A Proteome-Calibrated Mechanistic Digital Twin of the Glioblastoma Astrocyte Derived from Patient Multi-Omic Data

Digital twins of biological systems have been proposed as a framework for personalized medicine; however, existing implementations in oncology operate at the tissue or population scale with parameters derived from textbook kinetics rather than patient data. We describe a mechanistic digital twin of the glioblastoma (GBM) astrocyte in which all rate constants are calibrated to proteomic measurements from 50 wild-type GBM patients from the Clinical Proteomic Tumor Analysis Consortium (CPTAC, BCM source, n = 99 total). The twin comprises a 25-node ordinary differential equation (ODE) signaling network spanning the EGFR–MAPK, PI3K–AKT–mTOR, JAK–STAT, and cell cycle axes, a Hamiltonian covariance matrix constructed from 9,358 proteomic features (spectral concentration c = 0.168, 80% variance in 24 modes), and a gradient-based algebraic steady-state calibration procedure that achieves Spearman ρ = 1.000 between simulated and measured protein abundances across 10 observable signaling nodes in 3.6 seconds on an A100 GPU. Five clinically relevant genetic perturbations were implemented (EGFRvIII, PTEN loss, IDH1 R132H, PD-L1 upregulation, compound EGFRvIII + PTEN loss), and four drug classes were simulated (erlotinib, temsirolimus, atezolizumab, ivosidenib) at multiple doses. All six pre-specified biological validation checks passed, including PTEN-loss AKT escape from erlotinib via basal PI3K activity. Erlotinib dose–response curves reveal that EGFRvIII cells maintain pERK activity until approximately 1 µ M before collapsing, while PTEN-loss cells exhibit deeper suppression because the EGFR–MAPK axis becomes the sole remaining survival signal when PI3K is constitutively active. Perturbation directionality is mechanistically coherent across all genotype–drug combinations. The framework demonstrates that proteome-calibrated mechanistic cell digital twins are computationally tractable, biologically interpretable, and distinguishable from existing population-level statistical models in both their construction and their limitations. Code is available at https://github.com/radres2019/gbm-digital-twin .