Document resource
Background Understanding multicellular interactions within the tumor-immune microenvironment (TME) is essential for developing cancer therapies and precision medicine strategies. Tumors pose significant biological complexity due to patient and tissue heterogeneity, cell-cell interactions, and intricate signaling pathways. Macrophages are a heterogeneous cell population associated both positively and negatively with tumor progression and response to therapy, but the factors driving macrophage influence on the TME remain poorly understood. Spatial analysis of human tumors with machine learning offers a novel approach to deciphering the mechanisms underlying macrophage function.Methods We developed a spatial virtual cell foundation model, a custom multimodal transformer with 450M parameters, trained on CosMx spatial transcriptomics across over 40 million cells from 1399 non-small cell lung cancer (NSCLC) tumor resections. 1 2 This model allows the simulation of gene expression in ‘virtual cells’ placed within spatially resolved TMEs and unlocks the ability to conduct in silico, cell-type specific experiments. To identify local cell-cell interactions between macrophages and particular cell types, we designed virtual ‘clonal neighborhood’ simulations, in which a virtual macrophage is surrounded by digital replicates of real tumor cells or fibroblasts sampled from NSCLC patients’ TME. We simulated 500 tumor or fibroblast clonal neighborhoods for each patient, leading to inferred gene expression on a total of 873,000 virtual macrophages. This allowed us to identify tumor-intrinsic and fibroblast-mediated transcriptional programs spatially associated with macrophage state. Topic modeling was applied to interpret the inferred macrophage transcriptional states, and predictive modeling analyses were used to identify tumor and fibroblast gene expression programs associated with different states.Results Analysis of virtual macrophage revealed a diversity of transcriptional programs acquired from interacting with tumor cells or fibroblasts neighborhoods. A subset of these programs were associated with macrophage polarization towards immunosuppressive SPP1+ or immunogenic CXCL9+ states, known to correlate with patients‘ response to immunotherapy. We further identified programs in tumor cells and fibroblasts that correlate with these macrophage states. Specifically, our experiments unveiled collagen expressing fibroblasts promoting SPP1+ virtual macrophage phenotypes. This population of fibroblasts was enriched in KRAS STK11 mutant vs KRAS mutant tumors, suggesting genotype specific TME characteristics that may account for resistance to immune checkpoint therapy in KRAS STK11 mutant tumors.Conclusions Our spatial virtual cell foundation model approach uncovers novel multicellular processes within the NSCLC TME, and helps elucidate how spatially-resolved events shape macrophage polarization. This approach identifies actionable targets, offering new insights to enhance immunotherapy efficacy and reverse immunosuppressive tumor environments.References Lacey J Padron, et al. #1231 Foundation models of cell and tissue biology enabled by custom scaled data generation: insights from 1000 lung tumor samples. Journal for ImmunoTherapy of Cancer. 2024;12. Yubin Xie, et al. #3652 OCTO-virtual cell: a foundation model of cell and tissue spatial biology with application to patient stratification and target discovery. AACR. 2025.