At last week’s Microarray Data Analysis conference in Washington D.C., several speakers laid out what distinguishes a good microarray experiment from a bad one. Here is their collected advice:
• Consult a statistician before you even design your experiment. They have more to offer than just analysis tools.
• Formulate your hypothesis, even if you are going on a “fishing expedition.” What is the question you want to answer?
• Do a power analysis to determine the number of replicates (i.e. chips) you need to detect an effect. To estimate the effect size, you might want to run a pilot study first or obtain the estimate from previous similar experiments. Regardless of the power analysis results, obtain at least three replicates on different slides or chips.
• Find sources of technical variation before you embark on a hunt for biological effects. Some common sources are: target preparation and chip processing, day or time of day of the experiment, location, chip lots, scan order, or the researcher running the experiment.
• Standardize your protocols as much as possible, and train your technicians until they obtain consistent results.
• Randomize your variables: for example, don’t run all your treatment slides on one day and all your controls on the next; stagger chip lots across experiments if you need to.
• Choose the appropriate statistical tool for your analysis. Cluster analysis cannot answer all the questions.