Open imran-aifi opened 1 month ago
I can't say for sure what's going on, but I suspect some subset of cells has extremely low coverage (the median is pretty low here). Proseg infers cell boundaries using transcripts, so it may simply be doing a poor job because it lacks the data in cases where there are only a few transcripts. It may think these are so sparse it looks like background noise.
Do you know what fraction was empty in Xenium's segmentation by comparison?
Thanks for the point of reference for the low coverage!
Xenium's segmentation resulted in 4.5% empty. In Xenium Explorer looking at the unassigned (presumably empty) cells, they tend to be spatially located in sections of tissue that we know are low quality (high RBC count, etc).
I got some improvement after running proseg with --voxel-layers 1
where the % empty is now 10.3%, though those unassigned cells are more evenly distributed across the tissue.
From what I've gleaned, the starting point for the simulation is the initial nuclei assignments from 10X - perhaps some of these nuclei are misidentified, or have few/no transcripts assigned to them (transcripts for these nuclei are assigned to neighboring cells, or treated as background)?
When running proseg on Xenium data and following the Xenium Explorer tutorial, I get a high fraction of cells empty as reported by the xeniumranger QC report:
I am running proseg with
proseg --xenium --ncomponents 10
as the only parameters used. I'm wondering if you had recommended parameters, or guidance on whether to use--voxel-layers 1
to simplify the segmentation as we do not have serial sections for 3D spatial transcriptomics.