*2.2. Designing Neural Networks for Utilizing Enhancer and Promoter Features*

Our main goal was to construct predictive CNN models for detecting distal EPIs from sequence and epigenomic data, separately or combined. CNNs for sequence and epigenomic annotations were built separately, and we constructed another model to combine features from the two sources. In either sequence or epigenomics CNN, we considered two separate branches for enhancers and promoters respectively, then concatenated the features later. This followed from that enhancers and promoters are expected to have different sequence motifs or epigenomic profiles, as shown in previously designed CNNs [13,14,20].
