Skip to main content


Fig. 9 | BMC Bioinformatics

Fig. 9

From: Visualizing complex feature interactions and feature sharing in genomic deep neural networks

Fig. 9

Hierarchical clustering results of 919 biological targets using correlation of positive OFIV as distance metric. Each panel represents a cluster, in which the left matrix is the sub-matrix of the class similarity map in 2nd convolutional layer(see Fig. 7) among classes in the cluster, and the right matrix is the sub-matrix of label correlation between the classes. Each of the clusters consist of TFs that are known to be interacting, such as forming a complex or cohesin (c-Fos and JunD (b), SMC3 and Rad21 (a)), co-repression (KAP1 and ZNF263 (c)), competing (ELK1 and GABP (d) or known to be essential for each other to regulate transcription (EZH2, SUZ12 and H3K27me3 (f)). Cluster (e) consists of the subunits of Pol III (RPC155) and 2 essential transcription factors for Pol III : TFIIIB (BRF1/2 and BDP1 are subunits for TFIIIB) and TFIIIC. We show that even when the label correlation is not significant, our class similarity matrix can still capture the functional relevance of the interacting TFs

Back to article page