NIH and FDA's vision of personalized medicine involves a drug, and a companion biomarker test identifying the patient subgroup that the drug targets. Examples of personalized medicine approved by the FDA are surprisingly few. (Can you name the only 3 microarray devices approved by the FDA?) Yet, personalized medicine is beginning to take shape. After the FDA issued its VGDS (Voluntary Genomic Data Submission) draft guidance in 2005, drug companies have been routinely banking biological samples from clinical trials. In June 2010, FDA held public hearing on regulation of genetic prognostic/diagnostic tests. This seminar will indicate what the hundreds of PhDs with quantitative training working in the pharmaceutical industry can expect to be the skills that will be needed as medicine transitions from "on average" to "for individuals".
I will first give a solid mathematical foundation of multiple testing that is difficult to gather from literature. Then I will focus on two typically overlooked issues in testing of biomarkers. One issue is interpretation of an unconditional expectation error rate such as FDR, however it is controlled. The other issue, which has only recently come to light, is the ever popular permutation testing requires a strong assumption on the (unknown) joint distribution of biomarkers to control its error rate. These and other issues will be illustrated in the Genome-wide Association Studies (GWAS) setting.