Detection of drug effect in small groups using PET

DESCRIPTION (provided by applicant): The purpose of this proposal is to develop a tool, based on positron emission tomography (PET) imaging and advanced image-analysis methods, to address a critical need faced by the pharmaceutical industry. Evaluation of Pharmaceuticals that target the brain can be difficult and subjective; therefore, pharmaceutical companies often spend millions of dollars in clinical development of a compound that is ultimately dropped due to problems encountered in human testing. Thus, methods are needed which can help make a simple go/no-go decision about a candidate drug, so as to eliminate ineffective drugs early in their evaluation process. PET imaging with FDG promises to be an effective solution to this problem. Based on PET, we propose to develop a new product that will directly answer the basic go/no-go question of whether a drug merits further study. Specifically, we will investigate in Phase I whether it is possible to determine, using only around five to eight subjects, whether a candidate drug has any measurable metabolic effect on the brain. If an effect is found, then additional subjects can be scanned to obtain the full complement needed for more detailed characterization of the drug. If no effect is found, then the drug can be eliminated from further consideration, and resources can be applied to the next candidate. The technical challenge will be to determine whether optimized machine-learning techniques can permit this initial go/no-go decision to be made with a small number of PET scans. We will use prediction accuracy as a measure of the strength of drug effect, and determine whether this metric can be computed reliably in small groups of subjects. In recent academic work, we have demonstrated that prediction accuracy can be tremendously enhanced by using multivariate machine-learning methods. Our goal will be to bring these new approaches to bear, so as to obtain the greatest amount of information possible from the data at the least cost. The specific aims will be to: 1) Implement a hierarchy of learning algorithms ranked in terms of model complexity, from low to high, as follows: generalized likelihood ratio test (GLRT) with white noise assumption, canonical variates analysis (CVA), quadratic discriminate analysis (QDA), and relevance vector machines (RVM); 2) Use analytical statistical models, and the NPAIRS resampling framework, to estimate prediction accuracy for each method, which will be used as a measure of the strength of drug effect; and 3) Compute performance versus number of subjects N, spatial smoothing, and number of principal components used. The performance of optimized algorithmic configurations will reveal the feasibility of the proposed concept.