Relevance of UCell Score difference

[Rayan21100]

Hi,

Thanks again for this package !

I think my question is also linked to #45 and #49
I had a question about the interpretation when comparing two UCell Scores.I have spatial transcriptomic data and I'm interested in the link between the activation status of one subset of cells and its location in the tissue on 4 different patients (that have the same zones)
So, I computed a score on my cells of interested based on activation genes and I wanted to compare these scores across the different zones.
I aggregated them into pseudobulk values: for each patient × tumor zone I calculated the mean UCell score.
I used a GLMM (beta family) with:

• the mean UCell score as the outcome
• the zone as fixed effect
• the patient as a random effect
• the number of cells in the pseudobulk as precision weights

Finally, I used estimated marginal means (EMMs) to compute pairwise contrasts of UCell score between tumor zones on the original (0–1) scale.

Because pseudobulking strongly stabilizes variance, the effect sizes I observe are small (a difference of 0.004 or 0.008) but they are very significant.
My question is:
Are such small differences in pseudobulk UCell score biologically meaningful ? How can I know if a statistical difference is interesting ?

Thanks in advance !

[Rayan21100]

Hi again, I also have a question about the maxRank parameter, the lowest median gene of my Xenium spatial transcriptomic samples is 58 (the other samples are around 100-110) so I choose 60 as my maxRank value (total number of genes in the panel is 5050 genes).

However, two of the signatures I'm comparing have more than 60 genes (one is 66 and the other one is 273).

My idea was to keep the maxRank value at 60 for all other signatures but to choose 273 for these 2 signatures. I will never compare the signatures computed with maxRank = 60 with the ones computed with maxRank=273 but only the signatures computed with the same maxRank values. Does it make sense ? Isn't it a problem to have a maxRank of 273 with a median between 58-110 ? Thanks in advance

[mass-a]

Hi @Rayan21100
thanks for the good questions!
I think your approach for comparing UCell scores is reasonable. If you see consistent differences between tumor zones, even if the effect is small, it's fine to report that they are significantly different. I would be open about it and state that the difference is statistically different but the effect size is small. If you have large signatures, you could perhaps dig into individual genes and see whether there is a subset of genes that drive the separation, while the rest are unchanged?

As for the maxRank parameter, it does not make much sense to use signatures with more genes than the actual number of detected genes. That actually breaks some assumptions for the Mann Withney U, which is normalized by n*(n+1)/2, where n is the length of the signature. As you suggest, you could adjust the maxRank to the length of the signature. Otherwise, If you want to be consistent across signatures, another option would be to (randomly) subset the 273-gene signature to 60 genes. You could even do this multiple times and generate e.g. 10 random subsets of the signature, and then average+SD their UCell scores? this would also give you an idea about the robustness of the signature and of its UCell scores.

Best
-m