Welcome to the blog


My thoughts and ideas

Differential Expression | Griffith Lab

RNA-seq Bioinformatics

Introduction to bioinformatics for RNA sequence analysis

Differential Expression


Differential Expression mini lecture

If you would like a brief refresher on differential expression analysis, please refer to the mini lecture.

Ballgown DE Analysis

Use Ballgown to compare the UHR and HBR conditions. Refer to the Ballgown manual for a more detailed explanation:

Create and change to ballgown ref-only results directory:

mkdir -p $RNA_HOME/de/ballgown/ref_only/
cd $RNA_HOME/de/ballgown/ref_only/

Perform UHR vs. HBR comparison, using all replicates, for known (reference only mode) transcripts:

First, start an R session:


A separate R tutorial file has been provided below. Run the R commands in your R session.

###R code###

# load the required libraries

# Create phenotype data needed for ballgown analysis

# Load ballgown data structure and save it to a variable "bg"
bg = ballgown(samples=as.vector(pheno_data$path), pData=pheno_data)

# Display a description of this object

# Load all attributes including gene name
bg_table = texpr(bg, 'all')
bg_gene_names = unique(bg_table[, 9:10])

# Save the ballgown object to a file for later use
save(bg, file='bg.rda')

# Perform differential expression (DE) analysis with no filtering
results_transcripts = stattest(bg, feature="transcript", covariate="type", getFC=TRUE, meas="FPKM")
results_genes = stattest(bg, feature="gene", covariate="type", getFC=TRUE, meas="FPKM")
results_genes = merge(results_genes, bg_gene_names, by.x=c("id"), by.y=c("gene_id"))

# Save a tab delimited file for both the transcript and gene results
write.table(results_transcripts, "UHR_vs_HBR_transcript_results.tsv", sep="\t", quote=FALSE, row.names = FALSE)
write.table(results_genes, "UHR_vs_HBR_gene_results.tsv", sep="\t", quote=FALSE, row.names = FALSE)

# Filter low-abundance genes. Here we remove all transcripts with a variance across the samples of less than one
bg_filt = subset (bg,"rowVars(texpr(bg)) > 1", genomesubset=TRUE)

# Load all attributes including gene name
bg_filt_table = texpr(bg_filt , 'all')
bg_filt_gene_names = unique(bg_filt_table[, 9:10])

# Perform DE analysis now using the filtered data
results_transcripts = stattest(bg_filt, feature="transcript", covariate="type", getFC=TRUE, meas="FPKM")
results_genes = stattest(bg_filt, feature="gene", covariate="type", getFC=TRUE, meas="FPKM")
results_genes = merge(results_genes, bg_filt_gene_names, by.x=c("id"), by.y=c("gene_id"))

# Output the filtered list of genes and transcripts and save to tab delimited files
write.table(results_transcripts, "UHR_vs_HBR_transcript_results_filtered.tsv", sep="\t", quote=FALSE, row.names = FALSE)
write.table(results_genes, "UHR_vs_HBR_gene_results_filtered.tsv", sep="\t", quote=FALSE, row.names = FALSE)

# Identify the significant genes with p-value < 0.05
sig_transcripts = subset(results_transcripts, results_transcripts$pval<0.05)
sig_genes = subset(results_genes, results_genes$pval<0.05)

# Output the significant gene results to a pair of tab delimited files
write.table(sig_transcripts, "UHR_vs_HBR_transcript_results_sig.tsv", sep="\t", quote=FALSE, row.names = FALSE)
write.table(sig_genes, "UHR_vs_HBR_gene_results_sig.tsv", sep="\t", quote=FALSE, row.names = FALSE)

# Exit the R session

Once you have completed the Ballgown analysis in R, exit the R session and continue with the steps below. A copy of the above R code is located here.

What does the raw output from Ballgown look like?

head UHR_vs_HBR_gene_results.tsv

How many genes are there on this chromosome?

grep -v feature UHR_vs_HBR_gene_results.tsv | wc -l

How many passed filter in UHR or HBR?

grep -v feature UHR_vs_HBR_gene_results_filtered.tsv | wc -l

How many differentially expressed genes were found on this chromosome (p-value < 0.05)?

grep -v feature UHR_vs_HBR_gene_results_sig.tsv | wc -l

Display the top 20 DE genes. Look at some of those genes in IGV - do they make sense?

In the following commands we use grep -v feature to remove lines that contain “feature”. Then we use sort to sort the data in various ways. The k option specifies that we want to sort on a particular column (3 in this case which has the DE fold change values). The n option tells sort to sort numerically. The r option tells sort to reverse the sort.

grep -v feature UHR_vs_HBR_gene_results_sig.tsv | sort -rnk 3 | head -n 20 | column -t #Higher abundance in UHR
grep -v feature UHR_vs_HBR_gene_results_sig.tsv | sort -nk 3 | head -n 20 | column -t #Higher abundance in HBR

Save all genes with P<0.05 to a new file.

grep -v feature UHR_vs_HBR_gene_results_sig.tsv | cut -f 6 | sed 's/\"//g' > DE_genes.txt
head DE_genes.txt


Assignment: Use Ballgown to identify differentially expressed genes from the StringTie expression estimates (i.e., Ballgown table files) which you created in Practical Exercise 8.

  • Hint: Follow the example R code above.
  • Hint: You will need to change how the pheno_data object is created to point to the correct sample ids, type, and path to StringTie results files.
  • Hint: Make sure to save your ballgown data object to file (e.g., bg.rda) for use in subsequent practical exercises.
  • Hint: You may wish to save both a complete list of genes with differential expression results as well as a subset which are filtered and pass a significance test

Solution: When you are ready you can check your approach against the Solutions

ERCC DE Analysis

This section will compare the observed versus expected differential expression estimates for the ERCC spike-in RNAs:

cd $RNA_HOME/de/ballgown/ref_only
wget https://raw.githubusercontent.com/griffithlab/rnabio.org/master/assets/scripts/Tutorial_ERCC_DE.R
chmod +x Tutorial_ERCC_DE.R
./Tutorial_ERCC_DE.R $RNA_HOME/expression/htseq_counts/ERCC_Controls_Analysis.txt $RNA_HOME/de/ballgown/ref_only/UHR_vs_HBR_gene_results.tsv

View the results here:

  • http://YOUR_IP_ADDRESS/workspace/rnaseq/de/ballgown/ref_only/Tutorial_ERCC_DE.pdf

edgeR Analysis

In this tutorial you will:

  • Make use of the raw counts you generated above using htseq-count
  • edgeR is a bioconductor package designed specifically for differential expression of count-based RNA-seq data
  • This is an alternative to using stringtie/ballgown to find differentially expressed genes

First, create a directory for results:

mkdir -p de/htseq_counts
cd de/htseq_counts

Note that the htseq-count results provide counts for each gene but uses only the Ensembl Gene ID (e.g. ENSG00000054611). This is not very convenient for biological interpretation. This next step creates a mapping file that will help us translate from ENSG IDs to Symbols. It does this by parsing the GTF transcriptome file we got from Ensembl. That file contains both gene names and IDs. Unfortunately, this file is a bit complex to parse. Furthermore, it contains the ERCC transcripts, and these have their own naming convention which also complicated the parsing.

perl -ne 'if ($_ =~ /gene_id\s\"(ENSG\S+)\"\;/) { $id = $1; $name = undef; if ($_ =~ /gene_name\s\"(\S+)"\;/) { $name = $1; }; }; if ($id && $name) {print "$id\t$name\n";} if ($_=~/gene_id\s\"(ERCC\S+)\"/){print "$1\t$1\n";}' $RNA_REF_GTF | sort | uniq > ENSG_ID2Name.txt
head ENSG_ID2Name.txt

Determine the number of unique Ensembl Gene IDs and symbols. What does this tell you?

#count unique gene ids
cut -f 1 ENSG_ID2Name.txt | sort | uniq | wc -l
#count unique gene names
cut -f 2 ENSG_ID2Name.txt | sort | uniq | wc -l

#show the most repeated gene names
cut -f 2 ENSG_ID2Name.txt | sort | uniq -c | sort -r | head

Launch R:


R code has been provided below. If you wish to have a script with all of the code, it can be found here. Run the R commands below.

###R code###
#Set working directory where output will go
working_dir = "~/workspace/rnaseq/de/htseq_counts"

#Read in gene mapping
mapping=read.table("~/workspace/rnaseq/de/htseq_counts/ENSG_ID2Name.txt", header=FALSE, stringsAsFactors=FALSE, row.names=1)

# Read in count matrix
rawdata=read.table("~/workspace/rnaseq/expression/htseq_counts/gene_read_counts_table_all_final.tsv", header=TRUE, stringsAsFactors=FALSE, row.names=1)

# Check dimensions

# Require at least 1/6 of samples to have expressed count >= 10
sample_cutoff <- (1/6)
count_cutoff <- 10

#Define a function to calculate the fraction of values expressed above the count cutoff
getFE <- function(data,count_cutoff){
 FE <- (sum(data >= count_cutoff)/length(data))

#Apply the function to all genes, and filter out genes not meeting the sample cutoff
fraction_expressed <- apply(rawdata, 1, getFE, count_cutoff)
keep <- which(fraction_expressed >= sample_cutoff)
rawdata <- rawdata[keep,]

# Check dimensions again to see effect of filtering

# Running edgeR #

# load edgeR

# make class labels
class <- c( rep("UHR",3), rep("HBR",3) )

# Get common gene names

# Make DGEList object
y <- DGEList(counts=rawdata, genes=gene_annotations, group=class)

# TMM Normalization
y <- calcNormFactors(y)

# Estimate dispersion
y <- estimateCommonDisp(y, verbose=TRUE)
y <- estimateTagwiseDisp(y)

# Differential expression test
et <- exactTest(y)

# Extract raw counts to add back onto DE results
counts <- getCounts(y)

# Print top genes

# Print number of up/down significant genes at FDR = 0.05  significance level
summary(de <- decideTestsDGE(et, adjust.method="BH", p=.05))

#Get output with BH-adjusted FDR values - all genes, any p-value, unsorted
out <- topTags(et, n = "Inf", adjust.method="BH", sort.by="none", p.value=1)$table

#Add raw counts back onto results for convenience (make sure sort and total number of elements allows proper join)
out2 <- cbind(out, counts)

#Limit to significantly DE genes
out3 <- out2[as.logical(de),]

# Order by p-value
o <- order(et$table$PValue[as.logical(de)],decreasing=FALSE)
out4 <- out3[o,]

# Save table
write.table(out4, file="DE_genes.txt", quote=FALSE, row.names=FALSE, sep="\t")

#To exit R type the following

Once you have run the edgeR tutorial, compare the sigDE genes to those saved earlier from cuffdiff:

cat $RNA_HOME/de/ballgown/ref_only/DE_genes.txt
cat $RNA_HOME/de/htseq_counts/DE_genes.txt

Pull out the gene IDs

cd $RNA_HOME/de/

cut -f 1 $RNA_HOME/de/ballgown/ref_only/DE_genes.txt | sort  > ballgown_DE_gene_symbols.txt
cut -f 2 $RNA_HOME/de/htseq_counts/DE_genes.txt | sort | uniq | grep -v Gene_Name > htseq_counts_edgeR_DE_gene_symbols.txt

Visualize overlap with a venn diagram. This can be done with simple web tools like:

To get the two gene lists you could use cat to print out each list in your terminal and then copy/paste.

cat ballgown_DE_gene_symbols.txt
cat htseq_counts_edgeR_DE_gene_symbols.txt

Alternatively you could view both lists in a web browser as you have done with other files. These two files should be here:

  • http://YOUR_IP_ADDRESS/workspace/rnaseq/de/ballgown_DE_gene_symbols.txt
  • http://YOUR_IP_ADDRESS/workspace/rnaseq/de/htseq_counts_edgeR_DE_gene_symbols.txt
Example Venn Diagram (DE genes from StringTie/Ballgown vs HTSeq/EdgeR)

Expression | Griffith Lab

RNA-seq Bioinformatics

Introduction to bioinformatics for RNA sequence analysis



Expression mini lecture

If you would like a refresher on expression and abundance estimations, we have made a mini lecture.

Use Stringtie to generate expression estimates from the SAM/BAM files generated by HISAT2 in the previous module

Note on de novo transcript discovery and differential expression using Stringtie:

In this module, we will run Stringtie in ‘reference only’ mode. For simplicity and to reduce run time, it is sometimes useful to perform expression analysis with only known transcript models. However, Stringtie can predict the transcripts present in each library instead (by dropping the ‘-G’ option in stringtie commands as described in the next module). Stringtie will then assign arbitrary transcript IDs to each transcript assembled from the data and estimate expression for those transcripts. One complication with this method is that in each library a different set of transcripts is likely to be predicted for each library. There may be a lot of similarities but the number of transcripts and their exact structure will differ in the output files for each library. Before you can compare across libraries you therefore need to determine which transcripts correspond to each other across the libraries.

Stringtie basic usage:

    stringtie <aligned_reads.bam> [options]*

Extra options specified below:

mkdir -p expression/stringtie/ref_only/
cd expression/stringtie/ref_only/

stringtie --rf -p 8 -G $RNA_REF_GTF -e -B -o HBR_Rep1/transcripts.gtf -A HBR_Rep1/gene_abundances.tsv $RNA_ALIGN_DIR/HBR_Rep1.bam
stringtie --rf -p 8 -G $RNA_REF_GTF -e -B -o HBR_Rep2/transcripts.gtf -A HBR_Rep2/gene_abundances.tsv $RNA_ALIGN_DIR/HBR_Rep2.bam
stringtie --rf -p 8 -G $RNA_REF_GTF -e -B -o HBR_Rep3/transcripts.gtf -A HBR_Rep3/gene_abundances.tsv $RNA_ALIGN_DIR/HBR_Rep3.bam

stringtie --rf -p 8 -G $RNA_REF_GTF -e -B -o UHR_Rep1/transcripts.gtf -A UHR_Rep1/gene_abundances.tsv $RNA_ALIGN_DIR/UHR_Rep1.bam
stringtie --rf -p 8 -G $RNA_REF_GTF -e -B -o UHR_Rep2/transcripts.gtf -A UHR_Rep2/gene_abundances.tsv $RNA_ALIGN_DIR/UHR_Rep2.bam
stringtie --rf -p 8 -G $RNA_REF_GTF -e -B -o UHR_Rep3/transcripts.gtf -A UHR_Rep3/gene_abundances.tsv $RNA_ALIGN_DIR/UHR_Rep3.bam

What does the raw output from Stringtie look like? For details on the Stringtie output files refer to Stringtie manual (outputs section)

less -S UHR_Rep1/transcripts.gtf

View transcript records only and improve formatting

grep -v "^#" UHR_Rep1/transcripts.gtf | grep -w "transcript" | column -t | less -S

Limit the view to transcript records and their expression values (FPKM and TPM values)

awk '{if ($3=="transcript") print}' UHR_Rep1/transcripts.gtf | cut -f 1,4,9 | less -S

Press ‘q’ to exit the ‘less’ display

Gene and transcript level expression values can also be viewed in these two files:

column -t UHR_Rep1/t_data.ctab | less -S

less -S -x20 UHR_Rep1/gene_abundances.tsv

Create a tidy expression matrix files for the StringTie results. This will be done at both the gene and transcript level and also will take into account the various expression measures produced: coverage, FPKM, and TPM.

cd $RNA_HOME/expression/stringtie/ref_only/
wget https://raw.githubusercontent.com/griffithlab/rnabio.org/master/assets/scripts/stringtie_expression_matrix.pl
chmod +x stringtie_expression_matrix.pl

./stringtie_expression_matrix.pl --expression_metric=TPM --result_dirs='HBR_Rep1,HBR_Rep2,HBR_Rep3,UHR_Rep1,UHR_Rep2,UHR_Rep3' --transcript_matrix_file=transcript_tpm_all_samples.tsv --gene_matrix_file=gene_tpm_all_samples.tsv

./stringtie_expression_matrix.pl --expression_metric=FPKM --result_dirs='HBR_Rep1,HBR_Rep2,HBR_Rep3,UHR_Rep1,UHR_Rep2,UHR_Rep3' --transcript_matrix_file=transcript_fpkm_all_samples.tsv --gene_matrix_file=gene_fpkm_all_samples.tsv

./stringtie_expression_matrix.pl --expression_metric=Coverage --result_dirs='HBR_Rep1,HBR_Rep2,HBR_Rep3,UHR_Rep1,UHR_Rep2,UHR_Rep3' --transcript_matrix_file=transcript_coverage_all_samples.tsv --gene_matrix_file=gene_coverage_all_samples.tsv

column -t transcript_tpm_all_samples.tsv | less -S
column -t gene_tpm_all_samples.tsv | less -S

Later we will use these files to perform various comparisons of expression estimation tools (e.g. stringtie, kallisto, raw counts) and metrics (e.g. FPKM vs TPM).


Assignment: Use StringTie to Calculate transcript-level expression estimates for the alignments (bam files) you created in Practical Exercise 6.

Solution: When you are ready you can check your approach against the Solutions


For more on the differences between abundance estimates like FPKM and count data with HTSeq-count, see this mini lecture.


Run htseq-count on alignments instead to produce raw counts instead of FPKM/TPM values for differential expression analysis

Refer to the HTSeq documentation for a more detailed explanation:

htseq-count basic usage:

htseq-count [options] <sam_file> <gff_file>

Extra options specified below:

Run htseq-count and calculate gene-level counts:

mkdir -p expression/htseq_counts
cd expression/htseq_counts

htseq-count --format bam --order pos --mode intersection-strict --stranded reverse --minaqual 1 --type exon --idattr gene_id $RNA_ALIGN_DIR/UHR_Rep1.bam $RNA_REF_GTF > UHR_Rep1_gene.tsv
htseq-count --format bam --order pos --mode intersection-strict --stranded reverse --minaqual 1 --type exon --idattr gene_id $RNA_ALIGN_DIR/UHR_Rep2.bam $RNA_REF_GTF > UHR_Rep2_gene.tsv
htseq-count --format bam --order pos --mode intersection-strict --stranded reverse --minaqual 1 --type exon --idattr gene_id $RNA_ALIGN_DIR/UHR_Rep3.bam $RNA_REF_GTF > UHR_Rep3_gene.tsv

htseq-count --format bam --order pos --mode intersection-strict --stranded reverse --minaqual 1 --type exon --idattr gene_id $RNA_ALIGN_DIR/HBR_Rep1.bam $RNA_REF_GTF > HBR_Rep1_gene.tsv
htseq-count --format bam --order pos --mode intersection-strict --stranded reverse --minaqual 1 --type exon --idattr gene_id $RNA_ALIGN_DIR/HBR_Rep2.bam $RNA_REF_GTF > HBR_Rep2_gene.tsv
htseq-count --format bam --order pos --mode intersection-strict --stranded reverse --minaqual 1 --type exon --idattr gene_id $RNA_ALIGN_DIR/HBR_Rep3.bam $RNA_REF_GTF > HBR_Rep3_gene.tsv

Merge results files into a single matrix for use in edgeR. The following joins the results for each replicate together, adds a header, reformats the result as a tab delimited file, and shows you the first 10 lines of the resulting file :

cd $RNA_HOME/expression/htseq_counts/
join UHR_Rep1_gene.tsv UHR_Rep2_gene.tsv | join - UHR_Rep3_gene.tsv | join - HBR_Rep1_gene.tsv | join - HBR_Rep2_gene.tsv | join - HBR_Rep3_gene.tsv > gene_read_counts_table_all.tsv
echo "GeneID UHR_Rep1 UHR_Rep2 UHR_Rep3 HBR_Rep1 HBR_Rep2 HBR_Rep3" > header.txt
cat header.txt gene_read_counts_table_all.tsv | grep -v "__" | awk -v OFS="\t" '$1=$1' > gene_read_counts_table_all_final.tsv
rm -f gene_read_counts_table_all.tsv header.txt
head gene_read_counts_table_all_final.tsv | column -t

-grep -v "__" is being used to filter out the summary lines at the end of the files that ht-seq count gives to summarize reads that had no feature, were ambiguous, did not align at all, did not align due to poor alignment quality, or the alignment was not unique.

-awk -v OFS="\t" '$1=$1' is using awk to replace the single space characters that were in the concatenated version of our header.txt and gene_read_counts_table_all.tsv with a tab character. -v is used to reset the variable OFS, which stands for Output Field Separator. By default, this is a single space. By specifying OFS="\t", we are telling awk to replace the single space with a tab. The '$1=$1' tells awk to reevaluate the input using the new output variable.

ERCC expression analysis

Based on the above read counts, plot the linearity of the ERCC spike-in read counts versus the known concentration of the ERCC spike-in Mix. In this step we will first download a file describing the expected concentrations and fold-change differences for the ERCC spike-in reagent. Next we will use a Perl script to organize the ERCC expected values and our observed counts for each ERCC sequence. Finally, we will use an R script to produce an x-y scatter plot that compares the expected and observed values.

cd $RNA_HOME/expression/htseq_counts
wget http://genomedata.org/rnaseq-tutorial/ERCC_Controls_Analysis.txt
cat ERCC_Controls_Analysis.txt

wget https://github.com/griffithlab/rnabio.org/raw/master/assets/scripts/Tutorial_ERCC_expression.pl
chmod +x Tutorial_ERCC_expression.pl
cat $RNA_HOME/expression/htseq_counts/ercc_read_counts.tsv

wget https://github.com/griffithlab/rnabio.org/raw/master/assets/scripts/Tutorial_ERCC_expression.R
chmod +x Tutorial_ERCC_expression.R
./Tutorial_ERCC_expression.R ercc_read_counts.tsv

Now let’s also create a plot for the ERCC spike-in StringTie TPM estimates versus the known concentration of the ERCC spike-in Mix.

cd $RNA_HOME/expression/stringtie/ref_only
wget http://genomedata.org/rnaseq-tutorial/ERCC_Controls_Analysis.txt
cat ERCC_Controls_Analysis.txt

wget https://github.com/griffithlab/rnabio.org/raw/master/assets/scripts/Tutorial_ERCC_expression_tpm.pl
chmod +x Tutorial_ERCC_expression_tpm.pl
cat $RNA_HOME/expression/stringtie/ref_only/ercc_tpm.tsv

wget https://github.com/griffithlab/rnabio.org/raw/master/assets/scripts/Tutorial_ERCC_expression_tpm.R
chmod +x Tutorial_ERCC_expression_tpm.R
./Tutorial_ERCC_expression_tpm.R ercc_tpm.tsv

To view the resulting figures, navigate to the below URL replacing YOUR_IP_ADDRESS with your amazon instance IP address:

Which expression estimation (read counts or TPM values) are better representing the known/expected ERCC concentrations? Why?