K. Lindauer(1), A. Ahmad(1), F. Khalil(1), L. Kessler(2), A. Schwarz(2), L. von Wedel(1), S. Engelen(1)
(1)Grünenthal GmbH, Aachen, Germany, (2)Five1 GmbH, Heidelberg, Germany
Introduction
First in human (FIH) clinical trials aim to study the human pharmacology, tolerability and safety of an investigational medicinal product (IMP). The respective design of a FIH trial could vary, but comprises commonly a single ascending dose (SAD) and a multiple ascending dose (MAD). The highest priority of the FIH trial conduct is the safety and well-being of the recruited subjects [1]. Therefore dose decisions are carefully taken to assess and evaluate the potential risk for the subjects to be treated. The selection of the starting dose of the FIH clinical trial is supported by pre-clinical data and estimation of the exposure in human by state-of-the-art modelling. Although the further dose escalation steps are pre-defined, clinical data from previous cohorts have to be taken into account for the dose decision of the following cohort [1]. In general these decisions are informed by predictions obtained from pharmacokinetic (PK) models constructed using all available PK information.
PK models based on deep learning approach using neural ordinary differential equations (ODE) have been successfully developed to predict untested treatment regimens [2] and learning an unknown absorption process, while simultaneously estimating other PK parameters such as drug distribution and elimination [3]. This technique does not require a sophisticated mechanistic model development process, but learns from the available data to predict unknown treatment regimens.
Objectives
- Develop neural-ODE framework to support dose escalation decision in FIH trials
- Evaluate minimal required PK information for model development to accurately predict PK profile of next doses
Methods
A population PK model (zero and first order absorption, 2 compartment and first order elimination) was used to simulate a SAD dataset and to test the performance of the neural-ODE models. The dataset consists of 9 cohorts with different doses (5, 10, 20, 30, 45, 60, 75, 90 & 110mg) with 10 subjects in each cohort. The tmax is at about 1 to 2 h after dosing and t1/2 = 2.9h. Therefore the following PK sampling scheme was considered: 0.3, 0.5, 1.0, 1.5, 2.0, 4.0, 6.0, 8.0, 12.0, 14.0, 24.0 & 36.0h.
Two neural network approaches (model a[3] and b[2]) were used. The settings are loosely based on the publications. The hyperparameter settings were the following: lr=0.01, epochs=350 (model a) and lr=0.001, epochs=75(model b). No hyperparameter optimization was performed. Similar to the SAD trial design, we trained each model using all lower dose levels and predicted the population (model a) and individual (model b) response of one or two higher dose levels. We iterated the process from the lowest dose (5mg) to the second highest dose (90mg). The prediction quality was evaluated by goodness-of-fit estimates such as graphical display and r2 score, which indicates the degree of confidence of the predictions.
Results
The trained network has to predict the respective concentration profiles of unknown (higher) doses. The quantitative degree of confidence of the predicted population mean (model a) and individual (model b) profiles are shown in table 1. Regardless of how many dose levels were considered for the training dataset, the r2 score value was in the range between 0.64 and 0.71 (model a). In using the first dose cohort (5mg) only, the prediction of the population average response of the next two higher dose levels (10 and 20mg) is reasonably good (r2 = 0.69). For model b confidence in the predicted profiles is higher (0.84 – 0.95), if more data is used for training.
Table 1: Quantitative assessment of the degree of goodness-of-fit of the validation dataset.
|
Validation dataset |
110mg dose cohort |
90 and 110mg dose cohort |
75 and 90mg dose cohort |
60 and 75mg dose cohort |
45 and 60mg dose cohort |
30 and 45mg dose cohort |
20 and 30mg dose cohort |
10 and 20mg dose cohort |
|
r2 score of model a (population mean) |
0.64 |
0.63 |
0.63 |
0.68 |
0.71 |
0.71 |
0.67 |
0.69 |
|
r2 score of model b (individual prediction) |
0.95 |
0.84 |
0.86 |
0.92 |
0.84 |
0.64 |
0.82 |
0.45 |
|
Number of subjects in training dataset |
80 |
70 |
60 |
50 |
40 |
30 |
20 |
10 |
|
Number of subject in validation dataset |
10 |
20 |
20 |
20 |
20 |
20 |
20 |
20 |
Conclusions
The neural-ODE framework seems to be a promising tool to estimate the exposure of higher dose levels and thereby support dose escalation decisions during safety review meetings of FIH trials. However this approach has yet to be explored with more complex PK profiles and network optimization procedures.
References
[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 16. Feb 2024)
[2] 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.
[3] Valderrama D, Ponce-Bobadilla AV, Mensing S, Fröhlich H, Stodtmann S. Integrating machine learning with pharmacokinetic models: Benefits of scientific machine learning in adding neural networks components to existing PK models, CPT Pharmacometrics Syst Pharmacol 2024
Reference: PAGE 32 (2024) Abstr 10810 [www.page-meeting.org/?abstract=10810]
Poster: Methodology – AI/Machine Learning