zijianni
commented
10 months ago Hi @Streisenberg , are your findings all based on SpotClean processed data? If so, see my responses below:
- Filtering in CreateSlide Function: It is typically a good practice to filter out lowly expressed genes before downstream analyses since they usually do not capture much useful information. However, it's not a required step for SpotClean. SpotClean itself will select another set of highly-expressed and highly-variable genes (via
keepHighGene()) and perform decontamination on them. The remaining genes will simply be scaled up by the ratio between tissue UMI counts and total UMI counts for each gene. If you believe keeping more genes lead to more promising results, please use your own judgement and proceed accordingly. - Barcode Mismatch: For genes marked with "DEPRECATED", unfortunately I don't have full context on such set of genes. But I saw people filter them out before proceeding with downstream analysis.
- Issues with CytAssist Gene + Protein Assay: When we published this work, there was no 11x11mm slides, nor gene+protein assay. As a result, there is no guarantee that the model assumptions still hold for these new experiments. The 11x11mm gene expression slides should be fine, but I'm pretty sure we can't directly run SpotClean on gene+protein slides. The spot swapping effects we observed for gene UMI counts might not hold true for protein probes. The magnitude of protein expression could also be significantly different from gene expression, distorting the model estimations. As an alternative approach, you could consider running SpotClean on the matrix subset containing only gene features, and combine the output with the matrix subset with protein features.
Again, given that we haven't validated SpotClean on these new experiments, please rely on your own judgement for your analysis.
Hello and thank you again for this wonderful package,
During the release of the paper, you had mentioned that the package was only tested on 6.5x6.5mm Visium V1 slides. I have both Visium CytAssist Gene Expression (GEX) and Visium CytAssist Gene + Protein Expression (GEX/PEX) outputs from 11x11mm slides. I have a few questions about the downstream analysis outputs of these data:
Filtering in CreateSlide Function: When I use the CreateSlide function, roughly ~30,000 genes with average expressions below or equal to 0.1 get filtered out. When I reduce the gene_cutoff parameter to 0.01, around ~25,000 genes are filtered out, leaving between approximately 6,000 and 10,000 features. Is this number of features sufficient for downstream analyses, or should I reconsider the filtering process? Cause' when I compare scRNA-seq deconvolution methods using unfiltered and filtered data, the deconvolution from the unfiltered data seemed more consistent with the Immunohistochemistry (IHC) staining results. However, the unfiltered data appears noisy. Are there any different filtering methods you could recommend?
Barcode Mismatch: As I mentioned in #15 , there was a mismatch between my slide info object and raw data in terms of barcodes. I proceeded with my analysis by discarding barcodes not present in the raw matrix. However, when I run PCA, I find genes like “DEPRECATED-ENSG00000026297” as variable genes. Is this an expected behavior, or has a technical artifact occurred?
Issues with CytAssist Gene + Protein Assay: When I run the CytAssist Gene + Protein Assay (additionally to GEX, co-detect 35-Plex Immune Proteins with oligo-tagged antibody panel on same slide), it captures protein expressions similarly to gene expressions. However, this seems to lead the Principal Components (PCs) astray. PEX values average around 15,000 to 20,000, and when combined with GEX values, it appears to cause overrepresentation of these proteins, potentially affecting the downstream analysis.
Would appreciate any insights or guidance on these matters.
Thank you.