A non-parametric model for transcription factor binding sites

Abstract

We introduce a non-parametric representation of transcription factor binding sites which can model arbitrary dependencies between positions. As two parameters are varied, this representation smoothly interpolates between the empirical distribution of binding sites and the standard position-speci®c scoring matrix (PSSM). In a test of generalization to unseen binding sites using 10-fold cross-validation on known binding sites for 95 TRANSFAC transcription factors, this representation outperforms PSSMs on between 65 and 89 of the 95 transcription factors, depending on the choice of the two adjustable parameters. We also discuss how the nonparametric representation may be incorporated into frameworks for ®nding binding sites given only a collection of unaligned promoter regions.