Knowledge-primed neural networks enable biologically interpretable deep learning on single-cell sequencing data - Genome Biology

Tags

Top Highlights

• The resulting “GEN KPNN” integrates signaling pathways and gene-regulatory interactions from public databases into a single network that is directly useful for interpretable deep learning and compatible with a broad range of single-cell RNA-seq datasets.

• we connected each of these output nodes to all cell surface receptors, reflecting the concept that phenotypic cell states can be captured by the ability of individual cells to interact with their cellular environment.

• Different datasets were combined by counting the number of datasets that support a connection for each transcription factor/target gene pair.

• For each gene, the transcription factors with the largest weighted number of datasets supporting the connection to the gene were retained.

• extract a directed acyclic graph that connects the selected receptor(s) to all reachable transcription factors

• The resulting graphs are used for interpretable deep learning by reversing the cascade: