CellChat
CellChat copied to clipboard
jinworks
best practice for multiple samples
Hello!
I want to analyze a dataset with several samples. Should I run cellchat workflow one sample by one sample, or just merge all cells together and apply cellchat workflow as a single sample?
For the first one, I can extract inferred interactions from each sample and analyze them myself, the disadvantage is that I can not use the useful functions in CellChat. For the second one, I'm worried about the artificial interaction from different samples.
@songmingl We usually merge all cells together and apply cellchat workflow as a single sample. This is similar to clustering analysis using integration pipeline, which increase the power of single cell data analysis
@sqjin Thanks for the rapid reply. If merging all cells, how will CellChat process the sample information to escape artificial interactions? I mean, suppose there is an interaction A-B, A is expressed in sample A while B is expressed in sample B, will CellChat discard this kind of interaction?
@sqjin, thank you for a fantastic tool. I am also interested in @songmingl question being addressed.
One could hypothetically run CellChat per sample and then merge the replicates together by experimental group, however, I would find it difficult to believe that an interaction present only in one sample is meaningful. Additionally, the merging of datasets with uneven cell groups makes CellChat's functionality difficult to accomplish this. Any thoughts on how to address this?
@songmingl @alyxgray7 Under your assumption, CellChat does not perform further filtering step to exclude such interactions. Have you observe such case in your data? I do agree that it is meaningful to identify interactions that are present in at least two samples. We will address this potential issue.
@sqjin Thank you for your quick reply!
I ran group-wise and sample-wise CellChat runs with the same input cell groups in each dataset (i.e., I required >10 cells per cell type in each sample). When I compare the significant interactions (pvalue <0.05) identified in each group-wise and sample-wise analysis, on average there's 67% overlap across my three groups (see plots attached). Also, the group-wise analyses find on average 4-5% unique interactions that are not identified in any of the sample-wise runs which suggests there could be some artificial interactions being generated.
Group1: venndiagram.significant_shared_with_sample_analysis.pdf
Group2: venndiagram.significant_shared_with_sample_analysis.pdf
Group3: venndiagram.significant_shared_with_sample_analysis.pdf
@alyxgray7 @songmingl We updated the filterCommunication to enable identifying consistent signaling across samples. Please check it and let me know if it works for your data.
@sqjin I appreciate the quick fix turnaround! After updating CellChat and including samples within the metadata, I reran the group-wise analysis and required an interaction to be found in >= 2 samples/group. I'm finding that there is an overall reduction in the number of unique interactions with the updated function in the group-wise set, but there are still artificial L-R being found. Also, I should clarify that this is single-cell RNA-seq data, not spatial.
Group 1 (updated): venndiagram.significant_shared_with_sample_analysis.pdf
Group 2 (updated): venndiagram.significant_shared_with_sample_analysis.pdf
Group 3 (updated): venndiagram.significant_shared_with_sample_analysis.pdf