From Protein Structure to Function with Bioinformatics.pdf

11 Function Predictions of Structural Genomics Results 279determine how successful the structure-based methods in ProFunc were at predictingfunction. These 93 proteins of known function were submitted to the ProFuncserver and the top scoring matches from each method retrieved and stored. Theresults were then backdated to the deposition date for each structure to ensure thatwhat was being measured was how successful the server would have been had itbeen fully operational from the start. The resultant top hit for each method was thenmanually compared with the known function for each protein and a judgementmade as to whether the prediction was correct.The results from the study indicated that, of the methods available as part of theProFunc server, the fold recognition and “reverse template” approaches were themost successful with approximately 60% of the known functions identified correctly.Detailed investigation revealed that both of these methods often identify thesame function by matching to the same protein but cases could be found where onemethod succeeded where the other failed. This is due to the fact that the fold matchingis looking at the global similarity in compared proteins whereas the “reversetemplate” approach is a very local comparison. One major drawback identified inthe study was its inability to address the question of how structure-based approachescompare against sequence-based approaches. However, this is a generic problemwhich has not been adequately addressed in the literature due to the immense difficultiesinvolved in accurately rolling back the sequence databases, as well as patternsand profiles derived from them, to a particular date.In addition to the general comparison of methods Watson et al. (2007) presentedsome specific examples of function predictions, sometimes verified. An example ofsuccessful structure-based function prediction was the BioH protein fromEscherichia coli (Sanishvili et al. 2003). This protein was known to be involved inbiotin synthesis but no biochemical function had been assigned to it. Analysis ofthe structure using ProFunc returned a highly significant match (with an RMSD of0.28 Å) to an enzyme active-site template for the Ser-His-Asp catalytic triad of thelipases. Fold comparison using DALI indicated structural similarity to a largenumber of proteins with a variety of enzymatic functions although the sequenceidentity of these hits was low, ranging from 15–25%. The closest matches includeda bromoperoxidase (EC 1.11.1.10), an aminopeptidase (EC 3.4.11.5), two epoxidehydrolases (EC 3.3.2.3), two haloalkane dehalogenases (EC 3.8.1.5), and a lyase(EC 4.2.1.39). Only through extensive manual analysis of these enzymes and a literaturereview would it have become obvious that each of these contain a Ser-His-Aspcatalytic triad in their active sites. The enzyme active-site template search identifiedthe presence of the catalytic triad instantly. Experimental characterisation of thisprotein revealed it to be a novel carboxylesterase acting on short acyl chainsubstrates (Sanishvili et al. 2003).A further example illustrating how functional clues can be derived from thestructure involves a hypothetical protein (IsdG) from Staphylococcus aureus.Sequence-based analysis by methods in ProFunc revealed a variety of functions,including antibiotic biosynthesis monooxygenase, cysteine peptidase, oxidoreductase,methyltransferase, epimerase, transportation, possible RNA binding, andothers. When the structure was examined using the MSDfold/SSM service, the top