A structured foundation model framework for state-conditioned gene regulatory network inference in glioblastoma

Gene regulatory networks describe how regulators coordinate gene expression to shape cell identity, plasticity, and responses to environmental signals. In glioblastoma, these relationships vary across malignant programs and across interacting non-malignant cell types, making a single global network an inadequate description of tumor biology. At the same time, single-cell transcriptomic measurements are noisy and heterogeneous across studies, and direct causal supervision for regulatory edges is sparse and incomplete. These constraints motivate methods that can learn transferable representations from large-scale data while remaining explicit about what is supported by evidence and what remains a hypothesis.This work introduces a structured learning framework for estimating context-specific regulatory hypotheses in glioblastoma using a staged strategy. First, it learns reusable gene representations from a diverse compendium of brain tumor and brain single-cell datasets, guided by existing biological knowledge and molecular interaction resources. Second, it uses genetic perturbation measurements to calibrate a global edge scoring component that prioritizes regulator and target relationships consistent with observed responses to perturbation. Third, it projects this calibrated structure into a glioblastoma atlas to produce cell type specific regulatory activity estimates, along with interpretable regulator-centered target sets that can be summarized and tested for functional coherence.Across evaluations, the framework generalizes to regulators not seen during perturbation calibration and yields context-specific activity patterns that recover expected lineage regulators, highlight coherent functional programs, and show agreement with curated regulatory resources beyond simple network degree effects. Overall, the work provides a scalable and modular foundation for studying context dependent regulation in glioblastoma and for prioritizing experiments that can validate and refine proposed regulatory programs.