Mathilde Marchand (1), Antonio Gonçalves (1), Phyllis Chan (2), Jin Y Jin (2), Jérémie Guedj (3), René Bruno (4)
(1) Certara Strategic Consulting, Paris, France; (2) Clinical Pharmacology, Genentech, South San Francisco, California, USA; (3) Université de Paris, INSERM, IAME, Paris, France; (4) Clinical Pharmacology, Genentech-Roche, Marseille, France
Introduction: Tumor growth inhibition (TGI)-overall survival (OS) models have been shown to predict OS distributions and study outcomes (OS hazard ratio (HR)) in several cancers and settings [1]. Tumor growth rate (KG) using a biexponential TGI model has been proposed as an exploratory endpoint to support early decisions using a two-stage TGI-OS model [2, 3]. However, KG estimation and two-stage TGI-OS models may be prone to estimation bias. Joint TGI-OS modeling has been proposed as a better approach [4].
Objectives: To compare the performances of the two modeling approaches: joint and two-stage in the prediction of OS and HR using tumor dynamics in patients with metastatic non-small cell lung cancer (NSCLC) patients.
Methods: Pooled data from six clinical studies [5] were used to assess the two methods in 3699 NSCLC patients with non-missing baseline characteristics and a total of 21684 tumor size data (sum of longest diameters, SLD). Patients with no post-baseline scan were included.
In the two-stage approach individual parameters of the longitudinal process were estimated (using NONMEM™ version 7.3.0) followed by a parametric regression analysis (using R). The latter included the individual tumor growth rate (log(KG)) as well as the baseline covariates, hence leading to a TGI-OS model as previously described [2, 3]. The joint model consisted of the same two sub-models (using Monolix 2020R1), but all model parameters are estimated in one step, hence providing unbiased parameters of the longitudinal and survival process at stake [4]. Although it was admitted that KG was linked to OS in the two-stage approach [5, 6], additional metrics such as the current SLD (SLD(t)) and the SLD derivative (SLDslope(t)) were evaluated in the joint modelling approach. The estimates provided by the two approaches were compared as well as their prediction performances. Trials were replicated 500 times and OS distributions and HR comparing atezolizumab-containing arms to their respective controls (median and 95% prediction interval) were calculated.
Results:
For the two methods, KG estimates were very similar and OS data followed a log-logistic distribution. As previously observed, among all tested variables log(KG) was found to be the best predictor of OS [5, 6]. In both approaches, the same covariates were selected in the final TGI-OS model: patients with slower tumor growth, better ECOG, lower C-reactive protein, neutrophil to lymphocyte ratio, LDH, metastatic sites, no liver metastasis, positive PD-L1 status, 1st-line therapy, higher albumin, Asian, tended to have longer survival. The OS distributions were well captured for each treatment arm using both models. Finally, the predicted HR were similar with the two approaches and indicated good model performance across the 6 studies (Table below), with observed HR within the 95% prediction intervals for most of the studies. Eventually, joint modelling approach provided slightly larger prediction intervals.
|
Treatment (Study) |
N |
Observed HR |
Two-stage model Predicted HR (95% PI) |
Joint model Predicted HR (95% PI) |
|
Atezo vs Doce (OAK) |
364 / 373 |
0.76 |
0.89(0.79-0.99) |
0.87(0.75-1.0) |
|
Atezo vs Doce (POPLAR) |
134 / 127 |
0.66 |
0.84(0.69-1.05) |
0.85(0.66-1.1) |
|
Atezo+C/C+P vs C/C+P (IMpower132) |
160 / 140 |
0.83 |
0.85(0.72-1.02) |
0.82(0.68-0.99) |
|
Atezo+CnP vs CnP (IMpower130) |
437 / 214 |
0.85 |
0.86(0.66-1.11) |
0.87(0.67-1.14) |
|
Atezo+CnP vs CnP (IMpower131) |
317 / 309 |
0.88 |
0.78(0.67-0.91) |
0.81(0.67-0.97) |
|
Atezo+CP vs CP+B (IMpower150) |
384 / 370 |
0.88 |
0.90(0.78-1.05) |
0.88(0.74-1.03) |
|
Atezo+CP+B vs CP+B (IMpower150) |
370 / 370 |
0.80 |
0.82(0.69-0.95) |
0.79(0.66-0.94) |
Atezo, atezolizumab; Doce, docetaxel; C/C, cisplatin or carboplatin; P, pemetrexed; CnP, carboplatin + nab-paclitaxel; CP, carboplatin + paclitaxel; B, bevacizumab
Conclusions: Overall, the two approaches led to similar results in terms of impact of KG and baseline prognostic factors on OS, prediction of OS distributions, and finally prediction of HR (treatment effect) for all studies. Operating characteristics in support to early decisions [2, 3] are being evaluated for the joint modelling approach.
References:
[1] Bruno R et al. Clin Cancer Res (2020) 26, 1787-95.
[2] Bruno R et al. PAGE 29 (2021) Abstr 9679 [www.page-meeting.org/?abstract=9679].
[3] Bruno R et al. Proceedings: AACR Annual Meeting 2022, Abstract 5176.
[4] Desmée S et al. AAPS J. (2015) 17 691-9.
[5] Chan P et al. CPT PSP (2021) 10, 1171-82.
[6] Claret L et al. Clin Cancer Res (2018) 24, 3292-8.
Reference: PAGE 30 (2022) Abstr 9978 [www.page-meeting.org/?abstract=9978]
Poster: Drug/Disease Modelling - Oncology