A Modeling and Simulation Framework to Support Early Clinical Drug Development Decisions in Oncology

Abstract

The success rate of new molecules in oncology is the lowest in any therapeutic area. The high failure rate and cost associated with oncology drug development may reflect poor predictability of in vivo preclinical tumor xenograft models and lack of quantitative approaches to guide both preclinical and clinical development. Antitumor activity in early clinical studies is typically evaluated using objective response rate (ORR) or progression-free survival (PFS). However, these estimates in typical small, noncomparative phase 1 or phase 2 trials are generally imprecise and uninformative to efficiently support go—no go decisions and design of phase 3 clinical trials. The phase 3 trial failure rate is particularly high in oncology,1 and there clearly is a need for more quantitative approaches to improve the success rate of oncology drugs consistent with the FDA recent initiatives.2,3 To address these issues, we are developing a drug-disease modeling framework that has been successfully applied to predict expected clinical response and survival in cancer patients in a number of clinical settings. This modeling framework focuses on efficacy, and the core of this framework is constituted by an exposure-driven tumor growth inhibition (TGI) model that uses the full longitudinal tumor size data as opposed to categorizing data, as in the calculation of ORR. Change in tumor size from baseline is used as a biomarker of drug effect to predict survival in a drug-independent survival model. It is therefore an informative and predictive patient-level continuous end point that can be assessed in early phase 1 or 2 clinical studies. The proposed modeling framework (drug-specific TGI model coupled with drug-independent survival model) can enhance learning from early clinical studies compared with the traditional approach of estimating ORR and PFS. The TGI model4,5 accounts for the dynamics of tumor growth, antitumor drug effect, and resistance to drug effect. The model describes tumor size (ie, sum of the longest diameters of target lesions according to the RECIST [Response Evaluation Criteria in Solid Tumors Group] criteria) as a function of time and drug exposure. The model incorporates a first-order tumor growth rate, drug action on the tumor (ie, cell kill driven by drug exposure). In addition, a resistance process assuming exponential decrease in cell kill is incorporated to describe the regrowth of tumor apparent in some patients' profiles. The exposure metrics to drive drug effect can be either total dose or area under the concentration—time curve (AUC; estimated using a population pharmacokinetic model) at each of the dosing events. A full pharmacokinetic/pharmacodynamic model with drug concentration driving drug effect can also be implemented with the same results as with AUC6 when there is no schedule dependence. The model parameters are estimated in a nonlinear mixed-effect analysis. The survival model5 describes the survival time distribution as a function of covariates. The probability density function (eg, log normal, Weibull) that best describes the observed survival time is selected by using difference in log-likelihood and goodness-of-fit plots. The baseline prognostic factors are incorporated in the model (eg, tumor size, performance status) together with relative change in tumor size from baseline at the first posttreatment visit (typically end of cycle 2; ie, week 6–8 depending on the cycle duration) as a biomarker of drug effect. The survival model can be considered as a drug-independent model that relates a biomarker response (change in tumor size) to a clinical end point (survival time). This modeling framework was first used to predict survival in phase 3 trials of capecitabine plus docetaxel versus docetaxel in metastatic breast cancer (MBC)7 and of capecitabine versus 5-FU in colorectal cancer (CRC)5 based on capecitabine single-agent phase 2 data in the respective tumor types and historical standard of care (docetaxel and 5-FU) phase 3 data. Capecitabine response in the phase 3 studies was predicted from phase 2 capecitabine data by using drug-specific and disease-specific, drug-independent parameters and using patient characteristics as described in detail by Claret et al5 for the CRC case. In the MBC case, phase 3 study treatment effect in the combination arm was simulated assuming additive drug effects. The CRC model was subsequently extended to simulate dose response using a specific model to account for dose-reduction events and simulate dose intensity.8 The modeling framework was further evaluated based on phase 2 data of motesanib in thyroid cancer9 to simulate ORR and PFS using the combination of dose-reduction, TGI, and survival models. Dose-response simulations were performed to support end-of—phase 2 decisions for mosetanib in this specific tumor type. When tumor markers are available, they can be used as measures of tumor size and are amenable to the same modeling exercise (eg, using myeloma protein in multiple myeloma10,11). In a similar approach, a drug—disease model has been proposed using data from 3398 patients entered in four phase 3 pivotal studies in NSCLC12 that can also be used to simulate tumor response and survival (based on end-of—cycle 2 change in tumor size) for approved therapies and investigational treatments.13 A modeling framework to analyze phase 2 data from investigational treatments based on longitudinal tumor size data and to simulate expected clinical end points and phase 3 studies is proposed to support end-of—phase 2 decisions and phase 3 clinical trial design. This approach opens a new design paradigm for phase 2 studies based on change in tumor size rather than the typical clinical end points of ORR and PFS.14–16 Financial disclosure: The authors are employees of Pharsight (Bruno, Claret) or Amgen (Lu, Sun).