Normalization methods to combine scRNA-seq experiments with different sequencing depths

Question

I am training a classifier to identify a cell type in a particular state of activity using scRNA-seq. There is a large variation in the sequencing depth (reads average per cell) of the testing data (data I will be classifying not training or validating the classifier with) and my supervisor suggested binarizing the genes. I was wondering if there are other options I should consider. I was wondering if there are particular normalization methods for ML classifiers which would be best for this task.

Answer 1

Coincidentally, a recent preprint out of Lior Pacther's lab covers some of these details around scRNA-seq depth normalization.

They find that a "PFlog1pPF" normalization method is able to stabilize variance across genes while still reducing the impact of total cell depth on that variance. If I've read this preprint correctly, "PFlog1pPF" is essentially a transformation of the form $\frac{1}{N_c} \log \left(1 + \frac{k_g}{N_c} \right)$ where $N_c$ is the number of reads for cell $c$ and $k_g$ is the number of reads for feature $g$.

This preprint is mostly dedicated towards differential expression and identifying marker genes in certain cell populations, but that may still help you in your application.

Using a transformation like this (or others that are discussed in this preprint and its references) gives you more granularity and sensitivity for highly expressed genes than what binarizing the data will give you. And it has additional benefits like stabilized variance, which can cause problems in some downstream applications if you don't keep it under control. But starting with a binarized input like your supervisor suggested might be a simple first attempt to see if your idea has some potential in the first place.

Answer 2

I don't think you need to complicate the idea of normalisation by introducing machine learning classifiers as a necessary component. Normalisation is common when comparing different datasets for all differential analysis.

Given that you have single cell data, have a look at integration techniques in the Seurat workflows. These include normalisation as a component of integration:

https://satijalab.org/seurat/articles/integration_rpca.html