florekem
commented
2 hours ago According to:
Comparative analysis of antibody- and lipid-based multiplexing methods for single-cell RNA-seq https://pmc.ncbi.nlm.nih.gov/articles/PMC8851857/
Importantly, the tested demultiplexing functions [HTODemux and MUTIseqDemux] from Seurat use antibody- or lipid-derived hashtag oligo (HTO) expression data for cell line annotation
The problem with large proportion of Negative cells after demultiplexing 5` samples with Seurat's default algorithm and vignette (https://satijalab.org/seurat/articles/hashing_vignette.html) is probably not caused by my wrong approach with cellranger.
After I've obtained the same results even if two lanes (L001, L002) were used, I've tried modifying pattern of antibody capture from 5PNNNNNNNNNN(BC) - which I believe is the default one from 10x tutorial of analyzing 5` samples - to ^NNNNNNNNNN(BC)NNNNNNNNN, which I've found in one of the sample datasets published by 10x. It did not help at all. The default Seurat algorithm (HTODemux()) again produces huge numer of Negative cells. So it does not matter how strict the pattern is. On the other hand, the fact is, that the latter pattern led to double time of cellranger analysis. So if it produces the same output, it should be avoided. On the other hand, now I'm trying official (but unsupported) 10x way of demultiplexing (using cellranger, not Seurat). And in this vignette (https://www.10xgenomics.com/analysis-guides/demultiplexing-and-analyzing-5%E2%80%99-immune-profiling-libraries-pooled-with-hashtags) the ^NNNNNNNNNN(BC)NNNNNNNNN pattern is used. Interesting. I should also check the manuals of antidoby manufactuter, maybe there I'll find some interesting info?
As mentioned above, now I am trying to demultiplex samples using cellranger. This approach was not recommended to me, but now I think it is worth trying, as 5` v3 chemistry may change some things and Seurat didn't catch up yet (at least with the official algorithm).
Now, what still keeps me in a somewhat good mood, is Seurat's MULTIseqDemux() algorithm, which is way more capable of demultiplexig my samples. Sadly I can't find any information about it and I can't find any examples of using it. I want also to try other demultiplexing algos, not available in Seurat (I'll create separate issue for that).
Now that I'm thinking about it, HTODemux() is not doing well because it's designed for CMO's, and 5` chemistry is not using CMO's for multiplexing, but it uses hash antibodies instead. [Not sure what is the difference thou].What bothers me on the other hand is very similar number of both hashtaged cells (according to the cellranger web summary, ~18k cells - which is almost the total number of cells - has hashtag1 or hashtag2).