Prediction of protein-protein interaction inhibitors by chemoinformatics and machine learning methods

We describe a collection of structurally diverse inhibitors of protein-protein-interactions (PPIs). This collection is compared against the FDA drug database and a subset of the ZINC database by machine learning methods which rely on classical QSAR descriptors. We obtain a decision tree that contains three descriptors. Of particular importance is a constitutional descriptor related to molecular shape and size. Validation of the decision tree by various procedures indicates that it does not result from chance correlations and has predictive value. We conclude that constitutional descriptors may be valuable tools in the preselection of potential PPI inhibitors from compound databases.