ValueError: Expected parameter loc (Parameter of shape (1, 13866)) of distribution Normal(loc: torch.Size([1, 13866]), scale: torch.Size([1, 13866])) to satisfy the constraint Real(), but found invalid values

[hua1991]

Dear cell2location team,

I encountered an error when training Visium HD spatial transcriptome data. We first combined the bins into cells using a 10x nucleus segmentation strategy, and then used this h5ad file as the input file for the cell2location analysis. For the scRNA-seq data, we used the same tissue 5' scRNA-seq data as a reference, the “mod.train” step for the scRNA-seq data worked fine, but for the spatial data, it reported a “ValueError: Expected parameter loc (Parameter of shape (1, 13866)) of distribution Normal(loc: torch.Size([1, 13866]), scale: torch.Size([1, 13866])) to satisfy the constraint Real(), but found invalid values".

I checked the input spatial data and it was raw count data (integer) and not normalized. For the spatial model, I set “N_cells_per_location” to 1 and “detection_alpha” to 200. I don't know if these parameters make sense for the Visium HD nucleus segmentation data or if these values cause the above error. Could you help resolve this issue?

Best,
hua

[vitkl]

Hi @hua1991

When does this happen - immediately on training step 1 (likely issue with input data format like containing NA or 0s for all locations) or after some training (likely issue with numerical stability due to extreme values - both high and low)?

The offending parameters seem to be gene-specific tech effect difference m_g. I would recommend doing additional gene filtering based on VisiumHD - not just scRNA gene filtering shown in tutorials.

With Visium HD it could be important to normalise column and row effects before aggregating per cell (as done in this work https://github.com/Teichlab/bin2cell). That raises an issue of how to apply count methods such as cell2location and scVI. You probably need to do normalisation for column and row effects keeping the scale of values similar to the original data, then binarise by rounding to sampling integers using normalised values as Poisson mean. All these options need to be tested.