Statistical Perspective on Adaptive Strategy for Matching Targeted Therapies with the Patients Most Likely to Benefit

Value of I-SPY 2 may go beyond the clinical results described in the NEJM articles

As more new targets and anticancer drugs are discovered traditional statistical designs will be insufficient for matching patients with effective drugs. In a perspective article, published on 7 July 2016 in The New England Journal of Medicine, David Harrington and Giovanni Parmigiani from the Department of Biostatistics and Computational Biology, Dana–Farber Cancer Institute, and the Department of Biostatistics, Harvard T.H. Chan School of Public Health, both in Boston applaud the use of I-SPY 2 platform and urge continued innovation in trial design, especially in phase I and III settings.

They wrote that a value of I-SPY (Investigation of Serial Studies to Predict Your Therapeutic Response with Imaging and Molecular Analysis) 2, may go beyond the clinical results described by Rugo et al. and Park et al. in same issue of the NEJM. Adaptive multigroup trials such as I-SPY 2 have the potential to answer several questions simultaneously and more efficiently than traditionally designed trials.

I-SPY 2 platform is a promising adaptive strategy for matching targeted therapies for breast cancer with the patients most likely to benefit from them. I-SPY 2 identified veliparib with carboplatin in triple-negative breast cancer and neratinib in HER2–positive, hormone receptor–negative breast cancer that met prespecified criteria for testing in phase III trials.

The challenges in identifying successful targeted therapies in cancer are substantial. There are too many new drugs tested in traditionally designed phase II trials that test treatments one at a time in heterogeneous groups of patients and the signal of a treatment effect can be diluted in these heterogeneous groups.

Both, the US Food and Drug Administration and the European Medicines Agency have acknowledged that the commonly used clinical trial designs need some changes. However, oncology has been slow to adopt Bayesian designs. The authors of this perspectives article on statistics emphasized differences and similarities between Bayesian and frequentist approaches in clinical trials.

Main goal of Bayesian approach is to predict outcomes of future trials and absolute risk for future patients, while frequentist approach estimates population average effects. Regarding assumptions, Bayesian approach requires explicit specification of prior distributions of unknown population parameters, incorporates a priori knowledge and clinical judgment formally and may be sensitive to specification of prior distributions, while frequentist approach does not require explicit specification of prior distributions of unknown population parameters and incorporates a priori knowledge and clinical judgment informally. In term of interim monitoring, in Bayesian approach only the data actually obtained are relevant for final conclusions and whether or not a clinician examines accumulating evidence with the possibility of stopping the trial does not affect inference, while in frequentist approach both the data actually obtained and the probabilities of data not obtained are relevant for final conclusions and whether or not a clinician examines accumulating evidence with the possibility of stopping the trial does affect inference. Bayesian approach is often computationally complex and careful modelling often requires simulation-based calculations. Frequentist approach is often computationally simple, though careful modelling may require simulation-based calculations.

Underlined similarities between the two approaches are adaptive design, multistage trials, early stoppings and adaptive randomisation; regarding role of statistical judgment the options for data-driven analyses are available; skill and substance-area knowledge of the data analyst are important in drawing correct conclusions. In term of compatibility, it is feasible to combine a Bayesian design with a frequentist analysis or a frequentist design with a Bayesian analysis. Both approaches rely on prior knowledge and clinical judgment, though they incorporate them in different fashions.

The efficiency of multigroup early-phase trials has long been recognised, but I-SPY 2 differs in important ways from traditional early-phase trials. The design of the platform acknowledges the complexity of phase II testing in cancer. The platform may be an appealing setting for cooperation among pharmaceutical companies and academic investigators. It was designed in 2009 and it is still an early work in the setting of trial designs. However, the fundamental tenets behind it are important first steps toward the efficient use of clinical resources. There is much to be learned and despite some unresolved issues, I-SPY 2 is an important addition to the inventory of clinical trial designs in oncology, according to Harrington and Parmigiani.

 

Reference

Harrington D, Parmigiani G. I-SPY 2 — A Glimpse of the Future of Phase 2 Drug Development? N Engl J Med 2016; 375:7-9. DOI: 10.1056/NEJMp1602256