Regarding reproducing results for image based technologies using Delaunay neighborhood structure

[MohammadFaizIqbalFaiz]

Hi,

Thank you for developing such an interesting and intuitive spatial clustering algorithm. SpatialLeiden is very unique in its approach and achieves top-tier performance across many datasets belonging to different technologies. I was doing bench-marking of SpatialLeiden on datasets belonging to image-based technologies. I am encountering issues related to reproducing results when the model uses Delaunay neighborhood structure.

For example, on MERFISH dataset, I put the layer-ratio as 1:1 as recommended in the supplementary section of the paper and followed all the recommended steps as given in the usage section of the tool's webpage. But I am not able to reproduce the results. I am getting results close to reported for knn neighborhood structure but for the delaunay, I am getting results well below the reported results for NMI.

Can you please help me out?

Thanking you.

With regards,

Mohammad Faiz Iqbal Faiz

[niklasmueboe]

Hi, thanks for the interest in our package.

I am not 100% sure I understand the question correctly.

The layer-ratio depends on a few factors, e.g. number of neighbors per cell, distance between the neighbors (which depends on the cell density and therefore also e.g. the tissue), the method to calculate edge-weights in the connectivity graph. Therefore there is not a layer-ratio that works well always. Generally I'd recommend to determine a reasonable value visually i.e. run a clustering and try to evaluate whether the spatial continuity of the identified clusters matches your "expectation" (this may depend on the biological question you are trying to answer). Higher weights will make the clusters more spatially continuous while lower weights lead to more scattered clusters (for 0 you would get "normal" leiden clustering).

If you are interested in reproducing the results from the study I'd recommend having a look here.

Does this help? Otherwise let me know