Quantitative Calibration of a Spatial QSP Model Identifies Fibroblast Impact on HCC Immunotherapy
AbstractComputational models are increasingly used to predict treatment response and optimize cancer therapy strategies. Among these, quantitative systems pharmacology (QSP) models mechanistically simulate tumor progression and pharmacological interventions, enabling virtual clinical trials, model-informed drug development, and biomarker identification. Coupling QSP with an agent-based model yields a spatial QSP (spQSP) platform that captures tissue-level spatial organization of the tumor microenvironment (TME). However, parameterizing such models to represent tumor biology remains an open problem. In this study, we developed a calibration framework using the Approximate Bayesian Computation - Sequential Monte Carlo (ABC-SMC) approach to calibrate the spQSP model with a combination of clinical and spatial molecular data, reflecting the TME characteristics of human tumors. This calibration framework matches tumor architectures between spQSP model predictions and patient spatial molecular data by fitting statistical summaries of cellular neighborhoods. We demonstrate that model calibration using CODEX data from untreated HCC patients enables prediction of TME spatial molecular states in an independent cohort receiving immune-checkpoint inhibitor (ICI) and tyrosine kinase inhibitor (TKI) combination therapy. Finally, we identify spatial and non-spatial pretreatment biomarkers and assess their predictive power for therapeutic response. This workflow demonstrates how integrating spatial-omics with multiscale mechanistic models enables quantitative calibration, biological insight, and in silico biomarker discovery, providing a framework for personalized cancer therapy across tumor types. SignificanceDigital twins and computational models are increasingly used to simulate disease and guide therapy, but they often struggle to capture the immense complexity driving the spatiotemporal evolution of the TME. The challenge is compounded by clinical sample limitations, which typically provide measurements from only static snapshots of the TME for parameter estimation. We demonstrate how mechanistic modeling frameworks can overcome this limitation by enabling inference of spatiotemporal model parameters from static spatial data - an intractable task for purely data-driven approaches. Ultimately, our work presents a workflow that integrates spatial-omics with multiscale mechanistic models, enabling quantitative calibration, deeper biological insight, and in silico biomarker discovery.