Figure 3.

Motif recognizability and discovery performance on synthetic data in the (A) basic, (B) 3′UTR and (C) decoy experiments. Recognizability and discovery performance are measured by nucleotide-level MCC (nCC) as a function of different variates, summarized over the remaining variates. Recognizability (red) serves as reference. Signal-only learning (blue, BW in Figure 2) was performed with the Baum–Welch algorithm on the signal data only, and used as seeds the 8mers of degeneracy at most 2 that are most frequent in the signal data. Discriminative learning (green, MICO in Figure 2) used MICO as objective function for seed finding and HMM parameter optimization. Supplementary Figures S13–S15 give the measures sAP, sSn and sPPV.