EdgeR or DESeq2? Comparing the performance of differential expression tools

It seems like this discussion comes up a lot. Choosing a differential expression (DE) tool could change the results and conclusions of studies. How much depends on the strength of the data. Indeed this has been coveres by others already (here, here).

In this post I will compare these tools for a particular dataset that highlights the different ways these algorithms perform. So you can try it out at home, I've uploaded the code for this to GitHub here. To get it to work, clone the repository and open the Rmarkdown (.Rmd) file in Rstudio. You will need to install Bioconductor packages edgeR, DESeq2 and the CRAN package eulerr (for the nifty venn diagrams). Once that's done you can start working with the Rmd file. For an introduction to Rmd  read on here. Suffice to say that it's the most convenient way to generate a data report that includes code, results and descriptive text for background information, interpretation, references and the rest. To create the report in Rstudio, click the "knit" button and it will execute the code and you'll get a HTML report. If you're working in the R console, for example if you are connecting to a server via SSH, then you can generate the report using the command rmarkdown::render("myreport.Rmd") and you will get the output HTML file.

In case you're impatient and just want to see the result for yourself, I've hosted my output report at this link. For everyone else, here's the walkthrough. 

Multidimensional scaling analysis

Always do an MDS because it will give you an idea of the intra and inter group differences. As you can see here, there is a good measure of both that distinguishes this n=3 experiment. (The counts were derived from an ATAC-seq experiment).

DE analysis

Here is the summary of the results with the different tools:

edgeR (LRT) Up: 297 Down:589
edgeR (QL) Up:15 Down:204
DESeq2 Up:3 Down:113
Voom Up:100 Down:128

You can see the massive difference between them. In particular between edgeR LRT and DESeq. Let's take a look. The Euler package venn diagrams are quite easy to make and the areas of the shaded parts are proportional which is a nice touch.

I selected 9 genes that were significant with edgeR that were discordant with DESeq2 and created barcharts. Gene9565 in demonstrates that the algorithms deviate in how they treat outliers. 

Here are the stats:

EdgeR

              logFC   logCPM       LR       PValue        FDR dispersion

Gene9565 -0.7916072 2.291451 11.07955 0.0008728513 0.04036433   0.064547

               C1       C2       C3       T1       T2       T3

Gene9565 21.96188 14.17148 40.87624 24.08976 24.94944 20.96064


DESeq2

         baseMean log2FoldChange     lfcSE      stat    pvalue padj       C1

Gene9565  24.7313     -0.8373232 0.3698349 -2.264046 0.0235713   NA 7.203133

               C2      C3       T1       T2       T3

Gene9565 7.148011 7.06639 6.875975 6.935422 6.907739


The T2 sample is clearly a bit of an outlier here but it is being marked as statistically significant by edgeR and NA by DESeq2, because of the outlier detection. 

The next gene I want to highlight is Gene28045. 

EdgeR

               logFC   logCPM      LR       PValue        FDR dispersion

Gene28045 -0.3876685 5.166385 13.8639 0.0001965379 0.01514142 0.00657676

                C1       C2       C3       T1       T2       T3

Gene28045 148.1977 95.62844 275.8308 215.6399 223.3354 187.6296


DESeq2

          baseMean log2FoldChange     lfcSE      stat     pvalue      padj

Gene28045 183.5495     -0.4058583 0.1610077 -2.520739 0.01171087 0.4080693

                C1       C2       C3       T1       T2       T3

Gene28045 8.468181 8.443831 8.337265 8.230507 8.135978 8.113126


Clearly this is a problem because there's no way that these genes could be validated with qPCR and t-test and get a p-value anywhere near 0.05. Therefore I'm happy to recommend people not use edgeR LRT anymore and go with the more conservative DESeq2.

What about the other tools?

Based on twitter discussions I also assessed edger with the QL test, which gave results more similar to DESeq.
Voom-limma as well

One thing I'm curious about is how the results from DESeq2 are biased with very few upregulated genes, which makes sense because this is a knock-down experiment. However with Voom-Limma, there is roughly equal number of up and downregulated genes.

UpSet plots can be useful to visualise commonlalities between three or more lists. Let's see how it looks for the 4 methods above. Firstly for the upregulated genes:

Then for the downregulated genes:




Popular posts from this blog

Data analysis step 8: Pathway analysis with GSEA

Image
In our RNA-seq series so far we've performed differential analysis and generated some pretty graphs, showing thousands of differentially expressed genes after azacitidine treatment. In order to understand the biology underlying the differential gene expression profile, we need to perform pathway analysis. We use Gene Set Enrichment Analysis ( GSEA ) because it can detect pathway changes more sensitively and robustly than some methods. A 2013 paper compared a bunch of gene set analyses software with microarrays and is worth a look. Generate a rank file The rank file is a list of detected genes and a rank metric score. At the top of the list are genes with the "strongest" up-regulation, at the bottom of the list are the genes with the "strongest" down-regulation and the genes not changing are in the middle. The metric score I like to use is the sign of the fold change multiplied by the inverse of the p-value, although there may be better methods out there
Read more

Two subtle problems with over-representation analysis

Image
Over-representation analysis is a helpful and really frequently used technique for understanding trends from omics data and gene lists more generally. Just to demonstrate this, I tabulated the number of citations of some of the most popular web tools and packages, which reaches a massive 191k citations! But if you have used some of these tools, you will notice that they yield subtly different results. We were curiuous about that and ran a whole bunch of investigations into the internal workings of these tools, in particular clusterProfiler. We identified two subtle problems with clusterProfiler, which we unpack in detail in our new publication , but here I will give you a quick overview. Problem #1: The background problem The first one we call the “background problem,” because it involves the software eliminating large numbers of genes from the background list if they are not annotated as belonging to any category. This results in removing a large number of unannotated genes from the b
Read more

Uploading data to GEO - which method is faster?

If you have had to upload omics data to GEO before, you'll know it's a bit of a hassle and takes a long time. There are a few methods suggested by the GEO team if you are using the Unix command line: Using 'ncftp' ncftp set passive on set so-bufsize 33554432 open  ftp://geoftp:yourpasscode@ftp-private.ncbi.nlm.nih.gov cd uploads/your @mail.com_ yourfolder put -R Folder_with_submission_files Using 'lftp' lftp  ftp://geoftp:yourpasscode@ftp-private.ncbi.nlm.nih.gov cd uploads/ your @mail.com _ yourfolder mirror -R Folder_with_submission_files Using 'sftp' (expect slower transfer speeds since this method encrypts on-the-fly) sftp  geoftp @s ftp-private.ncbi.nlm.nih.gov password: yourpasscode cd uploads/ your @mail.com _ yourfolder mkdir new_geo_submission cd new_geo_submission put file_name Using 'ncftpput' (transfers from the command-line without entering an interactive shell) Usage example: ncftpput -F -R -z -u  geoftp  -p "yourpasscode"
Read more