Abstract
The availability of high-resolution genetic profiling raises the possibility, during the course of a drug development program, of discovering a subset of patients at particular risk of an adverse drug reaction who might be excluded from subsequent randomization into studies and identified as unsuitable for post-licensing use. Such methods depend on the estimation of the risk of adverse drug reactions for patients with differing genetic profiles followed by an assessment of the risks and benefits of their exposure to the drug. In this paper we explore the performance of a number alternative statistical methods for the estimation of risk in terms of the success of the subsequent exclusion rules. The approaches were evaluated using a single-nucleotide polymorphism dataset concerning HIV patients at risk of hypersensitivity to the drug abacavir. Overall we found that a method based on LASSO performed better than the alternatives that we studied, which included a decision-theoretic Bayesian approach, and that its performance suggested suitability for its prospective implementation.