Using genomics-informed agent-based models to understand neoplastic phenotype transitions in the human PDAC microenvironment

We present a novel integration of data-driven single-cell analysis with mechanistic mathematical modeling to predict the impact of cancer-associated fibroblasts (CAF) on pancreatic ductal adenocarcinoma (PDAC) invasion. Bioinformatic analyses of high-throughput single-cell and spatial molecular assays give us access to human tumor gene expression at the time of measurement, enabling direct characterization of the human tumor microenvironment (TME), complementing our understanding of cell behavior gained from preclinical models. Notably, applying our Bayesian non-negative matrix factorization algorithm CoGAPS to single-cell RNA-seq data from PDAC allowed us to associate a gene program of concurrent epithelial-to-mesenchymal transition (EMT) and inflammation in epithelial cells with the presence of CAFs, validated through organoid co-culture experiments. We hypothesize that this gene program represents a mechanism by which CAFs promote invasive behavior in neoplastic cells over time. Testing this hypothesis requires inference of temporal changes, which is not possible with single-timepoint data collection in genomics data. In contrast to data-driven bioinformatics, mechanistic mathematical modeling can capture processes that happen over time and make predictions about a system. Agent-based models (ABMs) are one form of such mechanistic mathematical models that are well suited for temporal modeling of cellular phenotypes. We developed new software for ABMs that abstract cells into software agents with individual states and their own rules of behavior from single-cell and spatial molecular assays for temporal inference. We apply this technique to understand how CAFs in the PDAC microenvironment support the series of phenotype transitions comprising tumor progression and invasion, by encoding the CAF-induced ability of tumor cells to migrate and a cessation of proliferation in tumor cells, induced by CAF-secreted, non-diffusable substrate as inferred from our genomics analysis.  While this provides a virtual framework to test the impact of varying CAF density on tumor cell invasion, it initially lacked integration with the spatial architecture of human tumors. We then apply the ABM to human PDAC spatial transcriptomics data to forecast tumor behavior over time in a spatially-informed setting, varying the hypothesized mechanisms. This revealed that a substrate-driven mechanism results in a  broad mesenchymal-like zone surrounding epithelial-like outgrowths, while a cell contact-mediated mechanism results in a single-cell-wide boundary surrounding more significant outgrowth. When we compared these to real tumor compositions, this favors the explanation that secreted CAF substrate drives EMT in PDAC neoplastic cells, which we then validate in organoids with CAF conditioned media. Our work demonstrates how integrating ABMs with genomic and spatial data enables testing of mechanistic hypotheses about tumor-stroma interactions and provides deeper insights into PDAC microenvironments, potentially informing strategies to intercept disease progression.