Sreenath M Krishnan1, Mélanie Wilbaux1, Amit Roy1, Niklas Korsbo1
1Pumas-AI Inc
Background and Objectives: Traditional pharmacokinetic (PK) models often encounter challenges in accurately characterizing drugs with atypical absorption kinetics [1], particularly those exhibiting complex or delayed-release mechanisms. Current methodologies rely on predefined absorption models, which limit flexibility in capturing unknown absorption patterns. To address this limitation, a hybrid PK modeling framework integrating neural networks with conventional compartmental models (neural ODE-based model) was developed to learn absorption processes directly from data while maintaining interpretability [2]. This study evaluates the predictive performance of this approach in characterizing the absorption profile of a trimodal-release formulation. Methods: Simulated Data: PK profiles of 100 virtual subjects receiving a daily dose of a oral trimodal-release formulation were simulated on Day 1 and Day 4, with sampling at 0.5, 1, 2, 3, 4, 6, 8, 12, 16, 20, and 24 hours post-dose. The PK profiles exhibited three distinct peaks corresponding to the following absorption phases: an immediate release (~35% of dose), a sustained release (~3–4 h delay), and a delayed release (~7–8 h post-dose). Clearance followed linear elimination from the central compartment, with a half-life of 4.5 hours. Model Evaluation: To assess the benefits of neural ODE-based models over traditional approaches, three distinct models were evaluated. First, a traditional PK model incorporating three transit compartments (Transit3), with estimation of the mean absorption time, was utilized as a baseline. Then, two different approaches were investigated for the neural ODE-based models, where drug distribution and elimination were modeled using a classical compartmental PK model. In the first neural ODE-based model, absorption was described using a neural network function (PK-NeuralODE) [2]. In the second approach, absorption was modeled as a distributed delay, with a neural network approximating the transfer function that governs the shape of that delay (PK-NeuralDelay). The PK-neuralODE model was trained to learn the absorption function while simultaneously estimating key PK parameters, including clearance (CL) and volume of distribution (V). Model development was based on 80 subjects, while the remaining 20 subjects were used for validation. The neuralODE models were benchmarked against Transit3. The validation dataset was used to assess individual prediction accuracy and marginal likelihood, which was derived from the dataset, the model, and its final estimates. Model evaluation criteria included predictive accuracy for individual predictions, exposure measures (AUC0-last and Cmax) and marginal likelihood. The proposed method was implemented in the Pumas/DeepPumas software suite (v2.6.0). [3, 4] Results: The trimodal absorption models were successfully implemented in Pumas and fitted using the different modeling approaches. Neural ODE-based models demonstrated greater flexibility and effectively captured the trimodal-release absorption kinetics, outperforming conventional PK models. The multiple peaks in absorption were more accurately captured at the individual level by the neural ODE-based models compared to the traditional approach. Both neural ODE-based models improved upon the classical transit compartment approach, with the PK-NeuralDelay model exhibiting superior performance compared to the PK-NeuralODE model. The marginal likelihood for the Transit3 model was -1771. Relative to this baseline, the PK-NeuralODE model improved the likelihood by 48 points, whereas the PK-NeuralDelay model achieved a substantial improvement of 196 points. Compared to Transit3, the neural ODE-based models demonstrated superior accuracy for AUC, with PK-NeuralODE exhibiting near-zero bias (-0.285% vs. 5.73%), while the PK-NeuralDelay showed better precision (21.9% vs. 25.4% CV). For Cmax, the convolution approach outperformed Transit3 in accuracy (+2.21% vs. -12.4% bias), followed by NNODE with a 9% CV. Based on the current study, PK-NeuralODE proved to be the most reliable for AUC predictions, whereas PK-NeuralDelay was preferable for Cmax accuracy. Conclusions: The integration of neural networks with semi-mechanistic PK models provides a robust framework for characterizing complex absorption processes, which can be applied to multi-modal oral formulations, long acting injectables and RNA-based therapeutics. This approach improves predictive accuracy in data-limited settings, enhances model flexibility while maintaining interpretability, and facilitates better treatment and disease management.
[1] Zhou, H. J. Clin. Pharmacol. (2003) 43: 211-227. [2] Valderrama et al., CPT Pharmacometrics Syst Pharmacol. (2024) 13(1) :41-53. [3] Korsbo et al., PAGE 31 (2023) Abstr 10522 [www.page-meeting.org/?abstract=10522] [4] https://deeppumas-docs.pumas.ai/
Reference: PAGE 33 (2025) Abstr 11554 [www.page-meeting.org/?abstract=11554]
Poster: Methodology – AI/Machine Learning