Skip to Main Content
In this paper we use a modular neural network to predict the molecular functions of yeast proteins. To solve this class problem, our proposed approach decomposes the original problem into a set of solvable 2-class subproblems using class information. Each 2-class problem has a set of positive and negative data. The yeast data is not equally distributed in function classes and hinders the learning of each neural network. We adopt a sampling strategy that generates a set of new class data to the subordinate class in order to balance the positive and negative data set. In data preparation, the biological concept of "guilt-by-interaction" is used for covering possible interaction partners among proteins of known functions. The proposed framework has been tested as a predictive model of yeast protein functions where the data source is stored in a relational database. In the experiments, the proposed system shows an average accuracy of 91.0% in the test set.