L. Lamontagne and M. Marchand (Eds.): Canadian AI 2006, LNAI 4013, pp. 110 – 121, 2006. © Springer-Verlag Berlin Heidelberg 2006 Bayesian Learning for Feed-Forward Neural Network with Application to Proteomic Data: The Glycosylation Sites Detection of the Epidermal Growth Factor-Like Proteins Associated with Cancer as a Case Study Alireza Shaneh 1 and Gregory Butler 1 1 Research Laboratory for Bioinformatics Technology, Department of Computer Science and Software Engineering, Concordia University, 1515 St-Catherine West, Montreal, Quebec H3G 2W1, Canada {a_dariss, gregb}@cs.concordia.ca http://www.cs.concordia.ca Abstract. There are some neural network applications in proteomics; however, design and use of a neural network depends on the nature of the problem and the dataset studied. Bayesian framework is a consistent learning paradigm for a feed-forward neural network to infer knowledge from experimental data. Bayesian regularization automates the process of learning by pruning the un- necessary weights of a feed-forward neural network, a technique of which has been shown in this paper and applied to detect the glycosylation sites in epi- dermal growth factor-like repeat proteins involving in cancer as a case study. After applying the Bayesian framework, the number of network parameters de- creased by 47.62%. The model performance comparing to One Step Secant method increased more than 34.92%. Bayesian learning produced more consis- tent outcomes than one step secant method did; however, it is computationally complex and slow, and the role of prior knowledge and its correlation with model selection should be further studied. 1 Introduction Proteomic data are huge, sparse and redundant, and dealing with these characteristics is a challenge and requires powerful methods to infer knowledge. Soft computing techniques have been extensively used to mine and extract as much necessary infor- mation as possible from a large set of protein sequences. Generally speaking, the soft computing methods used to solve optimization problems are divided into computa- tional, statistical, and metaheuristic frameworks. Computational approaches optimize the learning algorithm by predicting the future state of a solution based on the past evaluation of data in both supervised and unsupervised ways. Artificial neural net- works are obvious examples of such methods. Statistical learning theory emphasizes on the statistical methods used for automated learning; for example, kernel-based methods such as support vector machines [24] find an optimal separating hyperplane by mapping data to a higher dimensional search space. Metaheuristic algorithms are