Figure 3. Specificity is maintained even when controlling for homology in training and testing.

Different classifiers (blastp besthit as a purely homology-based system, Macrel, AMP Scanner v.2, and iAMP-2L) were trained with different proportions of non-AMPs:AMPs and tested on datasets which did not contain any homologous sequences in the training set (using an 80% identity cutoff). The results obtained in the homology-free datasets are showed in terms of accuracy (A), precision (B), Sensitivity (C), and Specificity (D).