AIMS: Typically, only a proportion of the patients suffering from common diseases respond to frequently prescribed drugs. Since the presence of drug nonresponders in pharmacogenetic studies can adversely affect statistical power we propose a method to restrict genetic tests to drug responders only. In this paper, we estimate drug nonresponse in a clinical trial for the asthma drug montelukast as either the result of an inactive genetic variant or the presence of subgroups of patients not responding to the drug. MATERIALS & METHODS: We propose finite mixture models where unobserved (latent) categorical variables represent either a drug responder or nonresponder class. Analytical results show this method can substantially improve power by testing for genetic variants only in the drug-responder class. We also demonstrate how, if appropriate, placebo data can be used to further increase power to detect genetic effects. RESULTS: It was estimated that only 25-30% of the subjects responded to the drug montelukast. Genetic-association tests confined to the responder group resulted in a substantial increase in explained genetic variance, between 10.3 and 13.2%, for four markers in the arachidonate 5-lipoxigenase (ALOX5) and cysteinyl leukotriene receptor 1 (CYSLTR1) genes. CONCLUSION: The presence of subgroups of patients that do not respond to the drug was an important reason for nonresponse. Additional analyses using finite mixture models in pharmacogenetic studies may provide insight into drug nonresponse and a better discrimination between true and false discoveries.
AIMS: Typically, only a proportion of the patients suffering from common diseases respond to frequently prescribed drugs. Since the presence of drug nonresponders in pharmacogenetic studies can adversely affect statistical power we propose a method to restrict genetic tests to drug responders only. In this paper, we estimate drug nonresponse in a clinical trial for the asthma drug montelukast as either the result of an inactive genetic variant or the presence of subgroups of patients not responding to the drug. MATERIALS & METHODS: We propose finite mixture models where unobserved (latent) categorical variables represent either a drug responder or nonresponder class. Analytical results show this method can substantially improve power by testing for genetic variants only in the drug-responder class. We also demonstrate how, if appropriate, placebo data can be used to further increase power to detect genetic effects. RESULTS: It was estimated that only 25-30% of the subjects responded to the drug montelukast. Genetic-association tests confined to the responder group resulted in a substantial increase in explained genetic variance, between 10.3 and 13.2%, for four markers in the arachidonate 5-lipoxigenase (ALOX5) and cysteinyl leukotriene receptor 1 (CYSLTR1) genes. CONCLUSION: The presence of subgroups of patients that do not respond to the drug was an important reason for nonresponse. Additional analyses using finite mixture models in pharmacogenetic studies may provide insight into drug nonresponse and a better discrimination between true and false discoveries.