A recent analysis by the Tufts Center for the Study of Drug Development estimates that the cost of developing a single new chemical entity (NCE) into a successful therapeutic agent is $802 million. This figure is largely dependent on the expense of investigating NCEs that ultimately fail to be approved for use: between 70 90% of NCEs do not achieve New Drug Application (NDA) approval, and many of these failures are identified during the later, more costly phases of drug development. The exponential growth in the number of putative NCEs as a result of combinatorial chemistry and high-throughput screening has only confounded this problem by significantly increasing the number of early-phase NCEs under consideration for further costly development in human clinical trials.
It is widely agreed upon that there are 3 major categories of reasons for drug failure: safety (toxicity), efficacy, and economics. This thesis is concerned with developing a Bayesian domain-knowledge probabilistic model (called Pharminator) to address the first two of these categories, with a goal of predicting clinical success of an NCE. Pharmacoeconomic modeling is a vastly different domain compared to Pharminator’s clinical trial domain, and is beyond the scope of this thesis. While several clinical predictive models have been described in the literature over the past 10 years, the ongoing costly failure rate in drug development warrants developing more reliable predictors of NCE clinical success. The number of NDA approvals in 2002 fell to a 5-year low of 18, compared to 30, 35, 27, and 24 in 1998, 1999, 2000, and 2001 respectively, despite rapidly increasing numbers of NCEs as a result of high-throughput screening and combinatorial chemistry. Therefore, previous decision models have had no apparent impact on this problem. The Pharminator model combines knowledge of drug development logistics, existing data on NCE attrition rates, and Bayesian decision theory in a manner that may improve upon the performance of previously described models.
The product of this model is an application to be used by drug development teams at the Phase I/Phase IIa time point for a given NCE that has passed the FDA Investigational New Drug (IND) screening process. The users are prompted to answer several key questions about the NCE. The answers provide information regarding which prior probabilities and conditional probability tables are to be used in the model, as well as the observed data upon which prediction will be made. The output is a numeric and graphical (distribution plot) report of the prior and posterior probability distributions for clinical success, safety and efficacy. The application is demonstrated on one fictional agent and one real agent designed to demonstrate key behaviors of the model. Retrospective and prospective testing and validation will continue beyond the completion of this thesis in order to optimize the performance of this model.