Superiority of artificial neural networks for a genetic classification procedure

Abstract

The correct classification of individuals is extremely important for the preservation of genetic variability and for maximization of yield in breeding programs using phenotypic traits and genetic markers. The Fisher and Anderson discriminant functions are commonly used multivariate statistical techniques for these situations, which allow for the allocation of an initially unknown individual to predefined groups. However, for higher levels of similarity, such as those found in backcrossed populations, these methods have proven to be inefficient. Recently, much research has been devoted to developing a new paradigm of computing known as artificial neural networks (ANNs), which can be used to solve many statistical problems, including classification problems. The aim of this study was to evaluate the feasibility of ANNs as an evaluation technique of genetic diversity by comparing their performance with that of traditional methods. The discriminant functions were equally ineffective in discriminating the populations, with error rates of 23-82%, thereby preventing the correct discrimination of individuals between populations. The ANN was effective in classifying populations with low and high differentiation, such as those derived from a genetic design established from backcrosses, even in cases of low differentiation of the data sets. The ANN appears to be a promising technique to solve classification problems, since the number of individuals classified incorrectly by the ANN was always lower than that of the discriminant functions. We envisage the potential relevant application of this improved procedure in the genomic classification of markers to distinguish between breeds and accessions.