import scgen
import scanpy as sc 

Reading the train data

train = sc.read("./tests/data/pancreas.h5ad",
                backup_url="https://www.dropbox.com/s/qj1jlm9w10wmt0u/pancreas.h5ad?dl=1")

We need two observation labels "batch" and "cell_type" for our batch_removal procedure. There exist a "batch" obs but no "cell_type", so we add it as a .obs of adata

train.obs["cell_type"] = train.obs["celltype"].tolist()

UMAP Visualization of the original uncorrected data

sc.pp.neighbors(train)
sc.tl.umap(train)

sc.pl.umap(train, color=["batch", "cell_type"], wspace=.5, frameon=False)

Creating and training the model

network = scgen.VAEArith(x_dimension= train.shape[1], model_path="./models/batch" )

We train the model for 100 epochs

network.train(train_data=train, n_epochs=100)

Batch Removal

Now, we pass our train (uncorected) adata and network object to batch_removal fucntion which returns adata of batch corrected data

corrected_adata =  scgen.batch_removal(network, train, batch_key="batch", cell_label_key="cell_type")

Visualization of the corrected data

sc.pp.neighbors(corrected_adata)
sc.tl.umap(corrected_adata)

WARNING: You’re trying to run this on 2448 dimensions of `.X`, if you really want this, set `use_rep='X'`.
         Falling back to preprocessing with `sc.pp.pca` and default params.

sc.pl.umap(corrected_adata, color=["batch", "cell_type"], wspace=.5, frameon=False)

Using uncorrected data

Note that original adata.raw for the adata.raw is saved to corrected_adata.raw and you can use that for fruther analaysis

corrected_adata.raw

<anndata.core.anndata.Raw at 0x7f4da4313390>

sc.pl.umap(corrected_adata, color=["TMEM66", "cell_type"], wspace=.5, frameon=False, use_raw=True)