cggnn workflow errors if no graphs are created from a specimen

[CarlinLiao]

When running on a study with very large and very small specimens (e.g., urothelial), sometimes specimens will be deemed too small to create a specimen for. When that happens, spt cggnn create-specimen-graphs will error because DGL will try to save an empty list of graphs.

A simple but insufficient fix is to change create-specimen-graphs so it doesn't try to save graphs when there are no graphs created from a specimen, but this results in a Nextflow error as it's expecting a graph data artifact to be created, causing an error.

Possible solutions:

• Move or duplicate the logic that decides if graphs are created from create-specimen-graphs upstream to prepare-graph-creation. Not my preferred strategy since deciding how many graphs are created happens concurrently with graph creation, but it's possible.
• Edit the Nextflow process so it knows that create-specimen-graphs may not create any output.

[jimmymathews]

I think I misunderstand something here. No matter how small, each specimen should not be "deemed too small" to create a graph. Surely it is easy enough to create 1 graph in that case. (Do we actually have such a case?)

[CarlinLiao]

This is intentional, as graphs are not created when specimens are too small, as they don't provide enough information to train a good model. This is the same reason why we don't run cggnn on datasets with smaller slides, like CyTOF or breast cancer IMC.

[jimmymathews]

None of the slides in the urothelial dataset are too small. The smallest that occurs is 2724 cells.

[jimmymathews]

No, we should still be able to run the pipeline in the presence of the "smaller slides" from those datasets. The performance is a separate matter, it does not justify restricting our implementation to the best case scenarios only.

[CarlinLiao]

I set the cells_per_slide_target (actually, this is a misnomer and should be changed to cells_per_ROI_target or something) that we've been using to 5000. If I recall correctly this was a value I arrived at after some observation of which datasets cg-gnn worked well for and which it didn't, after I switched from setting a predetermined ROI size to having it be determined dynamically by average cell density across specimens + a target number of cells I wanted to hit per ROI.

But yes, these are probably two separate issues:

1. the cggnn workflow errors if no ROIs are created from a specimen, which is intended functionality as-is (this issue)
2. should there be such a thing as a slide that's too small, and if so are the smallest slides in urothelial that small?

Let's open up a separate issue for the latter, or an email/in-person talking thread if we determine it's unrelated to the SPT codebase.