The accelerated growth of protein databases offers great possibilities for the study of protein function using sequence similarity and conservation. However, the huge number of sequences deposited in these databases requires new ways of analyzing and organizing the data. It is necessary to group the many very similar sequences, creating clusters with automated derived annotations useful to understand their function, evolution, and level of experimental evidence. We developed an algorithm called FastaHerder2, which can cluster any protein database, putting together very similar protein sequences based on near-full-length similarity and/or high threshold of sequence identity. We compressed 50 reference proteomes, along with the SwissProt database, which we could compress by 74.7%. The clustering algorithm was benchmarked using OrthoBench and compared with FASTA HERDER, a previous version of the algorithm, showing that FastaHerder2 can cluster a set of proteins yielding a high compression, with a lower error rate than its predecessor. We illustrate the use of FastaHerder2 to detect biologically relevant functional features in protein families. With our approach we seek to promote a modern view and usage of the protein sequence databases more appropriate to the postgenomic era.
Keywords: cluster analysis; clustering; computational biology; data mining; databases.