Fig. 2

Training and testing processes of the proposed method. The upper part illustrates the training process. The first-stage and second-stage staged heterogeneity learning models are trained on two data sets with heterogenous annotations. The lower part illustrates the testing process of the proposed method. An unsupervised clustering algorithm is involved in the testing process to recommend conformational B-cell epitopes. The shapes inside data sets denote patterns of propensities: PC–physico-chemical propensity, ASA–accessible surface area, SS–secondary structure, PSSM–position-specific scoring matrix