Estimating the contamination fraction without biological assumptions using Cell Ranger's default clustering results or ERCC spike-ins about soupx

[constantAmateur]

I suggest you have a look at the devel branch of the code which includes a routine to automate contamination estimation. This is still in the devel branch as it has not been widely tested, so I would appreciate your feedback on its performance if you do use it.

The automated workflow would be:

sc = load10X(dat)
contEst = autoEstCont(sc)
sc = setContamination(sc,contEst$rhoEst[1])
correctedCounts = adjustCounts(sc)

With extra steps if you load your data in a different way.

The alternative if you must have automation would be to set some fixed contamination fraction for all channels using setContamination. I would actually recommend trying to set this too high rather than too low (e.g 15-20%) as the biologically important things (the marker genes) are highly expressed relative to the background and basically never removed. This is a risky approach though, so if you're doing this you should make sure you see how your results depend on the constant value you use.

ERCC spike-ins are a potential solution to this, but in practice it is too hard to ensure that you load the right spike-in concentration. Too little spike in and you get no benefit, too much and you spend all your money sequencing spike-ins.

The isV3 tag doesn't do anything after the data is loaded. It's just used to use the file format approapriate to the cellranger version within load10X

[mesnger]

Hello, all.
I have quite a similar question as with vertesy.

I tried multiple gene list as "single list" or "combination of lists" for CalculateContaminationFraction and got quite a broad range of outputs (from 1.1% contamination rate to 10.44%).
I selected gene lists by differential_expression.csv provided by Cell Ranger and/or Seurat FindAllMarkers. But as the output varies greatly by each gene lists, I could not settle on one result.

Do you have any criteria for the best input gene? Following your tutorial, I assume that "Known Gene Marker " is your go to choice, but when it is not known, what gene should I prioritize on?

A. Genes "exclusively expressed" in a subset of clusters(other clusters have near 0 value)
B. Genes "highly expressed" in a subset of clusters
C. Genes with "highest difference" in expression value compared to other cluster
D. Genes with "high p-value" calculated by some DEG method

or is by considering all of above? (which might not lead to a sound result with real life data)

Thanks for making a great tool.
Hope to get a reply soon.