Elias Torstensson 1,2, Benjamin Olsson 1,2, Hongtao Yu 3, Morteza Chehreghani 2, Mikael Sunnåker 1
1 Clinical Pharmacology and Quantitative Pharmacology, Clinical Pharmacology and Safety Sciences, R&D, AstraZeneca (Gothenburg , Sweden), 2 Department of Computer Science and Engineering, Chalmers University of Technology (Gothenburg, Sweden), 3 Clinical Pharmacology and Quantitative Pharmacology, Clinical Pharmacology and Safety Sciences, R&D, AstraZeneca (Gaithersburg, US)
Objectives: Neural ODEs-based models [1] constitute a novel tool for PKPD modelling that utilize neural networks to describe longitudinal data and to make predictions. These models have demonstrated advantages in memory efficiency and the ability to adapt computation based on problem complexity, which makes them suitable for data from clinical trials. Neural ODEs have previously been employed in the pharmacometrics domain, e.g., for PK modelling [2]. A potential application of neural ODEs is longitudinal data for adverse events, which are commonly key for selecting Ph2b and Ph3 doses. The aim of this work is to investigate how well neural ODEs performs in such a setting.
Methods: A dataset was generated by simulating the frequency of mild or moderate nausea upon treatment with the experimental drug cotadutide, for different treatment schemes, from a previously developed Markov model [3]. The dataset was split into a training dataset comprised of three titration schemes, and a test dataset of eight additional datasets. The titration scheme differed in terms of starting dose, the frequency of data increases, and the rate of dose increases. A neural ODE model was implemented by encoding the cumulative doses into a two-dimensional latent space representing mild and moderate nausea. The resulting trajectory over time was then augmented by the time after last titration step and decoded into predictions of nausea. The neural ODE model was trained five times using early stopping resulting in five slightly different models. The final prediction, with uncertainty, was then calculated using the mean and standard deviation of the individual predictions.
Results: The neural ODE model could well describe the data in the training set for both severities of nausea. For a systematic evaluation and to infer the most informative titration schemes, the model was evaluated on training datasets for all treatment scheme combinations of exactly one weekly, one bi-weekly, and one monthly titration scheme. One optimal treatment combination was identified by comparing the MSE-loss after evaluating the model on different test datasets, and by visual inspection the model predictions appear to be in line with the test dataset.
Conclusions: We have shown that neural ODEs may be a valuable tool for predicting adverse events outcomes over time, which is commonly important for identification of the therapeutic window in drug development.
References:
1. R. T. Q. Chen, Y. Rubanova, J. Bettencourt, and D. Duvenaud, Neural ordinary differential equations, 2019. arXiv: 1806.07366 [cs.LG]. [Online]. Available:
https://arxiv.org/abs/1806.07366.
2. J. Lu, K. Deng, X. Zhang, G. Liu, and Y. Guan, “Neural-ode for pharmacokinetics modeling and its advantage to alternative machine learning models in predicting new dosing regimens,” iScience, vol. 24, p. 102 804, Jun. 2021. doi: 10.1016/j.isci.2021.102804
3. H. Yu, S. Ueckert, L. Zhou, et al., “Exposure-response modeling for nausea incidence for cotadutide using a Markov modeling approach,” CPT: Pharmacometrics Systems
Pharmacology, vol. 13, pp. 1582–1594, Jul. 2024. doi: 10.1002/psp4.13194.
Reference: PAGE 34 (2026) Abstr 12054 [www.page-meeting.org/?abstract=12054]
Poster: Methodology – AI/Machine Learning