E.coli peptides annotated as oxidised human/yeast peptides Quant LFQ ion DIA AIF

[vdemichev]

Describe the bug
PSMs annotated as oxidised human/yeast peptides likely come from E.coli peptides (with some different sequences?) and have the same distribution of ratios as E.coli peptides

To Reproduce
Download DIA-NN_20241216_121542
library(arrow) # R code
dia <- read_csv_arrow('C:/Temp/result_performance.csv')
hist(dia$log2_A_vs_B[dia$HUMAN == 'True' & !grepl('Oxidation',dia$precursor ion)],n=500)
hist(dia$log2_A_vs_B[dia$HUMAN == 'True' & grepl('Oxidation',dia$precursor ion)],n=500)

Expected behavior
Oxidised peptides correctly annotated

Screenshots
Non-oxidised:

Oxidised:

The bug does not seem to affect AlphaDIA results, but affects both DIA-NN and Spectronaut.

[RobbinBouwmeester]

Hi Vadim,

I can reproduce you figures, and indeed it looks strange. Now, with a very minimal code (less chance of having bugs on the raw input data) I still get the same figure though:

condition_A = ["LFQ_Orbitrap_AIF_Condition_A_Sample_Alpha_01", "LFQ_Orbitrap_AIF_Condition_A_Sample_Alpha_02", "LFQ_Orbitrap_AIF_Condition_A_Sample_Alpha_03"]
condition_B = ["LFQ_Orbitrap_AIF_Condition_B_Sample_Alpha_01", "LFQ_Orbitrap_AIF_Condition_B_Sample_Alpha_02", "LFQ_Orbitrap_AIF_Condition_B_Sample_Alpha_03"]

df_raw = pd.read_csv("input_file.txt",sep="\t")

human_mask = (df_raw["Protein.Names"].astype(str).str.contains("_HUMAN")) & \
             (~df_raw["Protein.Names"].astype(str).str.contains("_YEAST")) & \
             (~df_raw["Protein.Names"].astype(str).str.contains("_ECOLI"))

df_raw["Sequence_Charge"] = df_raw["Modified.Sequence"].astype(str) + "/" + df_raw["Precursor.Charge"].astype(str)

df_raw_pivoted = df_raw.loc[human_mask].pivot_table(index="Sequence_Charge",
                            columns="Run",
                            values="Precursor.Quantity",
                            aggfunc='mean',
                            fill_value=0)

oxid_mask = df_raw_pivoted.index.str.contains("(UniMod:35)")

df_raw_pivoted_filtered_oxid = df_raw_pivoted[((df_raw_pivoted != 0).all(axis=1)) & oxid_mask]
df_raw_pivoted_filtered_nonoxid = df_raw_pivoted[((df_raw_pivoted != 0).all(axis=1)) & ~oxid_mask]

df_raw_pivoted_filtered_oxid["mean_log2_condition_A"] = np.log2(df_raw_pivoted_filtered_oxid[condition_A]).mean(axis=1)
df_raw_pivoted_filtered_oxid["mean_log2_condition_B"] = np.log2(df_raw_pivoted_filtered_oxid[condition_B]).mean(axis=1)

df_raw_pivoted_filtered_nonoxid["mean_log2_condition_A"] = np.log2(df_raw_pivoted_filtered_nonoxid[condition_A]).mean(axis=1)
df_raw_pivoted_filtered_nonoxid["mean_log2_condition_B"] = np.log2(df_raw_pivoted_filtered_nonoxid[condition_B]).mean(axis=1)

plt.hist(df_raw_pivoted_filtered_oxid["mean_log2_condition_A"] - df_raw_pivoted_filtered_oxid["mean_log2_condition_B"], np.arange(-4,2,0.05), alpha=0.25)
plt.hist(df_raw_pivoted_filtered_nonoxid["mean_log2_condition_A"] - df_raw_pivoted_filtered_nonoxid["mean_log2_condition_B"], np.arange(-4,2,0.05), alpha=0.25)
plt.show()

[RobbinBouwmeester]

So only thing I can imagine is that the issue is in the fasta or something in DIANN. Do you agree? Or do you see any other potential issue?

[vdemichev]

that the issue is in the fasta or something in DIANN

Will need to be with both DIA-NN and Spectronaut... If not a mistake with a script, really wonder what's up, no clue so far.
Even if this were gross misidentification by both DIA-NN and Spectronaut, and the E.coli digest were overwhelmingly oxidised, I would expect it to be mixture of 1:1 and 0.5:1 IDs, not like this.

Would you please be able to share the original DIA-NN output?

[RobbinBouwmeester]

Sure, it is the input_file.txt from this zip:

https://proteobench.cubimed.rub.de/datasets/fdc2fffd886fe2a36722bfa85a418b09e6bc69a9/fdc2fffd886fe2a36722bfa85a418b09e6bc69a9_data.zip

[RobbinBouwmeester]

I blasted "AAAPAPEEEMDECEQALAAEPK" (AAAPAPEEEM(UniMod:35)DEC(UniMod:4)EQALAAEPK) and "YYADGEDAYAMK" (YYADGEDAYAM(UniMod:35)K/2), both came back as only a perfect match to "Homo sapiens", no perfect match to either E. coli or yeast...

[vdemichev]

This is contaminant peptides with M(ox):

Wonder if this relevant.

[RobbinBouwmeester]

FYI, @rodvrees analyzed some diaPASEF data yesterday and the pattern is not visible there. There is substantially less oxidation though.

[vdemichev]

AlphaDIA identifies 7362 oxidised human precursor, out of which 4866 also detected by DIA-NN. So overlap quite good. I picked a particular match which has same matched RTs in DIA-NN and AlphaDIA. Precursor is AAFTVSLDPGPLEQFPHSM(UniMod:35)EPQLR3 (Q9UKD2). Runs are ordered A1 - A3, B1 - B3:

Ms1.Area, DIA-NN: (raw integrated MS1 signal, no normalisation, etc):

Precursor.Quantity, DIA-NN (MS2-based, no normalisation):

Intensity, AlphaDIA:

RT, DIA-NN:

RT, AlphaDIA:

So raw signal as seen by DIA-NN is significantly up in B compared to A for this precursor. AlphaDIA, at the same RTs, sees no differential regulation. I just checked, apparently 'intensity' in AlphaDIA is the protein (not precursor!) intensity. Hence no regulation and hence also such good precision metrics.

[vdemichev]

Now, the above confirms:

• All three tools likely identify same oxidised precursors (overlap DIA-NN & AlphaDIA)
• DIA-NN and Spectronaut show 0.25:1 for oxidised human precursors

So either the E.coli digest is contaminated with oxidised human stuff (I am not sure how this can happen that just the oxidised stuff will be there?) or all three tools are wrong. While the latter is possible, this is very unlikely, I see 0.25:1 in DIA-NN also with filters that would guarantee high peptidoform confidence.

[RobbinBouwmeester]

Thank you Vadim, fully agree with your observations, we will have a look at everything. Including the problem of protein level abundances reported by AlphaDIA, indeed we should not take those values.

[mlocardpaulet]

Hey, this is really great that you spotted this and spent the time to look into it. Thank you so much.

1. we definitely need to change how we handle alphaDIA results and make sure that we take the precursor quant.
2. Regarding the sample contamination: should we ignore the human oxidised precursors when we calculate the metrics?
   Any other idea on how to proceed?

[vdemichev]

It's yeast too, which is very weird, I cannot think of any experimental explanation for this.

[mlocardpaulet]

We think that this comes from the sample. And it may be specific of this condition B alpha run. I'll have a look at other replicates and see if we couldn't use other runs. Keep you posted.

[mlocardpaulet]

I made a plot with only the precursors matched to a single species and no contaminant in the DIA AIF data:
I performed median centering per species after log2 transformation to align run signals
I calculated the median values of all precursors centered signal per species grouping as follows:
. precursors with min. one oxidation (blue)
. precursors with the same stripped sequence as the ones seen with min. one oxidation (red)
. all other precursors matching the same species (grey). These are roughly aligned because of the initial median centering (removes the effect of initial differences of quantities)
Here is what I get:

So it really is sample-dependent. And no idea of the origin.
We can select samples for which the effect is not as strong as with the ones originally picked (alpha), but there seem to be a difference between all the condition B and condition A samples.

[mlocardpaulet]

After discussion, we decided that we will remove this module, and make a very similar DIA module with Astral data (in the process of being acquired) to have next to the diaPASEF one.