Genome-Wide Association Studies (GWAS) face unique challenges in the era of big genomics data, particularly when dealing with ultra-high-dimensional datasets where the number of genetic features significantly exceeds the available samples. This paper introduces an extension to the feature selection methodology proposed by Mirzaei et al. (2020), specifically tailored to tackle the intricacies associated with ultra-high-dimensional GWAS data. Our extended approach enhances the original method by introducing a Frobenius norm penalty into the student network, augmenting its capacity to adapt to scenarios characterized by a multitude of features and limited samples. Operating seamlessly in both supervised and unsupervised settings, our method employs two key neural networks. The first leverages an autoencoder or supervised autoencoder for dimension reduction, extracting salient features from the ultra-high-dimensional genomic data. The second network, a regularized feed-forward model with a single hidden layer, is designed for precise feature selection. The introduction of the Frobenius norm penalty in the student network significantly boosts the method's resilience to the challenges posed by ultra-high-dimensional GWAS datasets. Experimental results showcase the efficacy of our approach in feature selection for GWAS data. The method not only handles the inherent complexities of ultra-high-dimensional settings but also demonstrates superior adaptability to the nuanced structures present in genomics data. The flexibility and versatility of our proposed methodology are underscored by its successful performance across a spectrum of experiments.