The Sheep and the Goats: Distinguishing transcriptional enhancers in a complex chromatin landscape

Predicting regulatory function of non-coding DNA using genomic information remains a major goal in genomics, and an important step in interpreting the cis-regulatory code. Regulatory capacity can be partially inferred from transcription factor occupancy, histone modifications, motif enrichment, and evolutionary conservation. However, combinations of these features in well-studied systems such as Drosophila have limited predictive accuracy. Here we examine the current limits of computational enhancer prediction by applying machine-learning methods to an extensive set of genomic features, validating predictions with the Fly Enhancer Resource, which characterized the transcriptional activity of approximately fifteen percent of the genome. Supervised machine learning trained on a range of genomic features identify active elements with a high degree of accuracy, but are less successful at distinguishing tissue-specific expression patterns. Consistent with previous observations of their widespread genomic interactions, many transcription factors were associated with enhancers not known to be direct functional targets. Interestingly, no single factor was necessary for enhancer identification, although binding by the 9pioneer9 transcription factor Zelda was the most predictive feature for enhancer activity. Using an increasing number of predictive features improved classification with diminishing returns. Thus, additional single-timepoint ChIP data may have only marginal utility for discerning true regulatory regions. On the other hand, spatially- and temporally-differentiated genomic features may provide more power for this type of computational enhancer identification. Inclusion of new types of information distinct from current chromatin-immunoprecipitation data may enable more precise identification of enhancers, and further insight into the features that distinguish their biological functions.