What is an appropriate cut off for p-values and log fold changes during the analysis of a time course global gene expression experiment using DESeq2?

Question

[mgranada3]

The experiment is performed on yeast in 4 different conditions over 5 time points, lag phase to late stationary, and ~5300 genes.

I am currently analyzing data at a cutoff of 0.01 p-value and +/- 1 LFC. But am still getting about 2000 differentially expressed genes which don't yield many GO term results.

Answer 1

[bk11]

Do you have any specific reason for choosing *p-value* instead of *padj*? I would not worry much about how many GO terms data will result. But, if you choose cut-off of padj < 0.05 and +/- 1 LFC should be good.

Answer 2

[kalavattam]

One should filter on the p-values following multiple test corrections, i.e., the "padj" values (a.k.a. the q-values).

It's tempting to filter on log2 fold change (LFC) thresholds following the assignment of p- and q-values; however, this is not what you want to do. Instead, the LFC threshold should be included in your tests as, for example, explained here. To filter on LFC after assigning corrected p-values is called post-hoc filtering, and it invalidates the p- and q-values from your initial tests. For more details, see here and here.

But am still getting about 2000 differentially expressed genes which don't yield many GO term results.

You should pick a fold change threshold (i.e., an effect size) that's meaningful for whatever particular biological system you're working with; then, do a log2 transformation of that value. The resulting value should then be supplied to the lfcThreshold argument of DESeq2::results(). In your case, perhaps a smaller effect size may be appropriate? For example, you could try an effect size of 1.5: log2(1.5) =~ 0.585: lfcThreshold = 0.585. Or you could try an even smaller effect size of 1.25: log2(1.25) =~ 0.322: lfcThreshold = 0.322.