Comparisons of predicted PBEMs based on various sizes of called peaks. A parallel ensemble learning and a serial computation of BayesPI2+ were applied on a portion of randomly selected (i.e. 5%, 10%, 25%, 50%) and all (i.e. 100%) called peaks, respectively. Sequence log representations of the predicted best PBEMs (100% called peaks) and the meta-PBEMs (i.e. 10 times random selection of called peaks) from various input sizes are displayed in the figure. The last column is the known ESR1 and SPIB binding motifs from JASPAR database.