*2.4. Epigenomics CNNs*

A 4-layer CNN (called basic CNN) and a 10-layer ResNet are shown in Figure 2, chosen by the validation performance to predict EPIs based on epigenomic data (Table 3 and Table S2). We also implemented a few other newer CNN models, including Inception and Capsule Networks because of their improved performance in image applications [29,30]. Likely due to limited epigenomic data, their results were less competitive than simpler CNNs; they provided no higher prediction power than random guessing (result not shown).


**Table 2.** Sequence model structures.

**Figure 2.** Epigenomics data preparation and CNN model structures. Different genomic annotation data type was denoted with different color. (**a**) Basic CNN structure; (**b**) ResNet CNN structure.
