Hello, We are investigating the knockdown effect of a specific long non-coding RNA lncRNA in cancer cell lines in an attempt to identify downstream target genes of this lncRNA. After knock-down(KD) we performed RNA-seq on KD vs wild type (WT) cells , and followed standard bioinformatics pipeline for quality checks and differential expression using DEseq2 (default parameters). We did not detect any deferentially expressed genes. There are several challenges with the current experiment: Target lncRNA is extremely lowly expressed (featureCounts raw counts =~ 12) The best achieved knockdown efficiency in wet-lab experiment was around 30% confirmed by QPCR. PCA plot shows some bacth effects between replicates (3 replicates for KD and WT) Correlation plots show 90% similarity between KD and WT samples I would like to find out if there is a way to detect the effect of knocking down a low expressed lncRNA on downstream targets where the expected number of deferentially expressed genes between KD and WT conditions is expected to be small. I am not sure if I am missing anything with DEseq2 parameters or if I should adopt a different approach for comparing the gene expression of the two conditions. Thank you
Well, not sure how to further help you from the bioinformatics side. You simply have no DEGs based on the current setup plus that potential batch effect. Probably have to go back to the lab and repeat the experiment. Was the gene that you checked by qPCR well-expressed or is it on the very left of the MA-plot? If so it is probably not too informative as you need either very large effects or many replicates to get lowly-expressed genes significant in RNA-seq. Best would probably to select qPCR genes based on their expression which should be moderate to high to have a decent chance to be significant in RNA-seq even at low replication numbers.
I agree with checking overall clustering, as well as increasing the FDR cutoff and/or filtering genes that were not expressed.
That said, you are sometimes not going to be able predict which method you need to use ahead of time. So, it sounds like there may be value in testing other methods (like edgeR, limma-voom, regular ANOVA, etc.). I have some notes on that here:
(however, that essentially says that you will eventually encounter a problem with a given method if you lock-down the methods and test enough experiments - so, perhaps just saying "it may be worth testing other methods" is sufficient)