Informing mechanistic rules of agent-based models with single-cell multi-omics

PROJECT SUMMARY Cancer is driven by interactions between diverse cell types and their tissue microenvironment. Emerging single- cell and spatial transcriptomic systems are mapping cancer tissues, in the process capturing the diversity of cell types and states and exposing the importance of spatial cell interactions in determining therapeutic response in individual patients. Multi-omics software developed in ITCR—including CoGAPS, projectR, SpaceMarkers, and Domino developed by our group—can analyze single-cell and spatial multi-omics data to infer cell types and phenotypes in the tumor microenvironment, identify which cells interact, and discover how cell-cell interactions drive molecular changes. However, these analyses yield static snapshots that cannot capture the dynamics of cancer ecosystems. Mathematical modeling can “fill in the gaps” between these snapshots, allowing teams to form hypotheses on how and why cells interact, “encode” their hypotheses as simulation rules, and perform “virtual experiments.” However, simulation rules and their parameters are difficult to match to genomic data. This proposal bridges the gap between bioinformatics and mathematical biology by merging our bioinformatics soft- ware for single-cell and spatial multi-omics data with PhysiCell, an agent-based modeling framework developed by our group to simulate the movement and interaction of many individual cell agents in virtual tissue environ- ments. The “glue” between these packages is a novel cell behavior grammar that “encodes” cell rules learned from high-throughput data as intuitive, interpretable hypothesis statements that can be automatically transformed into simulation code. In this proposal, we refine the cell behavior grammar while analyzing previously published cancer data to create digital “templates” for key cell types in cancer ecosystems, refine PhysiCell to import the templates, and create PhysiCell Cloud: a free, “zero-install” cloud resource to build, execute, and visualize can- cer models without writing computer code. We refine CoGAPS, SpaceMarkers, projectR, and Domino to learn cell behavior rules from spatial transcriptomics data and format them with the grammar, and extend PhysiCell to read cell types, positions, and rules stored in standard single-cell, spatial, and multi-omics classes. We develop sophisticated pipelines for PhysiCell models that can quantify model uncertainty, automatically fit models to tran- scriptomics data, and validate models on real world tumor datasets. We extend PhysiCell Cloud to a full-fledged science gateway that includes secure and searchable user storage, data structures and code (APIs) to connect PhysiCell Cloud to Python, R, and Bioconductor pipelines in ITCR, and a cost-free high-performance computing backend to seamlessly run large-scale model exploration and uncertainty quantification pipelines. Educational expertise and community feedback—including from an advisory board, annual training workshops, and daily classroom use—will drive usability refinements. Altogether, this approach will bridge bioinformatics and mathe- matical modeling to provide a comprehensive platform for patient-specific mechanistic tumor modeling, to enable future computationally-drien experimental design, virtual clinical trials, and digital twins research. ‘