Hi,

I have two small single cell datasets containing RNASeq data from 12, and 20 cells respectively, which were manually isolated.

Our hope is to potentially project this data onto another, annotated scRNA-seq dataset which should contain these cells, to see where our data falls into the other dataset and hopefully do some differential expression against other celtypes.

However, the first thing I'd need to do is to combine the two aforementioned datasets, which were isolated and sequenced at different times, as well as at different depths, and I was wondering what the best way to combine these datasets was.

I have used Salmon (separately for each dataset) for the quantification, followed by tximport with the hopes of later using this data as a SingleCellExperiment or Seurat object, in order to make use of some projection functions.

I searched this forum and found these two posts which seem relevant to my question:

Single cells batch effects

Pool Single-cells: Analysis

Would reading the data in from both salmon runs using tximport (with countsFromAbundance set to "lengthScaledTPM" as described here) and then using the txi$counts as the count matrix to create a SingleCellExperiment object, followed by limma batch correction be appropriate here?

Thanks,

Liam

I don't use salmon or tximport, but assuming you get some kind of count matrix from them, it should be simple enough to proceed:

1. Create a SingleCellExperiment object from the count matrix. I don't know if there's any consequences from running salmon separately for the two batches compared to running them all together; if there is, you should consider re-running salmon on all samples to avoid unnecessary batch effects arising from how you get from the reads to the counts.
2. Run normalize to obtain log-transformed normalized values. There are a few functions to compute size factors, but I'm not sure how useful they will be here. computeSumFactors really needs more cells (at least 100 to yield stable estimates), while computeSpikeFactors needs, well, spike-ins. If you run normalize without providing any size factors, it will compute size factors from the library sizes. This is probably your only option, as inaccurate as the library sizes are for estimating cell-specific bias. (Well, perhaps it's not so much of a problem if you only have cells of the same type, as there shouldn't be any composition biases.)
3. Run removeBatchEffect to remove the batch effect. This assumes that the composition of cells is the same between batches, which should be true of replicates. If this is not the case, you could consider using mnnCorrect, though you would have to turn down k considerably due to the low numbers of cells involved.

Combining your dataset with the published data is yet another kettle of fish. Assuming that your cell type exists in the published dataset, a simple nearest neighbors approach would probably be sufficient to identify the matching cells in the latter. As in, just identify the nearest neighbouring published cell(s) to each of your cells, and that is the putative corresponding population.

Finally,a direct DE analysis between your data and the published data would be suspect due to the batch effects involved. It would be a lot safer to use your data to identify the matching published cell type, and then do a within-batch DE analyses with that cell type.