possible reasons that very few significant DEGs from RNAseq

Question

[dew]

Dear all,

What are the possible reasons if the treatment vs control finds very very few significant DEGs(padj < 0.05, abs(log2FC) > 1) ?

The PCA plot also shows that the control group and treatment group samples are mixed together...

Thank you very much!

Answer 1

[ATpoint]

No DEGs in reality.

Low sample size and low depth.

Noise / confounders in data / poor library quality.

Flawed / wrong analysis and code, or suboptimal or underpowered (or wrong) experimental design during analysis.

That's the big 4. For a more elaborate data please add any detail on your data/samples/setup and analysis.

[matthew.pace]

To add some detail to the first point - its essential to validate your in silico results with in vitro ones. Did the cells change in morphology after treatment? Is there a good reason to believe that the treatment was indeed effective?

[dew]

Thanks a lot!

The experimental side states have found changed phenotype....

[dew]

Thanks a lot!

We have 3 replicates for control, and 3 replicates for treatment.

The 3 replicates are distant from each other in the 3D PCA, while for each replicate, the control and treatment are so close(seems the antibody effects are so small).

After DEseq2 and limma, and then the enrichment analysis, both pathways' annotations are very similar, and DEGs are all very very few(below 10) after padj < 0.05 and abs(log2FC) >1. Also checked several times of the code, and did not find an obvious bug to date....

[ATpoint]

What are you comparing? If you describe your setup a bit more one might come up with a more powerful design. Is this a paired analysis, so do you have like three donors and you expose cells from it to a treatment or mock?

Answer 2

[vj]

Given the technical side is very good, some more obvious conclusions are

1. Treatment did nothing significant.

2. Quite a bit of variability among the replicates.

Since PCA shows that control and treatment group are mixed together then I would say (1) is more likely than (2). Worth get the distance/similarity matrix between the samples in the analysis and performing a hierarchically clustered and heatmap. That would tell you if there is clear branching between treatment and control and also if the replicates are similar to each other within each condition.

[dew]

Thanks a lot!

The correlation among the samples is quite high as 0.5-0.7. Still, it's hard to understand the treatment and the control are so so close.

Thank you very much for your kind guidance!

[ATpoint]

Sample correlation hardly means anything in high-throughput data.