|
Abstract:
The sequence specificity of DNA-binding proteins is typically represented as a position weight matrix in which each base position
contributes independently to relative affinity. Assessment of the accuracy and broad applicability of this representation has been
limited by the lack of extensive DNA-binding data. However, new microarray techniques enable a broad comparison of both motif
representation and methods for motif discovery. Here, we consider the problem of accounting for all of the binding data in such
experiments, rather than the highest affinity binding data. We introduce the RankMotif++, an algorithm designed for finding motifs
whenever sequences are associated with a semi-quantitative measure of protein-DNA-binding affinity. RankMotif++ learns motif models
by maximizing the likelihood of a set of binding preferences under a probabilistic model of how sequence binding affinity
translates into binding preference observations. Because RankMotif++ makes few assumptions about the relationship between binding
affinity and the semi-quantitative readout, it is applicable to a wide variety of experimental assays of DNA-binding preference.
By several criteria, RankMotif++ predicts binding affinity better than two widely used motif finding algorithms (MDScan,
MatrixREDUCE) or more recently developed algorithms (PREGO, Seed and Wobble), and its performance is comparable to a motif model
that separately assigns affinities to 8-mers. Our results validate the PWM model and provide an approximation of the precision and
recall that can be expected in a genomic scan.