Fig. 1

Interpretable deep learning with knowledge-primed neural networks (KPNNs). Deep learning provides a powerful method for predicting cell states from gene expression profiles. However, generic artificial neural networks (ANNs, top row) are “black boxes” that provide little insight into the biology that underlies a successful prediction – for two reasons: (i) hidden nodes and edges in an ANN have no biological equivalent, which makes it difficult assign a biological interpretation to the weights of a fitted ANN model, and (ii) ANNs are inherently instable, and very different networks can achieve similar prediction performance. Knowledge-primed neural networks (KPNNs, bottom row) enable interpretable deep learning on biological networks by exploiting structural analogies between biological networks (such as the signaling pathways and gene-regulatory networks that regulate cell state) and the feed-forward neural networks used for deep learning. In KPNNs, each network node corresponds to a protein or a gene, and each edge corresponds to a potential regulatory relationship that has been observed and annotated in public databases. Weights within the KPNN are obtained by a deep learning method that has been optimized for interpretability, and the learned weights are interpreted as estimates of the regulatory importance of the corresponding signaling protein or transcription factor