On the necessity and biological significance of threshold-free regulon prediction outputs

The in silico prediction of cis-acting elements in a genome is an efficient way to quickly obtain an overview of the biological processes controlled by a trans-acting factor, and connections between regulatory networks. Several regulon prediction web tools are available, designed to identify DNA motifs predicted to be bound by transcription factors using position weight matrix-based algorithms. In this paper we expose and discuss the conflicting objectives of software creators (bioinformaticians) and software users (biologists), who aim for reliable and exhaustive prediction outputs, respectively. Software makers, concerned with providing tools that minimise the number of false positive hits, often impose a stringent threshold score for a sequence to be included in the list of the putative cis-acting sites. This rigidity eventually results in the identification of strongly reliable but largely straightforward sites, i.e. those associated with genes already anticipated to be targeted by the studied transcription factor. Importantly, this biased identification of strongly bound sequences contrasts with the biological reality where, in many circumstances, a weak DNA-protein interaction is required for the appropriate gene's expression. We show here a series of transcriptionally controlled systems involving weakly bound cis-acting elements that could never have been discovered because of the policy of preventing software users from modifying the screening parameters. Proposing only trustworthy prediction outputs thus prevents biologists from fully utilising their knowledge background and deciding to analyse statistically irrelevant hits that could nonetheless be potentially involved in subtle, unexpected, though essential cis-trans relationships.