Takayuki Katsube
Shionogi & Co., Ltd., Japan
Objectives: Machine leaning is an approach to provide a powerful computational efficiency and is expected to complement pharmacometric analyses in drug development [1]. In pharmacokinetic/pharmacodynamic analyses using time-course data, compartment model analysis is standardly used and work reasonably. However, this analysis would be time-consuming for model building in some cases where the numbers of subject, compartments, and/or covariates are large. Machine leaning approach may increase efficiency for the pharmacokinetic/pharmacodynamic analyses. The aim of this study was to apply machine leaning approach for model building and covariate selection in the pharmacokinetic/pharmacodynamic analysis of lusutrombopag, a thrombopoietin receptor agonist.
Methods: A machine learning approach, Gradient Boosting Decision Tree, using Python and LightGBM [2] was applied to platelet count data from thrombocytopenic patients with chronic liver disease after multiple doses of lusurombopag in clinical trials [3]. The learning data composed of 2520 platelet count data from the 253 thrombocytopenic patients and the test data of 1006 platelet count data from 94 thrombocytopenic patients were explored from time after initial dose (TAID), total cumulative area under the plasma drug concentration-time curve (AUC) of lusutrombopag during the study, Child-Pugh score, baseline platelet counts, age, sex, and ethnicity (Japanese or non-Japanese patients) using SHapley Additive exPlanation (SHAP) values [4]. The predictive performance of the selected model using LightGBM was compared with that of the pharmacokinetic/pharmacodynamic compartment model which had the same structure as the reported pharmacokinetic/pharmacodynamic model of lusutrombopag (3-compartment pharmacokinetic model and 5-compartment pharmacodynamic model) [3].
Results: The selected model using LightGBM described the platelet counts in the training data well. The absolute SHAP values from the selected model was the highest for TAID, followed in order by baseline platelet counts and total cumulative AUC of lusutrombopag, which were expected to be influential covariates based on the previous pharmacokinetic/pharmacodynamic modeling of lusutrombopag [3]. The predictive performance for platelet counts using the test data was comparable between the selected model from LightGBM and the pharmacokinetic/pharmacodynamic compartment model (mean absolute error and root mean squared error of 1.19 and 1.65 for the selected model from LightGBM and 1.07 and 1.53 for the pharmacokinetic/pharmacodynamic compartment model, respectively). The run time up to completion was very fast for training the model using LightGBM (a few seconds per run), which would be an advantage of the machine leaning approach.
Conclusion: The machine leaning was applied for the pharmacokinetic/pharmacodynamic modeling with covariate selection of lusutrombopag. The machine learning approach would be an alternative for pharmacokinetic/pharmacodynamic analyses and expected to provide a powerful computational efficiency.
References:
[1] McComb A et al. Br J Clin Pharmacol. 2022. 88:1482-99.
[2] Ke, G. et al. NIPS 2017. 3149–57.
[3] Katsube T et al. Clin Pharmacokinet. 2019. 58:1469-82.
[4] Lundberg SM et al. NIPS 2017. 4768–77.
Reference: PAGE 32 (2024) Abstr 10787 [www.page-meeting.org/?abstract=10787]
Poster: Methodology – AI/Machine Learning