Inaccurate pseudobulk deconvolution with raw counts

[emmanuelmekasha]

Hello,

Thank you very much for your package. My group is very interested in using this method in our work for neuroscience discovery. We performed some preliminary deconvolution benchmarks and observed some really weird phenomenons.

Using the ROSMAP brain dataset (scRNA-seq dataset), I generated a pseudobulk matrix by aggregating columns of the same celltype. I provided BayesPRISM with the scRNA matrix as a reference and the pseudobulk matrix as the objective to deconvolute. We found BayesPRISM was pretty inaccurate (~10% RMSE) with this simple deconvolution task while regression methods could predict the correct proportions.

Interestingly, when we provided BayesPRISM with the pseudobulk matrix as generated by Seurat's AggregateExpression method (which involves several normalization steps, including log normalization), BayesPRISM was right on the mark. In general, however, the documentation seems not to recommend normalization.

Help would be greatly appreciated. This seems to be really weird behavior from a robust method. I could elaborate on my code or any specifics.

Thank you.