Motivation: Locating transcription factor binding sites (motifs) is a key step in understanding gene regulation. Based on Tompa's benchmark study, the performance of current de novo motif finders is far from satisfactory (with sensitivity <or=0.222 and precision <or=0.307). The same study also shows that no motif finder performs consistently well over all datasets. Hence, it is not clear which finder one should use for a given dataset. To address this issue, a class of algorithms called ensemble methods have been proposed. Though the existing ensemble methods overall perform better than stand-alone motif finders, the improvement gained is not substantial. Our study reveals that these methods do not fully exploit the information obtained from the results of individual finders, resulting in minor improvement in sensitivity and poor precision.
Results: In this article, we identify several key observations on how to utilize the results from individual finders and design a novel ensemble method, MotifVoter, to predict the motifs and binding sites. Evaluations on 186 datasets show that MotifVoter can locate more than 95% of the binding sites found by its component motif finders. In terms of sensitivity and precision, MotifVoter outperforms stand-alone motif finders and ensemble methods significantly on Tompa's benchmark, Escherichia coli, and ChIP-Chip datasets. MotifVoter is available online via a web server with several biologist-friendly features.