Masato Fukae 1, Mohamed Tarek 2, Kazutaka Yoshihara 1
1 Daiichi Sankyo Co., Ltd. (Tokyo, Japan), 2 Pumas-AI Inc. (, USA)
Objectives:
Thrombocytopenia is one of the most common adverse events in oncology drug development, which is often a dose-limiting factor and determines dosing regimens. A longitudinal pharmacokinetic/pharmacodynamic (PK/PD) analysis using a nonlinear mixed-effects (NLME) model has played an integral role in modeling platelet count based on the mechanistic understanding of chemotherapy-induced myelosuppression. However, not every aspect of myelosuppression is fully understood, and uncommon platelet dynamics in which a typical longitudinal PK/PD model fails to capture the time course are frequently observed. In such a case, a novel hybrid methodology, scientific machine learning (SciML) using the DeepNLME technique, which combines mechanistic modeling, machine learning, and statistical modeling, is expected to adequately predict dynamics without compromising the mechanistic and individualizable nature of NLME models [1]. Milademetan is a selective small-molecule inhibitor of the murine double minute-2 and has been developed for advanced solid tumors or lymphoma. Milademetan induces thrombocytopenia with uncommon platelet profiles, leading to an intermittent dosing schedule of 3/14 in a phase 3 trial [2]. In the present study, the DeepNLME-based SciML approach was applied to predict platelet dynamics under milademetan treatment.
Methods:
The dataset included two studies, a first-in-human (FIH) study mainly in the US and a subsequent phase 1 study in Japan, and was split into training (the FIH study, N=102) and test (the Japan study, N=18) sets to assess generalizability. The training set was used to develop a model and estimate its parameters using both the traditional NLME and the DeepNLME approach. The myelosuppression model was used as the NLME with a slight time-dependent modification added to the feedback parameter [3]. For the DeepNLME approach, the feedback parameter was replaced by a neural network that used the ratio of platelet count over time to baseline, treatment duration, and a random effect (i.e., between-subject variability) as input variables, while everything else remained the same as in the NLME. Model performance was assessed primarily using the test set, with numerical measures such as the -2x log-likelihood and visual diagnostics. Finally, the models were interpreted by investigating the time course of each compartment and conducting a Shapley Additive exPlanations (SHAP) analysis. The first-order conditional estimation method was used for parameter estimation. Julia v1.10.8 and DeepPumas v0.8.1 were used throughout the analysis.
Results:
The dose levels in the training set ranged more widely than in the test set (15 to 340 mg vs. 60 to 120 mg), suggesting that prediction for the test set is likely feasible. The performance of NLME and DeepNLME was compared on the test set, and DeepNLME outperformed, with an improvement in the -2x log-likelihood of approximately 51, showing more adequate diagnostic plots in a quite flexible manner. The time course of each variable in a compartment suggested that the DeepNLME approach preserves the mechanistic nature of PK/PD modeling while acknowledging that a highly time-dependent feedback parameter can yield flexible predictions. The SHAP analysis of the neural network that replaced the feedback parameter indicated that the platelet ratio contributed most to describing the dynamics.
Conclusions:
Overall, the DeepNLME-based SciML model demonstrated the ability to flexibly predict the platelet time course, providing an alternative to the traditional NLME model for PK/PD modeling.
References:
[1] Rackauckas et al. (2020) Universal Differential Equations for Scientific Machine Learning arXiv:2001.04385
[2] Gounder et al. (2023) A First-in-Human Phase I Study of Milademetan, an MDM2 Inhibitor, in Patients With Advanced Liposarcoma, Solid Tumors, or Lymphomas J Clin Oncol 41, 1714-1724
[3] Friberg et al. (2002) Model of Chemotherapy-Induced Myelosuppression With Parameter Consistency Across Drugs J Clin Oncol 20, 4713-4721
Reference: PAGE 34 (2026) Abstr 12089 [www.page-meeting.org/?abstract=12089]
Poster: Drug/Disease Modelling - Oncology