Fig. 4

Gene attributions increase latent variable interpretability and differentiate TCR signaling pathway expression programs. Gene level attributions help to gain a deeper understanding of the expression programs represented by the latent variables of a biologically-constrained autoencoder trained on a dataset of Jurkat cells stimulated with anti-CD3 and anti-CD28 antibodies. a Pairwise correlations between the four latent variables in the pathway module corresponding to T cell receptor signaling. b–c Gene attribution dependence plots for the two most important genes, ranked by average magnitude gene attribution over all samples in the dataset, for TCR Signaling Latent Node 1. d–e Gene attribution dependence plots for the two most important genes, ranked by average magnitude gene attribution over all samples in the dataset, for TCR Signaling Latent Node 2. f Jurkat cells plotted by their embedding in the first two nodes of the TCR Signaling Pathway module. The first node, in which PTPRC and PTPN22 expression are not co-regulated, does not separate cells that have been stimulated by anti-CD3/anti-CD28 antibodies (Wilcoxon rank-sums test statistic \(= -0.95\), p value \(= 0.342\)), while the second node, in which PTPRC and PTPN22 expression levels are highly correlated, does separate cells that have been simulated by anti-CD3/anti-CD28 antibodies (Wilcoxon rank-sums test statistic \(= -6.27\), p value \(= 3.39 \times 10^{-10}\))