Use of ERCC spike-in for extremely different samples

[dfhdfh]

Hi, I would appreciate a little help on the following situation.
I have bacterial total RNA that I separate into several samples and want to sequence afterward. From my controls I know that each of those samples contains a very different part of the total RNA (e.g. some samples don't contain rRNA).
In order to still be able to normalize my sequencing results, I most likely cannot use the usual normalization methods that assume that most of the transcripts don't differ between samples.
Therefore, I think a spike-in is an option allowing me to normalize my data without the "help" of my actual sample. For this, I want to use the ERCC spike-in and use RUV for normalization afterward.

This is where I'm at at the moment. My current plan is to add the spike-in after RNA extraction (prior is not possible) and then continue with rRNA depletion, fragmentation and library prep (most likely using polyA-ligation). However, I'm not sure, how exactly the best way of adding the spike-in in this case would be. The "official" way is to measure the RNA concentration (which is not the precisest thing anyway) and then add a set amount of spike-in to each sample. As far as I understand, the reason to do so is to be able to calculate the dynamic range of your sequencing afterward. However, I'm not sure I'm interested in that. Is it possible to add the same amount of spike-in to each sample and then normalize according to the read numbers of the spike-in in each sample? What would be the (dis)advantage in doing so?

[jparsons]

A few points:

RUV really doesn't work well with ERCCs; it doesn't account for the fact that the controls have built-in differences. If you do want to use RUV for this type of purpose, you'll need to filter the ERCCs that you feed to RUV to only include the 1:1 component or only use a single pool of controls.

Your plan to use the spike-ins after extraction has some precedent, including the cell reference below and some of my work. I would recommend normalizing with a method similar to [www.cell.com/abstract/S0092-8674(12)01226-3].

The advantage of doing so would be that you should become aware of any differential enrichment of your samples between fractions [eg: there is more bacterial RNA sequenced from sample A compared to sample B]

However, the spike-ins have issues with rRNA depletion (if done by polyA beads) as well as awkward polyA tail lengths that may bias their abundances during library prep. Without carefully demonstrating that the spike-ins can properly normalize samples in your exact prep (which hasn't been done to my knowledge), I wouldn't recommend using them for quantitative assessment/normalization.

That said, in the few experiments I'm aware of where there are known differences in spike-in vs endogenous RNA, the spike-ins are able to recapitulate the correct differences effectively, so the biases they are affected by are likely uniform.

My recommendation would be to proceed with caution; use a few control samples to demonstrate that the normalization works.

[dfhdfh]

Thank you for your suggestions. PolyA beads won't be used for rRNA depletion since I have bacterial samples. RiboZero would be the way to go here. However, I am not convinced anymore if rRNA depletion is what I want to do anyway. This has two reasons: a) I can just sequence deeper to compensate for that and b) I am actually interested in the rRNA distribution since this is a good control. For the actual analysis I could then filter the rRNA reads and work without them (here, I am not too certain about compression effects that might arise, I have to read more on that).
Why do you think RUV is a bad idea? As far as I can tell from the Risso Nature Biotech paper, RUVg should be useful for normalization using negative control genes. In this case, I would only use one of the two ERCC sets in order to keep the distribution the same for all samples. Since I start with RNA extracted from X cells and then split over several samples, I naturally have differing RNA concentrations for all the samples. By spiking the same amount of ERCC into each sample, I want to keep the information about this difference in concentration so I can track differences in concentration for a single transcript over all my samples. In this way, it should be comparable to the Cell paper you mentionend.
Why wouldn't you use the ERCC for quantitative normalization?

Another option would be to add something like the ERDN (Locati et al., NAR, 2015) after fragmentation of my RNA. We tried something similar in an earlier experiment, which wasn't completely satisfying.

[jparsons]

My main reason why not to use the spike-ins for quantitative normalization is that I've never seen it demonstrated to work and that we are aware of biases that affect them - and these biases may vary from experiment to experiment. However, if you don't do rRNA depletion, the sources of bias are reduced to 'just' RT and PCR, which is a realm where the ERCCs are unlikely to differ from biological sequences (depending on your RT reaction type). I'd be comfortable with normalization if you're not doing rRNA depletion and you can sequence deep enough. I've avoided depletion in the past and had better than expected results.

As for RUV, on re-reading of the paper i might have taken away the wrong message initially ; the issues they report are primarily with other normalization methods using spike-ins, and RUVg appears to work well with controls given they're from the same pool.

[dfhdfh]

Since I had to use RiboZero anyways, I can't really see an issue with biases in rRNA depletion concerning ERCCs? Theoretically, they shouldn't bind to the probes.
The concern I have with RiboZero is that I'm not sure what exactly happens when you don't really have any rRNAs to start with. Meaning how many/which non-rRNA molecules will be bound and depleted when there's no rRNA to be bound. This will be true for some of my samples and I don't want to start and treat every sample differently.