Cellular proximity and conditional coordination in the colorectal and pancreatic tumor microenvironment: a patient-level analysis of multiplexed imaging

Abstract

Immune cells in the tumor microenvironment are organized, and in colorectal and pancreatic cancer their arrangement around the tumor epithelium carries prognostic information. Multiplexed tissue imaging records the position and phenotype of every cell in a section, and studies of such images now routinely report that two cell types are spatially associated. However, the summaries behind these reports answer different questions. A proximity statistic asks whether two cell types lie close together more often than expected by chance. A conditional network asks whether their local densities rise and fall together once the other cell types are held fixed. The two can disagree, because every cell type is abundant where tissue is cell-dense and because the images of one patient are repeated measurements of that patient. We asked how epithelial and immune cells are organized in two settings: colorectal cancer with a Crohn's-like reaction (CLR) or diffuse inflammatory infiltration (DII), in 140 CODEX images from 35 patients, and pancreatic ductal adenocarcinoma (PDAC) against its precursor, intraductal papillary mucinous neoplasm (IPMN), in 207 multiplex immunofluorescence images from 141 patients. We developed a patient-level analysis that keeps proximity, recurrent local composition and conditional coordination as separate quantities, tests conditional organization as a function of local epithelial density within each section, checks before every conditional estimate that the two cell types are disjoint sets of cells observed together in enough images, and calibrates each conclusion by patient-label permutation, cellularity adjustment under two definitions, spatial-scale perturbation and fully nested leave-one-patient-out prediction. In colorectal cancer, the conditional association between epithelial and myeloid density turned negative with local epithelial density about four times faster in DII than in CLR (difference $-0.0553$ per quartile; 95% CI, $-0.0833$ to $-0.0273$; patient-label familywise $P=0.006$), with no accompanying difference in proximity. In the pancreas, cytotoxic and helper T cells were closer together in PDAC than in IPMN at every scale examined ($q\leq0.023$), whereas their conditional association differed between the two diseases under only one of the two cellularity definitions. Recurrent neighborhoods carried the most patient-level information in colorectal cancer (AUC 0.794) and combined spatial features in the pancreas (AUC 0.771). Pancreatic epithelial--immune and regulatory T-cell quantities could not be estimated from these cores and are reported as such. The analysis is implemented in the R package spatialMIFtools, available at https://github.com/sagnikbhadury/spatialMIFtools.