Using supervised learning successful descriptors to perform protein structural classification through unsupervised learning

New protein sequences and structures are made available on an accelerated pace in public databases, which demands automatic reliable methods to make sense of this large volume of data. Many strategies for protein structural classification were proposed in the last years, using descriptors based on sequence and structure, as physicochemical properties of residues, solvent accessibility, and contact maps to perceive patterns that are able to characterize and predict protein structural families. In this paper, we assess whether it is possible to perform a structural classification of proteins through unsupervised learning using features successfully proposed in a well-established protein structural classifier. We conducted experiments using 5 clustering algorithms from different paradigms. The best result was achieved using agglomerative clustering with the Complete-Link algorithm (silhouette coefficient 0.843). Group labels were compared to protein superfamilies (CATH database) through the Fowlkes-Mallows Index to check if groups resulting from unsupervised learning correlated well with protein superfamilies. Our results show that even though successful, clustering analysis is not able to separate or classify the protein dataset in structural superfamilies in an efficient manner.