Reza Najafi Zare (1), David Busse (1), Ida Neldemo (2), Alejandro Perez Pitarch (1), Lena Friberg (2,3), Ali Farnoud* (4)
(1) Boehringer Ingelheim Pharma GmbH & Co. KG, Ingelheim, Germany, (2) Pharmetheus AB, Uppsala, Sweden, (3) Department of Pharmacy, Uppsala University, Uppsala, Sweden, (4) Boehringer Ingelheim Pharma GmbH & Co. KG, Biberach, Germany
Objectives: Thrombocytopenia, a frequent adverse event type in cancer therapy, requires clinical platelet management to avoid high-grade thrombocytopenia and to ensure treatment continuation. The semi-physiological model for myelosuppression [1] is frequently leveraged to characterize and predict myelosuppression (“standard model”), including platelet dynamics. Yet, this standard model may require time-intense model refinement (“refined model”) to reach adequate predictive performance, which may impede timely support of clinical drug development programs. Machine learning approaches, combining standard nonlinear mixed-effects (NLME) methods with neural networks (NN) promise to streamline model development in situations of incomplete mechanistic understanding or when further model refinement is required [2,3].
Based on the example of thrombocytopenia induced by brigimadlin, a highly potent, oral murine double minute 2 homolog-tumor protein 53 antagonist, this work aimed to:
- Integrate NN into the NLME method (“NN-assisted model”)
- Compare model predictivity between the standard, refined and NN-assisted myelosuppression model
Methods: Model development was based on platelet counts in ≥2nd line patients with locally advanced and metastatic solid tumors (n=82, NCT03449381). Patients received oral brigimadlin administrations (10-80 mg on day 1 q3w, day 1 and 8 q4w, or day 1 and 3 q4w).
The exposure-safety relationship was characterized by relating plasma exposure to platelet growth parameters of the myelosuppression model. The power function characterizing the feedback of circulating on proliferating platelets required further refinement (i.e., empirical description) to improve model performance (refined model). Standard model implementation and model refinement were performed in NONMEM 7.4 and both models were implemented in Pumas 2.5.0 to allow comparison to the NN-assisted model.
The conventional NLME-based approach of trial and error to refine the feedback function (i) lacked robustness and adaptability, particularly when additional patients were included in the dataset during the study and (ii) required significant time investment. NN were used as universal function approximators to identify an appropriate feedback function in DeepPumas (NN-assisted model). The NN simultaneously estimated inputs and explored its combinations to predict unknown terms. This dual optimization procedure involved the FOCE method maximizing the parameter estimations’ likelihood, while NN systematically searched the parameter space.
Results: The NN-assisted model reduced model development time by identifying a suitable feedback function and accomplishing parameter estimation in <2 h. Model execution times in Pumas 2.5.0 were approximately 15 min (refined model) and 60 min (NN-assisted model).
Comparison of model predictivity highlighted the NN-assisted model’s improved predictive performance over the standard and refined model. The visual predictive check (VPC) of the standard model indicated that the observed platelet nadir (t = 4-5 weeks) fell outside the 95% CI of the predicted 10th and 50th percentiles, with a relative difference of -48% and -5% (i.e., overprediction), respectively, when compared to the lower bound of the 95% CI around the 10th and 50th percentiles. In the refined myelosuppression model, the predictions generally conformed to the observation patterns across all percentiles. However, the observed platelet nadir fell outside the 95% CI of the predicted 10th percentile, exhibiting -20% relative difference (i.e., overprediction) between the 10th observed percentile and the lower bound of the corresponding CI. VPCs indicated best predictivity at the platelet nadir for the NN-assisted model: No observations were outside the 95% CI of the 10th, 50th, 90th percentiles and the overall trend in platelets over time was adequately described.
Conclusions: The NN-assisted myelosuppression model reduced model development time substantially while improving predictivity, compared to the standard and refined model. In contrast to the trial and error approach of refining the standard model, the application of NN enabled rapid and automated identification of the feedback function, enhancing its adaptability and generalizability when incorporating additional patient data. This work advocates for NN integration in complex pharmacometrics scenarios with incomplete mechanistic understanding or when further model refinement is required.
References:
[1] Friberg LE, Henningsson A, Maas H, Nguyen L, Karlsson MO. Model of chemotherapy-induced myelosuppression with parameter consistency across drugs. J Clin Oncol. 2002;20: 4713–4721. doi:10.1200/JCO.2002.02.140
[2] Farnoud A, Ohnmacht AJ, Meinel M, Menden MP. Can artificial intelligence accelerate preclinical drug discovery and precision medicine? Expert Opin Drug Discov. 2022;17: 661–665. doi:10.1080/17460441.2022.2090540
[3] Rackauckas C, Ma Y, Martensen J, Warner C, Zubov K, Supekar R, et al. Universal Differential Equations for Scientific Machine Learning. arXiv [cs.LG]. 2020. Available: http://arxiv.org/abs/2001.04385
Reference: PAGE 32 (2024) Abstr 11106 [www.page-meeting.org/?abstract=11106]
Poster: Methodology - New Modelling Approaches