) on the underlying level of field resistance. Plant Pathol 52:193–198CrossRefGoogle Scholar
 Utz HF, Melchinger AE (1994) Comparison of different approaches to interval mapping of quantitative trait loci. In: van Ooijen JW, Jansen J (eds) Biometrics in plant breeding: applications of molecular markers. CPRODLO, Netherlands, pp 195–204Google Scholar
 van Eck HJ, Jacobs JME, van Dijk J, Stieckema WJ, Jacobsen E (1993) Identification and mapping of three flower colour loci of potato (S. tuberosum L.) by RFLP analysis. Theor Appl Genet 86:295–300PubMedGoogle Scholar
 van Ooijen JW (2004) MapQTL^{®} 5, software for the mapping of quantitative trait loci in experimental populations. Kyazma BV, Wageningen, The NetherlandsGoogle Scholar
 van Os H, Andrzejewski S, Bakker E, Barrena I, Bryan GJ, Caromel B et al (2006) Construction of a 10,000marker ultradense genetic recombination map of potato: providing a framework for accelerated gene isolation and a genomewide physical map. Genetics 173:1075–1087PubMedCentralPubMedCrossRefGoogle Scholar
 Xu S (2003) Theoretical basis of the Beavis effect. Genetics 165:2259–2268PubMedCentralPubMedGoogle Scholar
Copyright information
Open AccessThis article is distributed under the terms of the Creative Commons Attribution License which permits any use, distribution, and reproduction in any medium, provided the original author(s) and the source are credited.