[Feature] Custom annotation

[sophiedoublier]

In the feature description for sopa, it is shown that we may use custom scripts for the automatic annotation portion of the pipeline.

However, I don't see how we can use another function other than higher_zscore or tangram without deeply modifying either the api, cli or the pipeline. Am I missing something?

[quentinblampey]

Hello @sophiedoublier, the "custom" annotation is only for API users, and can be done directly on the AnnData object (that can be found inside sdata.tables).

There is indeed nothing special to wrap a new annotation. So, what it means is simply that, since we use AnnData under the hood, any tool from the scverse ecosystem can be directly plugged in using the API.

What type of annotation would you like to use? And, do you use the API or CLI or pipeline?

[sophiedoublier]

Thank you for your response. I'm using the pipeline with multiplex
images as input. In the fluorescence-based annotation, I modified the config file to attribute multiple channels to a single cell type:
annotation:
method: fluorescence
args:
marker_cell_dict:
CD3e: T lymphocyte
CD4: T lymphocyte
CD8: T lymphocyte
CD45RO: T lymphocyte
CD278: T lymphocyte
CD31: Endothelial cell
...
While this approach allows for more complex annotations, I realize that assigning cell types based on the highest individual z-score may not fully capture the combinatorial nature of these markers. Would it be possible to implement a more sophisticated annotation method that considers combinations of markers? For example, creating a 'metachannel' that combines the z-scores of related markers (e.g., summing z-scores for all T lymphocyte markers) before cell type assignment? This could potentially provide a more accurate representation of cell types in complex tissues. I understand this might require modifications to the pipeline. If so, could you suggest the best way to implement such an approach using the API, or guide me on how to extend the current annotation method?"

[quentinblampey]

Thanks for the context!

Indeed the marker-based annotation is very simple, I totally understand that you could need something more sophisticated.
Here are some ideas, from most basic to more complex:

• After intensity preprocessing, use Scanpy score_genes to get a score per signature per cell. Then, for each cell, assign it to the cell type whose signature has the highest score.
• Run clustering based on all channels intensities per cell (e.g., Leiden), and show a heatmap of mean expression (per cluster per channel) to map each cluster to a cell-type
• Use a biology-driven model, such as Scyan (try using a high prior_std, because fluorescence data is less clearcut than mass cytometry expressions).

The thing is, since there is no strong consensus on cell-type annotation, I don't want to add too many methods in Sopa, because (i) most users end up using their own custom annotation anyway, and (ii) the main scope of Sopa is mostly segmentation/aggregation/scalability/visualization.

Don't hesitate if you have more questions!

[sophiedoublier]

Many thanks for your suggestions! I'll try.

[quentinblampey]

Closing the issue, don't hesitate to re-open after trying these approaches if you have additional questions :)