Classification of LTR Retrotransposons via Interaction Prediction

Transposable Elements (TEs) are genetic sequences that can relocate within the genome, thus promoting genetic diversity. Classifying TEs in eukaryotes involves a hierarchy formed by classes, subclasses, orders, superfamilies, families, and subfamilies. According to this taxonomy, LTR retrotransposons (LTR-RT) constitute an order. The primary objective of this study is to explore the classification of LTR retrotransposons at the superfamily level. This was achieved by predicting interactions between LTR-RT sequences and conserved protein domains using Predictive Bi-Clustering Trees (PBCTs). Two datasets were used to investigate the relationships among different superfamilies. The first one comprised LTR retrotransposon sequences assigned to Copia, Gypsy, and Bel-Pao superfamilies, whereas the second dataset included consensus sequences of the conserved domains for each superfamily. Therefore, the PBCT decision tree tests could relate to both sequence and class attributes. In the classification process, interaction is interpreted as either the presence or absence of a domain in a given LTR-RT sequence. Subsequently, this sequence is classified into the superfamily with the highest number of predicted domains. Precision-recall curves were adopted as evaluation metrics for the method, and its performance was compared to some of the most commonly used models in the task of transposable element classification. Experiments on D. melanogaster and A. thaliana showed that PBCTs are promising and comparable to other methods, especially in the classification of the Gypsy superfamily.