Time Series Transcriptomics of Human Embryonic Craniofacial Tissue

Tara N. Yankee & Dr. Justin Cotney

Showcasing RNA-seq data from 15 samples across six timepoints.

help

Experimental data

help

Row metadata

help

Quartile plots

help

Introduction

Principal component analysis is a dimension reduction technique designed to extract the strongest patterns from high-dimensional data. For more information refer to the references below, but essentially the objective is to boil the variation present in potentially tens of thousands of dimensions (e.g. genes) down to a limited number of 'components', each with a fixed contribution from all of the original dimensions. Each component then has 'loadings' representing the magnitude of these contributions from each dimension/gene.

Controls

Input matrix

The PCA is calculated here based on a matrix of features (e.g. genes, transcripts - the 'dimensions' of the data) vs samples. Controls for selecting rows and columns for the input matrix are provided under 'expression'. As with hierarchical clustering, patterns are often clearest when the more variable rows of this matrix are selected as input, so the 1000 most variant rows are selected by default.

PCA

The PCA controls principally serve to allow you to select which components are displayed. The first two or three (for 2D or 3D plots) components are shown by default, but in cases where the very strongest patterns are associated with unknown factors or experiment variables of lesser interest, you may wish to select other components.

Coloring the points by the variables associated with your samples (e.g. gender, sequencing lane etc) can reveal what factors have influenced the structure in the data. In many case it's expected that samples will group by a primary treament factor (disease vs normal etc). If the samples don't group as you expect then you're not likely to see the differential expression you may be looking for between groups. This can be for a number of resons, and it may be, for example, that you need to adjust for technical variables in your data.

Plots

Components plot

The components plot shows the first components plotted against one another, and should separate samples according to the strongest patterns in the data, hopefully (though not necessarily) by your main exerimental variables.

Loading plot

The loading plot shows which rows of the input matrix contributed most to each component, by default showing the union of the 10 strongest contributers to each component. For example, where a treatment has a strong impact on the expression of a set of genes, we expect that those genes will contribute strongly to the first few components. The nature of PCA is that components should be orthogonal to one another, the the strongest loadings at each component should be different to one another.

References

Useful beginner overviews of PCA

Plotting: the Plotly library

Sievert C, Parmer C, Hocking T, Chamberlain S, Ram K, Corvellec M and Despouy P (2015). plotly: Create Interactive Web Graphics via Plotly's JavaScript Graphing Library. http://CRAN.R-project.org/package=plotly

Loadings

help

Sample clustering dendrogram

help

Table of values by cluster

help

Fold change plot

help

MA plot

help

When looking for differences in replicate data between groups, it is necessary to examine both the magnitude of of the difference in mean value between groups as well as the variability within each group. A mean fold change of 2 is meaningless if points within each group regularly have fold changes much greater than that.

A volcano plot, and plotted on this page with Plotly [3], is designed to allow the rapid identification of large differences in replicate data [1,2]. The log(2) fold-change is shown on the x axis (preserving sign), and the -log(10) of a p-value is shown on the y axis. Points with high-magnitude changes between groups are shown at the extremes of the x axis, and where variability is low enough to make this difference statistically significant, they also appear at the extremes of the y axis, such that the top-left and top-right portions of the graph show the points of most interest.

Points above the specified thresholds of fold change and q value are shown in blue with mouse-over data available, while points below these thresholds are shown with inactive grey points. This is done for reasons of performance- labelling 10s of 1000s of points adds a lot of data and shows the browser.

Controls

Expression matrix

Only assays with associated data in the 'tests' slot of the input data structure are available. Where only a single assay has tests (p- and q- values), no matrix selection is provided.

Contrasts

The contrasts controls allow you to choose the comparison to be made, and the fold changes and q value threshold to use.

Scatter plot

This field group allows basic control on the size of the points, and whether the significant point are labelled (rarely a good idea).

Highlight points

It can often be useful to highlight sets of genes to illustrate trends. Use this set of filters to choose a set of genes to be highlighted by color on the plot.

References

[1] Cui, Xiangqin, and Gary A. Churchill. “Statistical tests for differential expression in cDNA microarray experiments. “ Genome Biol 4.4 (2003): 210.
[2] Li, Wentian. “Volcano plots in analyzing differential expressions with mRNA microarrays.“ Journal of bioinformatics and computational biology 10.06 (2012): 1231003.
[3] Sievert C, Parmer C, Hocking T, Chamberlain S, Ram K, Corvellec M and Despouy P (2015). plotly: Create Interactive Web Graphics via Plotly's JavaScript Graphing Library. R package version 2.0.16, http://CRAN.R-project.org/package=plotly.

Volcano plot

help

Introduction

This page leverages the excellent 'UpStart' set intersection of Lex, Gehlenborg et al [1], re-implemented for Shiny. Gene sets are derived from the supplied contrasts based on the supplied filters.

See the UpSet documentation for more information on interpretation.

Controls

Intersection

These controls dictate how the plotting is actually done. Intersections are complex with large numbers of sets, so you can restrict the number of contrasts considered. You'll also find that many intersections are too small to be useful, so you can change the number of intersections that will be displayed.

Intersection type

The standard behaviour of UpSet as published (and as presented by default here) is for the intersections to represent the areas of a Venn/Euler diagram. To see the consequence of this, imagine an intersection of three sets, A, B and C. The A/B area in a Venn diagram excludes the A/B/C area, so the A/B bar in an UpSet plot also excludes the A/B/C intersection. This ensures all the bars add up to the total number of entities (genes/transcripts in this case) and means that bars in the plot are exclusive, but the display can be counter-intuitive by e.g. excluding the entities from the A/B/C intersection from the A/B and B/C intersection bars.

To make the bars refelect the complete intersection of two sets, set intersection type to 'Complete', but realise that this will mean that the bars are no longer exclusive, and the values from all bars will add up to more than the total number of entities in the data.

Expression matrix

You may select any available matrix to use for generating comparisons, including the selection of a specified list of rows.

Contrasts

The choice of contrast will most tightly dictate the appearance of the plot. Choose the sample groups to compare, and the threshold on fold change (and p/q value where applicable) to apply.

Export

The plot is available for download via the provided button.

References

Lex and Gehlenborg (2014). Points of view: Sets and intersections. Nature Methods 11, 779 (2014). http://www.nature.com/nmeth/journal/v11/n8/abs/nmeth.3033.html
Gehlenborg N (2016). UpSetR: A More Scalable Alternative to Venn and Euler Diagrams for Visualizing Intersecting Sets. R package version 1.3.0, https://CRAN.R-project.org/package=UpSetR.