Use the qval, not the raw pvalue. The "minimum of 3 replicates" is a good general rule of thumb since you need that many to have a decent shot at measuring variance. I personally recommend at least 6 replicates, which happen to fit nicely on a single lane of a HiSeq if you have a standard two group comparison setup.

Imagine you're doing a statistical test on some data and you're 99.9% sure that it's correct. Then you can be pretty sure that what it tells you is right (the p value).

But if you do 1,000 tests, you're probably going to get one test that says something untrue (100% - 99.9%). If you do 30,000 tests then you're going to get a lot of false positive values by the end.

The q-value is a modified p-value that takes into account that you'll get some false positives based on how many tests you're doing. This is called a False Discovery Rate (FDR) and there are multiple ways of calculating it.

Long story short: Use the q-value, it reduces the number of false positives.