There is no standard way to do it. Obviously your method of feature selection is important, as it is undoubtedly already applying some kind of statistical test to your data. I gave a long answer on this topic, here: A: What is the best way to combine machine learning algorithms for feature selectio

However, my answer should not be regarded as the 'gold standard' by any means.

One could easily make the argument that 'unbiased' feature selection, i.e., performing feature selection on your entire unfiltered data, is best, whilst, in other situations, it may prove beneficial to first filter out genes based on some comparison, e.g., between cases and controls.

Kevin