Bootstrapping and normalization for enhanced evaluations of pairwise sequence comparison

The exponentially growing library of known protein sequences represents molecules connected by an intricate network of evolutionary and functional relationships. To reveal these relationships, virtually every molecular biology experiment incorporates computational sequence analysis. The workhorse methods for this task make alignments between two sequences to measure their similarity. Informed use of these methods, such as NCBI BLAST [1], WU-BLAST [2], FASTA [3] and SSEARCH, requires understanding of their effectiveness. To permit informed sequence analysis, we have assessed the effectiveness of modern versions of these algorithms using the trusted relationships among ASTRAL [4] sequences in the Structural Classification of Proteins [5] database classification of protein structures [6]. We have reduced database representation artifacts through the use of a normalization method that addresses the uneven distribution of superfamily sizes. To allow for more meaningful and interpretable comparisons of results, we have implemented a bootstrapping procedure. We find that the most difficult pairwise relations to detect are those between members of larger superfamilies, and our test set is biased toward these. However, even when results are normalized, most distant evolutionary relationships elude detection.