Xenium tutorial - downstream_analysis() bottleneck

[durr1602]

Hi, I've been trying to run the Xenium tutorial but ran into some issues. Most notably, the step running the downstream_analysis() function seems to be taking a very long time. The provided dataset is very big so that partly explains why, but I'd be curious to know the resources available on your end (and average runtime).

Also, I was confused by the following statement in the docstring:

(Note: This function executes default DE analysis as is mentioned in scanpy docs. For more flexibility, it is recommended for users to perform analysis over generated neighborhood scores manually with the help of scanpy")

Can I simply swap with a standard pipeline? If so, it's not clear to me why there would be operations on the pathways/gene sets and/or on the ligand-target pairs? It seems that the function combines multiple functional modules (pre-processing using scanpy, DE analysis, enrichment), which limits flexibility. For example, right now if one wants to specify igraph as flavor for the clustering, one needs to essentially modify the source code. Also, just looking at the code, it seems that there are several areas that would be computationally greedy (conversion to dataframes instead of working on the AnnData object, no vectorized operations,...)

Thanks for your help on this!

[narein97]

Hi,

Thank you for the detailed feedback — these are all very fair points!

The idea behind downstream_analysis() was to provide a simple, self-contained wrapper for users who just want to run Renoir end-to-end without worrying about the individual steps. You're right that it would be cleaner to separate the standard scanpy pipeline from the pathway/gene set cluster scoring into distinct functions, and that is something I plan to implement in the near future — it just isn't the immediate priority right now.

On the performance side, I agree there are several areas where the code could be made more efficient — avoiding unnecessary DataFrame conversions, vectorizing operations, and so on. These improvements will be rolled out gradually alongside minor bug fixes (if any), but again, are not the top priority at this stage.

As for runtime, all tutorials are currently designed to complete within an hour, with the Xenium tutorial naturally taking the longest given the dataset size. If it's running significantly beyond that on your end, please do let me know and I'd be happy to look into it.

Thank you for taking the time to raise these points!

[durr1602]

Thx for your reply! I'll adapt accordingly and will let you know if need be ;)