[DESeq2] time series strain-specific differences over time
2
7
Entering edit mode
john ▴ 130
@john-7466
Last seen 9.4 years ago
Germany

Hello guys,

I am doing some RNA-seq analysis using DESeq2 and following the vignette from here, where they use time points 0,15,30,60 ... 180 min. and two groups (strains)

http://www.bioconductor.org/help/workflows/rnaseqGene/

In the time series chapter it says:

"Genes with small p values from this test are those which, at one or more time points after time 0 showed a strain-specific effect."

As an example the following is shown:

## log2 fold change (MLE): strainmut.minute180 
## LRT p-value: '~ strain + minute + strain:minute' vs '~ strain + minute' 
## DataFrame with 4 rows and 7 columns
##               baseMean log2FoldChange     lfcSE      stat       pvalue         padj      symbol
##              <numeric>      <numeric> <numeric> <numeric>    <numeric>    <numeric> <character>
## SPBC2F12.09c  174.6712    -2.65763737 0.7498270  99.23199 7.671942e-20 5.184698e-16       atf21

Question 1)

As I am understanding it- this gene has a strain specific expression profile over time. This effect occures over the other time points (from 0 to 180).

If I had picked the strainmut.minute60  then we would have seen a list of DE genes from 0 to 60 minutes, right?

 

 

Just for clarification- here are the options of which one can choose:

resultsNames(ddsTC)

##  [1] "Intercept"           "strain_mut_vs_wt"    "minute_15_vs_0"      "minute_30_vs_0"     
##  [5] "minute_60_vs_0"      "minute_120_vs_0"     "minute_180_vs_0"     "strainmut.minute15" 
##  [9] "strainmut.minute30"  "strainmut.minute60"  "strainmut.minute120" "strainmut.minute180"

Question 2)

Imagine a gene that behaves in the same way in all timepoints (between the two groups) but time point 0 to 15.

Will this be detected by deseq?

 

 

Cheers,

john

deseq2 timecourse • 15k views
ADD COMMENT
10
Entering edit mode
@mikelove
Last seen 8 hours ago
United States

"If I had picked the strainmut.minute60  then we would have seen a list of DE genes from 0 to 60 minutes, right?"

This is a model with interaction terms, and this is an interaction term. .You can tell it is an interaction term because it contains the names of both variables strain and minute. So this term is a test for if the mut vs WT fold change is different at minute 60 than at minute 0. 

To generate the tables of log fold change of 60 minutes vs 0 minutes for the WT strain would be:

results(dds, name="minute_60_vs_0")

To generate the tables of log fold change of 60 minutes vs 0 minutes for the mut strain would be the sum of the WT term above and the interaction term which is an additional effect beyond the effect for the reference level (WT):

results(dds, contrast=list(c("minute_60_vs_0","strainmut.minute60"))

"Imagine a gene that behaves in the same way in all timepoints (between the two groups) but time point 0 to 15. Will this be detected by deseq?"

This is the description of an interaction term in the model. You can find this by testing the interaction term strainmut.minute15:

results(dds, name="strainmut.minute15")
ADD COMMENT
0
Entering edit mode

Dear Michael, many thanks for your help and time.

Do I understand correctly, that by default the results() function would use test LRT in this case. Hence, for each time point comparison the results fold change would change but the p-value is the same e.g. it is the p-value of the comparison across all time points.

results(dds, name="minute_60_vs_0")

However, in case I am not only interested in the log fold change between two time points but also the p-values of that comparison would it be correct to use the Wald test?

results(dds, name="minute_60_vs_0", test="Wald")

Kind regards, Florian

ADD REPLY
1
Entering edit mode

I will say it is possible but a bit cumbersome to perform an LRT for 60 vs 0. You would have to re-run DESeq() or minimally nbinomLRT with full and reduced as design matrices, where reduced removes the column associated with 60 vs 0 coefficient.

ADD REPLY
0
Entering edit mode

Dear Michael, thank you very much!

I think I do have a intuition of the problem. It is still correct to use the following to compare WT vs MUT at minute 15 controlling for baseline:

results(dds, name="strainmut.minute15", test="Wald") 

To compare time points within a strain I would then just subset the data by strain and compute a LRT with a time variable only. I could then ask for responder genes of a given strain over all time points and also do between time point comparisons using the Wald test as shown above. I think that way I do less harm.

ADD REPLY
1
Entering edit mode

Yes, this gives mut vs wt at minute 15 comparing to baseline.

ADD REPLY
0
Entering edit mode
john ▴ 130
@john-7466
Last seen 9.4 years ago
Germany

Hello Michael,

You said "So this term is a test for if the mut vs WT fold change is different at minute 60 than at minute 0. "

Alright. A gene that has a low pvalue in strainmut.minute60 AND 30 AND 120 AND 180 is then to be considered DE in all time points (compared to 0) between WT and MUT.

----------------

You said:

"To generate the tables of log fold change of 60 minutes vs 0 minutes for the WT strain would be:

results(dds, name="minute_60_vs_0")"

So this shows me genes that are DE between two time points in one group (WT). Not considering the MUT at all.

---------------

You said: "reference level (WT):"

Why is WT our reference? Where did we specify that WT is the reference and not MUT? Why this way?

 

-------------

 

By calling 

results(dds, name="strainmut.minute180")

Deseq finds this gene because it takes 0minutes and 180minutes into account. That's alright.

Is there a way to take all time points into account? In order to find a gene that is DE in all time points between two groups?

 

ADD COMMENT
2
Entering edit mode

See the DESeq2 vignette for details on how to tell R which level is the reference level for a factor variable.

In this case the reference level was set during object construction which can be followed at the fission package vignette.

Is there a way to take all time points into account? In order to find a gene that is DE in all time points between two groups?

If you want to find genes which have a significant mut vs wt fold change for each and every time point, I don't think there is a simple way to form this results table. You would have to test the fold change for mut vs WT at minute 0:

results(dds, name="strain_mut_vs_wt")

...and then test the fold change for mut vs wt at every other time point:

results(dds, contrast=list(c("strain_mut_vs_wt","strainmut.minute15")))

And you would then look for genes which have a small adjusted p-value in all of the results tables.

To clarify: the first comparison takes into account the difference at t0, and this is recommended. The second doesn't take into account the difference at t0, and this is not usually recommended.

ADD REPLY
0
Entering edit mode

So if we do

results(dds, contrast=list(c("strain_mut_vs_wt","strainmut.minute15")))

and then

results(dds, contrast=list(c("strain_mut_vs_wt","strainmut.minute30")))

and on and on and then just select the genes that have an adjusted pvalue less than some threshold for all these tests, would this give the correct significance? I mean all of these tests share strainmut.minute0 as the control so they are correlated. How can we model this correlation? Instead, is there a way to test for a composite hypothesis such as  \beta_1 >0 AND \beta_2 >0 where \beta_1 and \beta_2 are two different fold change effects with the same baseline?

ADD REPLY
1
Entering edit mode

This isn't a concern in linear models, think about the coefficient table that lm() returns for example when there are 3 or more factor levels. The assumption is that the data is not correlated conditioned on the design matrix X.

No, we don't have a way to test such composite alternative hypotheses.

ADD REPLY
0
Entering edit mode

I am interested in genes that are DE at each time point (i.e. ALL time points) with respect to 0. I ran some simulations for the lm model at http://rpubs.com/ysdel/corpvals and it seems that the tests are correlated. I don't know how to think about this correctly to find the significance levels.

ADD REPLY
1
Entering edit mode

I don't really follow your question, because I don't see where we make the assumption that test statistics from different contrasts would be independent.

Nevertheless, we don't have in DESeq2 a way for you to generate a p-value where the alternate hypothesis is that all time points after t0 have consistent sign of LFC compared to t0.

ADD REPLY
0
Entering edit mode

I'm sorry, I wasn't clear in what the problem is. Suppose I have results res1, res2, res3 from correlated tests. And I am interested in genes that satisfy the alternative hypothesis for all of these tests. So I will have to do

hits = (res1$padj<0.05) & (res2$padj<0.05) & (res3$padj<0.05)

[or, forgetting about multiple testing adjustments,

hits = (res1$p.value<0.05) & (res2$p.value<0.05) & (res3$p.value<0.05) ]

What is the statistical significance (in terms of FDR or Type I errors) of these hits? If I had a simple hypothesis, I know for (res1$padj<0.05) I will make a type I error 5% of the time.

I very much appreciate you taking the time to answer these questions.

ADD REPLY
1
Entering edit mode

You're correct that that statement doesn't have the same control as the tests we provide. We don't have functionality for this kind of combination in DESeq2. You could use pvalue combination methods.

ADD REPLY
0
Entering edit mode

Hi Michael,

I am analyzing RNA-Seq data from a time-course of a mutant strain and would like to ask you something related to this reply. 

Following what is described in the time-course vignette

- In order to calculate the fold change for mut vs wt at an individual time-point shouldn't we use the Wald test like in:

results(dds, name="strainmut.minute15", test="Wald") 

 

What is the difference between using the one above and the following contrast, which as you describe, "(...) test the fold change for mut vs wt at every other time point" (time-point 15 in this particular case):

results(dds, contrast=list(c("strain_mut_vs_wt","strainmut.minute15")))

 

Many thanks!

Miguel

ADD REPLY
2
Entering edit mode

"strainmut.minute15" is the difference between Mut vs WT at minute 15, controlling for baseline. If you add "strain_mut_vs_wt" to this, you get the LFC for Mutant vs WT at minute 15, not controlling for baseline. So the second one is the observed difference at minute 15 between the two groups (because you added in the change that was present at time=0).

ADD REPLY
0
Entering edit mode

Hi Michael, Just went through this post and I have a question. I have an experiment with 3 conditions (let's called them A/B/C) at 3 different time-points (0h/12h/24h) Following the tutorial I set

dds<-DESeqDataSetFromMatrix(countData = matrix, colData = colData, design = ~ condition + time + condition:time) dds <- DESeq(dds, test="LRT", reduced=~condition+time)

Thus, running

res<-results(dds, contrast=list(c("condition_A_vs_B", "conditionA.time12h")))

will give me the genes specifically DE at 12h, regarding if they are or are not DE at 0h I guess the same apply for 24h, that is, using

contrast=list(c("condition_A_vs_B", "conditionA.time24h")

will give me genes specifically DE at 24h regarding if they are or are not DE at 0h AND 12h, right?

Now, to have genes that are DE ONLY at 0h, should I just use "condition_A_vs_B" ? Moreover, what if I need to find genes that are DE in one condition (say condition A) between different time points (e.g. 12h vs 0h) ?

Thank you

Dave

ADD REPLY
1
Entering edit mode

This gives the condition effect at time 12 and 24 hrs, respectively.

If you want DE at a time point, and not DE at other time points, you can intersect gene lists.

In order to define "not DE", I recommend using lfcThreshold > 0 and altHypothesis="lessAbs".

ADD REPLY
0
Entering edit mode

Thank you Michael. Regarding the the first question I mispelled it, I meant "to have genes that are DE at time 0h" without "ONLY" Should then "condition_A_vs_B" works?

Regarding comparison of 2 condition at a specific time-point, If I would extract, from the whole expression matrix, a subset containing only samples at that specific time-point and run "standard" DESeq using contrast=c("condition", "A", "B") , would I have the same results?

Thank you again

Dave

ADD REPLY
1
Entering edit mode

Yes.

Similar but not same, see vignette FAQ.

ADD REPLY
0
Entering edit mode

Hi Michael,

I run the comparison of condition A vs B at all the 3 time-points:

res_0h_A_vs_B<-results(dds, name = "condition_A_vs_B")
res_0h_A_vs_B<-res_0h_A_vs_B[order(res_0h_A_vs_B$padj),]

res_24h_A_vs_B<-results(dds, contrast=list(c("condition_A_vs_B", "conditionA.time24h")))
res_24h_A_vs_B<-res_24h_A_vs_B[order(res_24h_A_vs_B$padj),]

res_48h_A_vs_B<-results(dds, contrast=list(c("condition_A_vs_B", "conditionA.time48h")))
res_48h_A_vs_B<-res_48h_A_vs_B[order(res_48h_A_vs_B$padj),]

Now, looking at the top genes, I get exactly the same values in all the 3 time-points, except for the log2FC and lfcSE which is odd...

For example:

@0h
    baseMean    log2FoldChange  lfcSE   stat    pvalue  padj
ENSG00000167723.15  700.416717013728    -2.17769696510155   0.273584252249565   101.361985430641    5.04493265941981e-21    7.37014212214639e-17

@12h
    baseMean    log2FoldChange  lfcSE   stat    pvalue  padj
ENSG00000167723.15  700.416717013728    0.163662883602135   0.278211085104576   101.361985430641    5.04493265941981e-21    7.37014212214639e-17

@24h
baseMean    log2FoldChange  lfcSE   stat    pvalue  padj
ENSG00000167723.15  700.416717013728    -0.386697187118573  0.28539382302352    101.361985430641    5.04493265941981e-21    7.37014212214639e-17

Any clue why I get the same values for baseMean/stat/pval/padj ?

Thank you

Dave

ADD REPLY
1
Entering edit mode

If you have specified an LRT, you'd need to use test="Wald" here to perform Wald tests of your contrasts.

ADD REPLY
0
Entering edit mode

Hi Michael, sorry to bother you but going back to this post I realized I still need an answer to my question:

"What if I need to find genes that are DE in one condition (say condition A) between different time points (e.g. 12h vs 0h) ?"

I would really appreciate if you could help me Thank you

Dave

ADD REPLY
0
Entering edit mode

Thank you Michael and sorry for my late reply but I got stuck on another project and forgot to reply. The workflow in the "Time course" experiment" it doesn't seem to answer my question (or at least I don't get it...). I read that the test showed in the text reports those genes that at some points after 0 show a strain specific effect. But how can I get genes that, in just one specific condition, change over time? Could I just use the read counts matrix of that specific condition in the 3 time points, use design = ~ time and then test between different time points? Thank you again

Dave

ADD REPLY
1
Entering edit mode

Start at the sentence: "Wald tests for the log2 fold changes at individual time points can be investigated using the test argument to results..."

For the reference condition you can just pull out results for e.g. minute_15_vs_0 with name (the vignette section on Interactions explains how this happens with a diagram).

For non-reference condition, you just add the main effect minute_15_vs_0 and the interaction term, e.g.

results(dds, contrast=list(c("minute_15_vs_0","strainmut.minute15")))
ADD REPLY
0
Entering edit mode

Got it Thank you very much

Dave

ADD REPLY
0
Entering edit mode

Hi Michael,

Just a quick question on the answer you gave for "Is there a way to take all time points into account? In order to find a gene that is DE in all time points between two groups?" 

I followed your tutorial on the fission data for time series experiments, and i thought that taking a summary of results > summary(resTC) after running the steps below would give an indication of the total genes differentially expressed across all time-points under consideration:

ddsTC <- DESeq(ddsTC, test="LRT", reduced = ~ strain + minute)
resTC <- results(ddsTC)
resTC$symbol <- mcols(ddsTC)$symbol
head(resTC[order(resTC$pvalue),],4)

Kindly clarify this part for me, because these are the same differentially expressed genes you then went on to use for construction of the heatmap

topGenes <- head(order(resTC$pvalue),40)

So which time-point are these representing??

Thanks,

Stan

ADD REPLY
2
Entering edit mode

The likelihood ratio test here is testing genes for condition-specific differences over time after time 0. Here's the relevant section from the workflow:

"The following chunk performs a likelihood ratio test, where we remove the strain-specific differences over time. Genes with small p values from this test are those which, at one or more time points after time 0 showed a strain-specific effect. Note therefore that this will not give small p values to genes which moved up or down over time in the same way in both strains."

 

 

ADD REPLY

Login before adding your answer.

Traffic: 695 users visited in the last hour
Help About
FAQ
Access RSS
API
Stats

Use of this site constitutes acceptance of our User Agreement and Privacy Policy.

Powered by the version 2.3.6