Volume 4 Supplement 1
Proceedings of the 13th European workshop on QTL mapping and marker assisted selection
Genomewide selection by mixed model ridge regression and extensions based on geostatistical models
 Torben SchulzStreeck^{1} and
 HansPeter Piepho^{1}Email author
DOI: 10.1186/175365614S1S8
© Piepho and SchulzStreeck; licensee BioMed Central Ltd. 2010
Published: 31 March 2010
Abstract
Background
The success of genomewide selection (GS) approaches will depend crucially on the availability of efficient and easytouse computational tools. Therefore, approaches that can be implemented using mixed models hold particular promise and deserve detailed study. A particular class of mixed models suitable for GS is given by geostatistical mixed models, when genetic distance is treated analogously to spatial distance in geostatistics.
Methods
We consider various spatial mixed models for use in GS. The analyses presented for the QTLMAS 2009 dataset pay particular attention to the modelling of residual errors as well as of polygenetic effects.
Results
It is shown that geostatistical models are viable alternatives to ridge regression, one of the common approaches to GS. Correlations between genomewide estimated breeding values and true breeding values were between 0.879 and 0.889. In the example considered, we did not find a large effect of the residual error variance modelling, largely because error variances were very small. A variance components model reflecting the pedigree of the crosses did not provide an improved fit.
Conclusions
We conclude that geostatistical models deserve further study as a tool to GS that is easily implemented in a mixed model package.
Background
Genomewide selection (GS) is a markerbased method that predicts breeding values on the basis of a large number of molecular markers, which typically cover the entire genome [1]. The idea is to estimate the effects of all genes or chromosomal segments simultaneously and to integrate these estimates in order to predict the total breeding value.
One basic approach for GS is ridge regression (RR) [1]. An interesting alternative to RR is to use spatial models [2] to model genetic correlation among relatives [3].
This study compares RR models and spatial models for estimating genomewide breeding values for the common dataset provided by the 13^{th} QTLMAS workshop. The focus is on methods that can be easily implemented using a standard mixed model package with facilities for spatial covariance structures.
Methods
Data
The dataset was simulated as part of the 13^{th} QTLMAS workshop (see [4] for details). Phenotypes of 1000 of 2025 individuals were recorded at five different times (0, 132, 265, 397 and 530), so there is a series of five repeated measurements for each phenotyped individual. Breeding values for the nonphenotyped individuals were to be predicted for time=600, which constitutes an extrapolation.
Extrapolation for time=600
would give a reasonable fit to the data, where y_{ it } is the trait value of the ith individual at time t (t = 0, 132, 256, 397, 530 or 600) and α_{ i }, β_{ i }, γ_{ i } are parameters pertaining to the ith individual. Observed data were modelled as y_{ it } = E( y_{ it } ) + ε_{ it }, assuming that errors ε_{ it } are independent and have constant variance. We initially fitted this model separately for each individual. Based on this analysis, we then obtained a pooled residual variance estimate from all individuals as = RSS /2n, where RSS is the pooled residual sum of squares and n is the number of phenotyped individuals. Each individual contributed two error degrees of freedom. Next, nonlinear regressions were rerun for all phenotyped individuals to predict y_{i600}, fixing the residual variance at the pooled estimate. Along with predictions the standard error was determined. The error variance of predicted values was estimated as the square of the standard error. This variance was subsequently regarded as a known withinindividual error variance for mixed model analyses.
Marker scoring
The marker covariate z_{ ik } for the ith genotype and the kth marker for biallelic markers with alleles A_{1}and A_{2}was set to 1 for A_{1}A_{1}, 1 for A_{2}A_{2}and 0 for A_{1}A_{2}and A_{2}A_{1}. Covariates were stored in a matrix Z = {z_{ ik } }.
Model
[2]. There were two options regarding the model for var(e_{ i }), the variance of e_{ i }. Either we fixed var(e_{ i }) at the squared standard errors of predictions of y_{i600} obtained from nonlinear regression and assumed heterogeneous residual variance, or we pooled e_{ i } with v_{ i } and thus implicitly included var(e_{ i }) in the variance structure for v_{ i } which is defined as var(v) = , where is the variance component for polygenic effects. It is common to model the polygenic effect by the relationship matrix. But in this case we assumed independent polygenic effects, because no pedigree information on the parents was available. It must be stressed that var(e_{ i }) is strictly a withinindividual error variance component that does not comprise betweenindividual error components. In the present application, these latter error components are effectively confounded with the polygenic effect v_{ i }. By contrast, in plant breeding applications, field replication would allow a separate assessment of betweenindividual error components. It is conjectured that explicit modelling of such an error component would be of benefit [2], because it is possible that the model captures part of the variance among individuals, which is nongenetic. For this reason it is advisable to generally obtain an independent estimate of error (as in [5]).
Genotypic covariance models of the form Γ = {f(d_{ ii' })}, where d is the Euclidean distance computed from marker data and θ is a parameter.
Name  Equation 

Linear 

Quadratic 

Power 

Exponential 

Gaussian 

Spherical 

We also considered an extended model , where Ω represents the covariance due to simple random effects, i.e., , where , and are the variance components for random effects of father and mother of crosses and of the crosses themselves, respectively, and V_{ f } , V_{ m } and V_{ c } are corresponding symmetric matrices of known constants and are the variance component for random effects of the individuals. In this case the polygenic variance is defined as , where v^{ ' }= v_{ 1, }v_{ 2 },...,v_{ G }. If Ω = 0 the model is equal to models with independent polygenic effects. When fitting the extended model we did not fix var(e_{ i }) at the squared standard errors of predictions of y_{i600}, therefore var(e_{ i }) is pooled with .
For each fitted model we obtained BLUPs of µ + h_{ i } corresponding to genomewide estimated breeding values (GEBV). For nonphenotyped individuals in the case of models with independent polygenic effects and for the extended model when Ω ≠ 0. The Pearson correlation between the GEBV and the fitted values y_{ i }_{600} were calculated. In addition, the Akaike Information Criterion (AIC) was recorded. Small values of AIC indicated a preferable model.
After the 13^{th} QTLMAS workshop the organizers reported the true breeding values (TBV). The TBV of the nonphenotyped individuals were compared to the GEBV by the Pearson correlation.
Software
The nonlinear regression was done by the NLMIXED procedure of the SAS System, while all mixed models were fitted by the REML method using the MIXED procedure of SAS.
Results
Analysis without fixing the residual variance
Model fits of different genetic covariance models and Pearson correlation between GEBV and fitted value and between GEBV and true breeding value (TBV).
Residual  Correlation  

Model for g_{ i }  AIC  variance^{§}  θ  Fitted value^{$}  TBV^{#} 
Independent  6789.1  51.89  
Ridge Regression (RR)  6418.5  28.17  0.734  0.889  
Spatial models  
Linear  6425.8  12.00  0.974  0.880  
Quadratic  6418.5  28.17  0.734  0.889  
Power  6428.9  12.16  0.99  0.974  0.879 
Exponential  6428.5  11.48  216.52  0.977  0.879 
Gaussian  6420.5  28.08  124.59  0.737  0.889 
Spherical  6427.8  11.96  959.97  0.974  0.880 
Model fits of different genetic covariance models with random effects for father and mother of crosses and for the crosses themselves and Pearson correlation between GEBV and fitted value and between GEBV and true breeding value (TBV).
Residual  Father^{&}  Mother^{&}  Correlation  

Model for g_{i}  AIC  variance^{§}  θ 

 Fitted value^{$}  TBV^{#} 
Independent  6605.1  40.77  7.16  6.16  0.481  0.649  
Ridge Regression (RR)  6420.2  28.18  0.61  0  0.734  0.889  
Spatial models  
Linear  6427.2  12.17  1.08  0  0.973  0.879  
Quadratic  6420.2  28.18  0.61  0  0.734  0.889  
Power  6430.3  12.35  0.99  1.20  0  0.973  0.878 
Exponential  6429.9  11.62  208.63  1.11  0  0.976  0.878 
Gaussian  6422.2  28.07  118.46  0.62  0  0.737  0.889 
Spherical  6429.2  12.17  802.96  1.09  0  0.973  0.879 
Analysis with fixing the residual variance
Model fits of different genetic covariance models. Residual variance var(e_{i}) fixed at value of squared standard error of y_{ i600 } and Pearson correlation between GEBV and fitted value and between GEBV and true breeding value (TBV).
Polygenic  Correlation  

Model for g_{ i }  AIC  variance  θ  Fitted value^{$}  TBV^{#} 
Independent  6789.1  51.80  
Ridge Regression (RR)  6418.5  28.09  0.734  0.889  
Spatial models  
Linear  6425.8  11.88  0.974  0.880  
Quadratic  6418.5  28.10  0.734  0.889  
Power  6428.8  10.38  0.99  0.981  0.878 
Exponential  6428.2  11.17  627.76  0.977  0.879 
Gaussian  6420.5  27.98  124.10  0.737  0.889 
Spherical  6427.8  11.87  959.00  0.975  0.880 
Discussion
There are only minor differences of the AIC values between RR and spatial models, like in [2]. Thus, some of the spatial models are viable alternatives to RR. Among the spatial models, the Gaussian model gave almost the same fit as RR. This can be explained by a Taylor expansion argument. The correlation function under the Gaussian model is exp(– d^{2} / θ^{ 2 }). When θ is large, then the exponent is close to zero and to first order we have , so the Gaussian model approaches the RR/quadratic model in this case, but when θ is small we expect different fits. It is noteworthy that the Gaussian model is essentially equivalent to reproducing kernel Hilbert spaces regression as proposed by [7] and also to least squares support vector machine (LSSVM) regression [8]. It may be conjectured that when inheritance is not merely additive it may be of particular importance to model the genetic covariance by some nonlinear spatial model such as an exponential or a Gaussian model.
In this study we have used AIC defined as AIC = 2 log(likelihood) + 2 * number of parameters as printed by mixed model packages such as the MIXED procedure of SAS. The models with the lowest AIC values showed the highest correlation between the GEBV and TBV. But the correlation between the model ranks produced by AIC and by the correlation between GEBV and TBV is not perfect. For smoothing methods modifications of the AIC have been proposed (e.g. corrected AIC (AIC_{C}: [9]) and different other criteria (e.g. generalized crossvalidation (GCV: [10]). The main difference to the AIC is that the complexity of the fitted model is calculated as the trace of the socalled smoother matrix tr(S_{ λ }) described in [11], which relates to the effective degrees of freedom of the fit. It is important to realize that GS may be regarded as a smoothing exercise that replaces observed data (adjusted genotype means) by smoothed fitted values. Thus, model selection criteria developed for smoothing can be a useful extension for selecting a preferable model in GS.
The comparison between GEBV and phenotypes is not a good indictor for accuracy of breeding values, when all individuals are involved in the prediction and no independent validation set is left (Tables 2, 3 and 4). Crossvalidation is one option to avoid this problem. The leaveoneout crossvalidation procedure is equivalent to the crossvalidation criterion, which is related to other selection criteria (AIC, AIC_{C} and GCV) [12]. Therefore one idea is to replace the crossvalidation procedure by model selection criteria, which would entail a considerable saving of computing time.
In the present study the data were simulated without polygenic effects. Nevertheless, it is prudent to cater for the case that the total genotypic variance can not be fully explained by the markers alone. We think that this unexplained part should be modelled by a polygenic effect.
We modelled the polygenic effect as independent. Alternatively, one can assume that the polygenic effect is correlated due to the pedigree. We modelled the pedigree of the crosses by variance components, however, reflecting the pedigree did not provide an improved fit in the present case.
We also think it is important to separate the polygenic effect from error in order to avoid overfitting [2]. In the present study, however, fixing the residual variance did not have much of an effect because essentially we only had a withinindividual error variance estimate. This ignored betweenindividual error variance, which is therefore expected to be confounded with the variance component for polygenic effects ( ). In plant breeding trials, where replication is available, one can separate polygenic effects from nongenetic betweenindividual errors. We expect that such separation will be crucial to the success of GS approaches in plant breeding.
Conclusions
Our study has shown that geostatistical models are viable alternatives to RR that deserve further study as a tool to GS. With respect to our analyses for the QTLMAS 2009 dataset, however, we prefer the RR/quadratic model without fixed residual variance for predicting GEBV.
Declarations
Acknowledgements
Two anonymous reviewers are thanked for constructive comments.
This article has been published as part of BMC Proceedings Volume 4 Supplement 1, 2009: Proceedings of 13th European workshop on QTL mapping and marker assisted selection.
The full contents of the supplement are available online at http://www.biomedcentral.com/17536561/4?issue=S1.
Authors’ Affiliations
References
 Meuwissen THE, Hayes BJ, Goddard ME: Prediction of total genetic value using genomewide dense marker maps. Genetics. 2001, 157: 18191829.PubMed CentralPubMed
 Piepho HP: Ridge regression and extensions for genomewide selection in maize. Crop Science. 2009, 49: 11651176. 10.2135/cropsci2008.10.0595.View Article
 Piepho HP, Möhring J, Melchinger AE, Büchse A: BLUP for phenotypic selection in plant breeding and variety testing. Euphytica. 2008, 161: 209228. 10.1007/s1068100794498.View Article
 Coster A, Bastiaansen J, Calus M, Maliepaard C, Bink M: QTLMAS 2009: Simulated Dataset. BMC Proc. 2010, 4 (Suppl 1): S310.1186/175365614S1S3.PubMed CentralView ArticlePubMed
 Bernardo R, Yu J: Prospects for genomewide selection for quantitative traits in maize. Crop Science. 2007, 47: 10821090. 10.2135/cropsci2006.11.0690.View Article
 Schabenberger O, Gotway CA: Statistical methods for spatial data analysis. 2005, CRC Press
 Gianola D, van Kaam JBCHM: Reproducing kernel Hilbert spaces regression methods for genomic assisted prediction of quantitative traits. Genetics. 2008, 178 (4): 22892303. 10.1534/genetics.107.084285.PubMed CentralView ArticlePubMed
 Suykens JAK, Gestel TV, de Brabanter J, de Moor B, Vandewalle J: Least squares support vector machines. 2002, World Scientific Publishers
 Hurvich CM, Simonoff JS, Tsai C: Smoothing parameter selection in nonparametric regression using an improved Akaike information criterion. Journal ofthe Royal Statistical Society Series B. 1998, 60: 27193. 10.1111/14679868.00125.View Article
 Craven P, Wahba G: Smoothing noisy data with spline functions: Estimating the correct degree of smoothing by the method of generalized crossvalidation. Numerische Mathematik. 1979, 31: 377403. 10.1007/BF01404567.View Article
 Ruppert D, Wand MP, Carroll RJ: Semiparametric regression. 2003, Cambridge Univ. PressView Article
 McQuarrie ADR, Tsai CL: Regression and time series model selection. 1998, World Scientific
Copyright
This article is published under license to BioMed Central Ltd. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.