search

JEFworks-Lab / STdeconvolve

Reference-free cell-type deconvolution of multi-cellular spatially resolved transcriptomics data
http://jef.works/STdeconvolve/
98 stars 12 forks source link

Combination of STdeconvolve with BayesSpace #23

Closed Acaro12 closed 1 year ago

Acaro12 commented 1 year ago

Hi bmill3r,

Thank you very much for this exciting package. It is very useful, especially when retrospectively analyzing archival tissue with Visium FFPE.

Since you feature the BayesSpace package in the Nat Com Paper (https://github.com/edward130603/BayesSpace), I was wondering if an upstream BayesSpace resolution enhancement would improve the performance of the STdeconvolve algorithm. I could imagine that STdeconvolve would in principle benefit from a higher (although predicted) resolution of the underlying gene expression matrix.

Do you think that this approach would be statistically sound or would it violate the underlying STdeconvolve assumptions?

Thank you very much for time in advance and thanks again for your great work.

bmill3r commented 1 year ago

Hi @Acaro12,

Thanks so much for the idea! I think in principle STdeconvolve would benefit from a higher resolution gene expression matrix, which could be provided by BayesSpace. However, the comparison between STdeconvolve and BayesSpace performed in the paper was to demonstrate that one of the assumptions made by BayesSpace, that spatially adjacent spots are correlated in terms of their transcriptional profiles, isn't necessarily true in all cases. Specifically, we looked at the thymus region of a mouse brain coronal section, which has high cell type heterogeneity, and found that this assumption led to worse predictions by BayesSpace. This was likely because BayesSpace uses spatially adjacent spots to determine the predicted transcriptional profiles of sub-spots. In cases like the thymus, where adjacent spots may actually contain different cell types, this assumption leads to incorrect transcriptional profile predictions. Now, I do think that in many cases the same cell types can be expected in adjacent spots and so the assumptions that BayesSpace makes are probably reasonable. But in situations where the underlying biology does not support this, BayesSpace may predict incorrect transcriptional profiles, which in turn would lead to less accurate predictions by STdeconvolve. Note that STdeconvolve currently treats each spot independently and does not take into account spatially adjacent spot information during the deconvolution.

Hope this helps and please let me know if you have any other questions, Brendan

Acaro12 commented 1 year ago

Dear Brendan,

Thank you so much for your detailed reply and sorry for the late reresponse. I know decided to use both methods independently from another and used the corrMatrix to compare their output. The conclusions I derive from them in my specific project seem to support each other, although, as expected they are both informative in their way.

Thanks again. I will open another issue not related to this one here. Best, Christoph