The latent space of Chrombus reflect unique epigenomic features of 3D-genome. Chrombus is trained on GM12878 dataset. Node embeddings were derived from the third (last) layer of edge convolution, which contained 32 dimensions (d1, d2, …, d32). (A) The node embedding correlates with input features. The heights of the bar plots indicate the Pearson’s correlation coefficients between the node embedding and each input node feature, with the color denoting the three normalization methods used. (B) Clustering of node embedding based on the first two kernel PCs reveal two types of segments. (C) Distribution of segment lengths for the type 1 and 2 segments. D: Interaction strengths within type-1, type-2, and between type-1 and type-2 are differently distributed. (E) Numbers of interactions within type-1, type-2, and between type-1 and type-2 are different. (F) DNase-I and POLR2A activities between the two types of segments show trivial differences. (G–I) CTCF, H3K4me3, and H3K27ac landscape at the 10 kb regions flanking the center show distinct patterns for the two types of segments.