Semi mechanistic joint modeling of tumor dynamics and PFS in advanced breast cancer: leveraging data from early amcenestrant phase I-II trials
Marc Cerou (1), Hoai-Thu Thai (1), Laure Deyme (2), Sophie Fliscounakis-Huynh (3), Sylvaine Cartot-Cotton (4), Christine Veyrat-Follet (1)
(1) Translational Disease Modelling Oncology, Sanofi, 91380 Chilly-Mazarin, France. (2) Modeling and Simulation, Sanofi, 34184 Montpellier, France (3) Translational Disease Modelling Oncology, Sanofi, on behalf IT&M Stats, 91380 Chilly-Mazarin, France (4) Pharmacokinetics Dynamics and Metabolism, Sanofi, 91380 Chilly-Mazarin, France
Introduction: Amcenestrant is an orally bioavailable selective estrogen receptor (ER) degrader developed for the treatment of ER+/HER2- advanced breast cancer. Despite a favorable safety profile and promising activity as monotherapy and in combination with palbociclib in the phase 1/2 AMEERA-1 (NCT03284957) and AMEERA-2 (NCT03816839) studies [1,2], amcenestrant failed to demonstrate progression-free survival (PFS) superiority over physician's choice endocrine monotherapy in the pivotal phase 2 AMEERA-3 (NCT04059484) study [3]. A TGI model [4] was initially constructed using tumor size (TS) data from 75 Japanese and non-Japanese patients from the AMEERA 1-2 studies who were treated with amcenestrant monotherapy. The model accounted for exposure to amcenestrant at doses ranging from 20 to 600 mg daily and a resistance process.
Objectives: The aims of this analysis were (1) to develop a joint model of TS and PFS using phase 1/2 data to further evaluate the exposure-response relationship of amcenestrant, and to identify baseline covariates influencing both TS & PFS, (2) to evaluate the model predictive performance based on data from the AMEERA-3 trial and (3) to refine the joint model using the pool data of AMEERA 1-2-3.
Methods: A learning dataset was defined based on the 75 patients selected in the AM 1-2 studies. A validation dataset was defined from the AM-3 trial, consisting of 98 patients with measurable target lesions. A calibration dataset was defined from the pool of AM 1-2-3 studies with a total of 214 patients, including 36 with non-measurable lesions. A classical model building strategy was used to develop the joint model, the main steps of which are: (1) develop a separate parametric time-to-event model with covariate selection for PFS (by SCM method), (2) identify a link function between TS and PFS when fitted simultaneously using the previously developed TGI model, and (3) remove non-statistically significant covariates using Wald test step by step.
Calibration was performed by repeating the covariate model building at each step of the joint model building. Model parameters were estimated using the SAEM algorithm implemented in Monolix2020R1, and simulations were performed using SimulX2020R1 and R version 3.6.1.
Results: The joint model developed on the learning dataset was composed of a PK model, a TGI model of sensitive and resistant cells for TS data where treatment effect (inhibition of growth rate of sensitive cells) was driven by amcenestrant concentration, and a Weibull proportional hazard model for PFS. The link function between TS & PFS was best characterized by the TS slope (sum of sensitive and resistant cells).
The joint model from the phase 1/2 data was able to predict the time course of TS and PFS profiles of the amcenestrant arm in the AMEERA-3 study (98 patients) overall and in most subpopulations. When analyzing the AMEERA 1-2-3 data, the calibrated model showed similar parameter estimates. The tumor growth rate (Kg) was estimated at 0.091 (day-1), the concentration in the effect compartment needed to achieve 50% of the inhibition of Kg (IC50) was estimated at 3.74 mol/L, the proportion of resistant cells was 66% and the coefficient effect of TS slope on PFS was 10.7. Compared to training model, IC50 was increased by 50% and time scale parameter for PFS distribution was increased by 44 %, while other parameters were similar with less than 10% of change. The significant baseline covariates in the joint model were age, menopausal status, number of organs with metastases, aspartate aminotransferase (AST) and liver metastases on tumor size kinetics, and AST and liver metastases on PFS. Patients with liver metastases or high AST levels tended to have faster tumor growth and a higher risk of progression. Other covariates had a limited impact on PFS.
Conclusions: In this retrospective analysis, we showed that the joint modelling framework with pharmacokinetics, TS dynamics and PFS, at early oncology development stage (phase 1/2 dose escalation/expansion) were able to predict the PFS outcome of the amcenestrant arm of the phase II in the overall population. Leveraging clinical data as early as possible in the development program and as efficiently as possible is key to optimize drug dosing in oncology development.
References:
- Bardia A et al. AMEERA-1 phase 1/2 study of amcenestrant, SAR439859, in postmenopausal women with ER-positive/HER2-negative advanced breast cancer. Nature Communications. 2022. https://doi.org/10.1038/s41467-022-31668-8.
- Kotani H et al. AMEERA-2: Phase 1 study of oral SERD amcenestrant (SAR439859) in Japanese women with ER+/HER2- advanced breast cancer. The Japanese Society of Medical Oncology Annual Meeting 2022.
- Tolaney S.M et al. AMEERA-3, a phase II study of amcenestrant versus endocrine treatment of physician’s choice in patients with endocrine-resistant ER+/HER2− advanced breast cancer. ESMO 2022.
- PAGE 30 (2022) Abstr 10015 [www.page-meeting.org/?abstract=10015]
