Exploiting genomic sequences for gene expression prediction

The increasing interest in the non-protein-coding portion of the genome is one of the most challenging aspects to comprehend the most basic interactions of our DNA, which are almost entirely unknown deeply. The primary purpose of this thesis is to exploit the active parts of the Promoters to predict the Gene Expression value related to the protein-coding genes of Lung Healthy tissues. Therefore, I implemented a Deep Learning approach based on a custom-designed Convolutional Neural Network (CNN). The promoters are empirically associated with their genes, defining a new method to merge the gene to its regulatory part. This thesis also investigates the possible influence of other regulatory elements that can affect the final expression value. The on-purpose-designed Neural Network has been trained on about 18'000 active promoters evaluating the gene Expression Level with a final value of 36% and 56,5% of the variation in actual data from genomic studies in Human healthy samples and in vitro cell lines cultures of Human samples, respectively. Another significant result is that the proposed approach has comparable outcomes with similar tasks but has the lowest complexity of the Neural Network designed to increase performances in terms of resources and time needed to be executed and trained.