PanIN and CAF Transitions in Pancreatic Carcinogenesis Revealed with Spatial Data Integration

Abstract

Spatial transcriptomics (ST) is a powerful approach for cancers molecular and cellular characterization. Pancreatic intraepithelial neoplasia (PanIN) is a pancreatic ductal adenocarcinoma (PDAC) premalignancy diagnosed from formalin-fixed and paraffin-embedded (FFPE) specimens limiting single-cell based investigations. We developed a new FFPE ST analysis protocol for PanINs complemented with novel transfer learning approaches. The first transfer learning approach, to assign cell types to ST spots and integrate the transcriptional signatures, shows that PanINs are surrounded by PDAC cancer associated fibroblasts (CAFs) subtypes, including the rare antigen-presenting CAFs. Furthermore, most PanINs are of the classical PDAC subtype while one sample expresses cancer stem cell markers. A second transfer learning approach, to integrate ST PanIN data with PDAC scRNA-seq data, identifies a shift between inflammatory and proliferative signaling as PanINs progress to PDAC. Our data support a model of inflammatory signaling and PanIN-CAF interactions promoting premalignancy progression and PDAC immunosuppressive characteristics.

Analysis of donor pancreata defines the transcriptomic signature and microenvironment of early pre-neoplastic pancreatic lesions

“…To ensure accurate assessment of PanIN burden, multiple samples were taken from the head, body and tail of each donor organ for histologic analysis. We observed no overt difference in prevalence by anatomic region; PanINs, when present, were typically multi-focal and widely dispersed throughout the gland, consistent with recent reports(36).…”

“…As a result, single cell studies of the human pancreas have included very few samples (75,76) or embryonic samples (77) and have focused on endocrine cells less susceptible to degradation. Previous autopsy studies on pancreata of patients deceased with no known pancreas pathology revealed frequent PanINs and KRAS mutations (10,(14)(15)(16)(17)(31)(32)(33)(34)(35)(36)78), but the samples mainly represented older individuals with limited transcriptional profiling performed. In recent years, the advent of single cell technologies has led to in-depth characterization of human pancreatic cancer and its accompanied tumor microenvironment (24,26,60,79,80).…”

“…While a similar prevalence was observed before, it was in the context of autopsy studies with patients in the 7 th and 8 th decades of life(14-17,78). Other investigators have detected a high prevalence of PanINs in the pancreas, but their studies were based on tumor adjacent normal areas(36). In contrast, our study indicate that PanIN are common in the pancreas of individual with no known pancreas pathology and occur early in life.…”

“…A key limitation is that control samples are largely constituted by adjacent normal pancreas; while lacking malignant tumor cells, these samples are often not histologically normal and present with extensive inflammation and other alterations. PanIN studies, while revealing that these lesions share a similar mutation profile as cancer samples (10,(31)(32)(33)(34)(35)(36), are largely conducted in tissue from cancer patients or patients with different pancreatic pathologies, based on paraffin-stored samples. Thus, our understanding of the microenvironment of the human pancreas and of early lesions is extremely limited.…”

Latent transcriptional programs reveal histology-encoded tumor features spanning tissue origins

“…Tumor histology can reveal additional prognostic information, including genomic features. Indeed, recent work used deep learning models to predict genomic alterations and gene expression data from tumor histology [19][20][21] ; other efforts combined genomics and histology to predict prognosis 22 , overall survival 23 , and, leveraging new spatial transcriptomics approaches, molecular heterogeneity of cell states 24 . While genomics and histology offer complementary, synergistic insights into latent tumor biology, clinical deployment of deep-learning-based histology biomarkers requires model transparency demonstrating the histologic features responsible for predictions 25 .…”