Instance-Level Spatial Transcriptomics Resolves TLS Functional States Predictive of Survival and Immunotherapy Response Across Cancer Types

Tertiary lymphoid structures (TLS) are ectopic lymphoid organs whose functional state, immunogenic or suppressive, determines their clinical impact, yet bulk transcriptomics cannot distinguish these states because critical spatial signals are lost when tissue is homogenised. We characterise TLS microenvironments at the instance level from spatial transcriptomics (ST), applied to 670 TLS instances detected by HookNet-TLS across 113 ST samples from 12 cancer types in the HEST-1k atlas. Feature vectors combine seven in-TLS gene module scores with five spatial neighbourhood scores within a 300 um microenvironmental ring. Leiden clustering of these 12 features resolves six functionally distinct TLS states (ARI = 0.83 +/- 0.14). Cluster C1, immunogenic TLS (n = 168), is enriched for germinal centre and plasma cell activity and independently predicts progression-free survival in TCGA-BRCA (n = 927, p = 0.031), overall survival in METABRIC (n = 1,980, p = 0.003) and TCGA-CESC (n = 304, p = 0.021), and pembrolizumab response in triple-negative breast cancer (p = 0.030). Cluster C4, myeloid-suppressive TLS (n = 70), is defined by neighbourhood myeloid enrichment and TGF-beta signalling, spatial features absent from bulk transcriptomic signatures, explaining its null survival signal. A morphological transfer experiment shows the C1/C4 distinction is recoverable from H&E alone: a Random Forest on nine aggregate features reaches AUROC = 0.744, and a two-region spatial graph attention network improves this to AUROC = 0.857. The framework establishes that TLS functional state is encoded in the spatial microenvironment rather than in aggregate marker abundance.