Exploratory CNV results and seeking general feedback on Wilms tumor annotation (SCPCP000014) #808
Replies: 4 comments 11 replies
-
Hi @JingxuanChen7, and thanks for filing this discussion topic. I hope that we are able to get some good feedback! First, I want to say that it looks like you have done a good amount of work for this initial exploration; we might want to capture some of this in the repository as a notebook, filed in an I think you may be right to conclude that There was one other method that we started to look at, but might deserve a longer look: SCEVAN You can see a bit about our experience with it here: #403 (comment); Note that one of the issues we encoutered, filtering duplicate gene names, does seem to have been fixed. I am not really sure what to do in the case where you do not have a good normal reference within the sample. It seems like you are using the compartment level annotations here: were there cells that you thought might be reliably normal with a more specific annotation? It may be that the answer is no! If the tumor sample is relatively pure, there may not be much of a baseline to compare against within the sample. I'd like to think there could be a good way to compare across samples to allow the incorporation of additional normal expression profiles, but I have not yet seen one that I can suggest. But even if copy number estimates were reliable, it might not be the best approach. As you note, aneuploidy may not be a universal or even particularly common attribute in Wilms tumor patients, so while we might regard the clear presence of a somatic CNV as denoting tumor cells, we would not want to assume that euploid cells are not tumor cells. Are there particular marker genes that you might be able to use to identify the Wilms tumor cells? For your other directions, I do have a couple of thoughts:
|
Beta Was this translation helpful? Give feedback.
-
Hi @jashapiro , thank you so much the prompt response!
Converting my code/result in this discussion to an Rmd notebook is a great idea, as the results shown here are most negative. I'll work on this in my next PR.
Thank you for letting me know that the "low quality warning" should be fine. Regarding copyKat output, I realized that I missed the heatmap, which seemed not as clear as the previous ewings output. No big CNV difference could be found between "aneuploid" and "diploid".
Really appreciate the suggestion on the new software! I will definitely look into it if time permits.
That's a very good point! In the preliminary analysis, I did try to find and marker genes either from literature ([1], [2], [3]) or inferring from DE gene analysis on available wilms tumor dataset ([4]), and try to explore with dotplots. However, it looks that a lot of the markers are either not expressed, or prevalently expressed among "nephrons". Also, my strategy to do preliminary dotplots was relied on clustering. Maybe I can get back to marker gene exploration after improving the clustering.
Thanks for pointing out normalization method here. |
Beta Was this translation helpful? Give feedback.
-
Thanks for taking the time to detail some of what you have explored here, @JingxuanChen7! I'm going to tag in @maud-p for this conversation too, since you are both working on different Wilms cell type annotation projects and are exploring similar approaches. @maud-p, some of the insights here are likely relevant to your analysis as well. Are there any specific similarities or differences you have noticed in your initial runs of CNV methods? |
Beta Was this translation helpful? Give feedback.
-
Adding @georgecresswell to the discussion, as he his much more experienced in Wilms tumor genetics than I 😃 Just some background, George is sitting in the office next door and I though might be more constructive asking my questions here than in private. What do you think about the infercnv profile? |
Beta Was this translation helpful? Give feedback.
-
Related proposed analysis: #628
Dear community, I'm opening this discussion in order to get some feedback on my preliminary results of CNV identification & general suggestions on next steps.
In my last PR (#784), I applied Seurat anchor transfer strategy to investigate the general cell type components in all 10 samples. Core result plots are accessible in the repo. In most samples, majority of cells are classified as
fetal_nephron
. In some samples,fetal_nephron
is even the only one compartment (such asSCPCL000847
,SCPCL000851
).In my proposal, I planned to identify tumor cells from normal cells, which could also help improve the annotation for normal cells. To start with, I tried
inferCNV
andcopyKat
using sampleSCPCL000850
, which shows more divergent cell compartments, and a relatively clearer annotation compared with others.Exploratory CNV results
Both strategies ask for a reference level/set of known normal cells. Here, I selected stroma compartment as the "normal" reference.
inferCNV outputs
To look more into the inferCNV results, I tried three ways to summarize the output.
If normal/tumor groups are clear, we should expect a bimodal distribution. However, we couldn't observe the bimodal distribution here. (reference = gray)
Based on the implementation in a previous module -
cell-type-ewings
, I also tried to calculate the number of CNV chrs & weighted means of scaled CNV proportion.I think no clear patterns can be observed, either. (reference = gray)
InferCNV also runs extremely slow in my case (~1hr)
CopyKat outputs
Also attach STDOUT message here. There seems a "low data quality" warning message?
STDOUT message shows plenty of `WARNING! NOT CONVERGENT!`, indicating a low confidence.
My questions/concerns
inferCNV
orCopyKat
could produce solid results to identify tumor cells. I was wondering if there were any issues with my workflow?fetal_nephron
is the only one compartment identified in some samples (such asSCPCL000847
,SCPCL000851
). In this case, it's pretty hard to specify a reference level for known normal cells. Identifying CNV in a merged dataset also seems to be pretty tricky. Any suggestions on such samples?Other possible next steps
SCTransform
+Azimuth
), it may be worthwhile to addSCTransform
and/orAzimuth
as functions to this module. However, since the rationale and reference database are same, I wouldn't expect major differences in the result. Any suggestions?Beta Was this translation helpful? Give feedback.
All reactions