potential cross validation of predicted doublets

[MichaelPeibo]

Hi, @swolock

Thanks for developing this package.

I wonder, is there any way to cross-validate the predicted doublets? In that way, it will give us a sense whether it is proper to define those cells are doublets or not?

[swolock]

Hi @MichaelPeibo,

Except for experimental methods for doublet detection (e.g. cell hashing), I don't know of any truly independent ways to validate the predicted doublets. That said, if you know good marker genes for the various cell types in your data set, we expect the Scrublet-predicted doublets to co-express marker genes of multiple cell types (for examples, see Figs 4G and 6D in the Scrublet paper). Furthermore, predicted doublets should, on average, have more total molecules (UMIs) detected than singlets, but this isn't always true (e.g., if one cell type has much less mRNA than others).

You could also try running other computational doublet prediction tools (DoubletFinder, DoubletDecon, DoubletDetection, and others that have appeared more recently). The ones I listed are mostly quite similar to Scrublet under the hood (less true for DoubletDecon), but they may give you additional confidence in the doublet predictions.

[MichaelPeibo]

Hi @swolock Thanks for reply. I checked counts of predicted doublet, shown below:

I tried also DoubletFinder, while, for same dataset, common doublet cells is few. And chosen pK in DoubletFinder seems to give rise to a higher rate of doublet.

I found in Fig4G, there are Marker Gene Co-Expression Score and Hybrid Doublet Score, would you mind adding reproducible code in the github?