Using protein-protein interactions data to improve predictions of the effect of aging-related genes on the longevity of model organisms

Abstract

This work focuses on improving the predictive performance on the task of classifying the e↵ect of aging-related genes on the longevity of model organisms. Usually, datasets for this problem are built using gene ontology features. We show that, by introducing a new set of features based on protein-protein interactions, we can improve the classification in a statistically significant way. Although beneficial, the introduction of these features brings a new problem for this task: performing classification based on uncertain information. This happens because the values of the new features only represent how certain we can be that an interaction exists between two proteins. To cope with this problem, we propose a novel similarity coe cient, based on the Jaccard index, and show that, by using it, we can benefit from the introduced uncertain data without adding too much complexity to the classification solution. Finally, we demonstrate how to achieve even better results on this task by performing a feature selection procedure