A few tips on training models

• We recommend you to set loss_fn = nb or zinb. These loss functions require the access to count and not normalized data. You need to have normalized log-transformed data in adata.X and raw count data in adata.raw.X. You also need to have normalization factors for each cell in adata.obs[scale_factors]. These normalization factors can be obtained with scanpy.pp.normalize_total or other normalization methods such as scran.

• If you don’t have access to count data and have normalized data then set loss_fn to sse or mse.

• If you want better separation of cell types you can increase the n_epochs. 100 epochs in most cases yield good quality but you can increase uo to 200. If some cell types are merged which should not be try to increase n_epochs and decrease alpha (see next tip).

• If you want to increase the mixing of the different batches then try to increase alpha when tou construct the the model. Maximum value of alpha can be 1. Increasing alpha will give you better mixing but it is a trade off! Increase alpha might also merge some small cell types or conditions. You can start with very small values (e.g 0.0001) and then increase that (0.001 ->0.005->0.01 and even 0.1 and finally 0.5).

• It is important to use highly variable genes for training. We recommend to use at least 2000 hvgs and if you have more complicated datasets, conditions then try to increase it to 5000 or so to include enough information for the model.

• Regarding architecture always try with the default one ([128,128], z_dimension`=10) and check the results. If you have more complicated data sets with many datasets and conditions and etc then you can increase the depth ([128,128,128] or [128,128,128,128]). According to our experiments small values of `z_dimension between 10 (default) and 20 are good.