Evaluation of the robustness of methods for model-based bioequivalence analysis for biosimilars: Clenoliximab as a case study
Mahmoud Ali Afifi (1,2), Oluwasegun Eniayewu (1,3), Verrah Akinyi Otiende (1,4), (Colin) Pillai Goonaseelan (1), Samer Mouksassi (2), Arne Ring (5,6)
(1) Africa Pharmacometrics Training Program (APT), Pharmacometrics Africa NPC, K45 Old Main Building, Groote Schuur Hospital, Cape Town, South Africa. (2) Integrated Drug Development, Certara, Princeton, NJ, United States. (3) Department of Pharmaceutical and Medicinal Chemistry, Faculty of Pharmaceutical Sciences, University of Ilorin, Nigeria. (4) Faculty of Biological and Physical Sciences, Tom Mboya University, Kenya. (5) University of the Free State, Dept. of Mathematical Statistics and Actuarial Science, Bloemfontein, South Africa. (6) Hexal/Sandoz, Biosimilars Biostatistics, Holzkirchen, Germany.
Objectives: Regulatory approval of biosimilars includes comparison of the pharmacokinetic (PK) profiles of the biosimilar and the innovator reference product [1]. The classical non-compartmental analysis (NCA) approach mandates rich sampling which increases the overall costs of biosimilar clinical development, while assessing biosimilarity using sparse sampling designs is known to be difficult and frequently inaccurate [2]. Previous reports indicate that the model-based methods have sufficient power to evaluate bioequivalence [2]. However, it is unclear how robust these approaches are against sparse sampling and model misspecification. This is a simulation-based study to assess the robustness of model-based bioequivalence (MBBE) method against sparse sampling designs and model misspecification using clenoliximab as a case study.
Methods: To evaluate the robustness, we chose a published population PK model for clenoliximab [3], a monoclonal antibody with the characteristics of a long half-life and non-linear elimination. A rich dataset was simulated using this model (i.e., 1000 participants for test and reference formulation and sampling each hour). The rich dataset was subsetted into 500 bootstrap datasets (sampling with replacement) based on the following scenarios: different numbers of individuals (N=15 till N to reach 90 % power, adding 5 individuals by formulation at each step),different formulations (Frel = 1.05, 1.1, and 1.25), and different sampling scenarios (n = rich, moderate, and sparse) with timepoints optimized using popED (R package).
The simulated data of all scenarios were fitted using the true model (clenoliximab model) and another two misspecified models that had either a misspecified linear elimination or a misspecified number of compartments (1CMT instead of 2CMT). PK metrics (Cmax and AUC0-inf) were computed for each scenario, and bioequivalence was assessed between the test and reference products. The standard thresholds (between 0.8 and 1.25) were used for establishing bioequivalence. MBBE was assessed by comparing the power of achieving bioequivalence between all models, and NCA over all scenarios. The power is defined as the percentage of bootstrap replicates that achieved bioequivalence over each scenario.
Results: PopED optimal sampling time points (n = 12) for the rich design were at (0, 1, 2, 96, 180, 200, 220, 330, 480, 1500, 2000, and 2016) hours and (n = 9) for the moderate design were at (0, 1, 2, 200, 220, 480, 2000, 1500, and 2016) hours. The sparse design was split into two groups, each with five sampling points: G1 = (0, 1, 2, 2000, and 2016) hours and G2 = (0, 2, 96, 480, and 2016) hours. The power curves of all models were higher than the traditional approach. Compared to the traditional NCA, the model-based approach (true model) requires about half the number of subjects to achieve the targeted power (80– 95%), with controlled Type I alpha error in both approaches. While in misspecified models, we noted inflation of the Type-I error. In MBBE, the power curve for the sparse designs was approximately the same as that for rich designs, while NCA was not applicable in sparse sampling designs.
Conclusions: The MBBE method showed good performance with sparse sampling designs and a reduced number of subjects. We noted inflation of Type-I error with model misspecification that might be mitigated by model averaging or careful model selection. We conclude that MBBE is a promising tool for bioequivalence assessment especially for biosimilarity assessment where data or a model from the innovator product would be available or more generally in situations where NCA is not feasible.
References:
- Kirchhoff, Carol F., et al. “Biosimilars: Key Regulatory Considerations and Similarity Assessment Tools.” Biotechnology and Bioengineering, vol. 114, no. 12, Dec. 2017, pp. 2696–705. PubMed, https://doi.org/10.1002/bit.26438.
- Hughes, Jim H., et al. “Comparison of Non-Compartmental and Mixed Effect Modelling Methods for Establishing Bioequivalence for the Case of Two Compartment Kinetics and Censored Concentrations.” Journal of Pharmacokinetics and Pharmacodynamics, vol. 44, no. 3, June 2017, pp. 233–44. PubMed, https://doi.org/10.1007/s10928-017-9511-7.
- Mould, D. R., et al. “A Population Pharmacokinetic-Pharmacodynamic Analysis of Single Doses of Clenoliximab in Patients with Rheumatoid Arthritis.” Clinical Pharmacology and Therapeutics, vol. 66, no. 3, Sept. 1999, pp. 246–57. PubMed, https://doi.org/10.1016/S0009-9236(99)70032-9.
