Abstract 756: Unveiling the spatial landscape of tumor and stroma heterogeneity in colorectal cancer with spatial transcriptomics and deep learning

Abstract Heterogeneity in colorectal cancer (CRC) manifests as inter- and intra-patient diversity in tumor and immune cell populations as well as varied patterns of tumor interaction with surrounding stroma. Spatial transcriptomics (ST) captures gene expression within the native spatial context of tissues, offering a high-resolution approach to dissect cancer molecular landscapes. Prior ST studies in CRC have been constrained by small sample sizes and tissue selection biases, reducing clinical relevance. Here we performed ST (10x Genomics Visium Spatial Gene Expression) on 40 retrospectively-collected CRC samples with available clinicopathological data (stages I-IV: 13, 8, 14, and 5 samples, respectively) and long-term clinical outcomes to define the molecular topography and spatial heterogeneity of tumors with their stromal context. Using ST data, we identified recurring spot-level (55-micron resolution) tumor and stromal sub-populations across the cohort, designating them as T1 (marker genes EPCAM, LGALS4, and FXYD3) and T2 (marker genes KRT8, KRT18, and KRT19) for tumor subtypes, and F1 (marker genes SPARC and LUM) and F2 (marker genes COL3A1 and PTGDS) for stromal subtypes. Comprehensive analysis of tumor/stromal pairs—including copy number variations (CNVs), cell type compositions, pathway activations, metabolic profiles, spatial distributions, and histological characteristics—revealed pathways enriched in T2 compared to T1, including epithelial-mesenchymal transition (EMT), angiogenesis, hypoxia response, TGF-beta signaling, and metabolic pathways supporting the energy and biosynthetic demands of aggressive tumor cells. We also identified recurrent spatially co-localized tumor and stromal sub-clonal pairs—T1&F2 and T2&F1—that consistently preserved their spatial architecture across samples. Notably, the prevalence of T1&F2 was significantly correlated with disease-specific survival (DSS) (log-rank test, p=0.05), highlighting the prognostic importance of tumor-microenvironment interactions. Using a transformer-based image feature encoder, we demonstrated that these clusters are morphologically distinguishable in hematoxylin and eosin (H&E)-stained images co-registered with the ST data, achieving an AUC (the Area Under the ROC Curve) of 0.84 for distinguishing T1 from T2 and an AUC of 0.75 for distinguishing F1 from F2. These clusters were also detected in H&E images from external datasets, including The Cancer Genome Atlas (TCGA, N=232) and an independent validation set (N=107). Deep learning models trained on TCGA data further revealed a trend where higher T2 prevalence or lower T1 prevalence correlated with shorter overall survival. This ST study establishes novel molecular subtypes and their spatial co-occurrence in CRC, linking spatial behaviors to clinically prognostic and H&E-based morphological characteristics. Citation Format: Jie Zhou, Todd B. Sheridan, Sergii Domanskyi, Stephanie L. Cowles, Sunghee Park, Ilham Putra, Olga Anczukow, Jeffrey H. Chuang, Jill C. Rubinstein. Unveiling the spatial landscape of tumor and stroma heterogeneity in colorectal cancer with spatial transcriptomics and deep learning [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2025; Part 1 (Regular Abstracts); 2025 Apr 25-30; Chicago, IL. Philadelphia (PA): AACR; Cancer Res 2025;85(8_Suppl_1):Abstract nr 756.