Fig. 4
DL-XAI clock reveals enrichment of gene modules, not retrievable with penalized regression. (a) PPI edges sorted (x-axis) by the fraction of meta-cells in which a given edge is selected among the top-100 predictive edges of a meta-cell (y-axis), for each cell-type separately: naive CD4 + T cells, CD4 + central memory T cells (CD4TCM), natural killer (NK) cells, and monocytes (CD14MC). The horizontal red dashed line demarcates the threshold for final selection of edges, while the vertical grey dashed line indicates the final number of selected edges. The number of selected edges and unique genes is given in each panel. (b) Violin plot comparing the statistical significance levels (-log10P-values) of Spearman correlations between a gene’s expression level and age, for the genes selected by PGExplainer (PGE) versus the complement of other genes not selected by PGE (Other). The P-value given in the plot is derived from a one-tailed Wilcoxon-rank sum test, comparing the two distributions. (c) Balloon plot representing enrichment of KEGG pathways among the top genes extracted by DL-XAI clock, the elastic net models with PPI restricted and all genes, the GLMgraph and RF models, in each cell type. P-values have been adjusted using the Benjamini-Hochberg (BH) procedure. (d) The number of KEGG enrichment terms for the five models and for each cell type.
