Cannot derive sufficient number of differentially expressed genes for reference

[kokonech]

Hi, I'm working on published neurobalstoma single cell dataset (10X v2). There are the following cell type counts:

Tumor cells      Immune cells Mesenchymal cells       Liver cells  Endothelial cells     Schwann cells
            59560              2407              1297               669.           559               277

Afterwards running DEG calling and conversion leads to exteremely small amount of DEGs (only 17), despite initial high numbers and relaxing filtering limits, leading to incorrect deconvolution results on WGBS /array data e.g. zero proprotion of tumor in control samples with single cell data. Is there some way to fix this/adjust with params? Below main code with comments:

targ.idx = as.factor(nbObj$annotation)
levels(targ.idx) = 1:6
alltypes= levels(as.factor(nbObj$annotation))
expref.o <- ConstExpRef(nbObj@assays$RNA@data,targ.idx,alltypes,markspecTH=rep(1,6)) 
# this count has nrow=163, even though initial DEGs minimum per cell type before filtering is 3349, maximum 13740
expref.m <- expref.o$ref$med 
tscm2.m <- ImputeDNAmRef(expref.m,db="SCM2",geneID="SYMBOL") 
trmap.m <- ImputeDNAmRef(expref.m,db="RMAP",geneID="SYMBOL")
# this is only 23, moreover multiple NA in final matrix
tref.m <- ConstMergedDNAmRef(tscm2.m,trmap.m) 

[aet21]

You don't need to find any DEGs.....We already provide a DNAm reference panel

for the respiratory epithelium and we used this one to study the neuroblastoma

DNAm dataset, as described in our Nat Methods 2022 paper. So I am not sure why

you are trying to build a new panel.......

In any case, if you are trying to build an mRNA panel and you can't find enough DEGs

then try relaxing the specificity score or significance thresholds, or try to decrease cellular

resolution to improve power.

A

[kokonech]

Thanks for detailed reply! We are using a custom reference panel to check specific properties of tumors , mainly tumor and non-tumor composition in bulk meth profiles, but not the cells-of-origin. So, do you mean to play with thresholds at the function wRPC() ? I tried this, but results demonstrated extermely small amount of tumor cells proprotion (10-30%), while in reality it should be up to 95% (we have single cell data from some meth profiles). Or do you mean custom relaxing and preparation of input matrix without using function ConstExpRef?

[aet21]

Well if you don't say how you built your customised panel, then I can't help much further, but the problem seems to stem from the way you have built your own panel. For a given tumor derived from normal cell-type "X", the proportion of cell-type "X" in a tumor will mostly be tumor X cells and not normal X cells. The normal cells that are more abundant would be the stromal and immune cells. So I suspect that in your panel you have references for tumor-X and normal-X besides all other normal cell-types, and that your panel is predicting abnormally low tumor-X and abnormally higher normal-X. But surely, your normal-X would be very low anyway?, so what you are actually measuring is tumor-X cells. Maybe I misunderstood your specific problem, but the issue stems from your panel. We applied the Resp Epith panel very successfully to our own NeuroBlastoma set and predictions concerning basal vs neuronal subtypes seemed to make sense.

[kokonech]

Sorry, maybe I misunderstood, but I followed strategy that we used typically when working on transcriptome deconvolution (here's our recent example study https://link.springer.com/article/10.1186/s40478-024-01921-w/figures/3 ). The reference is the combined single cell RNA-seq dataset from neuroblastoma tumors fully annotated (https://www.nature.com/articles/s41588-021-00806-1). We would like to use it as reference for methylation bulk profiles to track the assigned cell types proprotion to distinguish tumor vs non-tumor proprotion and further properties of the tumors. I'm using the full tumor scRNA-seq dataset as input (nbObj shown in code above). What could be potential problems in building panel?