UC Riverside

The traditional prescribing approaches in clinical therapies, such as “one drug fits all”, have limits due to consideration of drug effectiveness and safety. It is now well recognized that the Single Nucleotide Polymorphism (SNP) plays an important role in pharmacogenomics and personalized medicine by serving as the predictive biomarker for patient stratification and dose selection. Considering the economic efficiency of drug development, searching for the single best predictive SNP has just begun and has great potential. Current statistical methods for the best predictive biomarker selection rely on a variety of ranking procedures. However, these ordering approaches face three main issues: (1) they can potentially fail to distinguish predictive biomarkers from prognostic biomarkers; (2) the ordering is not necessarily correlated with the true significance, especially when the signal-to-noise is small; (3) such ranking approaches provide no control on the incorrect assertion in terms of the best predictive SNP detection.

In this paper, we propose to overcome the first issue by quantifying the predictive ability of each candidate SNP in two parallel approaches: (1) adjusting the model with Least Square mean (LSmean); (2) adjusting the data with Virtual Matching (VM) before doing estimation. For the selection of the best predictive SNP, we propose to apply the Multiple Comparisons with the Best (MCB) to the ranking and selection procedures. Specifically, we will do the indifference zone selection with MCB lower bounds; and the subset selection with MCB upper bounds (suppose a larger response is better). The probability that both inferences are correct is controlled. Simulation studies show that our proposed method is more reliable in distinguishing between predictive and prognostic effects, and has a larger chance to detect the true best predictive SNP than varieties of methods in both multiple testing and machine learning approaches. Furthermore, the adjusting-model approach is generally better than the adjusting-data approach when we assume there is one single predictive SNP; however, the latter one is easier to be extended to the scenario with multiple predictive SNPs.