Skip to main content
Fig. 5 | Genome Biology

Fig. 5

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

Fig. 5

KPNN analysis of cancer and immune cells. a Analysis of a large Human Cell Atlas (HCA) dataset comprising B cells, T cells, and monocytes obtained from bone marrow and cord blood. (Left) KPNNs were trained as multi-class predictors separately for immune cells from bone marrow and cord blood; the fitted models were compared by calculating differential node weights. (Middle) Top 10 hidden nodes with the most differential weights between bone marrow and cord blood. Gene set enrichments were calculated against the full list of hidden nodes that carried differential weights. (Right) Differential node weights of selected proteins are shown for illustration. Error bars indicate standard error of the mean. b Analysis of Langerhans cell histiocytosis (LCH). (Left) KPNNs were trained separately on single-cell RNA-seq data for LCH biopsies from bone (single-system LCH) and skin (multi-system LCH), distinguishing between progenitor-like and mature LCH cells. (Right) Volcano plot comparing differential node weights between KPNNs for bone vs skin. Significant nodes are highlighted in green, JAK-STAT proteins in purple. c Analysis of acute myeloid leukemia (AML). (Left) KPNNs were trained to distinguish between leukemic and normal cells at four stages of hematopoietic development based on single-cell RNA-seq data, and node weights of KPNNs trained on consecutive states were compared. (Middle) Number of differential nodes comparing consecutive states. (Right) Weights of selected nodes over the four stages of hematopoietic development. Error bars indicate standard error of the mean. d Analysis of glioblastoma. (Left) Four glioblastoma subtypes were arranged into quadrants as in the original publication. KPNNs were trained to distinguish pairs of glioblastoma subtypes based on single-cell RNA-seq data, and differential node weights were calculated for each comparison of trained KPNNs. (Right) Scatterplot showing differential node weights between glioblastoma subtypes. For better visualization, the axes were capped at a −log10(padj) value of 7.5, which affected TP53 and LEF1 (shown at the bottom left). Specific nodes of interest are highlighted in green

Back to article page