Abstract

Most spatial transcriptomic analyses of solid tumors focus on individual cell states or single-sample spatial domains rather than on recurrent multicellular tissue architectures shared across patients, and typically depend on predefined cell-type annotations or compartment definitions. We developed STORM (Spatial Topology analysis of Recurrent Motifs), an unsupervised graph-attention variational autoencoder that learns recurrent spatial motifs directly from cell-level graph structure and molecular profiles without cell-type labels, manual annotation, or predefined compartments. We applied to 32 Xenium sections from 16 patients with paired early-onset (EOCRC) and average-onset (AOCRC) colorectal cancer, STORM identified 10 recurrent motifs that self-organized into tumor-parenchymal, stromal, and immune macro-compartments. Among these, the Desmoplastic Fibrotic Barrier (DFB) motif, a CAF- and ECM-rich boundary architecture, was associated with restricted CD8 T-cell geodesic access to tumor cores independently of CD8 abundance, as demonstrated by abundance-normalized neighborhood enrichment statistics and within-sample mixed-effects models. EOCRC selectively amplified this barrier-exclusion architecture, exhibiting tighter tumor parenchyma, denser DFB shells, and a DFB-specific ECM activation program that yielded an age-specific prognostic signature in TCGA-COAD. Translation of motif macro-classes to H&E images via a Vision Transformer classifier produced an image-derived DFB-barrier composite that predicted overall survival in advanced-stage TCGA colorectal cancer. STORM provides an annotation-free framework for discovering recurrent spatial motifs and identifies a fibroblast barrier architecture whose topological association with immune exclusion is independent of effector abundance, amplified in early-onset disease, and translatable to a deployable pathology-based prognostic biomarker.