Questions regarding Symphony

[PauBadiaM]

Hi @cassandraburdziak and @hisplan,

First of all let me congratulate you for this fantastic work, Symphony is indeed an interesting method for GRN inference.
I've been reading through the original publication and the follow up manuscript and I've some questions regarding the method:

1. Is Symphony restricted to use bulk ATAC-seq or can it also leverage scATAC-seq? Since it deconvolutes bulk ATAC-seq samples into cell-type specific ATAC-seq profiles it should be possible, right? However, when I was checking the code I saw that the deconvolution step is incorporated into the main training loop, so I am not sure how I should adapt the code to use scATAC-seq data.
2. When you intersect peaks to genes's TSS, which window size did you use (X kbp upstream, Y kbp downstream of the TSS)? Also, which motif database did you use for FIMO? I was looking for this in both manuscripts but couldn't find it in the text nor in the code. Also, it would be good to add such processing scripts into the method for reproducibility and to facilitate the usage of the method to other users.
3. You mention in the methods section that Symphony can infer TF-Gene regulatory links, even when the TF motif is unknown. I guess this is thanks to the covariance propagation through the graph? Sorry but I struggle to understand this, since if a TF has no motif it will just not be included in the network, right? Can genes that are not characterized as TF regulate other genes in this model? Or did you mean that it will find indirect links from TFs that were not previously reported by motif analysis?
4. The current implementation of Symphony makes it hard to use it with other data, processing steps and documentation are missing. Are there any plans on wrapping it into a python package? Since it is a probabilistic model, it would also be cool to implement it using the scvi-tools framework, a package for probabilistic modeling and analysis of single-cell omics data from the scverse.

Thank you for your time!