Integrating Sample-Driven and Pattern-Driven Approaches in Motif Finding

Abstract

The problem of finding conserved motifs given a set of DNA sequences is among the most fundamental problems in computational biology, with important applications in locating regulatory sites from co-expressed genes. Traditionally, two classes of approaches are used to address the problem: sample-driven approaches focus on finding the locations of the motif instances directly, while pattern-driven approaches focus on finding a consensus string or a profile directly to model the motif. We propose an integrated approach by formulating the motif finding problem as the problem of finding large cliques in k-partite graphs, with the additional requirement that there exists a string s (which may not appear in the given sample) that is close to every motif instance included in such a clique. In this formulation, each clique represents the locations of the motif instances, while the existence of string s ensures that these instances are derived from a common motif pattern. The combined approach provides a better formulation to model motifs than using cliques alone, and the use of k-partite graphs allows the development of a fast and exact divide-and-conquer approach to handle the cases when almost every sequence contains a motif instance. Computational experiments show that this approach is feasible even on the most difficult motif finding problems of moderate size. When many sequences do not contain a motif instance, we complement the above approach by an optimized branch-and-bound algorithm that is much faster than standard clique finding approaches. We will discuss how to further generalize the formulation to better model biological reality.

This work was supported by NSF grant CCR-0311590.