Up and running with pcaExplorer

Federico Marini1,2* and Harald Binder1

1Institute of Medical Biostatistics, Epidemiology and Informatics (IMBEI), Mainz
2Center for Thrombosis and Hemostasis (CTH), Mainz

*marinif@uni-mainz.de

Package

pcaExplorer 2.18.0

knitr::opts_chunk$set(crop = NULL)

1 Setup

First things first: install pcaExplorer and load it into your R session. You should receive a message notification if this is completed without errors.

BiocManager::install("pcaExplorer")
library("pcaExplorer")

This document describes a use case for pcaExplorer, based on the dataset in the airway package. If this package is not available on your machine, please install it by executing:

BiocManager::install("airway")

This dataset consists of the gene-level expression measurements (as raw read counts) for an experiment where four different human airway smooth muscle cell lines are either treated with dexamethasone or left untreated.

2 Start exploring - the beauty of interactivity

To start the exploration, you just need the following lines:

library("pcaExplorer")
pcaExplorer()

The easiest way to explore the airway dataset is by clicking on the dedicated button in the Data Upload panel. This action will:

• load the airway package
• load the count matrix and the experimental metadata
• compose the dds object, normalize the expression values (using the robust method proposed by Anders and Huber in the original DESeq manuscript), and compute the variance stabilizing transformed expression values (stored in the dst object)
• retrieve the gene annotation information via the org.Hs.eg.db, adding gene symbols to the ENSEMBL ids - this step is optional, but recommended for more human-readable identifiers to be used.

If you want to load your expression data, please refer to the User Guide, which contains detailed information on the formats your data have to respect.

Once the preprocessing of the input is done, you should get a notification in the lower right corner that you’re all set. The whole preprocessing should take around 5-6 seconds (tested on a MacBook Pro, with i7 and 16 Gb RAM). You can check how each component looks like by clicking on its respective button, once they appeared in the lower half of the panel.

Figure 1: Overview of the Data Upload panel
After clicking on the ‘Load the demo airway data’ button, all widgets are automatically populated, and each data component (count matrix, experimental data, dds object, annotation) can be previewed in a modal window by clicking on its respective button.

You can proceed to explore the expression values of your dataset in the Counts Table tab. You can change the data type you are displaying between raw counts, normalized, or transformed, and plot their values in a scatterplot matrix to explore their sample-to-sample correlations. To try this, select for example “Normalized counts”, change the correlation coefficient to “spearman”, and click on the Run action button. The correlation values will also be displayed as a heatmap.

Figure 2: Screenshot of the sample to sample scatter plot matrix
The user can select the correlation method to use, the option to plot values on log2 scales, and the possibility to use a subset of genes (to obtain a quicker overview if many samples are provided).

Additional features, both for samples and for features, are displayed in the Data overview panel. A closer look at the metadata of the airway set highlights how each combination of cell type (cell) and dexamethasone treatment (dex) is represented by a single sequencing experiment. The 8 samples in the demo dataset are themselves a subsample of the full GEO record, namely the ones non treated with albuterol (alb column).

The relationship among samples can be seen in the sample-to-sample heatmap. For example, by selecting the Manhattan distance metric, it is evident how the samples cluster by dex treatment, yet they show a dendrogram structure that recalls the 4 different cell types used. The total sum of counts per sample is displayed as a bar plot.

Figure 3: Screenshot of the sample to sample heatmap
Selected is the Manhattan distance, but Euclidean and correlation-based distance are also provided as options. In this case, the user has also selected the dex and cell factors in the ‘Group/color by’ widget in the sidebar menu, and these covariates decorate the heatmap to facilitate identification of patterns.

Patterns can become clearer after selecting, in the App settings on the left, an experimental factor to group and color by: try selecting dex, for example. If more than one covariate is selected, the interaction between these will be taken as a grouping factor. To remove one, simply click on it to highlight and press the del or backspace key to delete it. Try doing so by also clicking on cell, and then removing dex afterwards.

Basic summary information is also displayed for the genes. In the count matrix provided, one can check how many genes were detected, by selecting a “Threshold on the row sums of the counts” or on the row means of the normalized counts (more stringent). For example, selecting 5 in both cases, only 24345 genes have a total number of counts, summed by row, and 17745 genes have more than 5 counts (normalized) on average.

Figure 4: Screenshot of the Basic Summary of the counts in the Data Overview panel
General information are provided, together with an overview on detected genes according to different filtering criteria.

The Samples View and the Genes View are the tabs where most results coming from Principal Component Analysis, either performed on the samples or on the genes, can be explored in depth. Assuming you selected cell in the “Group/color by” option on the left, the Samples PCA plot should clearly display how the cell type explain a considerable portion of the variability in the dataset (corresponding to the second PC). To check that dex treatment is the main source of variability, select that instead of cell.

Figure 5: The Samples View panel
Displayed are a PCA plot (left) and the corresponding scree plot (right), with the samples colored and labeled by cell type - separating on the second principal component.

The scree plot on the right shows how many components should be retained for a satisfactory reduced dimension view of the original set, with their eigenvalues from largest to smallest. To explore the PCs other than the first and the second one, you can just select them in the x-axis PC and y-axis PC widgets in the left sidebar.

Figure 6: PCA plot for the samples, colored by dexamethasone treatment
The dex factor is the main driver of the variability in the data, and samples separate nicely on the first principal component.

If you brush (left-click and hold) on the PCA plot, you can display a zoomed version of it in the frame below. If you suspect some samples might be outliers (this is not the case in the airway set, still), you can select them in the dedicated plot, and give a first check on how the remainder of the samples would look like. On the right side, you can quickly check which genes show the top and bottom loadings, split by principal component. First, change the value in the input widget to 20; then, select one of each list and try to check them in the Gene Finder tab; try for example with DUSP1, PER1, and DDX3Y.

Figure 7: Genes with highest loadings on the first and second principal components
The user can select how many top and bottom genes will be displayed, and the gene names are printed below each gene’s contribution on each PC.

While DUSP1 and PER1 clearly show a change in expression upon dexamethasone treatment (and indeed where reported among the well known glucocorticoid-responsive genes in the original publication of Himes et al., 2014), DDX3Y displays variability at the cell type level (select cell in the Group/color by widget): this gene is almost undetected in N061011 cells, and this high variance is what determines its high loading on the second principal component.

Figure 8: Plot of the gene expression levels of DUSP1
Points are split according to dex treatment, and both graphics and table are displayed.

Figure 9: Plot of the gene expression levels of PER1
Points are split according to dex treatment.

Figure 10: Plot of the gene expression levels of DDX3Y
Points are split according to cell type, as this gene was highly variable across this experimental factor - indeed, in one cell type it is barely detected.

You can see the single expression values in a table as well, and this information can be downloaded with a simple click.

Back to the Samples View, you can experiment with the number of top variable genes to see how the results of PCA are in this case robust to a wide range of this value - this might not be the case with other datasets, and the simplicity of interacting with these parameters makes it easy to iterate in the exploration steps.

Proceeding to the Genes View, you can see the dual of the Samples PCA: now the samples are displayed as arrows in the genes biplot, which can show which genes display a similar behaviour. You can capture this with a simple brushing action on the plot, and notice how their profiles throughout all samples are shown in the Profile explorer below; moreover, a static and an interactive heatmap, together with a table containing the underlying data, are generated in the rows below.