Error when using unmodified peptide to train and using modified peptides to test about deeplc

[RobbinBouwmeester]

Dear WeiqiangChen,

That is definitely possible! However, from the example you posted the column names are wrong. In this case the column "modification" should be ""modifications". Otherwise DeepLC is looking for a column that does not exist.

If you are interested in retraining a model definitely also keep an eye on this repo: https://github.com/RobbinBouwmeester/DeepLCRetrainer

Soon I will launch that code (with GUI) that should enable retraining/transfer learning in an easier way.

Hope that helped!

Kind regards,

Robbin

[WeiqiangChen]

Dear Robbin,

thanks for the reply. Now deepLC works.
I tried to train it with unmodified peptides from MaxQuant evidence.txt. Group_by modified_sequence, slice_max(order_by intensity, n = 1) to get the apex retention time for each modified_sequence. And the average(predicted_tr - tr) is 7.9mins.
seq,modifications,tr
AAAESIQMR,8|Oxidation,1353
AASVGPTMR,8|Oxidation,1264.26
ADLEMQIESLK,5|Oxidation,5267.34

I have tried to test also acetylated peptides. And the average(predicted_tr - tr) is now -20mins.
seq,modifications,tr
AARPLVTVYDEK,1|Acetyl,4367.64
ADFDTNPTSLYSIK,1|Acetyl,7029
AHIVQTHK,1|Acetyl,1314.48
AQHPLVQR,1|Acetyl,1989.42

Did I make some mistakes here?

[RobbinBouwmeester]

It could be that the current models you use are not able to extrapolate to your modifications. Could you try these models:

https://github.com/RobbinBouwmeester/DeepLCModels/blob/main/full_hc_mod_deeplc_train_filtered_1fd8363d9af9dcad3be7553c39396960.hdf5
https://github.com/RobbinBouwmeester/DeepLCModels/blob/main/full_hc_mod_deeplc_train_filtered_8c22d89667368f2f02ad996469ba157e.hdf5
https://github.com/RobbinBouwmeester/DeepLCModels/blob/main/full_hc_mod_deeplc_train_filtered_cb975cfdd4105f97efa0b3afffe075cc.hdf5

[WeiqiangChen]

Model2_[469ba157e.hdf5] got the best prediction for oxidated peptides when training with unmodified peptides.

All 3 models got bad predictions for acetylated peptides when training with unmodified peptides.

[RobbinBouwmeester]

I see... Now, are these n-terminal acetylated peptides? Officially we do not support terminal modifications. Although you can include them on the 0 or 1 position (it will default to the rest group of the first AA) it is likely to be suboptimal...

[WeiqiangChen]

Yes. These are acetylation(protein N term).

[RobbinBouwmeester]

Ok, that is likely to be the problem. I would recommend in that case to retrain a model with many acetylated termini, so it will "force fit" it into the current DeepLC. Feel free to contact me via e-mail (robbin.bouwmeester[at]ugent.be about the details). Can I close this issue now?