Fig. 4: Analysis of TLS parameters in different tumors.

TLS parameters were assessed in the Cancer Genome Atlas (TCGA) test set slides (n=767) if not otherwise specified. a Pairwise comparison of manually annotated versus HookNet-TLS predicted tertiary lymphoid structures (TLS) density in each image by paired Mann–Whitney U test. b TLS density compared across organs by Mann–Whitney U test obtained by manual or HookNet-TLS detection. Boxes in a and b span across the 25th and 75th percentiles, and whiskers (Tukey style) extend from the hinge to the largest value no further than 1.5 * interquartile range. c Median TLS density in each tumor type was used as a cutoff to define TLS-high and TLS-low tumors. Average TLS size was compared between the TLS density groups by Mann–Whitney U test. d The HookNet-TLS encoding features of all TLS and germinal centers (GC) ground truth annotations and predictions were extracted from all images and subjected to UMAP dimensionality reduction analysis. Data are displayed for ground truth annotations (dots) depicted for each tumor and each image dataset. e The HookNet-TLS encoding features of all TLS and GC ground truth annotations and predictions were extracted from all images and subjected to unsupervised clustering using the PhenoGraph algorithm (k = 50). Data are displayed for predictions (dots) obtained from the TCGA test set and dots are colored by their cluster assignment. f Fractions of HookNet-TLS predictions belonging to different PhenoGraph clusters were compared between TLS and GC classes by Chi square test. Clusters with significant enrichment in either TLS or GC class are marked by asterisks, respectively (Yates corrected p-values *<0.05, ***<0.000). Predictions obtained from all images (test+train+validation) were used in this analysis. g Representative visualizations of HookNet-TLS predicted TLS in the TCGA test set belonging to PhenoGraph (pg) clusters 1 and 7. A high-resolution central patch and a contextual level are shown for each prediction. Scale bars = 50 μm.