Automating Population Pharmacokinetic Model Development using Machine Learning
Sam Richardson* (1), Itziar Irurzun-Arana* (2), Andrzej Nowojewski (1), Diansong Zhou (3), Damilola Olabode (3), Jacob Leander (4), Weifeng Tang (5), Richard Dearden (1), Megan Gibbs (5)
* Equal contribution, (1) Imaging & Data Analytics, Clinical Pharmacology & Safety Sciences, BioPharmaceuticals R&D, AstraZeneca, Cambridge, UK, (2) Clinical Pharmacology & Quantitative Pharmacology, Clinical Pharmacology & Safety Sciences, BioPharmaceuticals R&D, AstraZeneca, Cambridge, UK, (3) Clinical Pharmacology & Quantitative Pharmacology, Clinical Pharmacology & Safety Sciences, BioPharmaceuticals R&D, AstraZeneca, Waltham, MA, USA, (4) Clinical Pharmacology & Quantitative Pharmacology, Clinical Pharmacology & Safety Sciences, BioPharmaceuticals R&D, AstraZeneca, Gothenburg, Sweden, (5) Clinical Pharmacology & Quantitative Pharmacology, Clinical Pharmacology & Safety Sciences, R&D BioPharmaceuticals, AstraZeneca, Gaithersburg, MD, USA
Population pharmacokinetic (popPK) models are used in decision-making throughout drug development to inform dose selection, clinical study design, and labelling claims. Development of these models can be time-consuming, requiring a lot of manual effort from pharmacometricians. Conventionally, models are built using a greedy stepwise approach which can result in non-optimal structure selection due to time restrictions limiting exploration and the lack of robustness to local minimums found in local search strategies [1]. Model structure development can be especially challenging for extravascular administered drugs as the parameter space is vast due to the potential for complex absorption behaviour.
Here we present a data-driven approach using machine learning models to automatically find a model structure to best fit phase 1 clinical data. To enable this, we have developed a parameter space capable of generating popPK models describing the time course of extravascular administered drugs. Additionally, we define a fitness function used to select the best performing model which also features plausible model parameters, mimicking the typical decision process of a pharmacometrician.
Objectives:
- Develop a generic approach for automatic discovery of popPK model structure.
- Evaluate the method on real and synthetic clinical data and monitor reproducibility of stochastic search strategy.
- Analyse the impact of the choice of penalty function and the use of global search algorithms in addition to local search.
Methods:
A model space was designed that includes >12K unique popPK models for extravascular administered drugs without active metabolites. These models were described by 17 features which detail various structural elements. A modified version of PyDarwin [2, 3] was used to search this model space for optimal popPK model structures using Bayesian optimisation with a random forest surrogate model followed by an exhaustive local search. Each candidate model was evaluated using NONMEM and ranked by a fitness score which was the sum of the NONMEM objective function value (OFV) and a custom penalty function. The penalty function comprised of two components. Firstly, an Akaike information criterion penalty, as used in prior studies [2], which penalized based on the number of model parameters. Secondly, an estimated parameter penalty designed to ensure the selection of plausible structures by penalizing for abnormal relative error and omega values. Evaluation was performed using a synthetic ribociclib dataset, as well as Phase 1 clinical datasets for camizestrant, osimertinib, olaparib, and tezepelumab. Searches were repeated 5 times to test reproducibility of the process and the overall top model from each repeat was compared to published model structures [4, 5, 6, 7, 8].
Results:
Automatically discovered structural models are presented for each dataset with comparison to published models. Structures generated automatically shared 15/17 structural features with published models on average. Differences found in the discovered structures led to improved fitness and reduced OFV values by 5% for the data tested. The average search time was less than 48 hours using 40 cores and 40GB with <3% of models tested in each search. To understand the impact of the global search step, searches were run using the local search only. These experiments were unable to identify the optimal model for one of the datasets demonstrating the potential for a global search to avoid local minimums. Ablation experiments demonstrated the importance of the choice of penalty function. With the estimated parameter penalty removed, structures with abnormal parameter values were generated for 5/5 datasets.
Conclusions:
We present an automated search approach for discovering popPK models which features a generic model space we developed for extravascular administered drugs and a penalty function we designed to select model structures that a human pharmacometrician would find plausible. We have evaluated our approach on real and synthetic clinical datasets where we were able to reproducibly generate models within 48 hours which were competitive with those manually developed. Analysis was provided which demonstrated the benefit of the penalty function and global search. Adoption of automatic search algorithms for PopPK development has the potential to free pharmacometricians from repetitive tasks, improve model quality, and speed up popPK analysis.
References:
[1] Hoos, H. H.; Stützle, T. Stochastic Local Search: Foundations and Applications; 2004.
[2] Sale, M.; Sherer, E. A. A Genetic Algorithm Based Global Search Strategy for Population Pharmacokinetic/Pharmacodynamic Model Selection. Br. J. Clin. Pharmacol. 2015, 79 (1), 28–39. https://doi.org/10.1111/bcp.12179.
[3] Li, X.; Sale, M.; Nieforth, K.; Bigos, K. L.; Craig, J.; Wang, F.; Feng, K.; Hu, M.; Bies, R.; Zhao, L. pyDarwin: A Machine Learning Enhanced Automated Nonlinear Mixed-Effect Model Selection Toolbox. Clin. Pharmacol. Ther. n/a (n/a). https://doi.org/10.1002/cpt.3114.
[4] Lu, Y.; Yang, S.; Ho, Y.-Y.; Ji, Y. Ribociclib Population Pharmacokinetics and Pharmacokinetic/Pharmacodynamic Analysis of Neutrophils in Cancer Patients. J. Clin. Pharmacol. 2021, 61 (8), 1054–1068. https://doi.org/10.1002/jcph.1856.
[5] Irurzun-Arana, I.; Sykes, A.; Bachina, S.; Brier, T.; Lindemann, J. P.; Klinowska, T. Population PK Modelling of Camizestrant, a next-Generation Oral Selective Estrogen Receptor Degrader, as Monotherapy and in Combination with Palbociclib in ER+/HER2- Advanced Breast Cancer. PAGE 31 2023.
[6] Zhou, D.; Li, J.; Bui, K.; Learoyd, M.; Berges, A.; Milenkova, T.; Al-Huniti, N.; Tomkinson, H.; Xu, H. Bridging Olaparib Capsule and Tablet Formulations Using Population Pharmacokinetic Meta-Analysis in Oncology Patients. Clin. Pharmacokinet. 2019, 58 (5), 615–625. https://doi.org/10.1007/s40262-018-0714-x.
[7] Brown, K.; Comisar, C.; Witjes, H.; Maringwa, J.; de Greef, R.; Vishwanathan, K.; Cantarini, M.; Cox, E. Population Pharmacokinetics and Exposure‐response of Osimertinib in Patients with Non‐small Cell Lung Cancer. Br. J. Clin. Pharmacol. 2017, 83 (6), 1216–1226. https://doi.org/10.1111/bcp.13223.
[8] Ly, N., Zheng, Y., Griffiths, J.M., van der Merwe, R., Agoram, B., Parnes, J.R. and Roskos, L. (2021), Pharmacokinetic and Pharmacodynamic Modeling of Tezepelumab to Guide Phase 3 Dose Selection for Patients With Severe Asthma. The Journal of Clinical Pharmacology, 61: 901-912. https://doi.org/10.1002/jcph.1803
Conflict of Interest statement:
All authors were employees of AstraZeneca at the time of study and may own stocks or stock options. AstraZeneca sponsored these studies and funded this work.
Funding information:
AstraZeneca sponsored all studies used in this work apart from NCT00972179 and NCT02512900 where Amgen and Medimune LLC were the sponsors respectively. Amgen was a collaborator for study D5180C00002 and Medimune LLC were a collaborator for study D5180C00003. AstraZeneca funded the research presented in this paper.