Prediction of Protein Function by Discriminant Analysis*

Approximately 53% of the protein sequences in the National Biomedical Research Foundation (NBRF) database can be allocated to one of 26 functional classes, each of which can be characterized by the joint occurrence of four or fewer attributes. The attributes reflect collective physicochemical properties of the sequences in a class, ranging from simple characteristics of composition, such as average hydrophobicity and net charge, to amphipathicity and the propensities of various residues to be in certain preferred configurations. In some, though not all instances, these variables can be related in a general way to topological or other structural features of the particular class they characterize. We show that the attributes permit 17 of the 26 groups to be filtered from all other proteins in the database with a misclassification error of less than 2%, and that the remaining 9 groups can be filtered with errors not exceeding 13%. Thus for a given functional class, the results point to the existence of relatively few characteristic variables which capture most of the intraclass similarity and interclass variability that is common and peculiar to members of that class.