UMAP shows scattered cells after pycisTopic LDA (SCENIC+ workflow)

[pooriass]

Hello,

I am running the pycisTopic workflow as part of SCENIC+ for my scATAC dataset(About 20,000 cells after QC). After running the run_cgs_model and proceeding with evaluate_models, UMAP and tSNE plotting, I noticed that my UMAP embedding looks very scattered and sparse. The cells don’t visually form clear clusters, although in tSNE I do see a bit more clustering.

Here is what I tried so far:

Ran the model with different parameters (alpha = 10 and 30 and 50, eta = 0.1 or 0.3). In some combinations, the model reached a relative plateau in the performance plot of evaluate_models function (mainly when alpha was either 10 or 30) , but the UMAP still looks scattered.

tSNE embeddings based on the same topic matrix look slightly more structured, but still not as clustered as I would expect.

I used the cell ranger output bed files and fragments files. I concatenated the bed files and made a consensus.bed, and fed this as region file to the make_cistopic_obj function( to further identify the clusters, the re-run the process with identified clusters, as explained here : #87)

plots:
plots_MODEL.pdf

My concern is whether this is caused by:

choice of number of topics or LDA hyperparameters,

data sparsity or low coverage cells,

the need for additional filtering, batch correction(I have run the harmony batch correction),

or if this behavior is expected depending on dataset heterogeneity.

Has anyone experienced similar scattered embeddings at this stage of the workflow? Are there recommended strategies (e.g. filtering topics, trying different embeddings, using Harmony on the topic matrix, or adjusting UMAP parameters) that usually help improve clustering?

Any guidance or suggestions would be greatly appreciated. Thank you for your time and for maintaining this great tool!