The seeding technique became central in the theory of sequence alignment and there are several efficient tools applying seeds to DNA homology search. Recently, a concept of subset seeds has been proposed for similarity search in protein sequences. We experimentally evaluate the applicability of subset seeds to protein homology search. We advocate the use of multiple subset seeds derived from a hierarchical tree of amino acid residues. Our method computes, by an evolutionary algorithm, seeds that are specifically designed for a given protein family. The representation of seeds by deterministic finite automata (DFAs) is developed and built into the NCBI-BLAST software. This extended tool, named SeedBLAST, is compared to the original NCBI-BLAST and PSI-BLAST on several protein families. Our results demonstrate a superiority of SeedBLAST in terms of efficiency, especially in the case of twilight zone hits. SeedBLAST is an open source software freely available http://bioputer.mimuw.edu.pl/papers/sblast . Supplementary material and user manual are also provided.