Senior Modeler Rosa & Co. LLC Estepona, Andalucia, Spain
Disclosure(s):
Renee Myers, BS: No financial relationships to disclose
Alvaro Ruiz-Martinez, PhD: No financial relationships to disclose
Objectives: Virtual patients and virtual populations (VPops) are used to explore clinical variability and uncertainty in Quantitative Systems Pharmacology (QSP) modeling. Although unweighted virtual patient cohorts are straightforward to generate using random sampling, development of virtual populations matching patient characteristics in a clinical trial is challenging and computationally intensive. To address these limitations, Rieger et al presented three algorithms (simulated annealing, genetic algorithms, and Metropolis Hastings) for improved generation of virtual populations [1]. Alternatively, parallel tempering, a well-established method for parameter estimation [2], offers a competitive approach for VPop generation, particularly for sampling of complex, high dimensional parameter spaces. This work describes an implementation of parallel tempering for virtual population generation using a published model and compares computational cost, efficiency, goodness of fit, and parameter diversity to simulated annealing (SA) and Metropolis Hasting (MH) methods.
Methods: A published MAPK signaling model [3], expanded to include a module describing mouse xenograft tumor growth, was used for evaluation of the three algorithms. A set of 14 parameters known to impact tumor growth were varied independently, and published mouse data containing bounded ranges for three treatment protocols (untreated, KRASi, and SHP2i) was used to calibrate the virtual populations [4]. Parallel tempering was performed using 4 parallel Markov Chain Monte Carlo (MCMC) simulations and uniform prior distributions for sampled parameters were used for sampling. SA was implemented using a cost function set to optimize between data bounds [1] and MH was implemented using a modified target distribution [1] assuming a uniform distribution of responses.
Results: All methods tested showed diversity in parameter values for the virtual populations generated, indicating that each method was capable of producing mechanistically distinct virtual patients. Parallel tempering showed improvements over both SA and MH for computational efficiency, producing virtual patients that did not require additional prevalence filtering following initial simulation. Goodness of fit was also evaluated for all methods assuming a uniform distribution for the data, and parallel tempering showed comparable results to SA and MH for populations of ~9000 virtual patients.
Citations: [1] Rieger T, et al. Improving the generation and selection of virtual populations in quantitative systems pharmacology models. Progress in Biophysics and Molecular Biology. (2018) [2] Gupta S, et al. Evaluation of Parallel Tempering to Accelerate Bayesian Parameter Estimation in Systems Biology. Proceedings—26th Euromicro International Conference on Parallel, Distributed, and Network-Based Processing.(2018) [3] Sayama H, et al. Virtual clinical trial simulations for a novel KRASG12C inhibitor (ASP2453) in non-small cell lung cancer. CPT Pharmacometrics Syst Pharmacol. (2021) [4] Sheehan R, et al. Parallel Tempering for Generation of Virtual Patients and Virtual Populations in QSP Models [Poster abstract]. ACoP2024. (2024)
