LncRNAs show lower expression in almost all cell types as compared to mRNAs [49]. Low and highly expressed promoters tend to have distinct epigenetic features [58], [59]. Sequence specific differences of high and low expressed genes have been discussed for years [60]. To avoid a potential bias of differentiating low and highly expressed promoters rather than lncRNA and protein-coding promoters, we compared only the lncRNAs and protein-coding genes with similar expression level in several cell types. Our model achieves over 81%, 80%, 81% and 80% accuracy for Gm12878, H1-hESC, K562 and HUVEC, respectively (Table 2) when expression levels were controlled for. If we exclude the CGIs and downstream regions from the consideration, the models still demonstrate over 76% accuracy. Importantly, the performance of the model has been increased as compared to the model with the same set of features, but without controlling for the expression levels (71%, see previous section, vs 76%). These results suggest that expression bias is very unlikely to contribute to the accuracy of the models.
