Differentially gene expression analysis using limma

Question

[fahim]

Hi

I analyzed a differentially gene expression analysis by limma. But when I run for top sig gene by giving this code - topSig <- top[top$adj.P.Val < 0.05 & abs(top$logFC) > 1, ] the result comes zero observation. So in that case what should I do. And will I take p value and logFC > 1 for the selection of sig gene?

[drew.b.ferrell]

What was the procedure you used to reach the list of differentially expressed genes? You could plot PCA/MDS for counts, and normalized counts. Limma doesn't work with counts data, but normalized expression. If you have large variation (which you will see in the PCA/MDS), and replicates, you can extend your procedure with voomWithQualityWeights.

You may want to find out why you do not have observations before you test for differential gene expression.

Try also using the plot = T option with voom to see mean-variance trends.

[fahim]

All the coding going well but when i tried to find top sign gene by the coding-topSig <- top[top$adj.P.Val < 0.05 & abs(top$logFC) > 1, ],i found no gene for signficant. But when i remove adj only keep p.value ,i found the signficant result.Will that result was correct or not ? Another thing i faced by analyzing another dataset and that was- by removing the log2FC i found the signficant result for the data set.This result also confused me to interpret.

[fahim]

So two of my data set which i should follow for differentially gene expression analysis,In one data set if i remove the log2FC i got the sign gene only by applying adj.p.value. Whereas in other data set by applying only p.value i got the result,there i have to avoid adj.p.value and log2FC

[Gordon Smyth]

There are many worked examples of differential expression analyses in the limma User's Guide and in online workflows. Just follow the examples.

Answer 1

[Gordon Smyth]

The fact is that limma has produced DE genes, but you have thrown them all out by requiring an additional ad hoc threshold on the fold-changes.

As the limma author, I strongly discourage people from applying fold-change thresholds on top of the limma FDR values. In the RNA-seq limma-voom context, the fold-change thresholds tend to prioritize low count genes instead of biologically significant genes. Fold-change thresholds would be useful if you were doing ordinary t-tests but, with limma empirical Bayes t-tests, the fold-change thresholds are unnecessary and actually counter-productive. Simulations show that fold-change thresholds frequently increase the true false discovery rate rather than decreasing it.

[Zhilong Jia]

which other parameter could be used to filter genes? using a more strict cut-off for p-value? will t be a good parameter to rank genes, like logFC? Thank you.

[Gordon Smyth]

limma already ranks genes for you with the topTable. There is no need for you to add on any ad hoc rules of your own.

If you have a really large number of DE genes and you want to work with a smaller list, then we recommend the treat function to prioritize genes with high-confidence large changes. See worked examples, e.g., https://bioconductor.org/packages/release/workflows/vignettes/RNAseq123/inst/doc/limmaWorkflow.html