I have been analyzing different online servers in predicting Bcell epitope for a protein. But each server giving different result i.e displaying epitopes at different positions. So I couldn't come to a conclusion to predict the B-cell epitope. There is no similarity between linear epitope and discontiunous epitope , so which I should consider for prediction?
Dear Sunmathi
It is important to have an idea of how the prediction tools are developed. In principle the process starts with 2 groups of experimentally verified biological data: (+ve data) eptiopes and (-ve data) non-epitopes peptides. Most of the currently available tools are built based on sequence data, i.e a collection of peptide sequences that were proven to act as epitopes and another group of peptide sequences that did not show any B cell epitope behivor. There are few tools that were built using structural data "conformational epitopes". The reason behind this is the fact that we have more knowledge of protein sequences than protein structures. In addition, the techniques for identifying (mapping) linear peptide epitopes are more advance, more practical and cheaper than the mapping of structural (conformational) epitopes.
In the second phase of developing these prediction tools, we need to use a statistical modeling system and define some features (sequence order, frequency of certain amino acids, or the physico-chemical properties of the studied peptides) in order to uncover the rule(s) that separate the two data sets (+ve VS -ve). We call this step "training on +ve and -ve data" and this is the step that leads to a model that is used later to design the tools that we use to predict whether or not a given sequence has an epitope.
It is important to stress that variations in performance among different tools that were developed for the same problem (such as B cell epitope) can be seen and these variations can be due to the size of the training data sets, the quality of the collected data, and the rationale in selecting features. For example if you build a tool on 10.000 data points (5000 epitope and 5000 non-epitope) you will have better prediction performance than using on 100 data points (50 and 50). The quality of the prediction models are expressed by a statistical measure called area under ROC curve "AUC" and this is scale to show the quality:
AUC 90-1 = excellent tool
AUC 80-.90 = good tool
AUC 70-.80 = fair tool
AUC 60-.70 = poor tool
AUC 50-.60 = fail "random tool"
Unfortunately, most of the present linear B cell epitope tools are have AUC ~ 60-65. On the other hand the tools to predict T cell epitopes of MHC class I have much better performance with AUC= 80-90.
Finally, my suggestion is to use a voting approach; that is to apply various tools and judge based on consensus results.
Good luck
Dear Sunmathi,
In general, prediction methods can be divided based on the level of input information to methods based on antigen sequence and methods based on 3-D structure of antigen. Structure-based methods significantly outperform sequence based methods.
It is generally believed that the most of identified linear antigenic determinants are parts of the conformational B-cell epitopes due to close contact in the three-dimensional conformation. The traditional antibody-ignored epitope prediction methods do not take into account the reconfiguration of epitope residues when an antigen is in complex with a specific antibody. The prediction based only on detection of exposed residues ignoring the residues buried in the spatial structure may affect the reliability of predictions.To reflect this biological reality, several prediction methods based on sequence or structure of interacting Ab and Ag have been introduced in the last few years (ASEP, PEASE, Bepar). The performance of antibody based prediction of epitopes is competitive, or even better, when compared with structure-based predictors.
Good luck
Lucia
- Badges
- Science topic
- Similar topics
- Biological Science
- Molecular Biology
- Biomolecules
- Proteins