Abstract
Similarity search in protein databases is one of the most essential issues in proteomics. With the growing number of experimentally solved protein structures, the focus shifted from sequence to structure.
The area of structure similarity forms a big challenge since even no standard definition of optimal similarity exists in the field. In this paper, we propose a protein structure similarity method called SProt.
SProt concentrates on high-quality modeling of local similarity in the process of feature extraction. SProt's features are based on spherical spatial neighborhood where similarity can be well defined.
On top of the partial local similarities, global measure assessing similarity to a pair of protein structures is built. SProt outperforms other methods in classification accuracy, while it is at least comparable to the best existing solutions in terms of precision-recall or quality of alignment.