Obtaining batch effect estimates at cluster level

[wmacnair]

Hi

First of all, thanks for developing and maintaining a fantastic tool 🥳 It's great to have a conceptually simple, sensibly motivated and fast tool to integrate single cell data iteratively.

I'm interested in understanding which genes Harmony identifies as being "batchy". From this tutorial, I have tried the following code

batch_adj_mat = Loadings(seu_obj, reduction = "pca") %*% t(hmny_obj$W)

which I think gives me the estimated effect on each gene, averaged for all cells in each cluster. Is that right?

batch_adj_mat is a matrix that is n_genes * n_samples. What I would like to get to is something like batch_adj_tensor that is n_genes * n_samples * n_clusters. I would then average this across some known celltype labels I have, so I would end up with n_genes * n_samples * n_celltypes. I'm curious about how much the Harmony adjustment differs across the different clusters.

Is this somehow achievable...? 😅 I'm hoping that all I have to do is add one line to what I used above, but I don't see an obvious way to do it so far...

Thanks!
Will

[pati-ni]

I just saw your mention about cell types and clusters. My previous comments are not relevant so I will delete them. Under normal operation, W is retaining information for the last cluster so the information you are looking is not there. I believe what you are asking can be found in a different function that does return a cell x gene x number of clusters, so you can project your own covariates.

To access this information you have to re-run the last part of the integration:

so you need to do the following: hmny_obj$moe_ridge_get_betas_cpp()

Hope this is more helpful than my previous responses