(M-073) A Scalable Cloud-Based QSP Modeling System for Virtual Patient Generation

David Hagen, PhD – Director of Software Engineering, Certara; Maciej Swat, PhD – Principal Scientist, Certara; Richard Matthews, PhD – Technical Architect, Certara; Andrzej Kierzek, PhD – Vice President, Certara; Joshua Apgar, PhD – VP, Global Head of ABM Scientific Affairs, Certara

Objectives:
This study focuses on enhancing the efficiency of generating virtual patient populations, a critical component in various Quantitative Systems Pharmacology (QSP) modeling workflows such as virtual clinical trial simulation, digital twin generation, characterizing biological variability, and biomarker generation. The primary objectives are to streamline the generation process, reduce computational costs, and improve the quality of the generated population.

Methods:
We developed a comprehensive cloud-based QSP modeling system that supports the entire virtual patient population workflow. This includes population parameter estimation, virtual population generation, and virtual population simulation. Our system integrates multiple algorithms for creating virtual patients, notably the Allen et. al. algorithm and the No-U-Turn (NUTS) algorithm, a Monte-Carlo-like sampling algorithm (1,2,3). In this work, we compare the populations generated, and computational performance of different algorithms across a representative sample of QSP and minimal-PBPK style models. We explore several changes to the proposal distribution produced in the Allen et. al. algorithm.

Results:
The implementation of the system demonstrated significant performance improvements. Specifically, the simulation times on a single thread and single worker were 30-40% faster than those of compiled SimBiology models for a selection of QSP and minimal-PBPK models. The cloud-based system can scale up to over 1000s of worker CPUs, leveraging cost-effective computational resources to rapidly generate virtual populations. Our approach achieved substantial enhancements in the speed and efficiency of generating virtual patient populations.

Conclusions:
The newly developed QSP modeling system markedly improves the generation and simulation of virtual patient populations, affirming the feasibility of scaling the process while managing computational costs effectively. Further exploration of modifications to the Allen et. al. algorithm showed potential for incremental improvements in population quality, computational resource utilization, and overall performance. We compare algorithmic approaches for their impact on the quality of population generated, efficient use of compute resources, and wall-clock performance.

Citations: [1] Abril-Pla, Oriol, et al. "PyMC: a modern, and comprehensive probabilistic programming framework in Python." PeerJ Computer Science 9 (2023): e1516.

[2] Allen, R. J., Theodore R. Rieger, and Cynthia J. Musante. "Efficient generation and selection of virtual populations in quantitative systems pharmacology models." CPT: pharmacometrics & systems pharmacology 5.3 (2016): 140-146.

[3] Rieger, Theodore R., et al. "Improving the generation and selection of virtual populations in quantitative systems pharmacology models." Progress in biophysics and molecular biology 139 (2018): 15-22.