PhD Thesis




Phases I are usually the first stage of testing a new drug involving human subjects. Phase I clinical trials evaluate the safety of the treatment and identify its side effects on patients with advanced cancer due to the harmfulness these treatments. The primary aim of Phase I in oncology is to determine the highest dose level with an acceptable toxicity rate of the new drug on a restricted number of patients, i.e. to select a dose level with a toxicity probability closest to a given target. This recommended dose level is called maximum tolerated dose. Phase I trials are sequential dose-escalation procedures.

In recent years, unlike standard chemotherapy, targeted therapies have emerged as another type of anti-cancer agents that interact with specific molecules involved in cancer spread rather than killing cancer and healthy cells. In this context, we have developed a Phase I/II dose-finding design in single-agent for molecularly targeted agents where the efficacy increases and can plateau. Our method focus on selecting the optimal dose, that is the dose associated with highest efficacy and if the plateau is reached the first dose on the plateau as it will be associated with the lowest toxicity. We used adaptive randomization in order to determine the plateau location. The proposed method gave good performance. We also extended this design on (1) unimodal relationships and (2) different biomaker’s groups leading to different optimal dose in each subgroup with shared toxicity.

Methods for single agent are not appropriate for combination phase I trials as they are not designed to take into account the multi-dimensionality. We studied several existing representative methods specifically designed for combination, and compared their performance. Based on an extensive simulation study, we have noticed that model-based methods seemed to perform better than algorithm-based methods in terms of the percentage of correct combination selections when targeting a single maximum tolerated dose at the end of the trial. All model-based methods have high operational characteristics and their performances were in general comparable. On this basis, our aim was then to propose our own innovative adaptive dose-finding design that would have good operational characteristics and in general would perform better than the existing designs. We proposed a Phase I dose-finding design for combination based on a logistic model with an interaction term. All the compared designs were efficient when the maximum tolerated doses were located on the same diagonal in the combination space, but the benefit of our method was that it was also efficient in other cases.

Finally a new challenge in cancer development is to combine both cytotoxic and molecularly targeted agent. Indeed, their action can be complementary, inactivate the cells or reduce cancer growth and killing cells, but also skirt drug resistance. When combining several agents, a possible synergistic effect on the efficacy is expected. We studied both toxicity and efficacy of the combination in a Bayesian Phase I/II design using the characteristics of each agent. Our goal was to maximize the efficacy while minimizing the toxicity under an acceptable threshold. We evaluated our design through a simulation study under various practical scenarios and observed that our design performed well by selecting the optimal combination with a high percentage. Nevertheless, the performance of the design highly decreased with the number of dose level of the molecularly targeted agent.

During these 3 years of PhD, we have proposed several adaptive early phase designs, either for molecularly targeted agents or combinations trials, to answer to the current need of practical statistical methods in oncology. Moreover, we have developed R packages to facilitate the access of these methods in the current practice.