Neural networks and dimensionality reduction to increase predictive efficiency for complex traits
The study of complex traits using large databases of molecular markers has reshaped genetic breeding programs as it allows the direct incorporation of information from a large number of molecular markers for the prediction of genomic values. However, the large number of markers can lead to problems of computational demand, multicollinearity, and dimensionality. We evaluated the use of Multilayer Perceptron Neural networks to resolve this problem and propose a new dimensionality reduction method called Probe Subset Selection Methodology, for the prediction of genetic values, in Genome Wide Selection studies. We used a simulated F1 population for 12 quantitative traits, including different modeling structures, average degrees of dominance and heritability. The Multilayer Perceptron Neural Networks, together with the proposed Probe Subset Selection Methodology, provided more accurate predictions than the RR-BLUP methodology and reduced the root mean square error from 577.249 to values below 24. The use of computational intelligence in breeding programs is a promising tool for prediction purposes, since epistasis and dominance were not limiting factors for the proposed Multilayer Perceptron Neural Network method.