An iterative phase I/II/I clinical trial design incorporating genomic biomarker information

The goal of a phase I cancer clinical trial is to determine the maximum tolerated dose (MTD) or the recommended phase II dose (RP2D) of a new drug that corresponds to some given acceptable rate of dose limiting toxicity. However, a common limitation to the generalizability of phase I and phase II clinical trials is high patient heterogeneity with respect to toxicity and efficacy. A poor estimate of the MTD may put patients at unprecedented risks by assigning subtherapeutic or excessively toxic doses. Ignoring patient heterogeneity with respect to efficacy may cause phase II trials to have extremely large false positive and false negative error rates within subpopulations. In either case, the large scale phase III clinical trial conducted with an inappropriate dose may put patients in the study at high risk, and doom a potential effective and safe treatment after substantial investment in its development.We propose to achieve this goal by using a novel compound iterative phase I/II/I clinical trial design which may be conceptualized into three stages. The first stage consists of obtaining an overall estimate of the MTD based on an initial phase I trial under the assumption of no patient heterogeneity. A phase II study is then conducted using the preliminary estimate of the MTD to estimate efficacy. The second stage focuses on the application of the predictive classifiers in early phases of clinical trials. Supervised learning algorithms (i.e., support vector machines (SVM)) are effective classification tools in high-dimensional microarray classification problems such as the development of genomic biomarkers from microarray data. We implemented the SVM binary classification algorithm using the approach of Xu (2008) for the development of predictive classifiers based on phase II trial efficacy data by comparing the gene-expression profiles of responders versus non-responders. Identification of patient subpopulations that might be expected to have distinct toxicity/efficacy profiles may thus be based on genomic biomarkers.The third stage of our proposed design incorporates the predictive classifier described above to estimate and refine the MTD through two key steps: (1) obtain an estimated MTD for each biomarker group by combining toxicity data from the first two stages and applying an estimation algorithm and (2) refine the MTD estimate by conducting a secondary phase I trial for each biomarker group. The secondary phase I trials account for patient heterogeneity by basing dose finding on toxicity while also accounting for each patient's genomic biomarker status. This provides a mechanism for incorporating genomic information that permits estimation of separate MTDs for the two subpopulations identified by the biomarker.We compared various MTD estimation methods, and based on the results of simulation studies recommended the use of the constrained logistic regression model to estimate biomarker group-specific MTDs prior to the secondary phase I studies. We investigated the impact of two adaptive design strategies in the secondary phase I trials across a range of dose-toxicity and dose-efficacy profiles.It is believed that most cancer clinical trials involve a heterogeneous group of patients at the molecular level. This heterogeneity is one of the reasons that not all patients with cancer respond to a given drug. In view of this patient heterogeneity, it is clear that a "one size fits all" approach may not be suitable in the drug development process. Therefore, it is important to be able to predict which patients are most likely to benefit from a new drug. This would not only save patients from unnecessary risk of toxicity but might facilitate their receiving beneficial treatment, and it would shorten the time required for drug development and lower associated costs. The goals of this research are two fold: (1) to investigate statistical issues involved in applying a genomic biomarker classification method to account for patient heterogeneity with respect to toxicity and efficacy response in early phase clinical trials, allowing for the possibility of differing treatment efficacy among subpopulations, and (2) developing a new design approach that incorporates high-dimensional genomic information into phase I/II clinical trials.Our results showed that by applying predictive classifiers to determine a MTD estimate for each subpopulation we can obtain an acceptably large value of probability of efficacy (i.e., improved response rates) while also controlling the probability of toxicity (i.e., increased safety) for each subpopulation. We concluded that the MTD estimates obtained from our proposed design are more accurate than would be expected from a single phase I trial. The results of this research provide important guidelines with respect to (1) incorporating genomic biomarker information from microarray data into early phase clinical trials to identify subpopulations and (2) use of predictive classifiers to obtain refined and improved estimates of MTDs for subpopulations.