Mengdi Tao, n/a: No relevant disclosure to display
Objectives: Quantitative Systems Pharmacology (QSP) is an emerging field that combines disease pathology with drug mechanisms of action allowing the generation of virtual patient twins, thereby accelerating and de-risking clinical trial development. QSP models utilize diverse datasets, including omics data, pre-clinical studies, clinical trials, and real-world evidence, which introduce significant development challenges due to data heterogeneity and the complexity of multi-scale model structures. This abstract introduces three innovative approaches aimed at developing a robust virtual twin workflow: 1) identifying appropriate calibration parameters for virtual twins; 2) assessing overfitting; 3) validating virtual twins with limited data.
Methods: The first approach employs a combination metric derived from local sensitivity analysis (LSA) index and a reliability classification score, modified from Braakman et al., 2019. LSA pinpoints parameters that significantly impact endpoints and biomarkers, while the reliability analysis evaluates the sources and quality of the parameter values, allowing the identification of candidate sensitive parameters for optimization. This contributed to the list of parameters used to generate virtual twins, along with known biological sources of variabilities and uncertainties. The second approach addresses potential model overfitting by comparing calibration results from different sequential optimization approaches and considering the stochastic nature of calibration algorithms. The third approach provides a validation framework when there is limited clinical data, in which case dividing the data into training and validation sets would prevent a meaningful analysis. A bootstrap approach of generating virtual populations based on the virtual twins was utilized to address this challenge.
Results: These approaches were successfully implemented on a QSP model used for regulatory interaction. Majority of the parameters in the model were deemed reliable, suggesting that the model is robust. The overfitting assessment further evaluated the robustness of the model, allowing the identification of the key model parameters with sufficient accuracy. The calibration quality of different sequential optimization approaches helped to assess the structural contribution of the model to the quality of the fits. The bootstrap-based virtual population validation supported the appropriate representation of model dynamics and increased its credibility even when the clinical data is limited.
Conclusions: The proposed approaches are important tools for enhancing virtual twin development with QSP models. Early identification of appropriate calibration parameters and recognition of overfitting risks are crucial for ensuring the model's quality and reliability. The validation exercise is a critical step to assess model performance and predictability, providing the level of confidence needed for QSP models to be applied in the proposed context.
Citations: Citations: [1] Braakman S, Gulati A, Tannenbaum S, Paxson R. Visualizing Parameter Source Reliability and Global Sensitivity for Quantitative Systems Pharmacology (QSP) Models, ACoP annual meeting, 2019; PII-145
