Computational Framework for Multi-Objective Optimization of Population Dosing Regimens
Sebastian Tandar1, Linda Aulin1,2, Coen van Hasselt1
1Leiden Academic Centre for Drug Research, 2Centre for Human Drug Research
Introduction Drug treatments often involve multiple competing objectives, such as balancing efficacy, toxicities, resistance or cost, making treatment optimization inherently complex[1]. This challenge is further compounded by interindividual variability (IIV), which can make standardized regimens suboptimal for certain patients[2]. While model-informed precision dosing enables patient-specific dosing[3], these approaches typically optimize only for a single objective, and for a single patient. Control theory-based methods offer alternatives to optimize dosing for individuals or populations using a typical individual as a surrogate[4], but currently do not yet incorporate patient variability and dosing intervals in the optimization strategy. In this study, we develop, evaluate, and apply a novel PK-PD-based framework to optimize dosing regimens for a patient population by incorporating IIV and patient covariate distributions. The framework supports objective weight assignment, enabling simultaneous, weighted, multi-objective optimization of dosing regimen to tailor to specific clinical priorities. Methods Optimization framework. Our framework uses optimization algorithms to identify an optimal treatment setting (e.g. dose amount, frequency, covariate-dependent dose). Each optimization iteration involves Monte Carlo simulations with a population PK-PD model to predict objective attainment for each patient in the simulated population. A logistic function converts attainment scores into continuous penalties, and the optimization algorithm minimizes the weighted sum of penalties across all treatment objectives. The efficiency of several optimization algorithms in identifying a (global) optimum was tested against exhaustive grid optimization in Case Study 1. The framework was built in R with a sample NONMEM implementation in Case Study 2. Case study 1: Eltrombobag. We applied the framework to optimize eltrombopag dose regimens to enhance platelet count in immune thrombocytopenic purpura (ITP) patients[5]. Treatment objectives were equally weighted and included: 1) maintaining steady-state platelet count within 50–200 GI/L, 2) keeping maximum platelet count <400 GI/L, 3) limiting daily eltrombopag dose to <200 mg[6], and (4) maintaining a minimum 70% time when platelet count is =50 GI/L. Case Study 2: Antibacterial combination therapy. optimized colistin (COL)-meropenem (MER) combination therapy for A. baumannii infections[7], [8], [9]. Objectives included: 1) limiting daily colistin and meropenem doses, 2) maintaining plasma drug concentrations below toxicity thresholds, 3) keeping bacterial density <107 CFU/mL, and 4) achieving an endpoint =2-log10 bacterial density reduction. To assess objective prioritization, the case was analyzed using equal weighting and a scenario with unequal weighting that prioritized bacterial density reduction and COL toxicity objectives. Results Case study 1: Eltrombobag. The standard (50 mg/day) and elevated standard (75 mg/day) eltrombopag doses achieved the target platelet range in 40.2% and 46.4% of the ITP population, respectively, with male patients exhibiting a markedly lower response. Dose optimization, incorporating adjusted dose amounts and sex-specific dosing, increased this target attainment by 1.35- and 1.17-fold compared to standard and elevated doses. Swarm-based algorithms approximated the global optima identified by the grid search (17,500 points) after evaluating ~1,200 points. Other algorithms were less effective and often sensitive to the starting point. Case Study 2: Antibacterial combination therapy. COL-MER combination therapy was optimized for an A. baumannii isolate resistant to standard monotherapies. Optimization with equal weight met the bacterial density target in 18% of the population. Prioritizing bacterial kill and COL toxicity increased the attainment of this target to 62%, with MER toxicity exceeding the threshold in 1% of cases. In NONMEM, similar weighting converged on a local optimum, raising target attainment to 50% but violating the MER toxicity threshold in all cases. Conclusion The presented novel framework provides a systematic approach to treatment optimization; leveraging population PK-PD models to identify a standardized population dosing regimen that can maximize the attainment of multiple treatment targets across a patient population while offering flexibility to prioritize selected treatment objectives.