Yomna M Nassar (1,2,3*), Jonas Schick (4), Charlotte Kloft (1), Sebastian Bossert (4), Alejandro Pérez Pitarch (3*)
(1) Department of Clinical Pharmacy and Biochemistry, Institute of Pharmacy, Freie Universitaet Berlin, Germany; (2) Graduate Research Training Program PharMetrX, Berlin/Potsdam, Germany; (3) Translational Medicine & Clinical Pharmacology, Boehringer Ingelheim Pharma GmbH & Co. KG, Ingelheim am Rhein, Germany; (4) Global Biostatistics & Data Sciences, Boehringer Ingelheim Pharma GmbH & Co. KG, Biberach, Germany. *Affiliation at the time of research
Background
In oncology, the FDA requests for more rational derivation of optimal doses [1]. A first step to dose optimisation is better leveraging of data from dose-finding trials and the identification of a dose-response (D-R) signal/shape [2]. The latter supports a broader understanding of the impact of different doses on efficacy and toxicity, and identifies the optimal dose for investigation in confirmatory trials. MCP-Mod is a model-based approach for the analysis of phase II dose-finding trials that gained the FDA’s qualification [3-6]. Lately, cLRT-Mod was proposed as an extension [7]. cLRT is based on the principles of NLME modelling and aims to better leverage longitudinal phase II trial data. Thus, it has shown higher power than MCP in detecting a D-R signal [7].
Our work aimed to evaluate the performance of cLRT vs MCP in the oncology setting by simulation-based explorations, comparing the cLRT step when using continuous sum of longest diameter (SLD) data to the MCP step when using the best SLD change from baseline as endpoint, under different D-R shapes, drug effects and study design variables.
Methods
Data were simulated using a previous model [8] which described the time course of SLD as a function of drug dose, given a pre-specified D-R shape (data-generating model, DGM). The model also considered disease progression, and probability of dropouts and deaths for a realistic scenario. A placebo-controlled, parallel group, dose-finding phase II study with balanced randomisation was simulated. Observations were scheduled every 7 days for 4 months. The simulation scenarios were based on:
- D-R shape: flat, linear, loglinear, Emax, sigmoidal
- Drug effect strength: strong, weak
- Study design variables
- Total number of patients: 24, 36
- Number of dose levels: 3, 4, 6
A set of candidate models (CM), to test against the simulated data, described the change in SLD; other events were considered as dropouts from the study (simplified version of DGM). CM included the same shapes as DGM; a flat shape (no D-R) was the reference.
The MCP analysis started with a pre-selected set of CM for which multiple contrast tests established evidence of a D-R, based on the observed study data [9]. The cLRT analysis estimated parameters by maximum likelihood (here: Laplace). Using the LRT between the best-fitting CM and the no D-R (flat) model—after fitting each CM to the data—the presence of a D-R was established by comparing the calculated LRT against the critical value obtained using the same procedure but under the null hypothesis (Ho), using placebo arm data [7]. Evaluation of cLRT and MCP was based on power, type I error, selection of the best-fitting CM, and identification of the true shape of the D-R DGM.
For cLRT, NONMEM® and PsN were used for simulations and (re-)estimations within an R framework. For each scenario, N=100 datasets were simulated and for each simulated dataset, the critical value was approximated from data generated under the Ho (M=100 simulations). The MCP step was performed with the DoseFinding R package.
Results
Generally, cLRT power was higher than MCP (89.8% vs 27.0%). Power was also higher for scenarios with a strong drug effect (cLRT: 96.9–100%, MCP: 17–52%) and a bigger number of patients (cLRTn=36>73.7%, cLRTn=24>66.3%; MCPn=36>12%, MCPn=24>10%). Increased power with fewer dose levels and more patients per dose level was observed with MCP but not cLRT. For all scenarios, type I error was inflated for cLRT (mean: 13%) but was controlled for MCP (<4%).
There was a tendency towards selection of the simplest (linear, loglinear) D-R shapes. Identification of the true shape of the D-R model was independent of the study design. Complex models (Emax, sigmoidal) had a weaker chance than simple models to be successfully detected.
Conclusion
Power was higher for cLRT vs MCP; however, in contrast to previous simulations [7], cLRT showed an inflated type I error. This could have been due to our small sample size, compared with [7, 10], and/or placebo model misspecifications and selection bias which may have impacted the reference distribution used to derive the critical value under the Ho [7, 10]. At a fixed sample size, cLRT power was not sensitive to different study designs. In cLRT, the need for a placebo group can be obtained, as an example, from data from very low dose levels gathered at earlier development stages and assuming irrelevant exposures. Investigation of the robustness of cLRT to model complexities and study variables, e.g. sample size, is needed to justify its computational expenses for simulations.
References
[1] Food and Drug Administration. Project Optimus: Reforming the dose optimization and dose selection paradigm in oncology. https://www.fda.gov/about-fda/oncology-center-excellence/project-optimus (Accessed on: 08.12.2023).
[2] European Medicines Agency. ICH Topic E 4: Dose Response Information to Support Drug Registration. https://www.ema.europa.eu/en/documents/scientific-guideline/ich-e-4-dose-response-information-support-drug-registration-step-5_en.pdf . (Accessed on: 08.12.2023).
[3] Food and Drug Administration. Drug Development Tools: Fit-for-Purpose Initiative. https://www.fda.gov/drugs/development-approval-process-drugs/drug-development-tools-fit-purpose-initiative . (Accessed on: 12.12.2023)
[4] S. S. Shord et al. CPT: Pharmacometrics & Systems Pharmacology 12: 1573 (2023).
[5] Food and Drug Administration. Statistical Review and Evaluation: Qualification of Statistical Approach. https://www.fda.gov/media/77169/download . (Accessed on: 08.12.2023).
[6] Food and Drug Administration. FDA Qualification of MCP-Mod Method. https://www.fda.gov/media/99313/download . (Accessed on: 07.12.2023).
[7] S. Buatois et al. Statistics in Medicine 40: 2435–2451 (2021).
[8] S. M. Krishnan et al. 30th PAGE meeting, abstract 10071. (2022). www.page-meeting.org/?abstract=10071
[9] F. Bretz et al. Biometrics 61: 738–748 (2005).
[10] E. Chasseloup et al. Pharmaceutics 15: 460 (2023).
Reference: PAGE 32 (2024) Abstr 10769 [www.page-meeting.org/?abstract=10769]
Poster: Methodology - Study Design