Abstract

Biological databases are sparse, huge and redundant. Therefore, knowledge inference from those databases needs a consistent approach. Widely accepted as a most complex process of protein modification, glycosylation has been the main focus in this study. In this process a simple chain of carbohydrates attaches to a target protein at a specific amino acid, so-called glycosylation site. Epidermal Growth Factor-Like (EGFL) repeats have been the target proteins of this study because of having a particular glycosylation process. Moreover, they may associate with many type of cancer as well as other diseases. The objective of this study was to detect and predict the number of glycosylation sites in EGFL protein sequences using feed-forward neural networks. Bayesian automated regularization was exploited to prune the unnecessary weights and biases of the feed-forward neural network. The result of applying eight learning algorithms showed that One Step Secant (OSS) learning algorithm is more reliable than the others in terms of the accuracy and performance as measured in this study. The Bayesian regularized neural network outperformed OSS method according to the employed assessment measures. Compared to the existing neural detectors, Bayesian automated learning could improve the consistency of the model by 39.48%. The concept of Reduction Factor was also introduced to determine the efficiency of Bayesian automated learning quantitatively. Glycobiologists can use and validate such connectionist models to choose and study on the selected EGF-like proteins which are associated with cell malignancy.