P1.108. Molecular Determinants of Tumor Morphology in Esophageal Adenocarcinoma

Abstract Topic Esophageal Cancer: Molecular Biology/Pathology Background Esophageal adenocarcinoma (EAC) is histologically graded as well, moderately, or poorly differentiated. Poorly differentiated tumours are associated with worse prognosis compared to well and moderately differentiated tumours. However, the molecular mechanisms driving these differences are poorly understood and their delineation may lead to improved biologically-informed patient stratification and therapeutic development. Methods We primarily analyzed two cohorts of laser-capture microdissected (LCM) EAC samples: a discovery cohort consisting of 73 untreated primary biopsies and a validation cohort of 34 metastatic and/or post-treated samples. Additional validation cohorts included an independent cohort from The Cancer Genome Atlas (TCGA; n=80), a spatial transcriptomics (Visium) cohort (n=6), and a patient-derived EAC organoids cohort (n=24). Consensus non-negative matrix factorization (NMF) was performed on the discovery cohort (K=9 components, 3 glandular, 2 transitional, 1 signet ring associated, and 3 others) and independently on TCGA (K=9, 3 glandular, 1, transitional, and 5 other). The discovery NMF was projected onto the validation cohort. Morphology scores were calculated by summing sample component weights within their predefined component groups and correlated with histologic quantification by a board-certified gastrointestinal pathologist. Results The discovery NMF model was validated in TCGA, where all three glandular signatures and one transitional signature were reproducible. The second transitional signature was confounded by squamous contamination in TCGA samples. The signet ring-associated component was not detected in TCGA and was not observed in corresponding slide images. NMF-derived morphology scores accurately predicted tumour histology (Figures 1,2), with mean absolute errors (percentage points) of 16.05%/14.89% (glandular), 13.59%/15.36% (transitional), and 8.18%/9.41% (signet ring) in discovery/validation cohorts, respectively; these errors seem reasonable given potential LCM bias and the inherent limitations of pathologist quantification. Signatures were further validated by spatial transcriptomics with pathologist annotations, multi-area within-slide LCM RNA-seq, and patient-derived EAC organoids. Importantly, the transitional and signet ring signatures were associated with worse prognosis compared to tumors with a glandular signature. Conclusion We report the first validated morpho-molecular classification of EAC. Our findings highlight the limitations of bulk RNA-seq, particularly the overinflation of poorly differentiated signatures due to the presence of squamous contamination and demonstrate the value of profiling following LCM. Ongoing studies are integrating whole-genome sequencing and clinical correlates to further define the biological and therapeutic implications of these morpho-molecular subtypes.