Questions about Scissor and SIDISH

[lxpsxx]

Hi SigBridgeR team,
I’m currently developing a similar tool, and I’m very lucky I found your project. Faced with such a perfect tool, I simply chose to give up, hahaha. The Tutorial and documentation are extremely clear and well organized, and I learned a lot from reading them. I also appreciate that the whole workflow is implemented directly in R—overall it’s much more complete than what I have built so far.
I have a few questions about Scissor and SIDISH and would really appreciate your advice:
For the Scissor implementation, when working with very large scRNA-seq datasets (e.g., >100k cells), have you ever encountered repeated errors? If so, do you have any suggestions on how to run it more stably?
For SIDISH, in the Phase 2 training step: according to the original deepcox description, phase2_lr should give higher precision, but I noticed the original author still keeps it at 1e-4. Is 1e-4 the recommended choice in practice?
SIDISH also seems not to output fitted-parameter or model-fit indicators. I made a small patch and found that without dropout it seems to overfit consistently. How do you usually handle this (e.g., dropout/regularization/early stopping), and what settings do you recommend?
Thank you again for your excellent work. I’m excited to start using SigBridgeR.

[Exceret]

Thank you for using SigBridgeR. We have benchmarked all algorithms except SIDISH on datasets with >100k cells and found that none of them ran properly—this may be a performance limitation of such algorithms. Therefore, we recommend splitting large datasets into smaller subsets for screening. SIDISH was only recently added to the package, and we are still tuning its parameters, so I apologize for not being able to answer your questions about it. You may want to consult the original author of SIDISH for more insights.

[lxpsxx]

Thank you so much for getting back to me. I completely understand the situation with SIDISH. Since it's a recent addition, I will reach out to the original author directly to discuss the parameter tuning and usage. Thanks for pointing me in the right direction!
In the description of your forked Scissor branch, are the WGCNA and pROC parts incorrect?
In any case, I’ve decided to stop doing the benchmarking myself for now and will wait for the official SigBridgeR release. Going forward, I’ll focus on developing my own phenotype-related algorithm, and I’ll try to keep the interface aligned with your framework.
If you have time, we could also communicate on Xiaohongshu（RUBBISH Bioinformatics）.