We have tried to sequence RNA from spent cell culture media using Smart-seq3. We have two groups across two time points, around 18 negative control samples which are just fresh media with no cells grown on them, and 4 water samples (RNA-free) which were subject to library preparation. Our idea was, we would use the negative control samples to identify and subtract background counts, and use Water samples to subtract any RNA signals generated by the reagents/contamination (As the water samples were processed in the same batch as rest of the samples).

However, from what I'm seeing this design seems rather unorthodox (but still logical to me).

We planned to perform the analysis with DESeq2 and other tools, however, from what I've read, the tools do not provide a provision to specify negative samples. What would be the most sensible way to approach this?

This is typically done with an interaction term (see vignette). That asks if comparison B vs A is larger in one treatment than in another treatment. See vignette for many details.

Making a library based on water was a waste of time and resources. No one does this. I don't think there is any reason to think that the noise is going to be the same across all samples, so subtracting a random noise amount from all the samples is only going to hurt.

Besides, likely the noise levels will be way way too high. The person loading the empty sample on the instrument will surely put a very high concentration onto the instrument when they see the library looks terrible.

Surely some person doing the benchwork warned you that the RNA looked terrible? And the library looked terrible? Garbage in...