Figure 4

Performance ROCs of different T4S effector prediction models. (A) Comparison of ‘Pos_Aac_SPB’, ‘Seq_Aac’, and ‘Pos_Aac_SPB + Seq_Aac’ models. ‘Pos_Aac_SPB’ only extracted the features of positive dataset. ‘Seq_Aac’ only learned sequence-based single-residue composition features. ‘Pos_Aac_SPB + Seq_Aac’ combined the features of ‘Pos_Aac_SPB’ and ‘Seq_Aac’. (B) Comparison of ‘Pos_Aac_SPB’, ‘Pos_Aac_BPB’, ‘Pos_Aac_SPB + Seq_Aac’ and ‘Pos_Aac,Sse,Acc’ models. ‘Pos_Aac_BPB’ model extracted the Aac features of both positive and negative datasets, while ‘Pos_Aac,Sse,Acc’ learned the joint position-specific Aac, Sse and Acc features. All comparisons were performed with a 5-fold cross-validation strategy.