Fig. 3

Expression reconstruction improves downstream analyses including cell identification, gene regulation, and trajectory inference. The cite-lq dataset was reconstructed using bulk-hq data and compared to ground truth CITE-seq (surface protein) information. The CITE-seq information was not used during training of DISCERN. A t-SNE visualization of CD2 (first row) and CD8A (second row) gene (first two columns) and protein (last column) expression. The first column depicts gene expression for uncorrected cite-lq, the second for reconstructed-hq, and the third protein surface expression ground truth information. Cell types commonly known to express these genes are highlighted with colored circles in the last column. B t-SNE visualization of CD4⁺ T cells in the cite-lq dataset. Cell types were assigned using louvain clustering on the reconstructed-hq data (see C) and show no clear clustering. C t-SNE and trajectory information of CD4⁺ T cell subtypes found by Slingshot analysis on reconstructed-hq data. While uncorrected data shows no clear cell type clustering (see B), reconstructed data shows a clear grouping of cell types. Trajectories were calculated using CD4_naive as starting point and TH2, TH17, TH1, Active_TREG, CD4_CM as endpoints. Lineage1 indicates TH1, Lineage2 TH17, Lineage3 Active_TREG, Lineage4 TH2, and Lineage5 Effector cell differentiation. D: Detection of regulons that are specific for CD4⁺ T cell subtypes using pySCENIC. The first column shows regulons found in the uncorrected cite-lq and the second column in reconstructed-hq data