Comparison of tumor dynamics models and tumor growth inhibition metrics in support of decisions in early Phase Ib/II clinical oncology studies
Mathilde Marchand (1), Antonio Gonçalves (1), Francois Mercier (2), Pascal Chanu (3), Jin Y Jin (4), René Bruno (5)
(1) Certara Strategic Consulting, Paris, France; (2) Clinical Pharmacology, Genentech-Roche, Basel, Switzerland; (3) Clinical Pharmacology, Genentech-Roche, Lyon, France; (4) Clinical Pharmacology, Genentech, South San Francisco, California, USA; (5) Clinical Pharmacology, Genentech-Roche, Marseille, France
Introduction/Objectives:
Model-derived Tumor Growth Inhibition (TGI) metrics derived may serve as predictors of overall survival (OS). However, challenges persist in accurately estimating effect size during early drug development, where follow-up, and tumor size measurements are limited. We demonstrated that effect size (geometric mean ratios: GMR) comparing experimental vs. control using TGI metrics estimated with the biexponential Stein model [2], including the tumor growth rate (KG), the tumor size ratio (TR) at week 6, 12 or 24 to baseline, and time of tumor growth (TTG) have superior operating characteristics (OCs) than RECIST endpoints to support early decision-making in randomized Phase Ib/II studies [1] .
To extend this work, various biexponential models relying on different hypotheses were used to estimate TGI parameters and metrics. Three TGI models were used to estimate model parameters as well as TGI metrics, namely: Stein [2], generalized Stein (gStein) with a mix of sensitive and resistant cells [2], and Claret simplified TGI (sTGI) with an exponential resistance on the shrinkage rate [3, 4]. We conducted a comparative analysis of the OCs using estimates derived from each model.
Methods:
In the IMpower150 study, first-line non-small cell lung cancer (NSCLC) patients were randomly assigned to atezolizumab, bevacizumab, carboplatin and paclitaxel (ABCP arm), or BCP (control). ABCP significantly prolonged both PFS and OS compared to BCP.
The resampling method described in [1] was applied to generate 500 datasets of 40 patients per arm with a 24 weeks follow up after the last patient included from IMpower150. Assuming the alternative hypothesis, we resampled from ABCP and BCP to mimic a positive study. While, under the null hypothesis, we resampled from the BCP control arm twice to simulate a negative study.
Stein, gStein, and sTGI models were fit using NONMEM [6] for each dataset. TGI parameters (growth and shrinkage rates) and metrics (TR at 6 to 24 weeks, at maximum shrinkage TRmax or TTG [4]), were then estimated from each patient's resampled tumor profile, summarized by subsampled dataset and arm. Effect size was expressed as the GMR of control/experimental for all metrics except TTG where the effect size was estimated as a hazard ratio (HR) between the two arms. Subsequently, the probability of the effect size being greater or lower than specified thresholds was determined for each scenario across the 500 replicates under either the alternative or null hypothesis. The estimated probabilities were visually represented as receiver operating characteristics like curves, illustrating the correct go rate (ABCP vs. BCP comparisons) versus the incorrect go rate (BCP vs. BCP comparisons) as previously done [1].
Results:
The model parameters could be accurately estimated for all three models using the complete IMpower150 study data. The sTGI model outperformed both gStein and Stein models, as evidenced by its lower objective function value (-68 and -1627, respectively). All models had similar and good visual predictive checks.
When these models were applied to the estimation of the 500 subsampled datasets of 40 patients per arm over 24 weeks follow-up, the estimated tumor growth rate with Stein (KG) had good OCs with correct go rate close to 80% and incorrect go rate <20% as previously shown [1], while the growth rate estimates with both sTGI and gStein had poor OCs with correct go rates <30%. In those two models, growth rates are not impacted by the treatment (natural growth in the sTGI, growth of the resistant cells in gStein). None of the other parameters (especially the shrinkage rates) had good OCs.
The derived TGI metrics TR24 and TRmax had good and similar OCs (around 80% correct go rates) whatever the model used, while TR12 and 6 were inferior. TTG had best correct go rate when estimated by sTGI (>80%), followed by Stein (<80%) and gStein (around 70%).
Conclusion:
The sTGI best fitted complete IMpower150 study data. When applied to subsampled datasets, the only parameter with good OCs was the KG estimated using the Stein model. Derived TGI metrics (TR24, TRmax or TTG) had good performance irrespective of the model with a preference for sTGI estimated TTG. Selecting the right TGI parameter/metric is critical for assessing treatment effects and decision-making in early oncology drug development.
References:
[1] Bruno, R. et al. Tumor dynamic model-based decision support for Phase Ib/II combination studies: A retrospective assessment based on resampling of the Phase III study IMpower150. Clin Cancer Res 29, 1047-1055 (2022).
[2] Stein, W. D. et al. Tumor Regression and Growth Rates Determined in Five Intramural NCI Prostate Cancer Trials: The Growth Rate Constant as an Indicator of Therapeutic Efficacy. Clin Cancer Res 17, 907–917 (2011).
[3] Claret L et al. Model-based prediction of phase III overall survival in colorectal cancer on the basis of phase II tumor dynamics. J Clin Oncol 27:4103-4108 (2009).
[4] Claret L et al. Evaluation of tumor-size response metrics to predict overall survival in Western and Chinese patients with first-line metastatic colorectal cancer. J Clin Oncol 31:2110-2114 (2013).
[5] Beal SL et al. 1989-2011. NONMEM Users Guides. Icon Development Solutions, Ellicott City, Maryland, USA.