Great read. I think similar model architectures to those you discussed could also help expand our understanding of how RNA/DNA sequence affects post-transcriptional regulation (e.g., splicing and turnover). Some great work in this space already, but a lot more to learn.

Alternative possibility put forth in our preprint is that s4U RNA is sticky and can sometimes get lost on plastic surfaces during RNA extraction, especially with adherent cell lines where you might be trizoling the cells directly on the plate. Scraping cells into low-bind tubes first may help

800 uM is pretty high (100 uM being "standard"), though I assume you opted for higher concentration given the shorter label time. You could do a simple TAMRA dot blot assay with a range of s4U concentrations to see if you still get good incorporation at much lower concentration and 30 min labeling.

Some cell lines (e.g. HeLa cells) uptake s4U with far greater vigor than others. This can cause physiological impacts due to s4U, and that can cause the dropout you are observing. In this case, you can often go orders of magnitude lower in s4U concentration and get robust data without these problems

Also responded on X, but here's the preprint and paper describing this phenomenon that I shared there:

Preprint: www.biorxiv.org/content/10.1...

Paper: academic.oup.com/nar/article/...

The bioinformatic causes aren't relevant here. Could thus be an RNA handling problem or cells are s4U sensitive.