The regulatory information for a eukaryotic gene is encoded in cis-regulatory modules. The binding sites for a set of interacting transcription factors have the tendency to colocalize to the same modules. Current de novo motif discovery methods do not take advantage of this knowledge. We propose a hierarchical mixture approach to model the cis-regulatory module structure. Based on the model, a new de novo motif-module discovery algorithm, CisModule, is developed for the Bayesian inference of module locations and within-module motif sites. Dynamic programminglike recursions are developed to reduce the computational complexity from exponential to linear in sequence length. By using both simulated and real data sets, we demonstrate that CisModule is not only accurate in predicting modules but also more sensitive in detecting motif patterns and binding sites than standard motif discovery methods are.T ranscription factors (TFs) regulate genes by binding to their recognition sites. The common pattern of the binding sites for a TF is called a motif, usually modeled by a position-specific weight matrix (PWM). Experimental methods such as DNase footprinting (1) and gel-mobility shift assay (2, 3) have allowed the determination of some binding sites for selected TFs. Because these procedures are time-consuming, several computational methods have been developed for de novo motif discovery, including progressive alignment (4, 5), the expectationmaximization algorithm (6, 7), the Gibbs sampler (8-12), word enumeration (13,14), and the dictionary model (15, 16). The propagation model (17) and the recursive Gibbs motif sampler (18) have been developed for locating multiple motifs simultaneously. In addition, methods also exist that combine motif discovery with gene expression data (19-21) or phylogenetic footprinting (22,23). These experimental and computational analyses have given us a good number of useful TF motifs. However, there are still many important TFs whose motifs remain to be characterized. What is more, molecular analyses have established that most eukaryotic genes are not controlled by a single site but by cis-regulatory modules (CRMs), each consisting of multiple TF-binding sites (TFBSs) that act in combination (24-27). It can be argued that motif discovery is but an intermediate step toward the characterization of CRMs. Current approaches on module prediction such as those based on logistic regression (28,29) or hidden Markov models (30, 31) depend on the availability of known motifs, i.e., PWMs for several TFs hypothesized to bind synergistically to regulatory modules. Clearly, we cannot apply these methods to the situations where no prior knowledge on the TFs is available, and in these cases we must resort to de novo motif discovery algorithms. We hypothesized that greater sensitivity and specificity can be achieved for motif discovery by considering the colocalization of different TFBSs and searched for modules and motifs simultaneously. It is clear that the task of module discovery and motif estimation is ...