Various tools have been developed to predict B-cell epitopes. We proposed a multistrategy approach by integrating two ensemble learning techniques, namely bagging and meta-decision tree, with a threshold-based cost-sensitive method. By exploiting the synergy among multiple retrainable inductive learners, it directly learns a tree-like classification architecture from the data, and is not limited by a prespecified structure. In addition, we introduced a new three-dimensional sphere-based structural feature to improve the window-based linear features for increased residue description. We performed independent and cross-validation tests, and compared with previous ensemble meta-learners and state-of-the-art B-cell epitope prediction tools using bound-state and unboundstate antigens. The results demonstrated the superior performance of the bagging meta-decision tree approach compared with single epitope predictors, and showed performance comparable to previous meta-learners. The new approach—requiring no predictions from other B-cell epitope tools—is more flexible and applicable than are previous meta-learners relying on specific pretrained B-cell epitope prediction tools.