Multivariate Statistical Machine Learning Methods for Genomic Prediction

A Comparison of Three Machine Learning Methods for Multivariate Genomic Prediction Using the Sparse Kernels Method (SKM) Library

“…Additionally, the SKM library allows enables design of cross-validation strategies (cv_random and cv_kfold) with simple commands that are in agreement with real prediction scenarios of interest for breeders. Additionally, the SKM library offers some functions for summary of the prediction accuracy (numeric_summary(), categorical_summary() and gs_summaries()), as well as many options for metrics (for example, Pearson’s correlation (Cor), mean square error (MSE), mean absolute error (MAE), root mean square error (RMSE), normalized root mean square error (NRMSE), coefficient of determination (R2) and mean arctangent absolute percentage error (MAAPE)) for continuous response variables and for categorical response variables (proportion of cases correctly classified (PCCC), kappa coefficient (Kappa), Brier score, sensitivity and specificity) [ 8 ].…”

“…We implemented sevenfold cross validation for each of the 6 datasets [ 8 ]. Therefore, we randomly divided the dataset into 7 subsets of similar size, using subsets as a training set and the remaining group as a test set until each of the 7 subsets played the role of test set once.…”

“…However, there are still many factors that need to be carefully considered for the successful practical application of GS. These factors include (1) the size and degree of relatedness of the training and testing sets, (2) the complexity of the trait that needs to be predicted, (3) the quality of the phenotypic and marker data available for the training process, (4) the goal of prediction (for example, predicting new lines in a complete year or location or predicting some lines that are missing in some environments but present in others) and (5) the statistical machine learning method that needs to be implemented [ 2 , 3 , 8 ].…”

“…Bayesian methods have the advantage that with default hyperparameters, they work well in many circumstances and are very robust in the face of prediction problems, even with small datasets, and can be used on a personal computer without the need for sophisticated servers with many cores. Additionally, conventional mixed models and Bayesian methods (GBLUP, BayesA, BayesB, BayesC and Bayesian Lasso) have recently been implemented successfully using kernels, which increase the power of these methods, as with the use of non-linear kernels, they are able to capture not only linear but also non-linear patterns more efficiently [ 5 , 8 ].…”

“…About the PLS method, few applications have been reported in the context of genomic prediction; however, this method has a long history of prediction in the chemometrics and biological sciences due to its ability to efficiently deal with datasets with correlated inputs and more independent variables than observations. This method requires tuning of the number of principal components to be used; however, because this is the only hyperparameter to tune, the selection of this optimal hyperparameter does not require expensive computation [ 8 ].…”