Paulo Paneque Galuzio1, Weirong Wang2, Shalla Hanson2
1Johnson & Johnson, 2Johnson & Johnson
Introduction: Quantitative Systems Pharmacology (QSP) is being increasingly adopted as a strategy for Model Informed Drug Development (MIDD). QSP uses high-dimensional, multi-scale models that leverage data from multiple sources, and isn’t always amenable to traditional Pharmacometrics-style model fitting strategies, due to the scale and complexity of the models. QSP models also often involve generation of virtual patients (VP) and virtual population (Vpop), and different methods for their calibration exist across the industry, many of which are computation intensive and time consuming. In this work we propose a workflow for the use of Bayesian optimization for QSP model calibration. The time efficiency features of Bayesian optimization make it a suitable method for accelerated model calibration, which is particularly useful in the early stages of model development. Objectives: Develop and evaluate the feasibility and limitations of Bayesian Optimization for QSP model calibration. Methods: A proof of concept of the Bayesian optimization method was established for a two-dimensional search space for a large, multi-scale QSP model in immuno-oncology. The model includes cell and antibody PBPK at the organ level, killing and activation at the cellular level, and receptor binding and internalization at the molecular level. The model was calibrated against clinical response data from a Phase 1 dose escalation study. A log-likelihood function of the response rates for the different cohorts was used as the optimization objective function. Distributed parameters of the generated virtual population were generated from pre-specified (known) log-normal distributions. For each objective function evaluation, cohorts of N randomly sampled virtual patients were used to estimate the model predicted probability of response per cohort. Results: Results were generated in a workstation with an Intel(R) Xeon(R) Platinum 8160T CPU @ 2.10GHz, 2095 Mhz, 24 Core(s), 48 Logical Processors, with 192GB of installed memory, running Matlab 2022b with 48 cores parallel pool. For a two-dimensional search space, and 22 cohorts simulated, using 100 VPs per cohort (leading to a total of 2200 model simulations per objective function evaluation) the Bayesian optimization achieves satisfactory convergence, with 5×48 iterations of the method, with run times in the order of ~1h. By reducing the number of iterations and the number of VPs per simulated cohort, runtimes can be reduced to as low as ~15min. The ability of Bayesian Optimization to handle non-deterministic objectives ensures good convergence properties even with reduced virtual populations. Additionally, the Bayesian optimizer can also constrain regions of the parameter space that fail to integrate successfully. Conclusions: Bayesian optimization was successfully used for QSP virtual population calibration with considerably small running times, which can help accelerate model development, especially in early stages of model development when multiple trials of model structure and comparison with data are necessary. Bayesian optimization is able to handle non-deterministic objectives, and to actively constrain regions of the parameter space that lead to unstable solutions, avoiding the selection of non-convergent parameters in the final Vpop. All these characteristics make Bayesian optimization suitable for QSP calibration. Its efficient execution also makes it a good candidate to be integrated into other more complex calibration procedures (replacing burn-in phases in Monte Carlo simulations, for example).
Reference: PAGE 33 (2025) Abstr 11610 [www.page-meeting.org/?abstract=11610]
Poster: Methodology - Estimation Methods