Interspecies integration

[gao-lab]

Interspecies integration

[Jeff1995]

Thanks for your interest in Cell BLAST and the detailed explanation!

Given the above information, I can think of two potential fixes:

1. One possibility is that the model has not fully converged. You may try increasing epoch number to larger values like 200 or 500. The model will trigger early stop if loss has converged.

2. In Cell BLAST we have a single tunable hyperparameter (denoted as lambda_b in the manuscript) that controls the strength of adversarial batch alignment. Larger values enforce stronger alignment, but also increase the risk of over-alignment. By default, we use lambda_b=0.01, which empirically produces good result in most cases.

   In this case, it seems that lambda_b=0.01 is insufficient for aligning datasets this diverse. My recommendation is to try increasing this hyperparameter to values like 0.1 or 1.0. That should produce better dataset mixing.

   The hyperparameter can be set like this:

model = cb.directi.fit_DIRECTi(combined_dataset, 
                               genes = var_genes_study, 
                               latent_dim=10, cat_dim=20, 
                               epoch=200, rmbatch_module_kwargs={"lambda_reg": 0.1},  # <- Changed here
                               batch_effect = ["study","species"],
                               path = model_dir
)

[bschilder]

Hi @Jeff1995 , thanks so much for the quick reply! These are all very helpful tips.

I tried setting lambda_reg to 0.1 and 1 as you suggested, but oddly this seems to have no effect on the results. I don't have a screenshot, but I think these plots are also identical to when I don't specify lambda_reg at all.

I even named the models differently and set reuse=False so that I wasn't accidentally plotting old results.

lambda_reg=0.1

lambda_reg=1

Do you have an idea what might be happening here?

Thanks,
Brian

[Jeff1995]

Well, that's weird... I have never seen anything like that. Maybe it's something in this data that triggered a bug in the model. Would you mind if you share the dataset you're using so I can have a closer look?

[bschilder]

Hi @Jeff1995 , sorry for the delay, had some issues finding a way to get my data shareable.

I think this should work now, but let me know if you have any issues.
https://drive.google.com/file/d/1hXXnpXiRp7evl727V3onkx0WmThGFFkx/view?usp=sharing

Thanks again,
Brian