This is a test version of Biostars. For the public version, visit https://www.biostars.org.
differential gene expression analysis (samples number)

Now I already did transcriptome analysis to get differential gene expression analysis between 2 groups (control and infected). enter image description here I have encountered an issue concerning sample (control_6h_4) being an outlier, which in turn affects the resulting number of DEG lists (just about 188 genes); for that reason, I think I will remove this sample and perform the analysis again using just 5 samples (3 infected + 2 control samples), which leads to an enhanced result of 500 genes to be differentially expressed.

Beside validating the selected genes using qPCR

My question is "Is it correct to do that concerning further publication?"

Advice, please!

thanks in advance

expression differential analysis gene
• 453 views • link
updated by swbarnes2 swbarnes2 • written by Hossam • Hossam

2 answers

It is highly problematic to remove a sample simply because you get more DE that way, without identifying a technical reason to justify why that sample is of lower quality.

On the other hand, you can use edgeR::voomLmFit with sample.weights=TRUE, which provides a statistically rigorous way to downweight outlier samples in RNA-seq DE analyses, without having to make subjective decisions about whether to remove samples. The approach is justifed and benchmarked by Liu et al (2015). The voomLmFit function is an update of the older voomWithQualityWeights function described in that paper.

This is a commonly asked question on the Bioconductor support forum.

Liu R, Holik AZ, Su S, Jansz N, Chen K, Leong HS, Blewitt ME, Asselin-Labat M-L, Smyth GK, Ritchie ME (2015). Why weight? Modelling sample and observational level variability improves power in RNA-seq analyses. Nucleic Acids Research 43, e97.

• 0 views • link
written by Gordon Smyth Gordon Smyth

I appreciate your priceless attention and advice.

However, may I justify as follows:

While initial sequencing included three control replicates, unsupervised PCA revealed that control_6h_4 acted as a severe biological outlier, exhibiting extreme transcriptomic deviation from the stable baseline of the other controls. Because Nile tilapia are an outbred species, high inherent genetic and phenotypic variance occasionally results in uncharacteristic baseline deviations due to natural individual physiology or subclinical stress. Technically, retaining this outlier would have artificially inflated the intra-group dispersion estimates within DESeq2, causing a severe loss of statistical power and an unacceptable increase in false negative rates (masking true biological signals). Therefore, in accordance with standard transcriptomic quality control for outbred animal models, this sample was excluded prior to DEG calling to ensure the control baseline accurately reflected a stable, healthy physiological state. The robust discovery of 439 highly significant DEGs, subsequently validated by our in vivo qRT-PCR data, confirms that this refined baseline yielded highly reliable biological insights.

So, what is your further advice?

Great thanks

• 0 views • link
written by Hossam • Hossam

No, you cannot justify removing an outlier in that way, despite all the strong adjectives you've packed into your paragraph. Outliers can only be removed because of technical causes, not simply because your species has high natural biological variability. Quite the opposite: if your species has "high inherent genetic and phenotypic variance" then you need to have enough biological replicates to capture and measure that variability.

• 0 views • link
written by Gordon Smyth Gordon Smyth

Can you look at the genes contributing to PC1 and determine a biological mechanism causing the problem? If you could say "This sample had high contamination of some other tissue's RNA", then you'd have a better argument for removing it.

You could also add PC1 as a batch effect.

• 0 views • link
written by swbarnes2 swbarnes2

Log in to answer this question.

More posts like this