Source: Dissertation Abstracts International, Volume: 78-07(E), Section: B.

Adviser: Andrew Phillips.

According to the World Health Organization, cancer is the leading cause of death worldwide, accounting for an estimated 7.6 million deaths in 2008. It is expected that in nearly all these cases, the function of the tumor suppressor protein p53 has been inactivated; in 10-20% of these cases, amplification or overexpression of p53's negative regulators MDM2 and MDMX appear to be responsible for abrogating the p53 pathway. Reactivation of p53 by inhibition of these negative regulators is an attractive approach to cancer therapy. While many compounds with nanomolar inhibitory constants are known for the p53-MDM2 interaction, MDMX still lacks such potent binders. Current research suggests that MDM2 and MDMX work cooperatively to inhibit p53 activity, indicating the need for potent p53-MDMX inhibitors to produce clinical efficacy.

This thesis presents the development of a statistical method for selecting hits from a high-throughput screen and its application in the identification of a novel series of p53-MDMX inhibitors. The statistical method was developed using a Bayesian approach to compare and contrast two populations. The model was evaluated based on its ability to recall active compounds compared to the industry standard z-score for 6 historical biophysical high-throughput screens. Additionally, the ability of the model to estimate the false negative rate was evaluated via a permutation test and a Bernoulli simulation. In all 6 screens, the Bayesian model was shown to have better recall of active compounds than the z-score and the Bayesian estimates were shown to be non-random and have reasonable agreement with a Bernoulli simulation in which the estimated activity was dependent on the experimental activity.

We hypothesized that the lack of p53-MDMX inhibitors stemmed in part from the z-score's insensitivity to weakly active compounds in primary high-throughput screens. To that end, 20,081 compounds were subjected to a fluorescent polarization based high-throughput screen. Both the Bayesian model and the z-score were used to select 199 compounds for retesting by concentration-response. Of the 199 tested compounds 37 were identified as active. The Bayesian model had better recall of these active compounds than the z-score. However, in the permutation test and Bernoulli simulation the Bayesian model failed to show better than random prediction of compound activity. This result appeared to be a defect with the high-throughput screen, and not the Bayesian model.

From the screen a set of 10 nicotinamides were identified with p53-MDMX inhibitory constants ranging from 8-100 muM. In the primary screen these compounds displayed weak signal. In a normal screening mode they may have been overlooked as the z-score does not recover any nicotinamides in the first 100 compounds. The Bayesian model on the other hand recalls 3 members within the first 100 compounds, and all 10 within the first 150.