A diffusion kernel-based approach for protein domain identification

It is almost half a century since the concept of protein domain, as compact and recurring units that are ableto fold and function independently, was introduced. Nevertheless, the inherent ambiguity of the definitionbesides the increasing number of newly solved structures keeps the accurate automated methods in highdemand. Contrary to the majority of the state-of-the-art methods, we employed enhanced measures ofproximity between amino acids rather than developing context-specific clustering algorithms. Here, thepower of kernel functions to separate structural domains in their corresponding Hilbert spaces is investigated.For this purpose, utilizing four different diffusion kernels on protein graphs, a novel pipeline for proteindomain assignment is developed. The result of the presented method on commonly used benchmark data setsshows a marginally better performance compared to the best available methods based on two differentmetrics. Moreover, by offering alternative partitionings, our method answers the problem of subjectivity inprotein domain definition. The high prediction accuracy of the approach reveals the diffusion kernels'potential to split entangled structures of complex proteins. In addition to out-competing other methods bymerely employing general (rather than context-specific) clustering algorithms, our pipeline provides theversatility to implement other graph node kernels that can potentially boost its performance.

Keywords:

Protein structure , Graph node kernel , Protein domain assignment , Clustering , Diffusion kernel