Supervised Classification Applied to Hot-Spot Detection in Protein-Protein Interfaces

The identification of protein complexes and interactions is crucial for the understanding of cellular organization and machinery. Due to the high difficulty in attaining experimental data about such an important subject, computational tools and methodologies are emerging as reliable alternatives. It is especially true that Machine-Learning (ML) algorithms hold an incredible promise for protein interaction research by identifying biological relevant patterns, which accelerates our knowledge of the functional mechanism of proteins within the cells. Over the last decades the improvement of a large number of computational techniques led to significant cost decreases and, also, increases in throughput by orders of magnitude. However, there is still room for improvement as their accuracy is still far from optimal. In this work, we have developed and applied computer modelling techniques that went beyond the current state-of-the-art, leading to quantitative and reliable molecular-level predictions of Hot-Spots at protein-protein complexes. We explored the feasibility of using ML in the HS detection and compared different classifiers as well as different preprocessing conditions. Based on this evaluation, we concluded that applying the C5.0 algorithm with minor class up-sampling leads to accurate results. The overall accuracy in an independent test set demonstrated to be 0.88. Due to the theme’s relevance to the large scientific community working on structural biology, we have assembled a freely available web-server that can be found at: http://milou.science.uu.nl/cgi/servicesdevel/SPOTON/spoton/