Determinants of antigenicity and specificity in immune response for protein sequences
1 Department of Biostatistics and Computational Biology, Dana-Farber Cancer Institute & Harvard School of Public Health, Boston 02115 MA, USA
2 State Key Laboratory of Software Development Environment, Beihang University, Beijing, 100191, China
3 Institute for Genomics and Systems Biology, The University of Chicago, Chicago 60637 IL, USA
BMC Bioinformatics 2011, 12:251 doi:10.1186/1471-2105-12-251Published: 21 June 2011
Target specific antibodies are pivotal for the design of vaccines, immunodiagnostic tests, studies on proteomics for cancer biomarker discovery, identification of protein-DNA and other interactions, and small and large biochemical assays. Therefore, it is important to understand the properties of protein sequences that are important for antigenicity and to identify small peptide epitopes and large regions in the linear sequence of the proteins whose utilization result in specific antibodies.
Our analysis using protein properties suggested that sequence composition combined with evolutionary information and predicted secondary structure, as well as solvent accessibility is sufficient to predict successful peptide epitopes. The antigenicity and the specificity in immune response were also found to depend on the epitope length. We trained the B-Cell Epitope Oracle (BEOracle), a support vector machine (SVM) classifier, for the identification of continuous B-Cell epitopes with these protein properties as learning features. The BEOracle achieved an F1-measure of 81.37% on a large validation set. The BEOracle classifier outperformed the classical methods based on propensity and sophisticated methods like BCPred and Bepipred for B-Cell epitope prediction. The BEOracle classifier also identified peptides for the ChIP-grade antibodies from the modENCODE/ENCODE projects with 96.88% accuracy. High BEOracle score for peptides showed some correlation with the antibody intensity on Immunofluorescence studies done on fly embryos. Finally, a second SVM classifier, the B-Cell Region Oracle (BROracle) was trained with the BEOracle scores as features to predict the performance of antibodies generated with large protein regions with high accuracy. The BROracle classifier achieved accuracies of 75.26-63.88% on a validation set with immunofluorescence, immunohistochemistry, protein arrays and western blot results from Protein Atlas database.
Together our results suggest that antigenicity is a local property of the protein sequences and that protein sequence properties of composition, secondary structure, solvent accessibility and evolutionary conservation are the determinants of antigenicity and specificity in immune response. Moreover, specificity in immune response could also be accurately predicted for large protein regions without the knowledge of the protein tertiary structure or the presence of discontinuous epitopes. The dataset prepared in this work and the classifier models are available for download at https://sites.google.com/site/oracleclassifiers/ webcite.