Protein sequence redundancy reduction: comparison of various methods

Non-redundant protein datasets are of utmost importance in bioinformatics. Constructing such datasets means removing protein sequences that overreach certain similarity thresholds. Several programs such as 'Decrease redundancy', 'cd-hit', 'Pisces', 'BlastClust' and 'SkipRedundant' are available. The issue that we focus on here is to what extent the non-redundant datasets produced by different programs are similar to each other. A systematic comparison of the features and of the outputs of these programs, by using subsets of the UniProt database, was performed and is described here. The results show high level of overlap between non-redundant datasets obtained with the same program fed with the same initial dataset but different percentage of identity threshold, and moderate levels of similarity between results obtained with different programs fed with the same initial dataset and the same percentage of identity threshold. We must be aware that some differences may arise and the use of more than one computer application is advisable.