Help explaining MDS plot of RNAseq?

Question

[Dayna]

Hello everybody

Can any one help me understand the MDS plot for two groups of samples like control vs. treated we generate from edgeR or Deseq2. I understand the figure as it separates the control from treated very well, but if we have the y-axis as the logFC, I don't get what the x-axis is? What is the x-axis in such 2D figure?

Thanks

Answer 1

[Kevin Blighe]

Hello, the y-axis in a MDS plot does not represent logFC, and neither does the x-axis.

Provided you have generated the plot according to the workflow, RNA-seq workflow: gene-level exploratory analysis and differential expression, on Bioconductor, your x- and y-axes will be representative of Euclidean distances between your samples. These Euclidean distances will be produced/repeated across multiple dimensions in different ways, but the standard way of representing MDS is to just plot the Euclidean distances from the first 2 dimensions, with x-axis being Dimension 1 and y-axis being Dimension 2. The dimensions are ordered based on how well they fit your samples (for further reading, take a look at the cmdscale Details section.

MDS is commonly used in genetic studies to find relationships between samples based on genotype, but, as you can see, it's also used in RNA-seq. For information on a related method, i.e., principal components analysis, please see my threads here:

• A: PCA in a RNA seq analysis
• A: PCA plot from read count matrix from RNA-Seq
• Produce PCA bi-plot for 1000 Genomes Phase III in VCF format

---

For what it's worth:

• A volcano plot represents logFC (x-axis) and negative log10 P- or adjusted P-value (y-axis)
• A MA plot represents average expression (x-axis) and logFc (y-axis)

[Dayna]

Thank a lot Kevin. Can you please give me an example of what could a dimension 1 and 2 be, just as an example to get the idea more?
thanks.

[Kevin Blighe]

Hi! You can say that each dimension represents similarity / dis-similarity between your samples of interest, but that's effectively it. They are not directly measuring any parameter in your data. Euclidean distances are abstract and unitless, and in your situation they are an abstraction of the values supplied in your input data-matrix, i.e., gene expression. You should see a very similar separation of your dataset by doing simple hierarchical clustering with Euclidean distance as the distance metric.

Think of it another way: given the expression levels of your input genes, how similar and dis-similar to each other are my samples based on the expression of these genes?

For downstream applications, I prefer to do PCA, as you can better quantify the similarities / dis-similarities with component loadings to each axis / 'dimension', and, thus, infer which specific genes are responsible for segregation along a particular axis.

Trust that further helps.

Kevin

[Dayna]

That's very helpful, thanks Kevin a lot !

[Kevin Blighe]

No problem - happy to have helped.