Comparison of PSI-BLAST and RankProp variants. Each figures plots the number of test set queries for which a given method achieves a ROC
score threshold. Figures (A), (B) and (C) use ROC50, ROC10 and ROC1 scores, respectively. Corresponding mean ROC
scores are given in Table 1. The variants of RankProp are described in the text. Previous work  used RankProp with σ = 100.