Protein Structure and Function: Application of Bioinformatics Methods - John Rigden 2014

Integrated servers for structure-based function prediction
ProKnow
Prediction success rate

Figure 10.3 illustrates a subset of the results for our example, the 2fck Structure. According to Dali, the folding type aligns almost exclusively with acetyltransferases. A BLAST search against the UniProt database also reveals A number of strong matches with acetyltransferases. Meanwhile, the RIGOR search suggests several false leads, such as a fibroblast growth factor, the lipid-binding protein lipovitellin, and an integrase. The PROSITE results yield even more dead ends, pointing to short motifs—two of which serve as phosphorylation sites and one as a myristoylation site (all annotated on the PROSITE website with the disclaimer: “In some cases, this entry may be disregarded by the program as being too nonspecific”). A DIP search returned no results. Nevertheless, the overwhelmingly strong prediction proved to be the correct one; namely, that the protein is an acetyltransferase. Consequently, the overall prediction in this case is correct.

Overall, the ProKnow server performs quite well. The authors tested it on a non-redundant set of Proteins with known Functions and found that approximately 70% of the functional annotations proved to be correct (Pal and Eisenberg 2005). Broader predictions (such as hydrolase) tend to be more accurate than specific ones (such as leucyl aminopeptidase). Prediction accuracy has been slightly enhanced by the recent Integration of the Prolinks database—which was absent in the initial version—and will continue to improve as Prolinks expands in size.