Outline second spatial submodule notebook

[sjspielman]

This issue tracks creating an outline for the 2nd spatial submodule notebook, the topic for which is spatial clustering/domain detection.

[sjspielman]

I've been playing around with a couple ideas for structuring the second notebook and corresponding exercise based on (but certainly not the same as) what was drafted here: https://github.com/AlexsLemonade/training-admin/issues/1453#issuecomment-3582038402

First, I do think it's going to be important to show deconvolution in the workshop. It has really emerged as a standard/core analysis for Visium data in the literature, so it will be useful for participants to learn about from us. In part because of this, I moved on from the prostate dataset we found in the linked issue to a dataset where I know we have a super solid reference for deconvolution - an osteo visium dataset (mouse, lung metastasis) from this paper we're familiar with:

I did some initial exploration (with super not-for-prod code, but it's sure good enough for an issue comment!) here, running deconvolution (shown is only on the coarse-grained labels from the OsteoCar atlas; the next level has like 20 labels but it works nicely there too) "regular" clustering, and banksy clustering and making a variety of heatmaps to compare (h/t claude for giving me good enough very quick heatmap code). I had originally wanted to include a comparison to bayesspace clustering here as well, but I paused on this - it requires an a priori number of clusters (they recommend using an elbow plot and when I ran this it gave me 11, about the same as banksy). It's also a really long runtime so it would have to be cooking show magic, so figured to bring it up first... this also means I haven't fully run it through yet!
See here: osteo_scran.nb.html

Not necessarily related to workshop but could be, mostly I'm just intrigued: I got curious, so I did this in duplicate where I instead normalized the data with spaNorm first rather than using the scuttle library size factors. Interestingly, the non-spatial clustering results on data normalized with spaNorm were super similar to Banksy. I haven't had a moment to deeply remind myself of their innards enough to know why precisely yet, but this is indeed interesting.
See here for the spanorm version (you'll notice that deconvolution results are the same - indeed this runs on the raw counts): osteo_spanorm.nb.html

As such, this is a loose set of exploratory analyses I'd like to include.

• non-spatial pca, clustering, umap (but only plot it once and then never again because coordinates are better)
• banksy pca, clustering, umap
  • potentially could vary parameters here for a mini exploration, but I think this is a rabbit hole we don't want similar to choosing k for SpotSweeper. So, stick to their parameter recommendations for visium data and take time to clearly explain them.
• compare different flavors of clustering to each other with heatmap
• potentially look for marker genes distinguishing spatial domains vs spatial clusters
  • look at banksy/regular clusters with significant overlap for this, run findmarkers
• Then, bring deconvolution into the mix as a complementary analysis. We can use the OsteoCAR atlas to perform deconvolution and compare cell type weights to clusters. I see this happening in one of two ways:
  • Break this out into a separate instruction notebook. This is similar to how we have 2 notebooks for filtering & normalization, and dimension reduction & clustering in the intro scRNA-seq module. This way we'd have two pretty targeted notebooks: one looking at spatially-informed clustering, and one looking at deconvolution
  • Use the associated exercise as an opportunity to teach deconvolution.
    • Alternatively for the exercise, we can do more of the same clustering things with the same dataset we use for the first exercise (ovarian carcinoma). As bonus learn-a-new-thing content (though I can't tell if I'm seriously leaning here) - they could also compare among scuttle vs spanorm, and non-spatial vs spatial clustering.
  • An implementation note: the deconvolution analysis will involve converting ensembl -> gene symbols since the reference object is Seurat, so this will take up some chunks and text in the notebook as well.

I'm curious to hear what others (currently that's just @jashapiro) think of these ideas! I am currently leaning towards the 2 instruction notebooks (but only 1 associated exercise notebook) concept, which keeps topics scoped while also flowing nicely together as common and related spatial analyses.

[sjspielman]

Tagging @allyhawkins now that it's next week

[sjspielman]

Noting that I ran this same (ish) exploratory notebook over the ovarian dataset we use for the first* exercise. This was a bit harder to find a usable reference for for deconvolution, but I had success with this GEO dataset as a reference.

Similar trend with the spanorm one but less pronounced than for the osteo dataset - when normalizing with spanorm, the "regular" clustering has more spatialosity (a technical term which I just made up).
ovarian-scran.nb.html
ovarian-spanorm.nb.html

[sjspielman]

We have discussed that we will proceed with two instruction notebooks for spatial clustering and deconvolution. We'll circle back to the exercise specifics here once these are outlined/drafted given that we're doing two notebooks in instruction.

[sjspielman]

Broken out into new issues #964 and #965