Effectiveness of combining different motif discovery programs. (A-C) The performance of each motif discovery program, applied to the Sandve et al. (2007) benchmark data set, was measured using the total number of true positive nucleotides (nTP, grey bars) and the total number of false positive nucleotides (nFP, black lines). Shown are scores for the three types of data sets that comprise the Sandve dataset: (A) synthetic (Algorithm Markov), (B) semi-synthetic (Algorithm Real), and (C) real promoters (Model Real). Shown are the scores of each standalone unfiltered program, as well as the scores after combining the outputs of the three programs without filtering (combined) or with filtering (combined filt). (D) The performance of each standalone program or the combined programs was compared using the average nucleotide sensitivity (nSn). Shown are the mean nSn scores for the synthetic data (AM: Algorithm Markov), semi-synthetic data (AR: Algorithm Real) and real data (MR: Model Real). The asterisks (***) indicate that the average nSn score of the combined filtered programs is statistically higher than the average nSn score using Weeder alone at p < 0.01. Each error bar represents the 95% mean confidence interval. (E) The partition of final true positives found by the three motif discovery tools after filtering is shown. Shared results are motif nucleotides retrieved by at least two of the standalone programs. Filtering and combining the standalone programs are the basis of Promzea.