Fig. 4: Neighborhood composition patterns reveal key drivers of GNN survival predictions.

a–c Heatmaps of test set neighborhoods partitions into 20 groups, g1-g20 (84,468 subgraphs each), ranked by GNN survival prediction (from unfavorable to favorable, bottom to top, a). The color scales indicate GNN prediction scores (blue - red) and relative enrichment or depletion of cell types (green - purple). Subdividing the most favorable neighborhoods in group g1 (magenta box, b), reveals elevated CD8+ and PD-1+ immune cell densities coincide with higher proportions of both PD-L1+ and PD-L1- tumor cells in the most favorable subgroup b1. In contrast, the least favorable neighborhoods, subgroup w1 (green box, c), show a general reduction in tumor cells, particularly PD-L1 + , and most immune cell populations, with slight enrichment of FOXP3+ cells. d, e Representative examples of cell-type distributions in selected neighborhoods from the most favorable (d) and least favorable (e) partitions illustrate the differential patterns of tumor and immune cell infiltration that underlie the GNN survival predictions. Cells outside the respective neighborhood graph are represented by white polygons in the Voronoi tessellations (upper rows) and with reduced opacity in the corresponding mIF section (bottom rows).