I have used the one implemented in mosaicmpi, because it is more up to date than the original. It’s fine but a little bit hacky if just using the cNMF functionality. If you find something else, let me know!

academic.oup.com/nar/article/...

What do you especially like about how they do modules? They use Hotspot, which a colleague recommended to me but I haven't yet tried out. I've been very happy with the results of cNMF (although a bit less happy with the implementations I've found... 😅)

I think that of you use modules that make sense to you given your understanding of the context, and somewhere note “other equally valid sets of modules are possible”, then they are fine and can be helpful 😊

However, I don’t think there is any way to show that the modules you have are correct / unique / objective. So I don’t think anyone should try 😅

What would ground truth look like for testing and validating this? I just think it’s impossible.

In principle, gene modules make sense to me: there will definitely be gene programs (i.e. sets of genes required for a specific function), and I think they are a more natural way to think about continua of cell states than v artificial clusters.

(I usually ignore claimed findings in papers like this... The data is so complex that understanding what is in it can take as long as generating it... Which is fine! It would be better for everyone if journals accepted such papers on the basis of good exp design and solid methods alone.)

I think with this kind of paper, the value is mainly in a huge and useful dataset that enables future work, rather than key insights from the paper itself.

So maybe good peer review is mainly checking their methods are ok?

But yes I feel drained just from skimming over the paper... 😅