Keqi Shi 1,2, Jens Borghardt 2, Niklas Kroemer 2, Sebastian Wicha 1
1 University of Hamburg (Hamburg, Germany), 2 Global Research DMPK, Global Drug Discovery Sciences, Boehringer Ingelheim Pharma GmbH & Co. KG (Biberach, Germany)
Introduction: Preclinical pharmacokinetic (PK) studies for new drug candidates are typically designed empirically and in a compound agnostic manner. Optimal experimental Design (OD) is well established within later stages of drug development [1]. However, in preclinical settings, the lack of robust prior knowledge remains a major limitation for informing optimal design strategies guiding PK study planning and potentially improving the quality and precision of PK parameter estimation.
Recent advances in machine learning (ML)-based prediction of preclinical PK parameters offer a potential way to overcome this limitation [2]. Therefore, the aim of this study was to evaluate the potential benefit of ML-informed OD on the identifiability of key PK parameters (CL&Vss) and recovery of PK profile
Methods: This retrospective analysis was based on intravenous PK data of about 2000 small‑molecule compounds alone and in cocktail-studies in rats. For each compound, the original study design, estimated compartmental PK parameters (used as “true” parameters), and corresponding ML‑predicted parameters were available [2]. The ML-predicted parameters were generated with model trained on data available before respective compound registration for PK for an external dataset that was not part of training dataset at the time.
The optimized experimental designs were generated by optimization of sampling times based on the original study layout (i.e. number of timepoints, number of animals) using the package PopED [3] in R. Two optimality criteria, D-optimality and ED-optimality were tested. It resulted in a total of five design groups:
1. Standard design (or base design)
2. D-optimized design with ML-predicted parameters
3. ED-optimized design with ML-predicted parameters
4. D-optimized design with true parameters
5. ED-optimized design with true parameters
All design approaches were evaluated using Stochastic Simulation and re-Estimation (SSE). Design performance was assessed based on:
1. Key parameter recovery, evaluating accuracy and precision of clearance (CL) and steady‑state volume of distribution (Vss). Recovery was quantified as the proportion of cases whose 95% confidence intervals for CL and Vss fell within predefined fold error thresholds.
2. PK profile recovery, assessing overall PK shape accuracy. This was quantified using the Curve Similarity Index (CSI), defined as the time‑integrated absolute difference between the true and re‑estimated concentration-time profiles over the first 24 hours. The performance of design approach is quantified by proportion of cases whose CSI not exceeding predefined fold error thresholds.
Robustness of results was assessed via two sensitivity analyses:
1.Residual error magnitude used during design optimization and SSE. Proportional and additive error levels ranged from 0.5× typical residual variability (15% proportional, 0.5 nM additive) to about 4× typical residual variability (121.8% proportional, 4.59 nM additive).
2. Sampling number, by altering the standard 8 sample IV design to include +1 or +2 samples, or -1 to -4 samples.
Results: Generally, the optimized designs using true parameters under D-optimality outperformed the standard design for both 1‑ and 2‑compartment models (improvements in proportion of qualified design up to ~20% and ~10%, respectively) and was considered as a benchmark.
When prior knowledge contained uncertainty (i.e., ML‑predicted parameters), ED‑optimality produced more robust designs than D‑optimality. ML‑informed optimal designs improved recovery performance by up to ~10% for 1‑compartment models but did not improve designs for 2‑compartment models. PK‑profile recovery via CSI supported these relative performance patterns.
Sensitivity analyses confirmed the results across varying numbers of sampling points and across different residual error settings. Notably, the benefit of optimal design became more pronounced when the number of samples was reduced compared to the standard design.
Conclusions: Given current ML prediction accuracy and already‑adequate standard preclinical designs, ML‑informed OD yields only marginal improvements (up to ~10%) for 1‑compartment models and offers no benefit for 2‑compartment models. These gains may not offset the additional logistical complexity of implementing ML‑informed OD in preclinical routine testing. The predominant limitation remains the precision and accuracy of ML‑predicted PK parameters. Further work is needed to define the gap, e.g., the level of ML improvement and technique required for ML‑informed OD to provide consistent and practical benefits.
References:
Acknowledgements:
Moritz Walter, Miha Skalic, Lina Humbeck, Christofer Tautermann, Cornelius Truenkle for their contribution on the ML predicted parameters and in vivo study insights.
Reference:
1. Mentré, F. et al. Current Use and Developments Needed for Optimal Design in Pharmacometrics: A Study Performed Among DDMoRe’s European Federation of Pharmaceutical Industries and Associations Members. CPT: Pharmacomet. Syst. Pharmacol. 2, 1–2 (2013).
2. Walter, M. et al. Predicting Pharmacokinetics in Rats Using Machine Learning: A Comparative Study Between Empirical, Compartmental, and PBPK ‐Based Approaches. Clinical and Translational Science 18, e70150 (2025).
3. Nyberg, J. et al. PopED: An extended, parallelized, nonlinear mixed effects models optimal design tool. Computer Methods and Programs in Biomedicine 108, 789–805 (2012).
Reference: PAGE 34 (2026) Abstr 12174 [www.page-meeting.org/?abstract=12174]
Poster: Methodology - Study Design