Principal Component Analysis for Prediction of O-Linked Glycosylation Sites in Protein by Multi-Layered Neural Networks

Glycosylation is one of the common post-translation modification of protein in eukaryotic cells. Conventional neural network methods have been applied to predict glycosylation sites in protein sequence and the prediction accuracy is dependent on the dimension of feature vector (length of protein sequence). Though the prediction accuracy can be improved by increasing the length of protein sequence, it is time-consuming. In this paper, we propose a novel approach which combines PCA with a multilayer neural network for efficient, and accurate prediction of O-glycosylation sites in protein. PCA is first used to extract main basis (subspace) of the protein sequence. The lower-dimensional projections on the subspace are used as features instead of the higher-dimensional protein sequences. Compared with conventional method (without PCA), our proposed method can significantly reduce the larger computation cost, while can keep the higher prediction accuracy.