search

bzhanglab / DeepRescore

DeepRescore: rescore PSMs leveraging deep learning-derived peptide features
5 stars 3 forks source link

Question regarding transfer learning approach #1

Open kevinkovalchik opened 3 years ago

kevinkovalchik commented 3 years ago

Hello!

Thanks for developing this tool. It is nice to see this rescoring approach finally packaged in an easily distributed and reproducible manner.

I have a question regarding the fragment ion intensity prediction model. Are you using the pre-trained model provided with pDeep2 as it is? Or is the model fine-tuned using the high-confidence PSMs (i.e. like AutoRT is) before predicting all the spectra?

Best wishes, Kevin

wenbostar commented 3 years ago

Hi Kevin,

Since MS/MS spectra generated from the same type of MS spectrometers or similar MS spectrometers with similar parameters are very similar, we just use the pre-trained models provided by pDeep2 for MS/MS spectrum prediction. Fine-tuning the pDeep2 models using the high-confidence PSMs may further improve the performance of rescoring. But we haven't tried that yet.

Bo

kevinkovalchik commented 3 years ago

Hi Bo,

Sorry for my super-slow response. I agree that instrument-specific differences wouldn't have a big impact on accuracy. I was thinking more about sample differences. My understanding is the pDeep models are trained on tryptic digests, and if someone is working with a different enzyme or with unspecific digests then the terminal peptides could be quite different from anything the model was trained on. From what I have seen models that predict fragment ions tend to generalize well enough to still do a good job, but it would be interesting to see if the transfer-training would help out.

Kevin