Han Zheng 1, Chen Wang 2, Jing Nie 2, Zhi Zhang 2, Lin Hu 2
1 Shanghai Bioguider Medical Technology Co., Ltd (Shanghai, China), 2 Abbisko Therapeutics (Shanghai, China)
Note: H. Zheng and C. Wang contributed equally to this research.
Introduction:
Maspectra [1] is a software platform for modeling and simulation in pharmacometrics. Its Python-based modeling language facilitates smooth integration with widely used scientific computing libraries and modern artificial intelligence (AI) frameworks, including PyTorch [2] and TensorFlow [3]. Maspectra supports established estimation methods for nonlinear mixed-effects modeling, including FOCE-I, LAPLCE, and SAEM. For convenience and computational efficiency, it provides built-in closed-form compartmental models while also accommodating user-specified ordinary differential equation (ODE) models using solvers such as DVERK and LSODA. While a previous simulation study [4] demonstrated high consistency in parameter estimates and standard errors between Maspectra and NONMEM across various model structures and data scenarios, a direct comparison between the two software using actual clinical data has not yet been conducted. In this study, we evaluated the performance of Maspectra and NONMEM using a population pharmacokinetic model for the FGFR4 inhibitor ABSK-011, currently under clinical development for advanced hepatocellular carcinoma.
Objectives:
Compare estimation performance of Maspectra and NONMEM based on a semi-mechanistic population pharmacokinetic model fitted to a clinical dataset.
Methods:
The population pharmacokinetic model for ABSK-011 was independently fitted in Maspectra and NONMEM using a dataset comprising 150 patients with hepatocellular carcinoma. In both software platforms, parameters were estimated using the FOCE-I method, employing identical model structures and initial parameter values. The pharmacokinetics of ABSK-011 after oral administration were described using a two-compartment model characterized by sequential zero- and first-order absorption, alongside nonlinear elimination at steady state. The comparative analysis evaluated fixed-effect parameter estimates, inter-individual variability, relative standard errors (RSEs), shrinkage, objective function values (OFVs), and goodness-of-fit diagnostics. All analyses were conducted using Maspectra v2.8.2 and NONMEM 7.5.
Results:
The parameter estimates obtained from Maspectra were highly consistent with those from NONMEM for the ABSK-011 population pharmacokinetic model. This consistency extended across fixed effects, inter-individual variability, residual error parameters, and shrinkage estimates. The relative standard errors were comparable between the two platforms, indicating similar uncertainty estimation. Furthermore, the objective function values were identical. Standard goodness-of-fit diagnostics showed similar patterns across both platforms, supporting the comparability of parameter estimation and model diagnostics.
Conclusion:
We evaluated the performance of Maspectra using a semi-mechanistic population pharmacokinetic model fitted to an actual clinical dataset. Despite the structural complexity of the model, Maspectra yielded parameter estimates and model diagnostics highly consistent with those obtained from NONMEM. This demonstrates its reliability as a viable alternative for nonlinear mixed-effects modeling. By incorporating data processing, model fitting, diagnostic tools, and simulation into a single platform, Maspectra establishes a unified workflow for pharmacometric analyses. This not only streamlines the modeling process but also makes it well-suited for future applications involving large language models (LLMs) and autonomous AI agents. Furthermore, its native Python foundation naturally aligns with major machine learning libraries. This makes it easier to combine traditional pharmacometrics with advanced AI methodologies, paving the way for hybrid models that can better predict pharmacokinetics and drug responses.
References:
[1] https://www.maspectra.com
[2] https://pytorch.org/
[3] https://www.tensorflow.org
[4] Xu et al. PAGE 31 (2023) Abstr 10676 [www.page-meeting.org/?abstract=10676]
Reference: PAGE 34 (2026) Abstr 12072 [www.page-meeting.org/?abstract=12072]
Poster: Methodology - Other topics