Klaus Lindauer1, Linda Kraske2, Ashar Ahmad1, Feras Khalil1, Anja Tündermann1, Lars von Wedel1, Edna Venneker1
1Grünenthal GmbH, 2Five1 GmbH
Introduction FIH clinical trials aim to study human pharmacology, tolerability and safety of an investigational medicinal product (IMP). The respective design of a FIH trial varies but comprises commonly a single ascending dose (SAD) part. The highest priority in conducting a FIH trial is ensuring the safety and well-being of the recruited subjects [1]. Therefore, dose decisions are made to evaluate the potential risk to subjects receiving treatment. The selection of the starting dose is supported by pre-clinical data and the estimation of human exposure through modelling. When deciding on the dose for the following cohort, clinical data from previous cohorts must be considered [1]. We have demonstrated that a deep learning approach using neural-ODE is suitable for predicting untested exposure in FIH trials [2]. Objectives Application of the neural-ODE framework to predict unknown exposure in FIH trial [3] and estimate the prediction error (PE) of maximal exposure (Cmax) and AUCt=0-24h Method The SAD dataset of the FIH trial [3] consists up to date of 5 cohorts with different doses (starting dose and applying the following dose escalation factors: 3, 2.5, 2.7, 2.0 and 1.5), each including 6 subjects on IMP. Pharmacokinetic (PK) samples were taken at the following time points: 0, 0.5, 1, 1.5, 2.5, 4, 5, 8, 10, 12, 14, 24, 28, 32, 38, 48 and 72 h post-dose. The neural-ODE model [4] was adapted for SAD data, optimized in terms of hyperparameters and extended by a transfer learning technique for the prediction of lower doses. The hyperparameter settings were optimized to a learning rate (lr) of 4.5e-4, a L2 regularization (l2) of 1e-6 and 100 epochs for the base model and a l2 of 1e-7 for a second training via transfer learning on the trial data. For the base model, simulated data derived from a preclinical allometric popPK model [5] was used. Following the SAD trial design, we retrained the base model with the encoder layers frozen using all lower dose levels to predict the individual response at the next planned higher dose. The model incorporated demographic characteristics as input features. To ensure the integrity of the study, subject characteristics were provided without affiliation and therefore assigned using the Cmax-bodyweight correlation. The PE [%] (100 abs(predicted – observed)/observed) was calculated for Cmax and AUCt=0-24h. Results The trained network predicted the respective concentration profiles of the next unknown dose cohort. The PE for Cmax and AUCt=0-24h is summarized in Table 1. For Cmax, the PE ranged from 4 to 50 % (mean) and 4 to 60 % (max), with an average mean PE of 27 % and an average maximum PE of 37 %. The PE for AUCt=0-24h ranged from 4 to 64 % (mean) and 18 to 44 % (max), with an average mean PE of 29 % and an average maximum PE of 28 %. With increasing dose levels, the neural-ODE PE of Cmax tended to increase, whereas the AUCt=0-24h PE showed a decreasing trend. Cohort;2;3;4;5; mean,max;mean,max;mean,max;mean,max; Cmax PE [%];4.2, 4.1;11.9, 28.2;42.7, 56.9;49.8, 59.3; AUCt=0-24h PE [%];64.1, 44.2;10.3, 24.3; 3.8, 18.3; 37.1, 27.2; No of subjects in training data, clinical trial data;6; 12; 18; 24; simulated data;110;110; 110; 110; No of subject in test data;6;6;6;6; Table 1: Estimated PE [%] for Cmax and AUCt=0-24h in the test dataset. The predicted values were derived from the planned higher-dose test data, using demographic characteristics from lower-dose cohorts before treatment began. The observed values represent the actual FIH data collected after treatment completion. Conclusions The neural-ODE framework demonstrates strong potential for predicting exposure at higher dose levels. This deep learning approach was validated in the FIH trial as an exploratory tool for forecasting individual PK profiles based on preclinical data, supporting dose escalation decisions in real time. As additional trial data became available, model performance improved, reducing bias in the simulated predictions. However, early-stage predictions were highly dependent on simulations and data augmentation due to limited clinical trial data. The increased PE of cohort 5 could be caused by differences in the intra-individual variability within each cohort. The results indicate that neural-ODE modeling can effectively capture dose-exposure relationships. Neural-ODE modeling proved to be a valuable, data-driven and time-efficient complement to popPK modeling in FIM trials, offering a dynamic approach to dose prediction with generally increasing accuracy as more data accumulates.
[1] Guideline on strategies to identify and mitigate risks for first-in-human and early clinical trials with investigational medicinal products (europa.eu). https://www.ema.europa.eu/en/documents/scientific-guideline/guideline-strategies-identify-and-mitigate-risks-first-human-and-early-clinical-trials-investigational-medicinal-products-revision-1_en.pdf last accessed 10. Mar. 2025 [2] K. Lindauer, A. Ahmad, F. Khalil, L. Kraske et al. Application of a neural-ODE framework to support dose escalation decision during safety review meetings in first-in-human clinical trial. PAGE 2024 last accessed 10. Mar. 2025 [3] A randomized, single-center, double-blind, placebo-controlled, first-in-human trial with single and multiple ascending doses to determine safety, tolerability, and pharmacokinetics of GRT7040 in healthy volunteers https://euclinicaltrials.eu/search-for-clinical-trials/?lang=en&EUCT=2024-512510-17-00 last accessed 10. Mar. 2025 [4] Lu J, Deng K, Zhang X, Liu G, Guan Y. Neural-ODE for pharmacokinetics modeling and its advantage to alternative machine learning models in predicting new dosing regimens. iScience. 2021;24:102804. [5] K. Lindauer, K. Pasikanti, F. Khalil, M. Gautrois, S. Engelen, B. Manning Allometric Scaling for Efficacious First-in-Human Dose Prediction of Pharmacologically-Active Parent and Metabolite Based on Rat, Dog and Monkey PK data. PAGE 31 (2023) Abstr 10540 [www.page-meeting.org/?abstract=10540] last accessed 10. Mar.. 2025
Reference: PAGE 33 (2025) Abstr 11543 [www.page-meeting.org/?abstract=11543]
Poster: Methodology – AI/Machine Learning