Fetal-like epithelial niches at the invasive margin to inform prognosis in colorectal cancer.

e15118 Background: The development of metastases remains a major unmet challenge in colorectal cancer (CRC), due to limited mechanistic understanding and a lack of effective preventive therapies. Adjuvant chemotherapy provides only modest benefit and is not tailored to individual tumor biology. Whereas neoadjuvant chemotherapy trials show limited impact on metastatic risk, prompting exploration of immunotherapy and targeted approaches aimed at controlling metastasis-initiating cells. Emerging evidence indicates that epithelial plasticity and fetal-like reprogramming enable tumor cells to persist at the invasive margin (IM) through interactions with the tumor microenvironment. Tissue-based machine learning on H&E whole-slide images (WSIs) allow identification of high-risk histologic regions at the IM. We and others have shown these tissue features are strongly prognostic, highlighting them as potential targets for drug development within tumors and circulation. In this study we aimed to define the spatial, biological, and histologic features of such epithelial niches in CRC. Methods: We analyzed 20 CRC scRNA-seq and 16 matched Xenium spatial transcriptomics datasets (322-gene colon panel plus 100 custom genes) to define cell states and map their spatial organization. INSIGHT, a graph neural network on paired H&E WSIs, identified histologic regions associated with recurrence risk and integrated spatial single-cell abundances. Results: A fetal-like ANXA1⁺ epithelial subtype, defined by loss of adult colonic identity and reactivation of regenerative programs, was identified by scRNA-seq and exhibited extracellular matrix remodeling and inflammatory pathway engagement. Spatial neighborhood analysis on Xenium defined ten niche types along the tumor core–to–IM axis. Fetal-like cells, mapped via single-cell–to–spatial label transfer, were enriched in IM-associated niches versus core (16.54 vs 5.22 cells/mm²; p < 1 × 10⁻³) and co-localized with SPP1⁺ macrophages. Spatial trajectory analysis revealed a gradual shift from adult stem-like to fetal-like programs toward the IM, with downregulation of canonical WNT and colon lineage genes (AXIN2, CDX2) and upregulation of fetal/matrix remodeling genes (ANXA1, MMP7) (|ρ| ≥ 0.1, FDR < 0.05). Integration with INSIGHT showed high-risk regions were enriched for fetal-like cells and SPP1⁺ macrophages at the IM, both positively associated with regional survival risk scores (β = 0.0033 and 0.0042; p < 10⁻⁴). Conclusions: Fetal-like epithelial niches at the CRC invasive margin, co-localized with SPP1⁺ macrophages, define high-risk tissue states. Integration with AI-based histologic analysis underscores the prognostic relevance of these niches and their potential as targets for tailored (neo)adjuvant interventions.