handle negative reactivity

[AMA-cs]

Hello @Psirving @VonBoss

I noticed that deltaSHAPE doesn’t handle negative reactivity values, which can impact both the smoothing and the detection of significant differences between conditions.

What’s the rationale behind keeping negative values in the calculation?

Negative values in reactivity profiles usually arise when the untreated condition shows a higher mutation rate than the treated one — potentially due to noise, sequencing bias, or low signal regions. Biologically, I’ve always understood negative SHAPE reactivity values to be non-interpretable or artifactual.

Given that, I’m curious about your thoughts on:

• Why negative values are retained?
• Whether clipping, masking, or ignoring them was considered?
• How their presence might affect the downstream interpretation of deltaSHAPE?

Would love to hear your insights.

Thanks!

[Psirving]

You're right, there is no such thing, physically, as "negative reactivity". These are noise or artifacts, but they can be useful in diagnosing issues with the experiment.

I think the thought process of the developer of deltaSHAPE was that reactivity normalization and quality control should be handled by ShapeMapper and by manual inspection prior to any downstream interpretation. ShapeMapper2 will mask high background mutation rate positions and will issue warnings to check log files if it sees any red flags. Negative reactivities should be few and modest. If not, something likely went wrong with probing and/or library prep. A few, modestly negative reactivity values should not have a large effect on deltaSHAPE results. One justification for keeping these values could be that "zero" values are not perfectly calibrated, i.e. it is physically possible that a nucleotide is less reactive than other nucleotides whose normalized reactivity value is "zero". In this case, there is important information in the negative value.

If you inspect the profiles and decide the data are high quality, but you identified spurious values that you don't believe reflect actual chemical reactivity (not only negative values), it seems justifiable to mask these positions. Another option is winsorization, meaning set negative values to zero. I don't love this as it kind of artificially changes the (lowercase) shape of the data.

If you have specific examples that you'd like to think through, I'd be happy to help.

-Patrick

[AMA-cs]

@Psirving Thank you so much for your quick response, and very helpful insights. This all makes sense now.

The quality of my samples is really high with < 1% of the nt with high background. However, I'm interested in detecting significantly differential sites in two conditions. For example:

So, around nt 200 (mainly after nt 200), deltaSHAPE showed significantly differential region, as the negative values in the top condition (control) became close to 0 in the bottom condition. However, I'm not sure what to think of this site! I even have another control samples and it showed similar negative reactivity in that spot.

I tried to change the negative values to 0, and re-generate deltaSHAPE. Thankfully the full profile isn't affected that much except the 200 site.

To me your point here makes sense:

One justification for keeping these values could be that "zero" values are not perfectly calibrated, i.e. it is physically possible that a nucleotide is less reactive than other nucleotides whose normalized reactivity value is "zero". In this case, there is important information in the negative value.

Yet, I don't know if I can comfortably say that this spot is detected as a significantly differential site.

[Psirving]

The top sample has large negative values, which makes me wonder what is causing this. I like to start by looking at the raw mutation rates for each sample, modified vs. untreated. This is the second panel of the profiles.pdf output from ShapeMapper. Can you attach these PDFs?

[AMA-cs]

Yes. I noticed that, even using another control sample. This region still has a lot of negative values.
FYI, I'm using the denatured sample to calculate the reactivity. Also, the previous screenshots are from the averaged reactivity across mutliple replicates.

To look at the mutation rates, I'll share the profiles for some of the replicates:
rep1:

rep2:

[Psirving]

Your replicates look great. It's hard to know without more info, but at this point I'm thinking it could be that modifications, structure, or sequence is affecting the enzyme's mutation rate, not chemical reactivity. I'll assume the rest of the RNA doesn't contain many negative reactivities and that your other samples look just as reproducible. Can you attach one of those profiles for comparison? I'm mostly wondering which treatment condition (untreated/denatured/modified) explains most of the difference in reactivity profiles.

[Psirving]

Mutation rates can also be effected in weird ways by alignment ambiguity. You can sometimes see this in the 3rd panel of the PDF. Can you attach the whole PDF for one representative replicate of each sample? You can drag-and-drop or email it to me.

[AMA-cs]

@Psirving I'm sorry for the late response! I appreciate your troubleshooting and insights.
Attached is the profile for one of my samples.
rep_profiles.pdf

This profile passed all the QC checks from Shapemapper2, and the background level is low. However, for the region right after tick2 there are many negative reactivity, and this region has higher signal untreated vs. modified unlike the rest of the profile. But I'm not completely sure if I should exclude this region from my my differential analysis. Especially that deltaSHAPE doesn't detect it as significantly differential when I replace the negative values with 0's.

Thanks again for your help

[Psirving]

This is a tricky situation and how you proceed from here is up to you. My opinion is that DeltaSHAPE has pointed to an interesting differential signal in your data, but it is not as simple as "differential SHAPE reactivity". Because of the negative reactivity values, you could easily justify ignoring this region in your DeltaSHAPE analysis. If the point of this analysis was curious exploration, I think this site is intriguing. I haven't seen a MaP signal behave like this. To get an interpretable result, you'd need to come up with a strong hypothesis and test it using a different method. Depending on your goals, this might be more of a distraction than it is worth.

A few things stand out to me so far:

1. The negative values are highly localized around the 2nd tick. Everywhere else has typical, quality reactivity rates. I don't think this site is artifactual or noisy.
2. The background mutation rate is elevated in this region. This can be caused by structure or modified nucleotides. I'm curious if the differential signal is driven by differences in the background mutation rates. That would suggest the RNAs in the two conditions are physically different.
3. There is differential RT processivity in this region. Blue and red switch downstream of the 2nd tick in the read depth panel. I'm not sure if this is important. I would follow up by looking in the mapped_depths.pdf.
4. I noticed from a previous comment that this is an unusually purine-rich region, which could suggest some functional relevance.

If you're interested in more of a collaboration to figure this out, feel free to email me.