Development positions malignant cellular states but does not explain their diversification

Epithelial cancers are often described as aberrant reactivations of embryonic or tissue-forming programs, but whether malignant cellular-state diversification is actually constrained by developmental trajectories remains unclear. Here, we present a quantitative framework to test this idea in pancreatic ductal adenocarcinoma (PDAC). Using representation learning on large-scale single-cell data, we build a reference space that captures the main axes of normal foregut and pancreatic epithelial variation. Malignant cells can be mapped into this space, showing that developmental biology helps interpret their identity. However, these coordinates account for only part of the variation. A substantial portion lay outside the reference space, and the residual component was structured rather than diffuse. Furthermore, displacements between cancer states showed weak alignment with canonical axes and did not correspond to simple progression along them, even when they retained measurable components within the broader subspace. Thus, developmental programs offer a useful coordinate system for describing where PDAC states sit relative to normal epithelia, but they do not explain the directions along which malignant states differ. Instead, malignant diversification follows additional structured, cancer-adaptive axes outside the dominant geometry of normal development. More generally, these results demonstrate how disease-state variation can be decomposed into shared coordinates of normal identity and distinct directions of pathological reprogramming.