search

mahmoodlab / HEST

HEST: Bringing Spatial Transcriptomics and Histopathology together - NeurIPS 2024
Other
161 stars 12 forks source link

Normalize adata (found only log1p) #15

Closed rushin682 closed 4 months ago

rushin682 commented 4 months ago

Hi Guillaume,

Thanks for the brilliant resource!

Regarding the normalization process during benchmarking (for the paper benchmarks), is the spot only logarithamized (log1p) or is it normalised (sc.pp.normalize_total()) + sc.pp.log1p?

I am asking because in the code here, it seems it's only log1p. Is this intentional?

Thanks for your help. Looking forward to using the benchmark!

pauldoucet commented 4 months ago

Hi, thanks for your interest in HEST-bench!

It is intentional indeed, we decided to only apply log1p.

We made our decision based on this paper, which states that traditional Single Cells normalization methods could have a negative impact when applied to Spatial Transcriptomics and should be adapted with caution.

This more recent article also confirms the hypothesis from the first paper.

guillaumejaume commented 4 months ago

Exploring additional normalization strategies and their downstream impact would be valuable. It seems no consensus has been reached yet. In addition, exploring the impact of the downstream classification model is also interesting. We explored random forest and ridge regression and observed large variations.

rushin682 commented 4 months ago

That makes so much sense! Thanks a lot for the resources!