Predicting catalytic resiudes

Determining the functions of proteins can be greatly facilitated if the functional roles of protein residues can be identified first. These functional residues are valuable clues for improving the quality of functional annotation. In this work, we focus on the task of predicting one such class of functional residues -- catalytic residues in enzymes. To this end, we investigate the utilities of novel machine learning algorithms for building classifiers automatically from annotated data sets. Such computational tools are essential in response to the explosive growth of available genome sequencing data. Our approaches hinge on several core intuitions: a residue that lies near other putative catalytic residues in the 3D structure is more likely to be catalytic. Therefore, in building classifiers, it is beneficial to consider not only features derived from the residue and its sequential neighbors but also features of other residues that are from its 3D structural neighborhood. In this work, we show empirically that our probabilistic models for predictions indeed attain significantly better performance by exploiting these structural cues.

Related publications