A Functional Spatial Analysis Platform for Discovery of Immunological Interactions Predictive of Low-Grade to High-Grade Transition of Pancreatic Intraductal Papillary Mucinous Neoplasms.

Abstract

Intraductal papillary mucinous neoplasms (IPMNs), critical precursors of the devastating tumor pancreatic ductal adenocarcinoma (PDAC), are poorly understood in the pancreatic cancer community. Researchers have shown that IPMN patients with high-grade dysplasia have a greater risk of subsequent development of PDAC in the remnant pancreas than do patients with low-grade dysplasia. In this study, we built a computational prediction model that encapsulates the spatial cellular interactions in IPMNs that play key roles in the transformation of low-grade IPMN cysts to high-grade cysts en route to PDAC. Using multiplex immunofluorescent images of IPMN cysts, we adopted algorithms from spatial statistics and functional data analysis to create metrics that summarize the spatial interactions in IPMNs. We showed that an ensemble of models learned using these spatial metrics can robustly predict, with high accuracy, (1) the dysplasia grade (low vs high grade) and (2) the risk of a low-grade cyst progressing to a high-grade cyst. We obtained high classification accuracies on both tasks, with areas under the curve of 0.81 (95% confidence interval: 0.71-0.9) for task 1 and 0.81 (95% confidence interval: 0.7-0.94) for task 2. To the best of our knowledge, this is the first application of an ensemble machine learning approach for discovering critical cellular spatial interactions in IPMNs using imaging data. We envision that our work can be used as a risk assessment tool for patients diagnosed with IPMNs and facilitate greater understanding and investigation of the cellular interactions that cause transition of IPMNs to PDAC.