Applying Celltypist to sparse Visium HD: unexpected predictions and advice?

[oghzzang]

Dear Celltypist team,

I trained a model using the GSE149614 dataset (public HCC scRNA-seq cohort).
During training I normalized counts to 1e4 UMIs per cell and applied log1p.

I then applied this model to HCC Visium HD data.
Following the tutorial, I normalized each bin so that total UMIs equal 1e4 and applied log1p.

In this tissue, hepatocytes should predominate and B cells should be very rare, but I’m seeing unexpected predictions (e.g., many B-cell calls).
Is Celltypist appropriate for very sparse spatial data like Visium HD bins? If not, what checks or settings would you recommend?

Any guidance would be greatly appreciated.
Thank you!

[ChuanXu1]

@oghzzang, how many genes do you have in your Visium HD data?

[oghzzang]

Dear ChuanXu1,

Thank you for your response.
The total number of detected genes across the dataset is approximately 17,000, while each bin (like spot in Visium v2) contains about 50 to 200 genes.

[ChuanXu1]

@oghzzang, since the training data usually includes seven thousand genes, while each bin contains fewer than 200, this discrepancy could potentially cause prediction issues. You can try curating a short list of expressed genes (500~1000 maybe) from Visium data, and then subset the (raw count) training & query data by adata_train = adata_train[:, selected_genes].copy() and adata_query= adata_query[:, selected_genes].copy(), followed by normalization of the two data subsets sc.pp.normalize_total(target_sum = 1e4) and sc.pp.log1p() . For training, you can disable the feature selection by adding feature_selection = False.

Please let me know how it works.