use of run-specific q-value filter instead of global q-value filterin

[JM-Bader]

I process a large cohort of thousands of samples with the DIANN GUI
It appears that this leads to a problem with the global FDR filtering as random decoys accumulate over the thousands of raw files and lead to significant FDR-based cutting of the final dataset.

Q1: Is it possible to only apply a run-specific FDR and disable the global FDR?

Q2: I then plan to apply a valid value filter to remove the excess of decoy 'one-hit wonders' (and their corresponding 'one-hit wonder' wrong targets). I guess that makes sense, doesn't it?

Q3: The github suggests that a run-specific FDR instead of global FDR is possible but I don't really get how to set this in the GUI with commands. How do I do it?

Q4: Should the global FDR-filtering be disabled on precursor, peptide, or protein level ? Or all of them (as accumulation of random decoys over thousands of runs can happen in all these categories)?

[vdemichev]

Hi Jakob,

What you observe is strange. Do you mean you get A LOT less precursors at 1% global FDR with 5000 runs than with, say, 100 runs? There are algorithmic measures in DIA-NN 1.8 and later specifically designed to make sure that this does not happen. How do the numbers you observe look like?

Q1: Without MBR there's no global filtering applied to the main report. With MBR you can relax it a bit indeed. To, say, 2% or 3% should be OKish.
Q2: What do you mean 'decoy' one-hit wonders? DIA-NN does not report decoy hits at all.
Q3: Change precursor FDR margin in the GUI to 2%, for example. This will set MBR global FDR threshold to 2%. Run-specific will also be 2% in the final report, but that you can always filter to 1% afterwards.
Q4: I would never disable it fully. Relaxing protein global FDR a bit for plasma might make sense. For tissues I don't think this is ever necessary.

Best,
Vadim

[JM-Bader]

or maybe make that a button in the GUI to select library grouping instead of UniProt ID

[JM-Bader]

Alternatively, --library-headers *,*,*,*,*,*,Protein.Group (lets DIA-NN know to use the Protein.Group column as the main protein ID column) might also work instead of editing the library file, but I haven't tested this for a while.

FYI: this command causes DIANN not to recognize proteins in the library properly (see top row for command, bottom for issue):

[JM-Bader]

Q2: FYI, copying the ProteinGroup column from the DIA-NN-generated library to overwrite the UniprotID column there expectedly works - in the sense that DIANN is able to read the file, and accordingly adjusts the number of proteins in the library. I set Protein inference to OFF and unticked heuristic protein inference.
Does it make a differnce if there is fasta file still loaded ? (For now I left it, so it is still specified). I guess at this point the fasta is irrelevant though. Correct?

[JM-Bader]

Q3: Independently, for the Re-search I plan to use the first pass results. Does it make a difference if MBR is enabled (which also leads to a second pass and it's results) ? i guess it does not make a difference for first pass results. Correct?

[JM-Bader]

FYI as feedback:
Correctly transforming the library (using Protein.Groups to overwrite UniprotID column in DIA-NN output lib) does lead to a noticeable difference (which is reasonable as well!):

blue: initial search using MSFragger lib
orange: re-search using output library of initial search (not transforming the lib)
green: re-search using output library of initial (transforming the lib) + Protein grouping disabled in the re-search
Calculation of numbers: filter main report tsv for 1% Global.PG.Q.Value (initial search) or 1% Lib.PG.Q.Value (re-searches). For initial search, take second pass results, for re-searches first pass results (i.e disabling MBR).

The green and blue overlap much more than the orange one.
So this library transformation is actually important in my dataset

[JM-Bader]

Slightly different topic here: All of the above is about proteins.
Here, how about precursors and fragments? (relevant to manual assembly of protein intensities with iQ etc.)

Q1:
Like for proteins, I would now expect in the main report:

• run-specific q-value
• Global.Q.Value
• Lib.Q.Value
• each for precursors and fragments.
  Do I find these 3 q-values there each?

Q2:
I presume the table is already filtered for run-specific q-value each but not regarding the two other q-values - each for fragments and precursors. Correct?

Q3:
Do I need to filter the main report before assembling protein intensities based on precursor and fragment information with iQ?
Should I use the Lib.Q.Val when the library is a DIA-NN output and the Global.Q.Value when the Library is not a DIA-NN output?

[JM-Bader]

Q1:
a) for potential quality control. b) My idea was to filter for well-identified fragments and also use these for protein quantification (rather than all fragments). Or at least check whether that change (all vs well-identified fragments) correlates with a difference in protein CVs etc. I see your point that this would maximally go half way as the q-value controls for identification but not quantification and second your experience that MaxLFQ is relatively stable towards wrong precursors.

a) So in essence, DIA-NN currently only uses precursor intensity for protein quantification ?
b) DIA-NN filters precursors for run-specific q-value (but not global precursor Q-value and lib Q-value) for i) reporting in the main tsv and ii) protein quantification - correct?
c) .... So, what happens if fragments are identified and quantified but the precursor is not detected (e.g. in a crowded MS1 space) ? As far as I remember, this is actually not uncommon

Q3: I see, your answer pertains to precursors only as currently DIANN MaxLFQ only uses precurso intensity and you recommend to do it likewise with iQ?

[vdemichev]

Q1a Yes, makes sense :) But not implemented yet
a Yes
b Yes, but MBR means automatic global q-value (called Lib.Q.Value in this case) filtering
c Quant is MS2 based in DIA-NN
Q3 Yes, but it's been reported that fragment-level MaxLFQ also works nicely

[JM-Bader]

Wait, I am confused now.
Q3-answer: DIA-NN uses MaxLFQ to quantify proteins based on precursor intensities
Q1c-answer: DIA-NN uses MS2 quantities (i.e. fragments, and not precursors) to quantify proteins
Or do you mean that in a first step precursor intensities are inferred from MS2 level intensities (=fragments) (rather than integrating the MS1 traces of precursors) ? And subsequently precursor intensities are fed into MaxLFQ?

[vdemichev]

Sorry, 'precursor' in DIA-NN means a charged peptide molecule species, which is quantified using MS/MS signal, i.e. fragment intensities

[JM-Bader]

ok, understood. I just incorrectly assumed that precursor intensity was inferred from MS1 traces of the precursors (which may have quite a lot missing values). Ok, things make sense now

[JM-Bader]

yet another aspect:
if I run a primary search (i.e. using a MSFragger library), I get both the main report and the library from that search.
Surprisingly, the library contains more unique Protein.Group entries than the main report filtered at 1% Global.PG.Q.Value. I then looked at the 'PGQValue' reported in the library file - which I presume is the Global.PG.Q.Value in the main report (Q1: correct?). Unexpectedly, the library contains entries with a PGQValue > 1%.
Q2: Why is that (lib entries with PG.Q.value > 1%)? Is there another PGQValue filter (e.g. 10%) or none at all? What is the point of having them in the library? Using a very relaxed filtering in the secondary search?
Q3: When using said library for a Re-Search, these entries will be removed on protein level (when using a PG.Lib.Q.Value filter < 1%), so they don't matter anyway, do they ? Or is there any chance they can become significant (i.e. included in the final search results) ? Hypothetically, if protein grouping is changed the precursors could be assigned to new groups - but that is not desired (because PG.Lib:Q.Value filtering then doesn't make sense anymore)... Or am I missing anything?
Q4: When comparing unique Protein.Group entries, the library still has more entries than the unfiltered (ie. not Global.PG.Q.Value-filtered) main report. Why?

[vdemichev]

MSFragger libs should be used with protein inference disabled in DIA-NN.
When DIA-NN generates a lib from DIA data, it's filtered using global precursor FDR. It's not filtered using any protein FDR as the user can always do such filtering after the analysis is finished.
Q3: sorry, got a bit lost about what's happening.

[JM-Bader]

Ok, I see the logic.
Re Q3: I guess I don't have to care about the number of proteins in the library. I just use overwrite the UniProtID column with the PG.Group column there and remove all proteins in the final search which have a Lib.PG.Q.Value > 1%. I definitely need to disable protein grouping and heuristics for the re-search using the DIA-NN lib.

[vdemichev]

Yes