Predictive and robust probe selection (PROPOSE) is an approach to select genes for FISH studies. It relies on a reference scRNA-seq dataset, and it applies deep learning to identify genes capable of reconstructing the genome-wide expression profile, or any other target of interest (e.g., transcriptomic cell types). During the selection process, PROPOSE binarizes gene expression levels to account for the domain shift between scRNA-seq and FISH expression counts.
See here for a demo notebook. Please email me for access to the data required to run the notebook, as it's too large to upload here.