Unexpected behavior of lfcShrink() function in DESeq2 analysis
1
0
Entering edit mode
r.r.bioinf • 0
@50c49d01
Last seen 7 weeks ago
Italy

I am currently conducting a differential expression analysis of small RNA-seq data using the DESeq2 package in R, and I have encountered unexpected behavior when using the lfcShrink() function to adjust fold-changes (FC). Specifically, I have noticed that in some cases, the lfcShrink() function appears to inflate the FC values.

I have an experimental design with 5 groups: a control group and 4 from different conditions. I filter the matrix of counts for each unique sequence so that I keep those sequences that have at least 5 counts in more than half of the samples in at least 1 group.

dds <- DESeqDataSetFromMatrix(countData = counts_table_final,
                                  colData = sampletable,
                                  design = ~ Group)
dds$Group <- relevel(dds$Group, ref ="C")
dds_2 <- DESeq(dds)

res_test <- results(dds_2, name ="Group_A2_vs_C", lfcThreshold = 0.585, alpha=0.05)
shrunk_test <- lfcShrink(dds_2, coef="Group_A2_vs_C", res=res_test)

res_test_tb <- res_test %>%
  data.frame() %>%
  rownames_to_column(var="seq") %>% 
  as_tibble()

res_test_tb$lfcShrunk <- shrunk_test$log2FoldChange
res_test_tb$lfcShrunkSE <- shrunk_test$lfcSE

In general, I see that the correction works as expected, especially by decreasing very extreme LFCs. But if I inspect the sequences I see that strange things happen, as some sequences (aprox 13%) present an LFC after shrinkage in absolute value higher than the original LFC. It happens mostly in sequences with low baseMean, that are NS and/or that we cannot annotate as endogenous sRNA (it is important to note that we perform the annotation after DE, and therefore many of those sequences may have no biological relevance).

Although in many cases the difference is not very large, I also observe some extreme cases, e.g., a sequence going from 5.9 to 17.37 LFC after adjustment (A2 vs C), here the counts:

enter image description here

In most sequences I have plotted where I see such radical changes they present very low baseMean and 0 counts in the controls. When the baseMean is higher and the controls present counts, the increase in LFC Shrunken with respect to the original LFC is more subtle.

Although these are mostly sequences of little relevance to my study, I would like to understand why this occurs.

I have carefully reviewed my data and the documentation for the DESeq2 package, but I have not been able to identify the underlying cause of this phenomenon. I suspect that it may be related to the distribution of expression values and the characteristics of the data.

DESeq2 SmallRNA • 669 views
ADD COMMENT
0
Entering edit mode
@mikelove
Last seen 16 hours ago
United States

The actual LFC is not defined here, it is some positive value vs 0, so +Inf. So the value provided by any method is a bit of a stand-in for: "cannot be estimated without deeper sequencing".

ADD COMMENT
0
Entering edit mode

Many thanks for the quick response!

I understand this for the extreme case I have presented, where the counts are 0 for controls. What happens is that I also see these strange trends in other sequences where in controls we have some counts. For example, this sequence goes from an LFC of 1.49 to 3.67. These are the counts: enter image description here

I don't know if somehow having a lot of very high LFCs because of what you comment could be affecting other more "normal" estimates. I guess a solution would be to apply stricter filtering of counts?

ADD REPLY
0
Entering edit mode

Have you made a scatterplot of the two LFC?

ADD REPLY
0
Entering edit mode

Yes, here they are.

Original LFCs: enter image description here

After shrinkage:

enter image description here

ADD REPLY
0
Entering edit mode

Sorry I meant x=regular, y=shrunken.

It strikes me that the first MA plot looks like there is a prior being used.

Can you print here what you see (the whole printout) when you print res_test to console?

ADD REPLY
0
Entering edit mode

Oh sorry, I thought you meant MAplots.

Here it is the print of res_test:

log2 fold change (MLE): Group A2 vs C 
Wald test p-value: Group A2 vs C 
DataFrame with 50042 rows and 6 columns
    baseMean log2FoldChange     lfcSE      stat     pvalue      padj
    <numeric>      <numeric> <numeric> <numeric>  <numeric> <numeric>
seq1    436439.4      1.8545515  0.537827  2.360521 0.01824930 0.0603510
seq2    407372.8     -0.7528780  0.542255 -0.309592 0.75687099 0.9716857
seq3    227136.1      2.0849424  0.505916  2.964805 0.00302875 0.0158557
seq4    105703.9      1.5603058  0.697369  1.398551 0.16194781 0.3110895
seq5    31456.8     -0.0156538  0.533533  0.000000 1.00000000 1.0000000
...                  ...            ...       ...       ...        ...       ...
seq50038    2.950406     -0.0235610   4.67351         0          1         1
seq50039    1.092356     -0.0221274   3.44701         0          1         1
seq50040    4.482006     -0.0226395   3.76443         0          1         1
seq50041    0.581461     -0.0233290   4.41936         0          1         1
seq50042    1.683103     -0.0222303   3.50389         0          1         1
ADD REPLY
0
Entering edit mode

Got it.

So a scatterplot of the two LFC against each other would be informative.

ADD REPLY
0
Entering edit mode

Perfect, here it is:

enter image description here

ADD REPLY
0
Entering edit mode

Can you repeat with the other two type options?

ADD REPLY
0
Entering edit mode

Of course

ashr:

enter image description here

normal:

enter image description here

ADD REPLY
0
Entering edit mode

So note that ashr receives the regular LFC as input (not the counts) and so it makes sense it is the most consistent with the LFC. It shrinks based on the SE and the distribution of observed LFC.

So the blob of genes with LFC ~20 is this one that's immeasurable. Shrinkage will bring them to 0 if the counts are scattered and near zero in the group that is not entirely 0's.

For the others, what is happening is likely that the counts are very heterogeneous across the groups, maybe some outliers in single groups and mostly zeros in the others.

You might chose to clean the data a bit more before modeling, e.g.:

keep those sequences that have at least 10 counts in 5 samples.

Finally, you could subset to two group comparisons, to better satisfy the modeling assumption of equal dispersion across the two groups.

ADD REPLY
0
Entering edit mode

Perfect, thank you very much!

Just one more question to make sure I got it right, you mean the counts matrix we give to DESeqDataSetFromMatrix() should be the one containing only the samples of the two groups we compare in each case? Filtered as you propose.

ADD REPLY
0
Entering edit mode

Yes that is what I meant

ADD REPLY

Login before adding your answer.

Traffic: 669 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