Computational classification of proteins using methods such as string kernels and Fisher-SVM has demonstrated great success. However, the resulting models do not offer an immediate interpretation of the underlying biological mechanisms. In particular; some recent studies have postulated the existence of a small subset of positions and residues in protein sequences may be sufficient to discriminate among different protein classes. In this work, we propose a hybrid setting for the classification task. A generative model is trained as a feature extractor, followed by a sparse classifier in the extracted feature space to determine the membership of the sequence, while discovering features relevant for classification. The set of sparse biologically motivated features together with the discriminative method offer the desired biological interpretability. We apply the proposed method to a widely used dataset and show that the peqormance of our models is comparable to that of the state-of-the-art methods. The resulting models use fewer than 10% of the original features. At the same time, the sets of critical features discovered by the model appear to be consistent with confirmed biological findings.