search

Danko-Lab / BayesPrism

A Fully Bayesian Inference of Tumor Microenvironment composition and gene expression
158 stars 46 forks source link

deconvolution of spatial data #6

Closed vravik closed 9 months ago

vravik commented 2 years ago

Hey guys,

Do you think BayesPrism can be used for deconvolution of spatial transcriptomics datasets, to get cell-type specific expression profiles?

dankoc commented 2 years ago

Yes, BayesPrism works great to integrate spatial data.

References along these lines include Tinyi's recent paper: https://www.sciencedirect.com/science/article/abs/pii/S1934590922002077

And work by Cornell colleagues: https://www.nature.com/articles/s42003-021-02810-x

On Thu, Jul 21, 2022 at 1:19 PM vravik @.***> wrote:

Hey guys,

Do you think BayesPrism can be used for deconvolution of spatial transcriptomics datasets, to get cell-type specific expression profiles?

— Reply to this email directly, view it on GitHub https://github.com/Danko-Lab/BayesPrism/issues/6, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAYUH7JNGC4DOMPBAIEDKG3VVGBA5ANCNFSM54IMXTWQ . You are receiving this because you are subscribed to this thread.Message ID: @.***>

-- Associate Professor Baker Institute for Animal Health College of Veterinary Medicine Cornell University

Website: http://www.dankolab.org E-mail: @.***

vravik commented 2 years ago

Thanks a lot Prof. Danko. This is going to help me greatly with literally every project I have! I also love the idea of embedding spots based on their cell type compositions in the Lymphatics paper. Does this produce a similar embedding as one based only on the bulk gene expression profile of the spots?

dankoc commented 2 years ago

Great - happy to hear it!

Thanks a lot Prof. Danko. This is going to help me greatly with literally every project I have! I also love the idea of embedding spots based on their cell type compositions in the Lymphatics paper. Does this produce a similar embedding as one based only on the bulk gene expression profile of the spots?

I would guess that using BayesPrism will pick up on somewhat different signals compared with just embedding the stops. Tinyi's experience on how different these are in practice will have to be the final word here though.

Best, Charles

— Reply to this email directly, view it on GitHub https://github.com/Danko-Lab/BayesPrism/issues/6#issuecomment-1205996328, or unsubscribe https://github.com/notifications/unsubscribe-auth/AAYUH7JKKYTFIKRB54EUH3LVXSBGZANCNFSM54IMXTWQ . You are receiving this because you commented.Message ID: @.***>

-- Associate Professor Baker Institute for Animal Health College of Veterinary Medicine Cornell University

Website: http://www.dankolab.org E-mail: @.***

tinyi commented 2 years ago

Hi. This is a good idea, but I have not tried. The rationale behind using cell type fraction as the input of SpaceFold ( https://github.com/dpeerlab/SpaceFold) is that we hypothesize that the fraction of each cell type contains sufficient information for inferring the relative coordinate. I think the raw expression may also do the work, but as dimension reduction, e.g. PCA, is required, the reduced dimension might be affected by the abundance of different cell types. That being said, the performance can be problem-dependent, and is worth giving a shot.

Best,

Tinyi

On Fri, Aug 5, 2022 at 9:23 AM Charles Danko @.***> wrote:

Great - happy to hear it!

Thanks a lot Prof. Danko. This is going to help me greatly with literally every project I have! I also love the idea of embedding spots based on their cell type compositions in the Lymphatics paper. Does this produce a similar embedding as one based only on the bulk gene expression profile of the spots?

I would guess that using BayesPrism will pick up on somewhat different signals compared with just embedding the stops. Tinyi's experience on how different these are in practice will have to be the final word here though.

Best, Charles

— Reply to this email directly, view it on GitHub < https://github.com/Danko-Lab/BayesPrism/issues/6#issuecomment-1205996328>, or unsubscribe < https://github.com/notifications/unsubscribe-auth/AAYUH7JKKYTFIKRB54EUH3LVXSBGZANCNFSM54IMXTWQ

. You are receiving this because you commented.Message ID: @.***>

-- Associate Professor Baker Institute for Animal Health College of Veterinary Medicine Cornell University

Website: http://www.dankolab.org E-mail: @.***

— Reply to this email directly, view it on GitHub https://github.com/Danko-Lab/BayesPrism/issues/6#issuecomment-1206455107, or unsubscribe https://github.com/notifications/unsubscribe-auth/AB4NHS2CXSAJLP5KLNPB4GTVXUIUVANCNFSM54IMXTWQ . You are receiving this because you are subscribed to this thread.Message ID: @.***>

vravik commented 2 years ago

Thank you. Tinyi. I have a related question about the single cell reference data. I have data from ~30 different patients that I'd like to use as input for deconvolution as this would be more representative of distinct cell states than using a single sample. However, there are going to be some batch effect variations between these populations. How would you recommend I process the single cell reference data before spot deconvolution?

tinyi commented 2 years ago

Thank you for your question. I have not extensively tested the integration of multiple single cell RNA-seq references. When cell types overlap between multiple references, you can label each batch as a cell state in the new.prism argument. My colleague has tried the negative binomial regression function provided by cell2location to correct the batches of scRNA-seq when deconvolving spatial transcriptomics, and had some success.

However, when there are cell types that do not exist in all reference datasets, it is generally difficult to correct the batch effects. You may either force some linear correction using the correction factor learned from overlapping cell types or see if there is other (non-linear) single cell data integration method that would allow you to project one dataset to the other at the count scale, i.e., non-log transformed / non-latent space. Perhaps there is a way of hacking into the scVI to change the batch labels so that all cells are projected to the same batch on the mean parameter of the final layer in the decoder, but I am not sure about this. Let me know if you have any success in finding the solution.

Best,

Tinyi

On Fri, Aug 19, 2022 at 4:06 PM vravik @.***> wrote:

Thank you. Tinyi. I have a related question about the single cell reference data. I have data from ~30 different patients that I'd like to use as input for deconvolution as this would be more representative of distinct cell states than using a single sample. However, there are going to be some batch effect variations between these populations. How would you recommend I process the single cell reference data before spot deconvolution?

— Reply to this email directly, view it on GitHub https://github.com/Danko-Lab/BayesPrism/issues/6#issuecomment-1221053622, or unsubscribe https://github.com/notifications/unsubscribe-auth/AB4NHS7EV4TMMFAMLGQERODVZ7SMFANCNFSM54IMXTWQ . You are receiving this because you commented.Message ID: @.***>

tinyi commented 2 years ago

Thank you for your interest in our work.

Yes. I have recently published a paper demonstrating that BayesPrism generalizes to Visium data. We recommend using the updated cell type fraction for Visium. https://doi.org/10.1016/j.stem.2022.05.007

Best,

Tinyi

On Thu, Jul 21, 2022 at 1:19 PM vravik @.***> wrote:

Hey guys,

Do you think BayesPrism can be used for deconvolution of spatial transcriptomics datasets, to get cell-type specific expression profiles?

— Reply to this email directly, view it on GitHub https://github.com/Danko-Lab/BayesPrism/issues/6, or unsubscribe https://github.com/notifications/unsubscribe-auth/AB4NHS5LKGVXDREO2FHONVLVVGBA5ANCNFSM54IMXTWQ . You are receiving this because you are subscribed to this thread.Message ID: @.***>