Hay, this question is sort of an extension of previous post. We have sequenced extra cellular RNA from culture media. Since the number of cells were limited, we did not have enough material for biological or technical replicates, so each sample has only been sequenced once, and there are essentially three groups and each group has 2 subgroups, main groups are control, group A and group B. I think the table below should help -

group subgroup Number of unique samples
Control Day 3 ~30
Control Day 5 ~5
Treatment_1 Day 3 ~30
Treatment_1 Day 5 ~50
Treatment_2 Day 3 ~40
Treatment_2 Day 5 ~20

We want to compare the gene expression (Or rather abundance of RNA) between samples based on the Day subgroup.

Since the same sample / culture batch has not been sequenced with replicates, and there is variance in number of samples per group, I was wondering if DESeq2 would be a valid option, as the documentation does mention that without replicates, the output from DESeq2 should be considered only for exploratory purpose.

If DESeq2 is not appropriate, which other tool would be better?

EDIT - I came across NOISeq which I will try once I get any output from DESEq2, if anyone has any other tools / papers which can potentially help me out, do let me know!


  • 1
    $\begingroup$ If I am understanding your question right, you are after Day 3 vs Day 5. In that sense you should be able to test this with a formula like ~ group + number-of-samples + subgroup. This will enable comparing "days" while controlling for the number of samples and treatment. You might also consider adding an interaction term group:subgroup if you think treatment (or lack of it) has differential effects on day 3 and day 5. $\endgroup$
    – haci
    Sep 26, 2023 at 8:38
  • $\begingroup$ We are after same day comparison between groups, sorry if I had not worded it correctly. $\endgroup$ Sep 26, 2023 at 8:43
  • 1
    $\begingroup$ When you are lacking replicates, most statistical tools would either complain or even refuse to run the test you would like to perform. Even when a stats tool gives you an output without replicates, you will not be able to generalize your results anyway. In your case, a gene that you get from such an analysis might be the result of biology or just some technical artefact. I don't think your questions can be answered with the data that you have in hand. $\endgroup$
    – haci
    Sep 26, 2023 at 8:57
  • $\begingroup$ I am aware of the caveats you've mentioned, I just want to do the best I can with the data we currently have. I came across NOISeq, but it doesn't seem to be in development any longer, but can work without replicates. $\endgroup$ Sep 26, 2023 at 9:29
  • 2
    $\begingroup$ It's very simple. You can fiddle with these unreplicated data and get unreliable results. That takes time and in the end will be unsatisfying. Or you repeat this experiment twice for n=3 and then have a triplicate. That's what I'd do. These rank-based analysis and custom playing with unreplicated data just never yield anything substantial. $\endgroup$
    – ATpoint
    Sep 26, 2023 at 10:11

1 Answer 1


Definitely a fan of NOIseq :-) (long story) ... there's criticism it lacks power. What you are looking at is the significance of the observation against Type 1 error. The general criticism of non-parametrics is that they have low power, so few results. Personally, I'd use it and given how widely used DSeq2 is perform the same analysis via Dseq2. I'm not a fan of parametrics in this context, but in this field they dominate and DSeq2 is the best. Thus I'd use a parametric (DSeq2) and non-parametric NOIseq. The justification is you are using one of each. Personally, I do neither, I use machine learning to direct the thresholds of a parametric - but thats outside the public domain.

The time series is more complicated. Its essentially a 2-way problem: within a time point and between time points. Binary log difference is a good analysis between two time points in this situation.

I assume there was treatment between day 3 and day 5 for the treatment samples but not the control.

The issue of replicates. Personally what I would attempt to take a short cut with replicates: if Treatment 1 and 2 have a shared impact on the sample, then treating them as replicates and compare them against the control would be the first stage of analytics. This way if you go to the expense of producing the replicates described here you can be sure you've got a result before conducting further wet-lab work.

Clearly I don't know what Treatment 1 nor Treatment 2 are, but they could be drugs attacking the same metabolic pathway. Basically to do this you need a biological rationale for combining Treatments 1 and 2 to be considered replicates. I've attempted to re-describe this approach to biological analytics below the "-----"

The post-hoc analysis is more complex. The post-hoc analysis is a database operation assessing the relative binary log fold difference between time points for treatment versus control.

@haci likely represent the orthodox approach (e.g. at review level). What I would attempt is to provide biological justification that Treatment 1 and 2 were replicates. That depends on precise justification, i.e. this time series models the impact of generic interruption of this generic class of drug on pathway X, i.e. something that biochemically combines treatment 1 and 2.

The other analysis is a within time point analysis and if you did that between Treatment 1 and Treatment 2 for day 3 and day 5 separately (2 analyses), this would be used as justification if there ain't much different within a time point. In addition if treatment is post day 3, then Treatment 1, 2, and control are replicates, i.e. a triplicate (that's unequivocal). That provides a basis for moving forward.

However, The interaction @haci is describing is likely a formal model because this is a two-way problem. I rarely disagree but you've got to be careful here because its a time-series which aren't independent events and most "interaction" based models require independence. Thus day 5 is a result of day 3, therefore they are not independent. I've encountered this as an editor and it can go very wrong. Summary analysis between time series - those are independent, analysis between data points within a time-series are not independent and interaction mixes those up.

Note Time series experiments are always complex and performing wet-lab replicates would keep things simple (but obviously more expensive). So, if you've good analytics support or you're an analyst then you probably do it without further wet-lab, otherwise send it back to the wet-lab and then follow the recipe.

  • 1
    $\begingroup$ Thanks for your input, I wouldn't mind listening to the NOISeq long story (Once I'm done with my analysis, but you're welcome to PM). The time series does complicate things and inherently Day 3 will be different than Day 5. There are a good number of samples for each group, but lets see what the analysis show. $\endgroup$ Sep 29, 2023 at 8:12

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.