Using Full Random Effects Models (FREM) in different software
Guibi Yu2, Jurgen Langenhorst1, Joakim Nyberg1, Niclas Jonsson1
1Pharmetheus AB, 2Uppsala University
Introduction: The Full Random Effect Modelling (FREM)[1] is an innovative approach for exploring covariate effects in mixed effect models. Unlike conventional methods, it treats covariates as observations and evaluates their effects through covariances between parameters and covariates. This approach mitigates issues arising from high correlations among covariates and is also robust to missing covariate data[2]. FREM requires software capable of estimating multiple response variables simultaneously and fixing different residual errors individually. Additionally, due to its unique structure, it needs preprocessing the data for estimation and post-processing for model interpretation and diagnosis. These steps have primarily been implemented for NONMEM[2] through Perl-speaks-NONMEM (PsN) and PMXFrem. However, similar tools are not available for other software. Objectives: ?To investigate to what extent FREM models can be estimated in Pumas (v. 2.6), Monolix (v. 2024R1) and nlmixr2 (v. 3.0.2) while maintaining performance comparable to NONMEM (v. 7.5.0). ?To develop data pre-processing and results post-processing tools for each software. Methods: We simulated a dataset using covariates (weight, height, age, body mass index, creatinine clearance, race and hepatic impairment according to the National Cancer Institute Organ Dysfunction Working Group (NCIODGW)), for a Phase ? study with 473 patients with mixed non steady-state (SS) and SS dosing and sparse sampling[3]. The simulation model was a one compartment model with sequential zero and first order absorption with NCIODGW, age, and weight affecting CL and weight affecting V. A richly sampled dataset was also simulated. The FREM model considered all covariates on CL and V. SAEM was the default estimation method. However, since Pumas does not support fixing of residual error terms with EM algorithms, the comparison between Pumas and NONMEM was based on FOCE. We also focused on FOCE for the nlmixr2 runs. The tested software, dataset and estimation method combinations are given below: NONMEM: sparse data (FOCE, SAEM), rich data (FOCE) Pumas: sparse data (FOCE) Monolix: sparse data (SAEM) nlmixr2: rich data (FOCE) The results from each software were transformed into NONMEM style output and further processed using PMXFrem[4] and PMXForest[5]. Forest plots were generated to illustrate the relative effects of covariates and to compare the results. The uncertainty used in the forest plots was estimated using parametric bootstraps. Results: The FREM models were successfully estimated in NONMEM, Pumas and Monolix. However, with nlmixr2 we have so far been unable to obtain results we are confident are correct. We observed that NONMEM, Pumas and Monolix produced similar point estimates of the “true” covariates relationships, with absolute relative differences from NONMEM within 8% and 5% for Pumas and Monolix, respectively. The estimates of uncertainty were more variable across the software. Pumas and NONMEM produced nearly identical OFVs, while Monolix's OFVs were lower than that from NONMEM. Also, new PMXFrem functions transforming data and models to FREM format and converting results to NONMEM format output for all software were successfully developed. Conclusion: We showed that FREM can be applied in Pumas and Monolix, achieving similar performance compared to NONMEM. Additionally, we developed practical tools to facilitate FREM implementation for users of these software.
[1]Yngman G, Bjugård Nyberg H, Nyberg J, Jonsson EN, Karlsson MO. An introduction to the full random effects model. CPT: Pharmacometrics & Systems Pharmacology.2022;11(2):149–60. [2]Nyberg J, Jonsson EN, Karlsson MO, Häggström J. Properties of the full random-effect modeling approach with missing covariate data. Statistics in Medicine.2023 Dec 21;43(5):935–52. [3]Jonsson EN, Nyberg J. Full random effects models (FREM): A practical usage guide. CPT Pharmacometrics Syst Pharmacol.2024 Aug;13(8):1297–308. [4]https://github.com/pharmetheus/PMXFrem [5]https://github.com/pharmetheus/PMXForest