This document describes a method for discovering composite motifs in DNA sequences. The method searches for overrepresented patterns representing transcription factor binding sites. It improves on previous methods by modeling motifs as modules that occur together, rather than as isolated patterns. The algorithm ranks predicted modules based on support, specificity and significance. It was shown to outperform other tools, particularly at realistic noise levels, due to its use of real DNA backgrounds and support-based scoring. Future work includes exploring the full Pareto front of optimal solutions and parameter interactions to improve predictions.