A comparison of random forests, boosting and support vector machines for genomic selection
 Joseph O Ogutu^{1}Email author,
 HansPeter Piepho^{1} and
 Torben SchulzStreeck^{1}
https://doi.org/10.1186/175365615S3S11
© Ogutu et al; licensee BioMed Central Ltd. 2011
Published: 27 May 2011
Abstract
Background
Genomic selection (GS) involves estimating breeding values using molecular markers spanning the entire genome. Accurate prediction of genomic breeding values (GEBVs) presents a central challenge to contemporary plant and animal breeders. The existence of a wide array of markerbased approaches for predicting breeding values makes it essential to evaluate and compare their relative predictive performances to identify approaches able to accurately predict breeding values. We evaluated the predictive accuracy of random forests (RF), stochastic gradient boosting (boosting) and support vector machines (SVMs) for predicting genomic breeding values using dense SNP markers and explored the utility of RF for ranking the predictive importance of markers for prescreening markers or discovering chromosomal locations of QTLs.
Methods
We predicted GEBVs for one quantitative trait in a dataset simulated for the QTLMAS 2010 workshop. Predictive accuracy was measured as the Pearson correlation between GEBVs and observed values using 5fold crossvalidation and between predicted and true breeding values. The importance of each marker was ranked using RF and plotted against the position of the marker and associated QTLs on one of five simulated chromosomes.
Results
The correlations between the predicted and true breeding values were 0.547 for boosting, 0.497 for SVMs, and 0.483 for RF, indicating better performance for boosting than for SVMs and RF.
Conclusions
Accuracy was highest for boosting, intermediate for SVMs and lowest for RF but differed little among the three methods and relative to ridge regression BLUP (RRBLUP).
Keywords
Background
Genomic selection is a method for estimating GEBVs using dense molecular markers spanning the entire genome [1]. Given the wide range of approaches for predicting GEBVs, it is important to evaluate their performance, pros and cons to identify those able to accurately predict GEBVs. Here, we compare predictive performances among three of the most powerful machine learning methods with demonstrated high predictive accuracies in many application domains, namely RF [2, 3]; boosting [5] and SVMs [5, 6] and with RRBLUP [7] for predicting breeding values for quantitative traits.
RF has several appealing properties that make it potentially attractive for GS [2, 4]: (i) the number of markers can far exceed that of observations, (ii) all markers, including those with weak effects, highly correlated and interacting markers have a chance to contribute to the model fit, (iii) complex interactions between markers can be easily accommodated, (iv) they can perform both simple and complex classification and regression accurately, (v) they often require modest finetuning of parameters and the default parameterization often performs well [2, 3], and (vi) they make no distributional assumptions about the predictor variables. Boosting is a stagewise additive model fitting procedure that can enhance the predictive performance of weak learning algorithms [5]. SVMs perform robustified regression using kernel functions of inner products of predictors [5].
We comparatively evaluated the predictive performance of the three machine learning methods and RRBLUP for estimating GEBVs using the common dataset simulated for the QTLMAS 2010 workshop. RF regression was used to rank the SNPs in terms of their predictive importance.
Methods
Data
The simulated data set contained 3226 individuals spanning five generations out of which 2326, constituting the first four generations, were phenotyped and genotyped for 10031 biallelic SNPs arrayed on a genome encompassing five chromosomes. The remaining 900 individuals, representing the fifth generation, had genomic but lacked phenotypic records on the single quantitative trait. The covariate for each genotype 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} orA_{2}A_{1}.
Random forests
where Ψ_{ b } characterizes the bth RF tree in terms of split variables, cutpoints at each node, and terminal node values.
We implemented RF in the R package randomForest with decision trees as base learners [3]. Following various recommendations [2, 3], we evaluated different combinations of the values of the number of trees to grow, ntree = {500, 1000, 2000}, the number of SNPs randomly selected at each tree node, mtry = {0.5, 1, 2} × the default value of mtry of sample size/3 for regression, and the minimum size of terminal nodes of trees, below which no split is attempted, nodesize = 1. The parameter configuration with the highest prediction accuracy was ntree =1000, mtry = 3000 and nodesize =1. We ranked SNPs by the relative importance of their contributions to predictive accuracy, quantified by how much prediction error increased when the observations left out of the bootstrap samples, the outofbag data for a SNP, were randomly permuted while data for all the other SNPs were left unchanged [2, 3].
Stochastic Gradient Boosting
where β_{ m }, m =1,2,…, M are the basis expansion coefficients, and b(x, γ) are simple functions of the multivariate argument x, with a set of parameters γ=(γ_{1},γ_{2},…,γ_{ M }).
We used regression trees as basis functions. Boosting regression trees involves generating a sequence of trees, each grown on the residuals of the previous tree [5, 9]. Prediction is accomplished by weighting the ensemble outputs of all the regression trees. We used stochastic gradient boosting, assuming the Gaussian distribution for minimizing squarederror loss in the R package gbm[9]. We determined the main tuning parameter, the optimal number of iterations (or trees), using an outofbag estimate of the improvement in predictive performance. This evaluates the reduction in deviance based on observations not used in selecting the next regression tree. The minimum number of observations in the trees' terminal nodes was set to 1, the shrinkage factor applied to each tree in the expansion to 0.001 and the fraction of the training set observations randomly selected to propose the next tree in the expansion to 0.5. With these settings boosting regression trees with at most 8way interactions between SNPs required 3656 iterations for the training dataset based on inspecting graphical plots of the outofbag change in squared error loss against the number of iterations [9].
Support Vector Machines (SVMs)
where are positive weights given to each observation and estimated from the data and the inner product kernel K(x_{ i },x_{ j }) is a N × N symmetric and positive definite matrix [5]. Typically only a subset of are nonzero, and the associated observations are called support vectors, hence the name support vector machines. Since the solution depends on the input values only through the inner products K(x_{ i },x_{ j }), a flexible fitting is achieved by transforming the crossproducts using the kernel function (K(x_{ i },x_{ j })) that alters how two observations are related to each other.
We used the εinsensitive SVM regression with a linear kernel to predict GEBVs in the R package e1071[8] with an insensitivity zone of ε = 10 and a regularization (cost) parameter (λ > 0) of λ = 0.001 determined by grid search.
Assessing prediction performance
We used 5fold crossvalidation and the Pearson correlation between the simulated values and predicted GEBVs from the validation set and between the predicted and true breeding values (TBVs) for the nonphenotyped individuals constituting the fifth generation to quantify the predictive accuracy of each method. The training and validation sets respectively contained 60 and 15 crosses and encompassed all phenotyped individuals except the 20 founders.
Results and discussion
Table 1
CV/TBV  Sample size  Random Forests  Boosting  Support Vector Machines  Ridge Regression BLUP  

Mean  Range  Mean  Range  Mean  Range  Mean  Range  Mean  Range  
CV  439  416514  0.466  0.3920.534  0.503  0.4310.567  0.503  0.4320.567  0.530  0.4510.620 
TBV  900  0.483  0.547  0.497  0.607 

Table 1
The two ensemble methods can accommodate complex relationships and interactions (epistasis), which is a potential advantage, but the simulated data did not display many such interactions. A few simulated interacting SNPs with large effects were ranked highly but not topranked by RF possibly because RF and boosting had to randomly subsample the 10031 predictors. Thus, it may happen that the SNPs closest to a QTL are not sufficiently frequently sampled, so that the signal of the QTL is captured by other more distant SNPs. Consequently, the signal of a QTL gets blurred relative to classical QTL mapping approaches, which always scan all the markers. This may be one reason that these methods may not perform as well as some other much simpler competitors (e.g., RRBLUP, BayesB). Nevertheless, for data with complex traits controlled by many genes that show epistatic interactions, the machine learning methods hold much promise and perhaps may even outperform BLUP. Not surprisingly, Moser et al. [10] found the accuracy of SVMs to be the highest among five methods (including BLUP) used to predict GEBVs of dairy bulls from empirical data.
Conclusions
Predictive accuracies of all three methods were remarkably similar, but boosting and SVMs performed somewhat better than RF. Although boosting was only slightly better than the other methods, it holds perhaps the greatest promise for GS because of its wide versatility, allowing it to assume simpler, faster and more interpretable forms, such as componentwise boosting, able to incorporate automatic predictor selection.
Declarations
Acknowledgements
The German Federal Ministry of Education and Research (BMBF) funded this research within the AgroClustEr “Synbreed – Synergistic plant and animal breeding” (Grant ID: 0315526).
This article has been published as part of BMC Proceedings Volume 5 Supplement 3, 2011: Proceedings of the 14th QTLMAS Workshop. The full contents of the supplement are available online at http://www.biomedcentral.com/17536561/5?issue=S3.
Authors’ Affiliations
References
 Meuwissen THE, Hayes BJ, Goddard ME: Prediction of total genetic value using genomewide dense marker maps. Genetics. 2001, 157: 18191829.PubMed CentralPubMedGoogle Scholar
 Breiman L: Random forests. Machine Learning. 2001, 45: 532. 10.1023/A:1010933404324.View ArticleGoogle Scholar
 Liaw A, Wiener M: Classification and regression by randomForest. R News. 2002, 2: 1822.Google Scholar
 Statnikov A, Wang L, Aliferis CF: A comprehensive comparison of random forests and support vector machines for microarraybased cancer classification. BMC Bioinformatics. 2008, 9: 319324. 10.1186/147121059319.PubMed CentralView ArticlePubMedGoogle Scholar
 Hastie TJ, Tibshirani R, Friedman J: The elements of statistical learning. 2009, New York: Springer, SecondView ArticleGoogle Scholar
 Bishop CM: Pattern recognition and machine learning. 2006, New York: SpringerGoogle Scholar
 Piepho HP: Ridge regression and extensions for genomewide selection in maize. Crop Science. 2009, 49: 11651176. 10.2135/cropsci2008.10.0595.View ArticleGoogle Scholar
 Dimitriadou E, Hornik K, Leisch K, Meyer D, Weingessel A: Misc Functions of the Department of Statistics (e1071), TU Wien. R package. version 1.524. Available at http://cran.rproject.org/web/packages/e1071/
 Ridgeway G: Gbm: Generalized boosted regression models. R package. version 1.63.1. Available at http://cran.rproject.org/web/packages/gbm/
 Moser G, Tier B, Crump RE, Khatkar MS, Raadsma HW: A comparison of five methods to predict genomic breeding values of dairy bulls from genomwide SNP markers. Genet Sel Evol. 2009, 31: 4156. 10.1186/129796864156.Google Scholar
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.