Fig. 4

Expression reconstruction improves COVID-19 cell type identification and allows for efficient disease severity prediction. Two COVID-19 blood datasets were reconstructed and analyzed. Hamburg covid-blood-lq and covid-lung-lq data was reconstructed using bulk-hq data, resulting in the respective -hq datasets. Similarly, Cambridge covid-blood-severity-lq data, which contains disease severity information, was reconstructed using bulk-hq data. A t-SNE representation of TH17 subclusters using reconstructed covid-blood-hq data. Clusters were defined using the Leiden clustering algorithm on CD4⁺ T cells. B t-SNE representation colored by expression of reconstructed genes distinguishing TH17_cluster1 and TH17_cluster2 cells. TH17_cluster1 displays a central memory and TH17_cluster2 a more activated phenotype. C Violin plots of expression levels for genes distinguishing TH17_cluster1 (C1) and TH17_cluster2 (C2) cells before (covid-blood-lq) and after (covid-blood-hq) reconstruction with DISCERN. D Rare and unexpected cell types found in the reconstructed covid-blood-hq data with covid-blood-severity and bulk data. Cytotoxic CD4⁺ T cells (CD4_cytotoxic) are displayed in green, CD8⁺ Tc2 helper cells (CD8_Tc2) in blue, and all other cells in gray color. E t-SNE representation of key marker genes in covid-blood-hq data for CD4_cytotoxic and CD8_Tc2 cells displayed in D. F Confusion matrix predicted on collection site Ncl with classifier trained on Cambridge data and vise versa for disease severity prediction using GBM classifiers trained on fractions of five T cell types (CD4_CM, CD4_cytotoxic, CD4_naive, CD8_EM, CD8_effector) using reconstructed covid-blood-severity-hq data. Category “critical” was combined with “severe” and “asymptomatic” with “mild”. G ROC curve of the GBM predictions outlined in F using reconstructed (blue color) covid-blood-severity-hq (CD4_CM, CD4_cytotoxic, CD4_naive, CD8_EM, CD8_effector) and published T cell information from uncorrected (yellow color) data (CD4.CM, CD4.Tfh, CD8.EM, NKT, Treg). Confidence intervals (color shades) indicate one standard deviation