Results
Using two of the comprehensive databases currently available for protein-protein interaction and protein modification site data (PDB and PhosphoSitePlus, respectively), we created new databases that map PTMs to their locations inside or outside of PPIRs. The mapped PTMs represented only 5% of all known PTMs. In order to predict localization within or outside of PPIRs for the vast majority of PTMs, a machine learning strategy was used to generate predictive models from these databases. First, the three mapped PTM databases (for acetylation, phosphorylation, and ubiquitylation) which had sufficient numbers of modification sites for generating models were encoded numerically using a specific subset of the AAindex relevant to protein structure. The support vector machine (SVM) was employed to perform classification tasks and model refinement via sequential feature selection procedures and rank based optimization, which
preserved information that would be lost with many other methods of optimization. The resulting models yielded high overall predictive performance as judged by a combined performance score (CPS). Among the multiple properties of amino acids that were used in the classification tasks, hydrophobicity was found to contribute substantially to the performance of the final predictive models. Compared to the other classifiers we also evaluated, the SVM provided the best performance overall.