Skip to main content
Fig. 2 | Genome Biology

Fig. 2

From: Knowledge-primed neural networks enable biologically interpretable deep learning on single-cell sequencing data

Fig. 2

Comparative structural network analysis of KPNNs and ANNs. The TCR KPNN is compared to a fully connected ANN (fANN) with the same number of nodes and the same median depth as the KPNN, and to sparse ANNs (sANNs) where edges were randomly removed to match the edge number of the KPNN while retaining an intact network (results are shown for 50 random sANNs). a Average distance to the input nodes for all hidden nodes in the fANN (top) and KPNN (bottom). b Cumulative distribution of the outdegree of hidden nodes in the KPNN, the fANN, and averaged across the sANNs. c Assessment of network sensitivity to fragmentation upon removal of important edges in the KPNN, the fANN, and the sANNs. Edges were iteratively removed based on importance measured by their network betweenness value. d Cumulative distribution of reachability, which measures the number of input nodes each hidden node can connect to, shown separately for the KPNN, the fANN, and averaged across the sANNs

Back to article page