Which cutoffs to adjust for more sensitive splitting cells?

[samuelikim]

Hi, I'm running FastReseg on tissues that are tumors infiltrated with T cells. I would like FastReseg to be sensitive to splitting so that it is more likely to split (as often, the segmentation is picking up tumor T cell combinations). My question is, which cutoff should I be adjusting -- is it the cutoff_spatialMerge, the svmClass_score_cutoff? Sorry, total noob at this. Thanks.

Sam

[github-actions]

Thank you for contacting us about our tools! To receive faster assistance, kindly email support.spatial@bruker.com with detailed information about your issue. If applicable, attach a screenshot of any encountered errors and include a copy of the modified script in Notepad. Our customer support team will help facilitate a review and resolution of the issue.

Thank you for choosing Bruker Spatial Biology,
Bruker Spatial Biology Dev Team

[lidanwu]

Hi @samuelikim, it sounds like you want more separation between T cells and tumor cells. There are several knobs you can tune about this.

• If you have done cell typing on your data, you can subset the single-cell expression counts and cell typing clust to the cells that you're most confident about their expression purity (say by circling on UMAP); and then feed that data to the pipeline (runPreprocess()) to get a cleaner separation between those cell types when scoring each transcript.
• If you adjust flagCell_lrtest_cutoff to higher value, say 3, you would include more cells with intermediate spatial dependency in their expression to transcript-level analysis.
• Besides, you can adjust the arguments associated with SVM to make it more sensitive to local expression changes when flagging transcripts, e.g. move positive svmClass_score_cutoff at -1.5 or even -1; smaller gamma at 0.1 (default is svm_args =list(kernel = "radial", scale = FALSE, gamma = 0.4)).

cutoff_spatialMerge is a physical constraint in candidate merging events, won't affect your outcomes in splitting.

Please try the impact of above parameters with a smaller subset of your data (say 1 FOV) to determine what are best for your data.