Predicting pathogenic single nucleotide variants (SNVs) in non-coding regions of the human genome presents a significant challenge for the extreme class imbalance between pathogenic “positive” variants and physiological “negative” ones, since most machine learning methods are biased toward predicting negative examples. We designed two “block-shaped” tabular-DNN architectures: a Modular Block-Deep Neural Network (MoB-DNN) and a tabular Residual Network (T-ResNet), able to address the class imbalance problem through a mini-batch balancing strategy. We employed a hierarchical optimization approach to efficiently tune hyper-parameters related to training procedure, architecture, batch size, and mini-batch balancing ratio. Our experimental results demonstrate that T-ResNet outperforms and MoB-DNN shows competitive performance with a state-of-the-art hyper-ensemble method, suggesting that residual connections provide significant advantages for capturing complex patterns in non coding regions of the human genome.