Discriminant analysis assigns objects to one of several classes on the basis of attributes which characterize the objects. The success of classification depends on the selection of discriminatory attributes and on the choice of an assignment rule. In this paper we focus on the latter and discuss ways to obtain nonlinear classification rules through maximum likelihood, canonical components and projection pursuit. We use both linear and nonlinear methods to classify proteins into three secondary structural types: alpha, beta, and mixed alpha and beta or irregular. Using simple attributes, dependent on amino acid properties, we show that the rate of incorrect classification can be decreased by more than 15% when nonlinear methods are used.