Applicability to cancer data

[Thapeachydude]

Hello,

would you have any suggestions on how to best apply spectra to data sets containing cancer cells?
I would be interested in applying your tool to a single-cell RNA-Seq data set consisting of primarily malignant B-cells.
Similar to this publication we see a strong separation of the malignant cells by patient, whereas non-malignant cells group together.
While I think it should be possible to "decompose" the expression signature of the malignant cells into factors, I was wondering on how to best feed them into the model (i.e. it is very difficult to assign cell type labels, beyond the most basic ones).

Any insights would be much appreciated.
I'm mainly used to the Seurat and SingleCellExperiment classes, so codable examples would be extremely helpful. : )

Cheers!

[OMIC-coding]

Same question (in scanpy pipeline), waiting for your reply~

Best regards~

[wallet-maker]

Hi, please excuse the late reply. We have used Spectra on some solid tumor datasets. It works fine but of course there are a lot of patient-specific factors. I think these strong differences between patients are likely biological although processing artifacts might contribute also especially in solid tumors.

Regarding hematological malignancy I believe it could be very interesting but we have not tested Spectra on such datasets. My intuition would be defining the malignant population as a separate cell type to obtain malignancy-specific programs. In case of strong inter-patient/batch differences the global factors should still capture some joint information in the malignant cells while the malignant-specific factors will capture most of the batch differences. Course cell type labels are usually fine. In fact if the cell type labels get too granular there might not be enough training data per cell type.

Please see our tutorial Google collab for the general Spectra workflow in python.