Transfer Learning allowed?

[lliligabriel]

Greetings Dr. Gabriels again:)

I saw the thread #227 from before. I wonder whether MS2pip supports transfer learning in any way? Not sure if building a xgboost model from scratch is the best idea for my purpose.

Thank you so much!

[RalfG]

Hi @lliligabriel,

I am not sure if transfer learning is natively supported by XGBoost. We have never tried it. What is the specific application domain you are aiming at?

Best,
Ralf

[lliligabriel]

Oof thanks! you are right - forgot this is an xgboost model.
I tested the performance of ms2pip on a dataset composed of chemically synthesized peptides as inputs and data processed & fdr-controlled at q< 0.01 (spectrum level, peptide level, and protein level). That is, the data shall be relatively "clean." The performance is fine but I wonder whether providing our own inputs can boost the performance a bit and make downstream rescoring easier, as I am not sure whether differences due to enzymatic processing methods or machine setup can impact the prediction.

Thanks!

[RalfG]

Indeed, both peptide properties and instrument can influence fragmentation patterns. Our most recent HCD (Orbitrap) and timsTOF models have been trained on both tryptic and non-tryptic peptides. Note that the training data did not contain any amino acid modifications beyond cysteine carabamidomethylation and methionine oxidation.

The correlations you're showing here are a bit below expectation. You can see expected correlation distributions for various peptide types here: https://academic.oup.com/view-large/figure/409531239/gkad335fig1.jpg