cell2location underestimated cell types density and variability in Visium FFPE samples

[PietroAndrei]

Hi there,

I recently started using cell2location to analyse Visium data coming from a collection of FFPE colorectal cancer (CRC) samples. These samples are extremely heterogenous in terms of tissue composition, meaning that each sample will include both tumour epithelium and stromal regions with different features.
I do not have a matched scRNAseq dataset available, so I used a public CRC single cell dataset (https://singlecell.broadinstitute.org/single_cell/study/SCP1162/human-colon-cancer-atlas-c295) to estimate the reference cell type signatures. The catalogue of cell subtypes is quite large (> 80 cell subpopulations), so I am currently using a mixed-level annotation (i.e. high granularity for immune cell types, low granularity for epithelial/stromal cell types), for a total of 51 cell types.

At the time of using cell2location on my Visium samples, I merged all the samples together to perform a single run, with N_cells_per_location = 15 and detection_alpha=20.

By checking cell2location results, it looks to me that the estimated cell abundances (the sum of all estimated cell abundances across all the cell types for each spot) are actually way lower than our estimation based on manual cell count:

Even when I considered only epithelial cells (represented by only one reference signature), the estimated abundances across different tissue types (based on manual annotation of the spots) seem kind of flattened, and this happens for most of the cell types that I am considering in my analysis:

The data quality of Visium samples is a bit mixed: for 2-3 samples I had to discard half of the total tissue spots available (~ 3000 per sample) due to low genes or count coverage (< 500 counts OR < 300 genes detected), with a median number of ~2000 genes per spot detected considering all the samples. However, I didn't notice any major difference between bad and good samples in terms of cell type estimation quality.

I guess my questions are:

1. Are cell2location estimated abundances supposed to be somehow comparable with the real number of cells within Visium spots?
2. Would it be useful to shorten reference signatures by removing genes with the lowest variability across the cell types? I am currently considering ~13000 genes expressed in both single cells and Visium samples, but a lot of them show little or no variation in terms of expression across all the cell types
3. Do you think that, by running cell2location on spots coming from each tissue type separately, I could obtain more accurate estimations? I noticed that cell type proportions are actually more reliable than the original cell abundances, but even cell proportions suffer from the same signal flattening observed using the estimated abundances.
4. Considering the mixed quality of my Visium samples, would it be better to use a less granular cell type annotation?

Thank you for your help :)

Pietro