Overview of concepts of the methods that this Single-Cell Differential Expression (SCDE) tool uses?

Question

I'm unsure about how this tool works, but as I understand it, it tries to use Bayesian probability in order to differentiate the two different groups from each other as well as some other methods.

Right now as I understand it; this R package does this:

1. Takes in a 2D matrix of samples vs. features, with each number representing the amount of appearances of a feature in a sample.
2. Receives another vector of keys that determine which group each sample belongs to.
3. Performs some sort of analysis in order to get correlation between each sample and their respective groups.
4. Transforms this data in order to determine some sort of version of variance or "differential expression scalar number" that ranks the features' differential quality between the two groups.
5. Leaves the final resulting data open so that we can extract the list of features that are the most different between the two samples.

Did I get that right? I'd like to understand the steps between 3 and 5. I think the concepts are outlined in the tutorial quite clearly but I don't have the direct stats know how to understand it. I think it's quite easy to read if you try to look at the variable names vs. their explanations. Sorry if this is poorly written.

http://pklab.med.harvard.edu/scde/Tutorials/

Answer 1

I'm a member of the Kharchenko lab, which developed this original SCDE package, and the author on the updated SCDE package: http://hms-dbmi.github.io/scde/ A fellow lab member saw this post and pointed me to it so hopefully I can help address your questions.

In general, your understanding is correct that SCDE uses Bayesian probability to identify features (such as gene expression) that differentiate two groups (such as two groups of single cells).

However, this method is very specific to single cell RNA-seq and the technical artifacts that arise from performing single cell RNA sequencing. Specifically, to capture drop-out events (where a gene is truly present but not detected) and PCR amplification biases that are common to single cell RNA-seq data, we model each cell as a mixture model of a Poisson and Negative Binomial respectively. These models can then be used to derive failure rates, get incorporated into the differential expression analysis, etc.

So to revise your outline of how SCDE works, it would be more apt to say that this R package:

1. Takes in a 2D matrix of cells vs. gene expression counts.
2. Receives another vector of keys that determine which group each cell belongs to.
3. Performs pair-wise cross-fitting between each cell in their respective groups to determine a set of robustly expressed genes that are then used to fit the aforementioned Negative Binomial and Poisson mixture error models.
4. Transforms the data in order to determine variance between the two groups that takes into consideration these error models in order to identify highly variable or differentially expressed genes between the two groups.
5. Leaves the final resulting data open so that we can extract the list of genes that are the most truly (as opposed to being driven by technical noise) different between the two groups.

Hope this helps!

Best,
Jean