Hello,
This is my first post on the scverse forum - I could not find a way to tag decoupler here, but I hope my questions will still reach the right people!
I have a few questions related to gene set quantification, particularly in bulk RNASeq.
In the original decoupleR method, it was mentioned that quantification could be performed using 2 types of input: a normalized expression matrix or the results of dea ( Pathway activity inference in bulk RNA-seq • decoupleR ). However, it seems that the use of dea results is recommended in the newer versions ( Bulk Enrichment Analysis — decoupler ).
Could you please explain or point me to the reasons of this preference?
I found it convenient to obtain pathway quantification per patient and then test for differences between groups. For example, we may have clinical parameters such as smoking or alcohol consumption, for which we don’t want to perform DEA but rather look at specific pathways, like inflammation.
I have been quantifying PROGENy pathways from a bulk RNASeq dataset using both inputs, following the tutorial in the decoupleR Github. The results are pretty different: the pathways showing the most differences using the dea inputs are not the same as the pathways that show significant group differences (wilcoxon tests) when using normalized expression as input.
Has this discrepancy been observed already?
Thanks in advance for your help.
Best,
Jane
Hi,
I’ll add a decoupler tag and pinged the right person to answer this. Thank you!
1 Like
Dear @merlevede,
Thanks for asking. Both approaches, computing enrichment scores at the observation (sample) level or at the contrast (DEA) level, are valid, but the choice depends on your study design and end goal.
If your main goal is to identify differences between groups, I’d recommend performing DEA first and then computing enrichment scores from the differential statistics. This approach already accounts for the group contrast and allows you to include covariates (e.g., smoking, alcohol consumption) directly in the model, something that simple group-wise tests like Wilcoxon cannot handle.
On the other hand, if your variable of interest is continuous (for example, age, a pseudotime trajectory, or spatial gradients), it’s more appropriate to compute enrichment scores at the observation level and then use statistical models that handle continuous predictors (e.g., linear regression, correlation, or machine learning models like XGBoost). This lets you capture gradual or non-linear effects without forcing artificial groupings.
I hope this helps clarify the rationale behind both approaches and why the results may differ. Please feel free to follow up if anything is unclear or if you’d like to discuss your setup in more detail.
Best,
Pau
Dear @PauBadiaM,
Thank you for your insightful reply. This explanation helped me to clarify when to use each approach.
I still have one doubt. Let’s consider the case of a NSCLC dataset with the 2 main histologies and variable PFS (high/low). We might want to look at deregulations induced by histology and deregulations observed between PFS high and low.
From your explanation, it seems clear that we should compute enrichment scores from the DEA results for these comparisons. However, we might also want to investigate whether smokers show deregulations in inflammation pathways for example. In this case, would you recommend performing a DEA between smokers groups and look at the enrichment scores at the contrast level or relying on the enrichment scores computed at the sample level?
I am currently working on 2 cohorts of renal cancers cohorts, where I compared the deregulated pathways obtained using both approaches: sample level and contrast level.
- For the first cohort, when looking at the PFS status, I found 2 deregulated pathways at the sample level. The most deregulated pathway was the same in both approaches but the second one did not match.
- When analyzing another tumor characteristic, I got 5 deregulated pathways at the sample level. The most deregulated pathway (“by far” from the others) was not retrieved at the contrast level. Only one of them overlapped and was the most deregulated one in the contrast level results.
- For the second cohort, focusing on progression, I found 3 deregulated pathways at the sample level. Again, the most deregulated pathway (“by far” from the others) was not retrieved at the contrast level and only one of them overlapped, which was not the most deregulated one in the contrast level results.
Do you generally expect such discrepancies between the 2 approaches?
Finally, I have a question regarding interpretation: which range of scores can be considered significant? I am looking at the mlm results particularly, where the highest scores I saw range between 5 and 15 depending on the analyses.
Thank you again for your help and for taking the time to clarify these points.
Jane