Improving promoter prediction using multiple instance learning

Abstract

Promoter prediction is a well known, but challenging problem in the field of computational biology. Eukaryotic promoter prediction, an important step in the elucidation of transcriptional control networks and gene finding, is frustrated by the complex nature of promoters themselves. Within this paper we explore a representational scheme that describes promoters based on a variable number of salient binding sites within them. The multiple instance learning paradigm is used to allow these variable length instances to be reasoned about in a supervised learning context. We demonstrate that the procedure performs reasonably on its own, and allows for a significant increase in predictive accuracy when combined with physico-chemical promoter prediction.