Identification of rare celltypes

[Elhl93]

Hi,

I really like the idea behind DESC, great work!

We have a single-cell dataset, which we harmonized using harmony. As a test I wanted to compare wether we can identify similar or the same celltypes when processing the data with DESC.

However, the algorithm with the settings I use does not seem to identify rare celltypes. In my previous clustering with harmony & scanpy I obtained 13 clusters, ranging from 35% to 0.4% of total cell abundance. The markers of the rare clusters are in line with known/validated markers, so I have high confidence that the rare clusters are true. I see this behaviour of celltypes with an abundance of <1%. As some celltypes are very similar, I added more hidden layers to improve the model performance (my dataset has in total 80k cells). Increasing the batch size did not help to solve the problem.

desc.train(adata, dims=[adata.shape[1], 64, 64, 32, 32], tol=0.005, n_neighbors=20,
batch_size=256, louvain_resolution=[0.5],
save_dir="test1“, do_tsne=False, learning_rate=300,
do_umap=True, num_Cores_tsne=4,
save_encoder_weights=True, pretrain_epochs=100)

Happy to hear your thoughts on this!