In this study, our goal was to computationally test the hypothesis that the global transcriptional regulatory programs of lncRNA genes and protein-coding genes are different. We set this problem within the framework of machine learning classification of promoters of these two broad gene classes. Previous studies [18]–[21] used support vector machines to distinguish non-coding RNAs (ncRNAs) from mRNAs, whereas experimental approaches including RiboSeq [22] and mass spectrometry [23] have documented that lncRNAs possess a low affinity for ribosomes and are rarely translated, but no comparable efforts have been devoted to comparing lncRNA and protein-coding gene promoters. Recently Lv et al. [24] used chromatin modification and genomic features to distinguish lncRNAs from protein-coding genes, while a statistical approach [25] singled out H3R2me1 as a distinctive histone mark between protein-coding genes and lncRNAs. Here, we interrogated multiple computational and empirical sources of regulatory information at promoters on a genome-wide scale. We found genetic and epigenetic signatures unique to protein-coding and lncRNA genes, respectively. These divergent promoter grammars may help to explain the observed differential and highly tissue- and condition-specific transcriptional regulation of
