Pancreatic ductal cells play important roles in both normal pancreatic function and disease. These cells can change their identity in response to injury, inflammation, or metabolic stress, and these changes may contribute to tissue repair, diabetes–associated pancreatic remodeling, pancreatitis, or pancreatic cancer. Understanding how these changes appear in human tissue is important for identifying early signs of pancreatic disease and for improving experimental models of pancreas biology.
Most studies define cell states by measuring gene activity. However, cells and tissues also contain visual information: their shape, organization, position, and relationship with neighboring cells can reflect underlying biological states. Recent advances in artificial intelligence and computer vision now make it possible to quantify these tissue features from digital pathology images in a systematic and reproducible way.
In this project, we will use machine learning to analyze existing digital pathology images from the nPOD collection and our collection, focusing on pancreatic tissue architecture and, where possible, ductal and exocrine regions. The project does not require new nPOD tissue, serial sections, or nPOD–generated transcriptomic data. Instead, image–derived features from nPOD samples will be interpreted using molecular reference datasets already generated by our group and collaborators, including human pancreatic single–cell and multiomic data.
In addition to the nPOD collection, the study will include an independent image–based cohort of approximately 80pancreatic samples available through our collaboration with Eduard Montanya. This cohort comprises individualswith obesity, including individuals with T2D and T1D and individuals without diabetes. We also have access to pancreatitis and PDAC samples through our collaboration with Juli Busquets (head of General and Digestive System Surgery at the Bellvitge Hospital). Our own cohort will be used to evaluate the external reproducibility of image–derived features and the consistency of disease associated morphological differences. Because these histological images are not all paired with the molecular reference datasets, this cohort will not be used for direct image–omics association or individual–level molecular–state prediction.
We will also use pancreatic organoid models generated in our laboratory. These organoids can be exposed to defined experimental conditions that mimic stress, inflammation, or disease–related perturbations. By comparing image–based features from organoids with those observed in human pancreatic tissue, we aim to evaluate how well these models reproduce human tissue states and to identify morphological patterns associated with ductal plasticity and disease–related remodeling.
With this project, we aim to transform pancreatic tissue morphology into a quantitative, information–rich readout ofdisease–relevant biology and to identify candidate morphological features associated with early or preclinical tissue alterations.
