ADaMO: End-to-end automation of Pharmacometric modelling in drug development, from dataset building to output generation
Joćo A. Abrantes (1), Kevin Smart (1), Claire Petry (1), Felix Jaminion (1), Guillaume Schmitt (1), Nicolas Frey (1), Clarisse Chavanne (1), Simon Buatois (1)
(1) Roche Pharma Research and Early Development, Pharmaceutical Sciences, Roche Innovation Center, Basel, Switzerland
Objectives: The Automated Datasets, Models and Outputs (ADaMO) initiative seeks to automate and streamline pharmacometric processes used for informing decision-making in drug development. The manual creation of datasets for modelling, development of population models and respective reports are complex, arduous and time-consuming processes. In reality, the time spent in these activities tends to be longer than the time invested in using the models to inform the clinical development of compounds.
Methods: The ADaMO initiative has been developed in a validated computational environment [1, 2] where data, models and outputs are generated and stored. It consists of three pillars that cover the end-to-end Pharmacometric modelling process: pillar I, automated dataset request and creation; pillar II, automated model building, and; pillar III, automated output creation.
Results:
Pillar I: automated dataset request and creation. A final dataset for modelling is produced using scripts generated automatically. The process starts with the user completing a data request form in a user-friendly platform, indicating the variables to be included in the dataset (e.g. demographic variables, subjects and dependent variable with definition of units). After reviewing the request, a data scientist triggers the automatic generation of SAS scripts in a validated GCP compliant environment. If needed, the data scientist can update those scripts based on the available data; then the scripts are executed to create the requested dataset. Finally, the dataset undergoes QC and the user is automatically notified that the dataset has been created. Therefore, automation simplifies the process of dataset request and creation, providing a gain of time as the scripts generated are reusable, which will speed-up future requests on the same studies.
Pillar II: automated model building. A pharmacometric model is developed based on the dataset generated in pillar I. A heuristic algorithm is being developed to create an initial population PK model and optimise its components (structural, inter-individual variability, residual error and covariate sub-models) with the goal of reaching a final PK model developed in NONMEM [3]. For this purpose, open-source packages have been built to generate the initial NONMEM control stream (R package assemblerr), manipulate the model, run the model, process the results and select the best models (Python package Pharmpy, and R package PharmR) [4,5]. The optimization of each model sub-component can be done on a stand-alone basis, but the tool is most powerful when used in a fully automated setting.
Pillar III: automated output creation. A report is produced for the final model produced in Pillar II using NONMEM. An R library and a series of Rmarkdown scripts were developed to generate a modelling report using model metadata and a few key pieces of information inputted by the user. The final report (in word or html format), contains the figures and tables necessary to describe the input dataset and model, including patient demographic and treatment assignment tables, summary of covariates, model parameter table, goodness-of-fit plots and visual predictive checks.
Conclusions: A solution to automate the end-to-end Pharmacometric modelling has been developed. The current scope of ADaMO includes classical PK models for a single entity (pillar II), but pillars I and III already go beyond this scope, supporting advanced PK and PKPD analyses. While the entire benefit of the solution will only be gained through the integration of all three pillars in one computational environment, it is possible to use each component on its own depending on the stage of the project. Automation boosts quality, reproducibility and frees time for pharmacometricians to use the model outcomes to influence clinical trials, address more complex modelling tasks and increase the number of supported projects.
Acknowledgement: This work was supported by F. Hoffmann-La Roche Ltd., Basel, Switzerland. The authors would like to acknowledge the contributions of the following individuals: Wolfgang Schwarzenbrunner and Christian Flandorfer from Scinteco (pillar I); Prof. Mats Karlsson, Prof. Andrew Hooker, Alzahra Hamdan, Osama Qutishat, Rikard Nordgren, Sebastian Ueckert, Shijun Wang, Simon J. Carter, Stella Belin, Tianwu Yang, Xiaomei Chen and Zhe Huang from the Department of Pharmacy, Uppsala University, Sweden (pillar II), and; Michael Hackl from Scinteco (pillar III). In addition, the authors would like to thank Henry Hofmann for project management, and Emilie Schindler, Julie Cabanel, Sébastien Jolivet, Sylvie Retout, Valérie Cosson, and Vincent Buchheit for their assistance with testing.
References:
[1] IMPROVE – MODELING AND SIMULATION PLATFORM. PAGE 25 (2016) Abstr 6073 [www.page-meeting.org/?abstract=6073]
[2] Usage of a validated environment to support modeling and simulation activities. Buchheit et al. ACoP11. 2020.
[3] Development of a tool for fully automatic model development (AMD). Chen et al. PAGE 2022 (submitted).
[4] Pharmpy and assemblerr - Two novel tools to simplify the model building process in NONMEM. PAGE 29 (2021) Abstr 9656 [www.page-meeting.org/?abstract=9656].
[5] Pharmpy: a versatile open-source library for pharmacometrics. Nordgren et al. PAGE 2022 (submitted).
