In measuring gene expression of qPCR, we usually does it this way
Where GeneX is some gene of which we need to decide.
I know there scRNAseq are usually normalize using TPM,FPKM, Deseq method etc. But my question is whether in single-cell RNAseq we need to divide again a gene expression against another gene expression (as base).
I would say normalization to a single gene or very few genes is not recommended. The power of methods such as TMM (edgeR), RLE (DESeq2), the deconvolution in scran is that a pool of reference genes is used to accurately derive size factors to normalize the raw counts. For an extensive discussion on the above methods with its pro/cons in the single-cell context I recommend Lun et al. (2016):
https://genomebiology.biomedcentral.com/articles/10.1186/s13059-016-0947-7
Single genes are suspectible to technical variation, in bulk and especially in single-cell contexts. In single-cell data basically each gene could suffer from a technical dropout. I see sometimes genes like GAPDH having a count of zero in some cells. It would therefore be required to a) use a larger number of genes to properly define a baseline, which is what the above methods do empirically and b) come up with a strategy to account for technical zeros.
=> The scran method especially addresses this in single-cell data by a pooling / summing strategy to avoid zeros, see the paper for details.
Use of this site constitutes acceptance of our User Agreement and Privacy Policy.
version 2.3.6
I'll just add that normalisation by a single housekeeping gene also performs pretty poorly in qPCR. I've had many occasions where a result that doesn't make sense in qPCR was because the housekeeping gene had changed.
The superiority of using a panel of genes in qPCR was demonstrated decades ago, and the geNorm method to normalise expression using them proposed. Unfortunately it never caught on because of the extra work involved in producing the data.
@ATpoint how about normalization of a single gene by another GeneX after taking the mean from all cells? Also not advisable?
I do not kow, never did that. In general, I'd say the more custom it gets the more unreliable it becomes. Why not just using established tools rather than reinventing the wheel? It is probably this "garden of forked paths" you enter here when introducing too many custom approaches. Keep in mind that this method you propose keeps you from using basically any of the standard differential analysis tools which yet again do their own normalization, so you would then have different sets of normalized counts for the same dataset, which again forks another path. THerefore I would probably recommend against it. Run it through scran and use the logcounts it spills out unless there is very good reason to not do that.
Some of the standard tools will allow you to nominal a panel of genes to base the normalization on, rather than using the entire transcriptome. I don't thin they would work with one gene, but you could probably nominal a few 10s. You would then be able to use that tool for DE etc.
However, I'm not saying this is a good idea, only that its possible. Unless you have a very good reason, the data will almost certainly be worse than doing it the normal way.