How to concatenate anndata properly?

I have 4 samples data trying to do batch effect cleaning but when I try to concatenate them it comes out like this
sc.pp.filter_genes(adata1, min_cells =1)
sc.pp.filter_genes(adata2, min_cells =1)
sc.pp.filter_genes(adata3, min_cells =1)
sc.pp.filter_genes(adata4, min_cells =1)
adata = adata1.concatenate(adata2, adata3, adata4, batch_key='Sample')
AnnData object with n_obs × n_vars = 55320 × 20678
obs: 'Sample'
var: 'gene_ids', 'feature_types', 'genome', 'n_cells-0', 'n_cells-1', 'n_cells-2', 'n_cells-3'

the n_cells comes out all separate so should I be concatenating the dataset before doing filtration/doublet removal/etc?

Hi @cookiemonster,

As a general rule, you should always perform quality control (QC) for each sample individually first. Each sample can be quite different from the rest, requiring different filtering thresholds, for example at the doublet score level. Once you have performed QC at the sample level, you can merge the samples into a single object using the concatenate method. By the way, I would recommend to add join='outer' to the concatenation because otherwise you might lose quite some genes (by default is set to inner).

Regarding the n_cells problem, is this related to your previous question?

If yes, it is very weird that you store this information into the var attribute of your adata object, since there it should only store metadata for your features (in this case genes). If what you want is to obtain the number of cells per sample and cell type, you can do it after concatenating your samples instead of doing it before:

adata.obs[["Sample", "louvain"]].value_counts().reset_index()

Since this results in a dataframe that its dimensions are not number of cells x samples (obs) nor number of cells x genes (var), if you want to store it you could do it in the uns attribute:

adata.uns['n_cells'] = adata.obs[["Sample", "louvain"]].value_counts().reset_index()

Hope this is helpful!

Stronly agree with @PauBadiaM about doing QC (especially doublet detection) per sample first – assuming your inputs are divided like this to start.

But I would prefer anndata.concat is over AnnData.concatenate. With anndata.concat, you would not get the n_cells-{n} columns. Exact behavior here is defined by the merge argument).

1 Like