Linear Epitope Prediction Using Statistics of Amino Acid Segments and SVM

Abstract

B-cell epitopes play an important role for developing synthetic peptide vaccines and inducing antibody responses. Applying biological experiments for epitope identification is time consuming and demands a lot of experimental resources. Therefore, it is useful and challenging to develop a linear epitope prediction system with high precision rates based on the computational technologies. In this paper, a combinatorial method to improve physico-chemical property based linear epitope prediction systems is proposed. Here, a collected set of verified epitope and non-epitope segments ranging from 2 to 4 amino acids were trained and applied as the features of SVM (Support Vector Machine). With the combination of physico-chemical characteristics and SVM classifier, the performance can be effectively improved. Ten HIV related antigens were adopted for system evaluation, and the results showed that the proposed method decreased the average sensitivity with 8.7% while the average specificity was increased by 22%. The average accuracy was increased by 9.6% and the average positive predictive value was increased by 12.7%. In comparison with the well-known BepiPred system, experimental results have shown that the proposed system outperforms BepiPred system in all respects, including a higher average sensitivity of 3.2%, specificity of 15.5%, accuracy of 8.8%, and positive predictive value of 12.9%.