Senior Director of Data Science InsightRX, United States
Objectives: Model-informed precision dosing (MIPD) leverages predictive models and patient biomarkers to tailor doses to individual patients. MIPD has shown promise in disease areas ranging from antibiotics to transplants to autoimmune disease; however, the success of a given MIPD protocol depends on a variety of factors, such as biomarker sample timing or frequency, dose selection algorithm, and model suitability1. The large number of MIPD protocol parameters that can be optimized prohibits prospective clinical trials for data-driven protocol development.
Simulation permits investigation of protocol parameters towards successful implementation of MIPD in a trial or in routine clinical practice, but existing tools are cumbersome to use. Software such as NONMEM or nlmixr2 emphasize model development, and expect predefined dosing regimens during simulations, rather than doses that change in response to simulated outcomes. Sample optimization algorithms emphasize information gain, and do not assess attainment of pharmacokinetic (PK) or pharmacodynamic (PD) targets. Here, we create an R package that provides tools for simulating MIPD treatment courses to assess the relationship between protocol parameters and attainment of PK/PD targets.
Methods: The mipdtrial R package was designed to take as input a predefined tabular data set of synthetic patient covariates, since this data structure is typically available for other simulation work. PK/PD model simulations are performed using the open-source package PKPDsim, chosen because it was designed to support MIPD and includes a variety of commonly used model structures and tools for flexibly implementing user-defined models. Maximum a posteriori (MAP) Bayesian estimation is performed with the open-source package PKPDmap, selected for its support of PKPDsim models. The mipdtrial R package presents a range of composable functions intended for use within an R script or notebook, with the goal of allowing analysts to focus on protocol parameters like sample timing or initial dosing nomogram instead of writing code.
Results: Three proof-of-concept experimental designs were investigated. First, the impact of model misspecification on attainment of cumulative area under the curve (cAUC) in busulfan MIPD for bone marrow conditioning was investigated. Using a 1-compartment model to estimate individual PK parameters and a 2-compartment model to simulate “true” concentration-time curves resulted in 100% estimated target attainment, but only 86% true target attainment. Second, the impact of sample timing was investigated. Reducing sample count from 4 samples to 3 reduced true target attainment to 82%, which may be an acceptable trade-off. Finally, MAP and non-compartmental analysis (NCA) approaches were compared, with NCA resulting in lower target attainment by 27 percentage points.
Conclusion: The mipdtrial package facilitates development of a MIPD protocol using simulation.
Citations: [1] Hughes, Long-Boyle & Keizer. "Maximum a posteriori Bayesian methods out-perform non-compartmental analysis for busulfan precision dosing." Journal of Pharmacokinetics and Pharmacodynamics. (2024).
