Ideas for analyzing phosphoproteomics DIA data using the DIANN software.

[BLACKWEI1]

Hello everyone! I'm really excited that DIANN can process phosphoproteomics DIA data and generate tables of phosphorylation sites. However, after looking into it, I found that the algorithm used by DIANN is not similar to PhosphoRS used by PD, SP, MaxQuant, etc. Top1 seems much simpler and more straightforward, but it lacks dynamic site localization and model statistics. The industry default score cutoff is 0.75, while DIANN generates 99.tsv and 99tsv files. I wonder if it would be possible for future software updates to switch to the more universally accepted PhosphoRS, which would make data analysis, interpretation, and publication much more convenient. What do you all think about this?

[vdemichev]

Hi,

I found that the algorithm used by DIANN is not similar to PhosphoRS

DIA-NN uses an advanced proprietary algorithm for localisation.

Top1 seems much simpler and more straightforward

We suspect it's more robust than alternatives. We do not know any benchmarks that would compare quantification methods in a situation where multiple peptidoforms matching same sequence exist and could be differentially regulated. However, the site_report introduced in DIA-NN 2.2 allows you to (i) filter phosphosites in any way you like and (ii) quantify them with any algorithm, based on MS1 or MS2 information or both. That is, the user has full control now.

The industry default score cutoff is 0.75

That's because if one wants to sell an instrument or software, it's beneficial to show good numbers, and hence it's very tempting to use the 'class I site' definition using 0.75 cutoff, first introduced in 2006 for small-scale DDA data (https://www.sciencedirect.com/science/article/pii/S0092867406012748), to justify very relaxed confidence filtering. Likewise, it's easier to publish a paper that shows good numbers at the expense of identification confidence. As you can imagine, what was appropriate back in 2006, as identifying any meaningful number of sites at better confidence used to be impossible, is hardly OK now, considering much better instrument capabilities and software algorithms as well as large experiment size (if one has 100 runs and 75% site confidence filter, what proportion of sites have nonsense quantities in their quantitative profiles across 100 samples?). We therefore suggest using at least 90% cutoff. But with the new site_report can now use any.

it lacks dynamic site localization and model statistics

Not sure what this refers to?

switch to the more universally accepted PhosphoRS

I think this is a localisation algorithm, not quantification?

Best,
Vadim