At large institutions dedicated to clinical research in cancer a large number of new agents or new combinations of anticancer agents undergo evaluation for activity. The process is typically carried out through separate phase II studies with only informal learning carried out between studies--even if the studies draw patients with similar disease characteristics. There is a need for a more systematic and rational approach to the whole phase II testing or screening process to allow for more efficient study design and greater learning about cancer treatment and what works or does not work. This application describes research building on work first proposed by Yao, Begg, and Livingston for evaluating anti-cancer vaccines, in which the entire process is considered a large enterprise within which multiple agents, introduced by some mechanism, undergo screening for activity and either progress to further testing or are discarded. By sharing as much information as can be shared between new agents via mathematical models that incorporate drug-specific common elements, research proposed in this application envisions broadening this process from one geared to vaccine trials to the larger phase II testing programs that exist at large biomedical research centers. The ultimate goal is to learn as much as possible about the agents and the patients, so as to maximize each patient's chances of benefit while minimizing the risk of detrimental side effects. Research in the initial pilot phase (R21) will provide a proof of concept for strategies to overcome the prohibitive computational challenges involved in carrying out a decision theoretic solution to the problem. The first specific aim develops a simulation based approach for sequential design. The second specific aim applies the developed algorithm in a highly stylized version of the drug screening problem. Milestones are set to define the targeted research goals in an easily verifiable manner. Research in the following extended development phase (R33) targets three specific aims. The first specific aim proposes to develop non-sequential policies to solve the sequential decision problem of evaluating an sequence of phase II trials. Policies are defined in terms of fixed decision boundaries, allowing optimization up-front, without the need for backward induction. Finding the optimal policy is a challenging high dimensional stochastic optimization problem. The second specific aim is the construction of new hybrid algorithms which combine the parsimony and robustness of non-sequential policies with the flexibility of unconstrained sequential solutions by dynamic programming. The third specific aim targets the extensions of the underlying probability model needed to accommodate a realistic application to continuous drug screening. These extensions pose challenging research problems related to modeling ordinal responses. multiple and delayed outcomes. and repeated longitudinal measurements.
