Prediction of enzyme catalytic sites on protein using a graph kernel method

Abstract

Structural Genomics projects are producing structural data for proteins at an unprecedented speed. The functions of many of these protein structures are still unknown. To decipher the functions of these proteins and identify functional sites on their structures have become an urgent task. In this study, we developed an innovative graph method to represent protein surface based on how amino acid residues contact with each other. Then, we implemented a shortest-path graph kernel method to measure the similarities between graphs. We tried three variants of the nearest neighbor method to predict enzyme catalytic sites using the similarity measurement given by the shortest-path graph kernel. The prediction methods were evaluated using the leave-one-out cross validation. The methods achieved accuracy as high as 77.1%. We sorted all examples in the order of decreasing prediction scores. The results revealed that the positive examples (catalytic site residues) were associated with higher prediction scores and they were enriched in the region of top 10 percentile. Our results showed that the proposed methods were able to capture the structural similarity between enzyme catalytic sites and would provide a useful tool for catalytic site prediction.