Tutorial

Tutorial: RNA-Seq analysis part IV: Spikes and qual-

ity control
June 21, 2011

CLC bio
Finlandsgade 10-12 8200 Aarhus N Denmark
Telephone: +45 70 22 55 09 Fax: +45 70 22 55 19
www.clcbio.com [email protected]
 Tutorial: RNA-Seq analysis part IV: Spikes and quality control

Tutorial: RNA-Seq analysis part IV: Spikes and quality control

This tutorial is the fourth part of a series of tutorials about RNA-Seq analysis. We continue
working with the data set introduced in the first tutorial, although in this tutorial we are no longer
considering only a subset of the data.
In this tutorial we will focus on quality control. First, we will examine spike-ins in the data set,
and second you will learn about general quality control tools.

Inspecting the spike reads

Tutorial

The data set includes six samples that come from three groups corresponding to the three
different types of tissue. Within each group of samples, one sample has six spike-in genes,
where controlled amounts of mRNA from Arabidopsis and phage lambda templates have been
introduced in the sample prior to sequencing. We will now inspect the six spike-in genes and
check if they are expressed in the spiked samples only as we would expect.
For this and the rest of the tutorial, we will work on the full data set. The data set is so large
that on a powerful eight-core workstation computer with 32 GB RAM, it takes more than an hour
to run one sample. Therefore, we have performed the analysis beforehand and included the final
result with the files you downloaded in the first tutorial. The full RNA-Seq results with mappings
of all the mouse genes take up too much space, so they are not included. Instead, we have set
up experiments with all six samples.
Open the experiment in the RPKM folder and type "spike" in the filter to only include the six spike
genes (when we ran the full analysis, we made an artificial sequence called "Spike chromosome"
containing the six spike genes). Next, click the IQR column header to sort the genes according
to the interquartile range of expression. You should now have a view similar to figure 1.

Figure 1: Comparing the expression of the spikes.

The important thing to check here is that we see expression of the spike genes in the spike
samples but not in the rest. The six spike transcripts were added in different concentrations to
the spike samples, and this is why we see the varying levels of expression.

P. 2
 Tutorial: RNA-Seq analysis part IV: Spikes and quality control

Select all six rows, and switch to the scatter plot ( ) to visualize this trend. Choose to compare
expression value of Brain spikes and Brain, and you will be able to see some red points in the
scatter plot. Change the Dot type to Dot in the Side Panel to see the red dots more clearly. The
scatter plot is shown in figure 2.
Tutorial

Figure 2: Comparing the expression of the spikes in a scatter plot.

At first glance, you only see one clear outlier at the bottom, whereas the other red dot (or rather,
'dots') seem to have more similar expression in the two samples (close to zero in both). But if
you Zoom in ( ) on these dots, you can see that 3 of them are outliers, although to a smaller
extent (see figure 3).

Checking within and between group variability

We have now confirmed that the spike controls look fine (you can check the other two groups in
the scatter plot if you like). The next step in our quality control efforts is to check whether the
overall variability of the samples reflect their grouping. In other words we want the samples from
the same group to be relatively homogenous and distinguishable from the samples of the other
groups.

Box plot
To examine and compare the overall distribution of the expression values in the samples you
may use a Box plot ( ). Box plots can be used to get an overall impression of the locations of
the distributions, and to some extend the spread of the distributions.

P. 3
 Tutorial: RNA-Seq analysis part IV: Spikes and quality control
Tutorial

Figure 3: Comparing the expression of the spikes in a scatter plot, zoomed in.

First, create a box plot based on the RPKM-based experiment:

Toolbox | Expression Analysis ( ) | Quality Control | Create Box Plot ( )
Select the experiment that is based on RPKM expression values and click Next and Finish.
The box plot is shown in figure 4.

Figure 4: A box plot of the 6 samples in the experiment, colored by group.

The median and mean lines (you can display the mean in the Side Panel) show the median and
mean expression values in the samples and the boxes extend from the p'th to the 100-p'th
percentile of the sets of expression values in the samples. Thus, the box plot view is rather

P. 4
 Tutorial: RNA-Seq analysis part IV: Spikes and quality control

sensitive to the choice of the percentile value, and you may get a better impression of how the
distributions compare by trying different percentiles.
The distributions of the expression values are dominated by a lot of really small values and much
fewer but much larger values. To diminish the effect on the box plots of the few very high values,
you can square root transform the values and create box plots for the transformed values. First,
transform the values in the experiment:
Toolbox | Expression Analysis ( ) | Transformation and Normalization | Transform
( )
Tutorial

Select the experiment, click Next and select Square root transformation. Click Finish. Now,
create a new box plot, but this time make sure to select Transformed expression values in the
second step. Figure 5 shows the box plot before and after square root transformation.

Figure 5: To the left: original box plot based on RPKM values. To the right: box plot based on
square root transformed expression values.

After square root transformation, the distributions appear more a bit more similar.
Now, create similar box plots for the experiment based on total exon reads. You can see the
result in figure 6 that again shows the box plot before and after square root transformation, this
time based on total exon reads rather than RPKM.
Overall, the box plots indicate that the locations of the distributions of the expression values
in the samples are similar both for RPKM and for Total Exon Reads, but there is considerable
difference in the spread of the values. The distributions of RPKM values for samples for the
same tissue are highly similar. This is expected, as the samples are technical replicates and the
sample size is factored out in the RPKM (see part III of the tutorials). The high variability of the
'Total exon reads' counts is obviously related to the numbers of (mapped) reads in the samples.

Principal component analysis (PCA)

We continue to work with the experiment based on RPKM expression values. Next, we perform a
Principal Component Analysis (PCA):
Toolbox | Expression Analysis ( ) | Quality Control | Principal Component Analysis
( )
Select the experiment, click Next and Finish. This will create a PCA plot as shown in figure 7.

P. 5
 Tutorial: RNA-Seq analysis part IV: Spikes and quality control
Tutorial

Figure 6: To the left: original box plot based on total exon reads. To the right: box plot based on
square root transformed expression values.

Figure 7: A principal component analysis colored by group.

The plot shows the projection of the samples onto the two-dimensional space spanned by the
first and second principal component. (These are the orthogonal directions in which the data
exhibits the largest and second-largest variability).
The dots are colored according to the groups, and they group very nicely in the plot (the two red
red dots are on top of each other).
You can display the names of the samples in the plot using the settings under Dot properties in
the Side Panel to the right of the view (see the result in figure 8).
This PCA was done on RPKM-based expression values. As you saw in the previous tutorial, the
RPKM normalizes for the sample size. In order to show the importance of doing this for this kind
of analysis, perform a new PCA on the experiment located in the total exon reads folder.
The result is shown in figure 9.
Although the replicates still group together, the grouping is not near as clear as in figure 8 which
uses the RPKM-based expression values.

P. 6
 Tutorial: RNA-Seq analysis part IV: Spikes and quality control
Tutorial

Figure 8: Displaying the sample names.

Figure 9: PCA plot using total exon reads.

Hierarchical clustering
In order to complement the principal component analysis, we will also do a hierarchical clustering
of the samples to see if the samples cluster in the groups we expect:
Toolbox | Expression Analysis ( ) | Quality Control | Hierarchical Clustering of
Samples ( )
Select the experiment from the RPKM folder and click Next. Leave the parameters at their default
and click Finish.
This will display a heat map showing the clustering of samples at the bottom (see figure 10).
The replicates cluster nicely together as expected, and it looks like the pattern of expression is
more similar between brain and liver than between muscle and the other two groups.
Note that the heat map is not a new element to be stored in the Navigation Area - it is just

P. 7
 Tutorial: RNA-Seq analysis part IV: Spikes and quality control
Tutorial

Figure 10: Sample clustering. Adjusting the settings in the Side Panel to put the sample names at
the bottom.

another way of looking at the experiment (note the buttons to switch between different views in
figure 11.

Figure 11: Different views on an experiment.

As a conclusion to this fourth tutorial on RNA-Seq, we can confirm that the spike signals are clear
and unambiguous, and we can conclude that both the PCA and hierarchical clustering shows
that the replicates are fairly homogenous. Just as in the previous tutorial, we can see the effect
of using RPKM over total exon reads as expression measure. When doing this kind of quality
control, it may be advisory to use the RPKM expression value as this is implicitly standardized
between samples (see tutorial part II). Alternatively, you can choose to use the total exon reads
counts and then normalize them before performing quality control (you find the Normalize ( )
tool here: Toolbox -> Expression Analysis -> Transformation and Normalization).
One of the reasons why the samples cluster so nicely is that there is no biological variation in
the samples. The two samples in each group are technical replicates, so what we are really
measuring is the quality of the method. In experiments with biological replicates, you would
expect to see much more intra-group variability.

P. 8