MS/MS Spectrum Prediction for Modified Peptides Using pDeep2 Trained by Transfer Learning.

Area of Science:

• Proteomics
• Computational Biology
• Mass Spectrometry

Background:

• Bottom-up proteomics using tandem mass spectrometry (MS/MS) is crucial for analyzing post-translational modifications (PTMs).
• Accurate prediction of theoretical MS/MS spectra for modified peptides is essential for identifying PTM-containing peptides and localizing modified residues.
• Current methods face challenges in PTM analysis due to the complexity of biological mixtures.

Purpose of the Study:

• To develop a deep-learning model, pDeep2, for accurate prediction of MS/MS spectra of peptides with post-translational modifications.
• To improve the identification and localization of PTMs in complex proteomic samples.
• To leverage transfer learning for efficient training of the model with limited PTM data.

Main Methods:

• Development of the pDeep2 model utilizing deep learning architectures.
• Application of transfer learning techniques to train the pDeep2 model on benchmark PTM datasets.
• Validation using public synthetic phosphopeptide and 21 synthetic PTM datasets from ProteomeTools.

Main Results:

• The pDeep2 model trained with transfer learning achieved high accuracy, with Pearson correlation coefficients exceeding 0.9 (>80% of cases).
• Transfer learning significantly improved model performance compared to training without it.
• Accurate prediction of fragment ion intensities, including neutral loss ions (e.g., phosphoric acid loss), enhanced discrimination of true modified residues.

Conclusions:

• The pDeep2 model offers a significant advancement in predicting MS/MS spectra for PTM analysis.
• Transfer learning is an effective strategy for training accurate PTM prediction models with limited data.
• Improved spectral prediction, particularly for neutral loss ions, enhances the reliability of PTM identification and localization in proteomics.