Abstract
<jats:p>Most pancreatic ductal adenocarcinoma (PDAC) single-cell and spatial studies analyze one cohort or platform, obscuring recurrent biology. We assembled a human pancreas single-cell and single-nucleus reference of 1,186,130 cells from 19 studies and interpreted 176 Visium sections comprising 458,877 spots across non-diseased pancreas, chronic pancreatitis, PanIN, IPMN, primary PDAC and metastasis. Marker-supported labels were used after two RNA-based copy-number callers failed known-diploid controls. BANKSY domains, two reference-mapping methods and sample-level analyses resolved a cross-sectional epithelial axis extending from acinar-rich to malignant tissue. Five trajectory algorithms recovered similar ordering on a shared embedding; their consensus was interpreted as transformation-associated, not temporal or clonal. Malignant regions were globally segregated from fibroblast and myeloid compartments. Signed-distance analysis refined this pattern into a malignant core, a CAF/myeloid surround beginning at the tumour boundary and a more distal lymphoid compartment. Candidate extracellular-matrix communication, led by COLLAGEN, LAMININ and FN1, concentrated at the interface. Changes were reproduced in six patient-matched Normal-tumour pairs using exact patient-level tests. Visium HD resolved the same organization at single-cell resolution and showed that 8-um bins distorted immune-adjacency estimates. Xenium also revealed recurrent neighbourhoods but sample-specific stromal boundaries. We provide a confound-aware framework for identifying recurrent epithelial and microenvironmental organization in PDAC.</jats:p>