Feature selection is very important procedure in many pattern recognition problems. It is effective in reducing dimensionality, removing irrelevant data, and increasing accuracy of a classifier. In our previous work we propose a classifier combining the support vector machine (SVM) classifier with regularized discriminant analysis (RDA) classifier used to protein fold recognition problem. However high dimensionality of the feature vectors and small number of samples in the training data set caused that the problem is ill-posed for an RDA classifier and the feature selection is crucible for the accuracy of the classifier. In this paper we propose a simple and effective algorithm based on the class similarity which solves our problem and helps us to achieve very good acuracy on a real-world data set.