Having a combination of cite-seq and gene expression data as input

[jtheorell]

Dear Will,
I currently work with a dataset containing both gene expression data and surface marker information, using cite-seq. Now I wonder if you think there is any inherent problem with using all this data as input for the glm-pca-analysis? I have 9 surface markers, separating all major cell subsets, and I believe that including them would "weight" the analysis to make sure that for example CD4 and CD8 cells are readily separated, but if I by doing this violate some data distribution assumptions, etc, then of course I should avoid it.
Best regards
Jakob

[willtownes]

Hi Jakob,
Good to hear from you! You can try GLM-PCA and see what happens, but I suspect it will have problems with the protein data as it has different noise properties. The way GLM-PCA is set up you have to use the same likelihood for all the features which may be inappropriate for CITE-seq. If you are familiar with python I would recommend using @adamgayoso 's totalVI instead of GLM-PCA for this kind of data. The idea behind totalVI is to have a negative binomial (NB) likelihood for the gene expression, and a two-component mixture of NBs for the protein (a low mean "background" component and a high mean "foreground" component). They use nonlinear dimension reduction with variational autoencoders but it may be possible to modify some of the settings to get linear dimension reduction similar to GLM-PCA. There may be some other tools like it in R but I'm not familiar with them. If you find something that works let me know!

[jtheorell]

Thank you so much for your reply, Will! I have some other reasons too not to bring in the cite-seq data into this specific part of the analysis, so this decides it. But I will definitely check out totalVI for the future!
All the best
Jakob