SELF-ORGANAZING MAPS FOR ASSESSING GENETIC DIVERSITY AND TRAIT IMPORTANCE IN TROPICAL MAIZE HYBRIDS

We used an unsupervised neural network, Self-Organizing Maps (SOM), to assess genetic diversity and traits importance for 123 commercial hybrids of tropical maize cropped under normal and stress N input. We evaluated traits related to nitrogen use and efficiency, grain yield components, and plant architecture over two summer seasons. The likelihood ratio test identified genetic variability for nearly all traits, except for N remobilized at each N input. Random forest was consistent with SOM in identifying important traits in the population context. However, SOM shows the patterns of traits in each neuron (group of hybrids). In the SOM neighborhood topology, it was possible to find neurons with high grain yield that exhibit different patterns of trait importance. We suggested that hybrids located in distant neurons from each other and exhibiting favorable traits are complementary. SOM proved to be efficient in assessing the genetic diversity of hybrids, identifying trait importance and in the organization of the studied germplasm.