Prediction of protein translation initiation site from the perspective of imbalanced classes

The correct prediction of protein translation initiation from messenger RNA (mRNA) is an important activity for genomic annotation. This problem is known to be highly imbalanced, since each molecule has a single mRNA translation initiation site and several others which are not initiator AUGs. In this context, the present work has focused on undersampling methods for balancing class distribution proposed to solve such problem. Here, we present the results obtained from the Condensed Nearest Neighbor Rule (CNN), Tomek, and random sampling methods already known in the literature, and introduce a new one for balancing, namely C-Blocks, which is based on clustering. Using the concept of Bagging [1], the results of these classifiers is combined to obtain a final ranking of the sequence.