Prediction of Transposon Insertion Sites in the Human Genome Using Deep Learning Techniques

Transposable elements (TEs) account for approximately 44% of the human genome but less than.05% are known to be active. In contrast, these TEs are highly mobile in the genome which can result genomic rearrangement and several associated human diseases. It is notoriously difficult to predict the properties of DNA sequences at prospective insertion sites that make them good elements. This study proposes a deep learning model that leverages DNA properties to predict potential transposon insertion sites in nucleotide sequences, focusing on the active L1 and Alu elements identified in human chromosome 3 using RepeatMasker. Positive datasets comprise 30 bp upstream sequences of detected L1 and Alu elements, while negative datasets are randomly selected 30 bp sequences without these elements. Employing DNASCANNER v2, we computed 159 DNA properties encompassing nucleotide composition, structural traits, physicochemical attributes, thermodynamics, and mechanics. The data was partitioned into training, validation, and testing sets, scaled, and normalized. Fully Connected Layers were utilized to construct the deep learning models, which underwent hyperparameter tuning and training on the training subset. Evaluation metrics like sensitivity, precision, accuracy, recall, specificity, and AUC assessed model performance. The proposed Alu_L1_TIS DL model aims to provide a versatile predictive tool for identifying potential Alu or L1 TIS in the human genome. The web server for TIS predictions can be accessed at https://dnasv2.kamalrawal.in/. This research contributes to a better understanding of the role of transposable elements in human diseases and genomic instability.