Motivation: Meaningful scores to assess protein structure similarity are essential to decipher protein structure and sequence evolution. The mining of the increasing number of protein structures requires fast and accurate similarity measures with statistical significance. Whereas numerous approaches have been proposed for protein domains as a whole, the focus is progressively moving to a more local level of structure analysis for which similarity measurement still remains without any satisfactory answer.
Results: We introduce a new score based on Binet-Cauchy kernel. It is normalized and bounded between 1-maximal similarity that implies exactly the same conformations for protein fragments-and -1-mirror image conformations, the unrelated conformations having a null mean score. This allows for the search of both similar and mirror conformations. In addition, such score addresses two major issue of the widely used root mean square deviation (RMSD). First, it achieves length independent statistics even for short fragments. Second, it shows better performance in the discrimination of medium range RMSD values. Being simpler and faster to compute than the RMSD, it also provides the means for large-scale mining of protein structures.
Availability and implementation: The computer software implementing the score is available at http://bioserv.rpbs.univ-paris-diderot.fr/BCscore/
Supplementary information: Supplementary data are available at Bioinformatics online.