K-mer Motif Multinomial Mixtures, a scalable framework for multiple motif discovery

1 Motivation The advent of inexpensive high-throughput sequencing (HTS) places new demands on motif discovery algorithms. To confront the challenges and embrace the opportunities presented by the growing wealth of information tied up in HTS datasets, we developed K-mer motif multinomial mixtures (KMMMs), a flexible class of Bayesian models for identifying multiple motifs in sequence sets using K-mer tables. Advantages of this framework are inference with time and space complexities that only scale with K, and the ability to be incorporated into larger Bayesian models. Results We derived a class of probabilistic models of K-mer tables generated from sequence containing multiple motifs. KMMMs model the K-mer table as a multinomial mixture, with motif and background components, which are distributions over K-mers overlapping with each of the latent motifs and over K-mers that do not overlap with any motif, respectively. The framework casts motif discovery as a posterior inference problem, and we present several approximate inference methods that provide accurate reconstructions of motifs in synthetic data. Finally we apply the method to discover motifs in DNAse hypersensitive sites and ChIP-seq peaks obtained from the ENCODE project.