Predicting the druggability of protein-protein interactions based on sequence and structure features of active pockets.

Protein-protein interactions (PPIs) are becoming highly attractive targets for drug discovery. Motivated by the rapid accumulation of PPI data in public database and the success stories concerning the targeting of PPIs, a machine-learning method based on sequence and structure properties was developed to access the druggability of PPIs. Here, a comprehensive non-redundant set of 34 druggable and 122 less druggable PPIs were firstly presented from the perspective of pockets. When tested by outer 5-fold cross-validation, the most representative model in discriminating the druggable PPIs from the less-druggable ones yielded an average accuracy of 88.24% (sensitivity of 82.38% and specificity of 92.00%). Moreover, a promising result was also obtained for the independent test set. Compared to other methods, the method gives a comparative performance, which is most likely due to the construction of a training set that encompasses less druggable PPIs and also the information of active pockets that have evolved to bind a natural ligand.