Extended Data Fig. 6: Tumour transcriptomic footprint of HLAprLOW.
a, Averaged feature importance (Gini) of top 100 genes in ML-model b, Kaplan-Meier curves of PFS in Leuven cohort, by tLHP. c, ORR stratified by tLHP. Two-sided p-value by Fisher’s Exact test. d, e, Forestplot displaying UVA correlation of tLHP, as continuous signature and by optimal cutoff, with PFS (d) and OS (e) in Leuven cohort. f-i, Forestplots displaying UVA and MVA (adjusted for age/IMDC) correlation of tLHP with PFS (f–g) and OS (h-i) within Leuven cohort treatment subgroups. j, Boxplots of tLHP by IMDC (n independent patients). Two-sided p-value by Kruskal-Wallis test. k-l, Fuhrman tumour grade (k) and sarcomatoid differentiation (l) stratified by tLHP. Two-sided p-value by Fisher’s Exact test. m, Violinplot showing tLHP expression in cells by disease stage. Two-sided p-value by one-way ANOVA, effect size by Eta-squared. n, Venn diagram showing tLHP and ILS overlap. o, ORR stratified by tLHP. Two-sided p-value by Fisher’s Exact test. p, Forestplot displaying UVA correlation of tLHP with PFS in IMmotion150 subgroups. q, Kaplan-Meier curves showing PFS in Javelin101 sunitinib arm by tLHP. r, Forestplot displaying UVA correlation of tLHP (optimal statistical cutoff in combined cohort) with OS by cancer type in CRI iAtlas. s, Forestplot displaying correlation of tLHP (optimal statistical cutoff in combined cohort) with OS in CRI iAtlas melanoma subgroup. Ipilimumab+pembrolizumab treated patients are not displayed as the statistical model could not be constructed due to insufficient events. For b and q, HR, 95%CI and two-sided p-values by Cox proportional hazard regression (high vs. low) from UVA and MVA [adjusted for age/IMDC (Leuven RWD cohort) or age/sex (Javelin101 cohort)]. For boxplots in j and m, boxes represent median (centre) and first/third quartile (bottom and top, resp.) values; whiskers show extreme values within 1,5x interquartile range (IQR). Outliers extending beyond 1.5x IQR above/below median are plotted individually. Forestplots in d-i, p, r and s show HR, 95%CI (centre, error bar, resp.) and p-values as calculated through Cox proportional hazard regression (high vs. low). P-values unadjusted for multiple testing.
