Anna Lyubetskaya, Brian Rabe, Andrew J. Kavran, Yulong Bai, Andrew Fisher, Alba Font-Tello, Anne Lewin, Hannah A. Pliner, Yunfan Fan, Lauren Giampapa, Yelena Cheng, Chao Dai, Ruifeng Hu, Tom Lila, Alexandre P. Alloy, Mike Mason, Constance Brett, Todd Brett, Fayaz Seifuddin, Steven Vasquez Grinnell, Sören Richard Stahlschmidt, Yilin Zhao, Ryan Golhar, Isaac Neuhaus, Daniel Carrera, Carlos Rios, Pradeep Kar, Abhishek Shukla, Rachael Bashford-Rogers, Matthew J. Meyer, Enas Abu Shah, Lara Heij, S. Sivakumar, Jimena Trillo-Tinoco, Benjamin J. Chen, Konstantinos J. Mavrakis, Eugene Drokhlyansky
Pancreatic ductal adenocarcinoma is heterogeneous, with low tumor purity, a prominent microenvironment, and complex architecture, which preclude the identification of shared tumor-intrinsic and stromal biology within and across patients. We overcame these challenges by achieving necessary resolution and context through the application of complementary genomics, pathology, and machine-learning approaches to characterize primary untreated tumors from 39 patients. We captured 340,000 spatial low-bulk and 530,000 spatial single-cell transcriptomes and observed a spectrum of classical-to-basal tumor subtypes present within all patients. We found that each subtype has distinct regulators, stromal neighborhoods, microenvironment, extracellular matrix, and histology corresponding to multiple immunosuppressive and therapy resistance mechanisms. We defined key tumor heterogeneity features, including the presence of mixed KRAS mutations and tertiary lymphoid structures, identifying biomarkers that distinguish the latter from lymph nodes. Lastly, by leveraging patient, cell, and mouse data, we determined which aspects of tumor biology are recapitulated in bulk datasets and reductionist models.